text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'It has been recently proved that a slab of negative refractive index material acts as a perfect lens in that it makes accessible the sub-wavelength image information contained in the evanescent modes of a source. Here we elaborate on perfect lens solutions to spherical shells of negative refractive material where magnification of the near-field images becomes possible. The negative refractive materials then need to be spatially dispersive with $\varepsilon(r) \sim 1/r$ and $\mu(r)\sim 1/r$. We concentrate on lens-like solutions for the extreme near-field limit. Then the conditions for the TM and TE polarized modes become independent of $\mu$ and $\varepsilon$ respectively.'
author:
- 'S. Anantha Ramakrishna'
- 'J.B. Pendry'
title: A spherical perfect lens
---
Introduction
============
The possibility of a perfect lens [@pendry_PRL00] whose resolution is not limited by the classical diffraction limit has been subject to intense debate by the scientific community during the past two years. This perfect lens could be realised by using a slab of material with $\varepsilon = \mu = -1$ where $\varepsilon$ is the dielectric constant and the $\mu$ is the magnetic permeability. Veselago had observed[@veselago] that such a material would have a negative refractive index of $n = -\sqrt{\varepsilon
\mu} = -1$ (the negative sign of the square root needs to be chosen by requirements of causality), and a slab of such a material would act as a lens in that it would refocus incident rays from a point source on one side into a point on the other side of the slab (See Fig. 1). Due to the unavailability of materials with simultaneously negative $\varepsilon$ and $\mu$, negative refractive index remained an academic curiosity until recently when it became possible to fabricate structured meta-materials that have negative $\epsilon$ and $\mu$ [@smith00; @smith01; @eleftheriades]. Most of the negative refractive materials(NRM), so far, consist of interleaving arrays of thin metallic wires (that provide negative $\varepsilon$ [@pendry96]) and metallic split-ring resonators (that provide negative $\mu$ [@pendryIEEE]). Although some initial concerns were expressed [@garcia_OL] that the observed effects in these experiments were dominated by absorption, the recent experiments of Refs.[@parazzoli; @houck; @grbic; @sridhar] have confirmed that negative refractive materials are today’s reality.
{width="8cm"}
It was demonstrated by one of us that the NRM slab acts a lens not only for the propagating waves (for which the ray analysis of Veselago is valid) but also for the evanescent near-field radiation [@pendry_PRL00]. This phenomenon of perfect lensing becomes possible due to the surface plasmon states [@raether] that reside on the surfaces of the NRM slab which restore the amplitudes of the decaying evanescent waves [@pendry_PRL00; @drs_APL03; @haldane; @gomez_PRL; @sar_JMO02; @xsrao_PRB]. Indeed, it has been confirmed by numerical (FDTD) simulations that an incident pulse is temporarily trapped at the interfaces for a considerable time [@foteinopolou]. For a detailed description of the perfect slab lens, we refer the reader to Ref. [@pendry_PRL00; @sar_JMO02; @pendry_JPC02]. The ‘perfectness’ of the perfect lens is limited only by the extent to which the constituent NRM are perfect with the specified material parameters. Absorption in the NRM and deviations of the material parameters from the resonant surface plasmon conditions of the perfect lens causes significant degradation of the subwavelength resolution possible [@drs_APL03; @platzmann_APL; @zye_PRB; @fang_APL]. We have suggested some possible measures to ameliorate this degradation of the lens resolution by stratifying the lens medium [@sar_JMO03] and introducing optical gain into the system [@sar_PRBr03].
The image formed by the NRM slab lens is identical to the object and hence there is no magnification in the image. Lenses are mostly used to produce magnified or demagnified images and the lack of any magnification is a great restriction on the slab lens on which most of the attention in the literature has been focussed. The slab lens is invariant in the transverse directions and conserves the parallel component of the wave-vector. To cause magnification this tranverse invariance will have to be broken and curved surfaces necessarily have to be involved. The perfect lens effect is dependent on the near-degeneracy of the surface plasmon resonances to amplify the near-field, and curved surfaces in general have completely different surface plasmon spectrum [@klimov]. It was recently pointed out by us that a family of near-field lenses (in the quasi-static approximation) in two-dimensions can be generated by a conformal mapping of the slab lens [@pendry_JPC02]. Thus a cylinderical annulus with dielectric constant $\varepsilon = -1$ was shown to have a lens-like property of projecting in and out images of charge distributions. Similiarly in Ref. [@pendry_OE03] and [@pendry_JPC03], it was shown how a general method of co-ordinate transformations could be used to map the perfect slab lens solution for the Maxwell’s equations into a variety of situations including the cylinderical and spherical geometries respectively.
In this paper, we elaborate on the perfect lens solutions in the spherical geometry and show that media with spatially dispersive dielectric constant $\varepsilon(\vec{r})/sim 1/r$ and magnetic permeability $\mu(\vec{r})/sim 1/r$ can be used to fabricate a spherical perfect lens that can magnify the near-field images as well. In section-2 of this paper, we will present these perfect lens solutions of the Maxwell’s equations for the spherical geometry. In section-3, we will examine the solutions in the extreme near-field limit or the quasi-static limit which is useful when the lengthscales in the problem are all much smaller than a wavelength. Then the requirements for TM and TE polarizations depend only on $\varepsilon \sim 1/r^2$ or $\mu \sim 1/r^2$ respectively. This is useful at frequencies where we are able to generate structures with only one of $\varepsilon$ or $\mu$ negative. We will investigate the effects of dissipation in the NRM and point out the connections to the 1-D slab lens solutions. We will present our concluding remarks in Section-4.
A perfect spherical lens
========================
Consider a spherically symmetric system shown in Fig. 2 consisting of a spherical shell of NRM with the dielectric constant $\varepsilon_{-}(r)$ and $\mu_-(r)$ imbedded in a positive refractive material with $\varepsilon_+(r)$ and $\mu_+(r)$. First of all we will find the general solutions to the field equations with spatially inhomogeneous material parameters: $$\begin{aligned}
\nabla & \times & \mathbf{E} = i \omega\mu_0 \mu(\mathbf{r}) \mathbf{H} ,
~~~~~~~~~~
\nabla \times \mathbf{H} = i \omega\varepsilon_0 \varepsilon
(\mathbf{r}) \mathbf{E} \\
\nabla & \cdot & \mathbf{D} = 0 , ~~~~~~~~~~ \nabla \cdot \mathbf{H} = 0, \\
\mathbf{D} &=& \varepsilon(\mathbf{r}) \mathbf{E}, ~~~~~~~~~~ \mathbf{B} =
\mu(\mathbf{r})
\mathbf{H}.\end{aligned}$$
{width="8cm"}
Under these circumstances of spherical symmetry, it is sufficient to specify the quantities $(\mathbf{r}\cdot\mathbf{E})$ and $(\mathbf{r}\cdot\mathbf{H})$ which will constitute a full solution to the problem. Let us now look at the TM polarised modes $\mathbf{r}\cdot\mathbf{H} = 0$, implying that only the electric fields have a radial component $E_r$. Operating on Eqn. (1) by $\nabla$, we have $$\begin{aligned}
\nabla\times\nabla\times\mathbf{E} &=& i \omega\mu_0 \nabla\times \left[ \mu(
\mathbf{r}) \mathbf{H} \right], \nonumber \\
&=& \frac{\omega^2}{c^2} \mu(\mathbf{r}) \varepsilon(\mathbf{r}) \mathbf{E} +
i \omega \frac{\nabla \mu(\mathbf{r})}{\mu(\mathbf{r})} \times
\nabla \times \mathbf{E}.\end{aligned}$$ Using Eqns.(2) and (3) we have $$\nabla \cdot \mathbf{D} = \nabla \cdot \left[ \varepsilon(\mathbf{r})
\mathbf{E}\right] = \nabla\varepsilon(\mathbf{r}) \cdot \mathbf{E} +
\varepsilon(\mathbf{r})~ \nabla \cdot \mathbf{E} = 0,$$ and if we assume $\varepsilon(\mathbf{r}) = \varepsilon(r)$ and $\mu(
\mathbf{r}) = \mu(r)$, we have $$\nabla \cdot \mathbf{E} = -\frac{\varepsilon'(r)}{r \varepsilon(r)}
\mathbf{r} \cdot \mathbf{E}= -\frac{\varepsilon'(r)}{r \varepsilon(r)} (rE_r).$$ We note the following identities for later use: $$\nabla\times\nabla\times\mathbf{E} = \nabla (\nabla\cdot \mathbf{E}) -
\nabla^2 \mathbf{E},$$ $$\nabla^2(\mathbf{r}\cdot \mathbf{E}) = \mathbf{r}\cdot \nabla^2\mathbf{E}
+ 2 \nabla\cdot \mathbf{E},$$ and using Eqn. (6) we also note that $$\begin{aligned}
\mathbf{r}\cdot\nabla (\nabla\cdot \mathbf{E}) &=& \mathbf{r}\cdot\nabla
\left( -\frac{\varepsilon'(r)}{\varepsilon(r)} E_r \right), \nonumber \\
&=& -r \frac{\partial}{\partial r} \left( \frac{\varepsilon'(r)}{\varepsilon(r)}
E_r \right), \nonumber \\
&=& -\frac{\partial}{\partial r} \left( \frac{\varepsilon'(r)}{\varepsilon(r)}
(rE_r) \right) + \left( \frac{\varepsilon'(r)}{\varepsilon(r)} E_r \right).\end{aligned}$$ We now take a dot product of $\mathbf{r}$ with Eqn. (4), and use the Eqns. (6),(7), (8) and (9) to get an equation for $(rE_r)$ as: $$\nabla^2 (rE_r) + \frac{\partial}{\partial r} \left[ \frac{\varepsilon'(r)}
{ \varepsilon(r)}(rE_r) \right] + \frac{\varepsilon'(r)}{r \varepsilon(r)}
(rE_r) + \varepsilon(r) \mu(r) \frac{\omega^2}{c^2} (rE_r) = 0 .$$ This equation is separable and the spherical harmonics are a solution to the angular part. Hence the solution is $(rE_r) = U(r) Y_{lm}(\theta,\phi)$ where the radial part $U(r)$ satisfies $$\frac{1}{r^2} \frac{\partial}{\partial r} \left( r^2 \frac{\partial U}{\partial
r} \right) - \frac{l(l+1)}{r^2} U + \frac{\partial}{\partial r} \left[
\frac{\varepsilon'(r)} { \varepsilon(r)}U \right] +
\frac{\varepsilon'(r)}{r
\varepsilon(r)} U + \varepsilon(r) \mu(r) \frac{\omega^2}{c^2} U = 0.$$ If we choose $\varepsilon(r) = \alpha r^p$ and $\mu(r) = \beta r^q$, we can have a solution $U(r) \sim r^n$ and we get $$[n(n+1) - l(l+1) + p(n-1) + p] r^{n-2} + \alpha\beta\omega^2/c^2 r^{p+q+n}
= 0,$$ implying $p+q=-2$ and $$n_\pm=1/2 \left[ -(p+1) \pm \sqrt{(p+1)^2 + 4l(l+1)-4 \alpha\beta\omega^2/c^2}
\right].$$ Hence the general solution can be written as $$E_r(\mathbf{r}) = \sum_{l,m} \left[ n_+A_{lm} r^{n_+ -1} +
n_- B_{lm} r^{n_- -1}
\right] Y_{lm}(\theta,\phi),$$ and $$\mathbf{H}(\mathbf{r}) = \sum_{l,m}\left[ A_{lm} r^{n_+} + B_{lm} r^{n_-
} \right] \mathbf{X}_{lm}(\theta,\phi),$$ where the vector spherical harmonic $$\mathbf{X}_{lm}(\theta,\phi) \equiv \mathbf{L} Y_{lm}(\theta,\phi) = \frac{1}{i}
(\mathbf{r}\times \nabla) Y_{lm}(\theta,\phi) .$$ A similiar solution can be obtained for the TE modes with $\mathbf{r}\cdot
\mathbf{E} = 0$.
Now assuming an arbitrary source at $r=a_0$, we can now write down the electric fields of the TM modes in the different regions for the negative spherical shell of Fig. 2 as $$\begin{aligned}
\mathbf{E}^{(1)}(\mathbf{r}) &=& \sum_{l,m}\left[ n_+ A_{lm}^{(1)} r^{n_+ -1}
+ n_- B_{lm}^{(1)} r^{n_- -1} \right] Y_{lm}(\theta,\phi), ~~~~~~~~~~
a_0<r<a_1,\\
\mathbf{E}^{(2)}(\mathbf{r}) &=& \sum_{l,m}\left[ n_+ A_{lm}^{(2)} r^{n_+ -1}
+ n_- B_{lm}^{(2)} r^{n_- -1} \right] Y_{lm}(\theta,\phi), ~~~~~~~~~~
a_1<r<a_2,\\
\mathbf{E}^{(3)}(\mathbf{r}) &=& \sum_{l,m}\left[ n_+ A_{lm}^{(3)} r^{n_+ -1}
+ n_- B_{lm}^{(3)} r^{n_- -1} \right] Y_{lm}(\theta,\phi), ~~~~~~~~~~
a_2<r<\infty,\end{aligned}$$ and similiarly for the magnetic fields. Note that the $B_{lm}^{(1)}$ correspond to the field components of the source located at $r=a_0$. For causal solutions $A_{lm}^{(3)} = 0$. Now the tangential components of the magnetic fields and the normal components of the displacement fields have to be continuous across the interfaces. Under the conditions $p = -1$, $q =-1$, $\varepsilon_+(a_1) =
-\varepsilon_-(a_1)$ and $\varepsilon_+(a_2) = -\varepsilon_-(a_2)$, we have $$\begin{aligned}
A_{lm}^{(1)} &=& 0,\\
A_{lm}^{(2)} &=&
\left(\frac{1}{a_1^2}\right)^{\sqrt{l(l+1)-\alpha\beta\omega^2/c^2}}
B_{lm}^{(1)}, \\
B_{lm}^{(2)} &=& 0,\\
B_{lm}^{(3)} &=&
\left(\frac{a_2^2}{a_1^2}\right)^{\sqrt{l(l+1)-\alpha\beta\omega^2/c^2}}
B_{lm}^{(1)}.\end{aligned}$$ The lens-like property of the system becomes clear by writing the field outside the spherical shell as $$E_r^{(3)} = \frac{1}{r} \left[\frac{a_2^2}{a_1^2}
r\right]^{\sqrt{l(l+1)-\alpha\beta\omega^2/c^2}} B_{lm}^{(1)}
Y_{lm}(\theta,\phi).$$ Hence apart from a scaling factor of $1/r$, the fields on the sphere $r=a_3=(a_2^2/a_1^2)a_0$ are identical to the fields on the sphere $r =a_0$. There is also a spatial magnification in the image by a factor of $a_2^2/a_1^2$.
Let us note a couple of points about the above perfect lens solutions in the spherical geometry. First, for $r>a_3$, i.e. points outside the image surface the fields appear as if the source were located on the spherical image surface ($r = a_3$). However, this is not true for points $a_2<r<a_3$ within the image surface. Second, given that $\varepsilon_-(a_2) = -\varepsilon_+(a_2)$, we have the perfect lens solutions if and only if $n_+ = - n_-$ which implies that $p = -1$ in Eqn. (13). Although the solutions given by Eqn. (14) occur in any medium with $\varepsilon \mu \sim 1/r^2$, the perfect lens solutions only occurs for $\varepsilon \sim\mu \sim 1/r$. Here we have written down the solutions for the TM modes. The solutions for the TE modes can be similiarly obtained.
The spherical near-field lens
=============================
As it has been pointed out in the previous section, the ‘power’ solutions are good for any $\varepsilon(r) \sim r^p$ and $\mu(r) \sim r^q$ such that $p+q
=-2$. However the perfect lens solutions for the Maxwell’s equations result only for the single case of $p = q =-1$. In the quasi-static limit of $\omega \rightarrow 0$ and $l \gg |p|,~|q|$, we can relax this condition. In particular, by setting $\varepsilon(r) \sim 1/r^2$ and $\mu(r)=$ constant, we can have a perfect lens for the TM modes alone. Similiarly, we can have a perfect lens for the TE polarization by having $\mu \sim 1/r^2$ and $ \varepsilon =$ constant.
This extreme near-field limit is both important and valid for situations when all lengthscales in the problem are much smaller than a wavelength of the radiation. This becomes useful at frequencies where we can only generate media with either negative $\varepsilon$ and positive $\mu$, or, negative $\mu$ and positive $\varepsilon$. Examples are the silver slab lens at optical frequencies [@pendry_PRL00], the imeta-materials (Swiss rolls) used for MRI at radio-frequencies[@wiltshire]. Particularly at radio- and microwave frequencies, we currently can practically engineer the required meta-materials with spatially dispersive characteristics at the corresponding length scales. Further it also lifts the restriction that the system has to have a spatially dispersive material parameters even outside the spherical shell of NRM. In this section we will work in this extreme near field limit. Then it is sufficient to solve the Laplace equation and we present lens-like solutions to the Laplace equation below.
Consider the spherical shell in Fig.2 to be filled with a material with $\epsilon_2(r)\sim -C/r^2$ with the inner and outer regions filled with constant dielectrics of $\varepsilon_1$ and $\varepsilon_3$ respectively. Let $\mu =1 $ everywhere. Now place a charge $+q$ at the centre of the concentric spheres and a charge $-q$ at a distance $a_0$ from the centre inside region-1. We will consider the z-axis to be along the dipole axis and make use of the azimuthal symmetry here, although it is clear that our results do not depend on any such assumption of azimuthal symmetry. Thus all our charge and their images will now lie along the Z-axis.
Now we will calculate the potentials in the three regions that satisfies the Laplace equation and the continuity conditions at the interfaces. The potential in region-1($r<a_1$) can be calculated to be (using the azimuthal symmetry): $$V_1(\mathbf{r}) = \frac{-q}{4\pi\epsilon_1} \sum_{l=1}^{\infty} \left[
A_{1l} r^l P_l(\cos\theta) + \frac{a_0^l}{r^{l+1}} P_{l}(\cos\theta) \right]$$ Note that the second term in the above expansion arises due to the dipole within the sphere. It can be shown (see appendix-1), that the general form of the potential in region-2 ($a_1<r<a_2$), where the dielectric constant varies as $1/r^2$ is $$V_2(\mathbf{r}) = \frac{-q}{4\pi\epsilon_0} \sum_{l=1}^{\infty} \left[
A_{2l} r^{(l+1)} P_l(\cos\theta) + \frac{B_{2l}}{r^{l}} P_{l}(\cos\theta)
\right].$$ In region-3($r>a_2$), the potential is given by: $$V_3(\mathbf{r}) = \frac{-q}{4\pi\epsilon_3} \sum_{l=1}^{\infty} \left[
\frac{B_{3l}}{r^{l+1}} P_{l}(\cos\theta) \right].$$ Now we must match the potentials at the interfaces at $r=a_1$ and $r=a_2$ (Put $\epsilon_0 = 1$) to determine the $A$ and $B$ coefficients. The conditions of continuity of the potential and the normal component of $\vec{D}$ at the interfaces are $$\begin{aligned}
V_1(a_1) = V_2(a_1) &,&~~~~~~~~~~V_2(a_2) = V_3(a_2), \\
\epsilon_1 \frac{\partial V_1(a_1)}{\partial r} = \epsilon_2
\frac{\partial V_2(a_1)}{\partial r} &,& ~~~~~~~~~ \epsilon_2 \frac{\partial
V_2(a_2)}{\partial r} = \epsilon_3 \frac{\partial V_3(a_2)}{\partial r} \end{aligned}$$ We determine the coefficients from these conditions to be: $$\begin{aligned}
A_{1l} &=& \frac{(l+1)a_0^l \{[l\epsilon_2(a_2) - (l+1)
\epsilon_3][\epsilon_1+\epsilon_2(a_1)]
- [\epsilon_2(a_2)+\epsilon_3]
[(l+1)\epsilon_1-l\epsilon_2(a_1)] \frac{a_2^{2l+1}}{a_1^{2l+1}}\} }
{l(l+1)[\epsilon_1+\epsilon_2(a_1)][\epsilon_2(a_2)+\epsilon_3] a_2^{2l+1}+
[l\epsilon_1 - (l+1) \epsilon_2(a_1)][l \epsilon_2(a_2) - (l+1)\epsilon_3]
a_1^{2l+1}},\\
A_{2l} &=& \frac{(2l+1)[l \epsilon_2(a_2) - (l+1)\epsilon_3]a_0^l a_1^{-1}}
{l(l+1)[\epsilon_1+\epsilon_2(a_1)][\epsilon_2(a_2)+\epsilon_3] a_2^{2l+1}+
[l\epsilon_1 - (l+1) \epsilon_2(a_1)][l \epsilon_2(a_2) - (l+1)\epsilon_3]
a_1^{2l+1}}, \\
B_{2l} &=& \frac{(2l+1)(l+1) [\epsilon_2(a_2) + \epsilon_3] a_2^l a_0^l
a_1^{-(l+2)}}
{l(l+1)[\epsilon_1+\epsilon_2(a_1)][\epsilon_2(a_2)+\epsilon_3] a_2^{2l+1}+
[l\epsilon_1 - (l+1) \epsilon_2(a_1)][l \epsilon_2(a_2) - (l+1)\epsilon_3]
a_1^{2l+1}}, \\
B_{3l} &=& \frac{(2l+1)^2 \epsilon_3\epsilon_2(a_2) a_0^l a_2^{2(l+1)} a_1^{-1}}
{l(l+1)[\epsilon_1+\epsilon_2(a_1)][\epsilon_2(a_2)+\epsilon_3] a_2^{2l+1}+
[l\epsilon_1 - (l+1) \epsilon_2(a_1)][l \epsilon_2(a_2) - (l+1)\epsilon_3]
a_1^{2l+1}}. \end{aligned}$$ Under the perfect lens conditions $$\varepsilon_2(a_1) = -\varepsilon_1, ~~~~\mathrm{and}~~~~
\varepsilon_2(a_2) = -\varepsilon_3,$$ we have: $$\begin{aligned}
A_{1l} &=& 0, \\
A_{2l} &=& \frac{1}{\varepsilon_1} \frac{a_0^l}{a_1^{2(l+1)}},\\
B_{2l} &=& 0,\\
B_{3l} &=& \frac{\varepsilon_3}{\varepsilon_1} \left(\frac{a_2}{a_1}
\right)^{2(l+1)}
a_0^l. \end{aligned}$$ Hence the potential outside the spherical shell for $r>a_2$ is $$V_3(\vec{r}) = \frac{-q}{4\pi\varepsilon_3} \sum_{l=1}^\infty
\frac{\varepsilon_3}{\varepsilon_1} \left(\frac{a_2}{a_1} \right) ^{2(l+1)}
\frac{a_0^l}{r^{l+1}},$$ which is the potential of a dipole with the positive charge at the origin and the negative charge at $a_3$, where $$a_3 = \left( \frac{a_2}{a_1}\right) ^2 a_0,$$ and of strength $$q_2 = \frac{\varepsilon_3}{\varepsilon_1} \left(\frac{a_2}{a_1}\right)^2 q = q,$$ as $\varepsilon_3/\varepsilon_1 = (a_1/a_2)^2$. Thus, on one side of the image (the region $r > a_3$) the fields of a point charge located at $a_3$ are reproduced. However it should be pointed out that there is no physical charge in the image location and, the fields on the other side of the image (i.e. in the region $a_2< r < a_3$) do not converge to the fields of the object and cannot do so in the absence of a charge in the image. Further there is no change in the strength of the charge either. There is a magnification in the image formed by a factor of $(a_2/a_1)^2$.
Now let us consider the case of a point source placed at $a_3$ in the outer region. Again assuming the z-axis to pass through $a_3$, we can write the potentials in the three regions as $$\begin{aligned}
V_1(\mathbf{r})&=& \frac{+q}{4\pi\varepsilon_1}\sum_{l=0}^{\infty}
A_{1l}r^l P_l(\cos\theta)~~~~~~~~~~~~~~ \forall ~~~~~~r<a_1, \\
V_2(\mathbf{r})&=& \frac{+q}{4\pi}\sum_{l=0}^{\infty} \left[A_{2l}r^{l+1}+
\frac{B_{2l}}{r^l}\right] P_l(\cos\theta)~~~~~~~~~~~~~~ \forall ~~~~~~a_1<r<a_2, \\
V_3(\mathbf{r})&=& \frac{+q}{4\pi\varepsilon_3}\sum_{l=0}^{\infty}
\left[ \frac{r^l}{a_3^{l+1}} + \frac{B_{3l}}{r^{l+1}}\right] P_l(\cos\theta),
~~~~~~~~~~~~~~\forall ~~~~~~a_2<r<a_3, \end{aligned}$$ where the first term in $V_3(\vec{r})$ comes from the point source at $a_3$. Now applying the conditions of continuity at the interfaces, we can similiarly obtain for the coefficients as before. In the limiting case of $\varepsilon_2(a_1) = -\varepsilon_1$ and $\varepsilon_2(a_2) = -\varepsilon_3$, we have $$\begin{aligned}
A_{1l} &=& \frac{\varepsilon_1}{\varepsilon_3} \left(\frac{a_2}{a_1}\right)^{2l}
\frac{1}{a_3^{l+1}}, \\
A_{2l} &=& 0, \\
B_{2l} &=& \frac{1}{\varepsilon_3} \frac{a_2^{2l}}{a_3^{2l+1}}, \\
B_{3l} &=& 0.\end{aligned}$$ Hence the potential inside the inner sphere is $$V_1(\mathbf{r}) = \frac{q}{4\pi\varepsilon_1} \sum_{l=0}^{\infty}
\frac{\varepsilon_1}{\varepsilon_3} \left(\frac{a_2}{a_1}\right)^{2l}
\frac{r^l}{a_3^{l+1}} P_l(\cos\theta),$$ i.e., that of a point charge of strength $q_1 = (\varepsilon_1/\varepsilon_3)
(a_1/a_2)^2 = q$ at $a_0 = a_3 (a_1/a_2)^2$. As before, for the inner region of $r < a_0$, the system behaves as if there were a single charge of strength $q$ located at $r=a_0$. Thus the shell has a lens-like action. We note that there is a demagnification of $(a_1/a_2)^2$ in this case.
Similiarities to the 1-D slab lens
----------------------------------
Let us point out the similiarities to the planar slab lens. In both cases, the electromagnetic field grows in amplitude across the negative medium when the perfect lens conditions are satisfied at the interfaces: as an exponential($\exp [+k_x z]$) in the planar lens and as a power of the radial distance $r^l$ in the spherical lens. The decaying solution away from the source is absent in the negative medium in both cases. Further, when the perfect lens conditions are matched at both the interfaces, there is no reflected wave in both the planar slab as well as the spherical lens: i.e. the impedance matching is perfect as well. In addition this mapping preserves the strength of the charge.
The key differences, however, are the different dielectric constants on either sides of the spherical shell of the negative medium. This is a direct consequence of the spatial $1/r^2$ dependence of the negative dielectric constant which relates the two positive dielectric constants to be $\epsilon_1
= (a_1/a_2)^2 \epsilon_3$. But this need not be a particular restriction as we can use the ideas of the asymmetric lens to terminate the different positive media at some radii beyond[@sar_JMO02]. The net result is that the image can now be magnified (or demagnified) when the image of the charge (source) is projected out of (or into) the spherical shell, which is true in the 2D cylinderical lens as well [@pendry_JPC02].
Possibility of the asymmetric lens
----------------------------------
In the case of a planar slab, it was possible to have the perfect lens effect by satisfying the required conditions at any one interface - not necessarily at both interfaces[@sar_JMO02]. Particularly , in the limit of very large parallel wave vectors the lensing is indeed perfect, although the image intensity differed from the source by a constant factor. Similiarly let us now investigate the effects of having the perfect lens conditions in the case of the spherical lens at only one of the interfaces.
Let us consider first, the case of projecting out the image of a point source from inside the spherical shell to outside and enable the perfect lens conditions only at the outer interface $\epsilon_2(r=a_2) = -\epsilon_3$ and have an arbitrary $\epsilon_1$. Now the $A$ and $B$ coefficients come out to be $$\begin{aligned}
A_{1l} &=& \frac{a_0^l}{a_1^{2l+1}} \frac{(l+1)
[\varepsilon_1+\varepsilon_2(a_1)]}
{l\varepsilon_1 - (l+1)\varepsilon_2(a_1)} \\
A_{2l} &=& \frac{(2l+1) a_0^l}{l\varepsilon_1 -
(l+1)\varepsilon_2(a_1) a_1^{2l+2}} \\
B_{2l} &=& 0 \\
B_{3l} &=& \frac{(2l+1) \varepsilon_3 a_0^{l}}{l
\varepsilon_1(a_1/a_2)^2 +
(l+1)\varepsilon_3} \left( \frac{a_2}{a_1} \right)^{2l}\end{aligned}$$ Only the growing solution within the negative spherical shell remains. The coefficient of the decaying solution ($B_{2l}$) remains strictly zero. Thus, amplification of the decaying field at least is possible in this case as well. But there is a finite reflectivity in this case. However, the solution outside for $r>a_2$ is not the exact image field of the point source as the coefficient $B_{3l}$ has an extra dependence on $l$ through the dependence on the dielectric constants. Moreover, the process does not preserve the strength of the charge due to the different dielectric constants involved. This should be compared to the solution of the planar asymmetric slab lens where, at least in the electrostastic limit, the system behaved as a perfect lens. In this case, the system behaves as a spherical asymmetric perfect lens only in the limit of large $l \rightarrow \infty$. The solution outside the spherical shell is the same when we meet the perfect lens condition on the inner interface – just as in the case of the planar slab lens. However, the reflection coefficient is again non-zero, but different to the earlier case. In either case, the fields are largest at the interface where one meets the perfect lens conditions or the interface on which the surface plasmons are excited.
Effects of dissipation
----------------------
Media with negative real part of the dielectric constant are absorptive (as all metals are), and hence we can write the dielectric constant $\varepsilon(r) =
C/r^2 + i \varepsilon_i(r)$ (Note that $\varepsilon_i \sim 1/r^2$ as well for us to be able to write the solution in the following form). Consider the first case of projecting out the image of a dipole located within the spherical shell where the potential outside the shell is given by Equation (27) and $B_{3l}$ is given by Equation (33). When we have a dissipative negative medium and have the perfect lens conditions at the interfaces on the real parts of the dielectric constant alone, $\varepsilon_2(a_1) = -\varepsilon_1 + i \varepsilon_i (a_1)$ and $\varepsilon_2(a_2) = -\varepsilon_3+i\varepsilon_i (a_2)$. In parallel with the case of the planar lens, we note that the denominator of $B_{3l}$ consists of two terms, one containing a power of the (smaller) radius $a_1$ and the other containing a power of the (larger) radius $a_2$. Crucially the amplification of the evanescent fields depends on the possibility that the smaller power dominates by making the coefficient of the larger term as close as possible to zero. The presence of the imaginary part of the dielectric constant would not allow the coefficient to be zero and the image restoration is good only as long as the term containing $a_1$ dominates in the denominator of $B_{3l}$, i.e., $$l(l+1) \varepsilon_i(a_1)\varepsilon_i(a_2) a_2^{2l+1}
\ll [(2l+1)\varepsilon_1 -i(l+1)
\varepsilon_i(a_1)][-(2l+1)\varepsilon_3+i\varepsilon_i(a_2)]a_1^{2l+1}$$ Hence a useful estimate of the extent of image resolution can be obtained by noting the multipole $l$ for which the two terms in the denominator are approximately equal[@drs_APL03]. We obtain for this value: $$l_{\mathrm{max}} \simeq \frac{ \ln \left\{ 3\varepsilon_1 \varepsilon_3 /
[\varepsilon_i(a_1)\varepsilon_i(a_2)] \right\}}{2 \ln \left( a_2/a_1
\right) }.$$ Higher order multipoles are essentially unresolved in the image. We can similiarly obtain the same criterion by considering the second case of tranferring the image of a charge located outside the spherical shell into the inner region. Again, we can consider the effects of deviating from the perfect lens conditions on the real part of the dielectric constant as well and obtain a similiar limit for the image resolution due to those deviations.
Conclusions
===========
In conclusion, we have presented a spherical perfect lens which enables magnification of the near-field images. The perfect lens solution requires media with $\varepsilon(r) \sim 1/r$ and $\mu(r) \sim 1/r$ and the conditions $\varepsilon_-(a_{1,2}) = -\varepsilon_+(a_{1,2})$ and $\mu_-(a_{1,2}) =
-\mu_+(a_{1,2})$ at the interfaces of the spherical shell of the NRM. We have shown that in the quasi-static limit of small frequencies ($\omega
\rightarrow 0$) and high-order multipoles $l \gg |p|$, this condition can be relaxed and the two polarizations (TE and TM modes) decouple. Thus a shell with negative dielectric constant $\varepsilon_-(r) \sim -1/r^2$ with $\mu =
\mathrm{constant}$ can act as a near-field lens for the TM polarization while $\mu_-(r) \sim -1/r^2$ with $\varepsilon_-(r) = \mathrm{constant}$ acts as a near-field lens for the TE modes. We have shown that dissipation in the lens material, however, prevents good resolution of higher order multipoles. Thus while the near-field lenses work best for the higher order multipoles, dissipation cuts-off the higher order multipoles. Further the spherical lens works in the asymmetric mode only in the limit of high order multipoles. Thus one has to find an intermediate regime where dissipation does not wipe out the near-field image information and yet the meta-materials work. This is the design challenge involving these near-field lenses.
Appendix: Solution of the Laplace equation in a spatially varying medium {#appendix-solution-of-the-laplace-equation-in-a-spatially-varying-medium .unnumbered}
========================================================================
We have to solve the Maxwell’s equations in material media $$\nabla \cdot\mathbf{D} = 0,~~~~\Rightarrow~~~~~ \nabla(\varepsilon
\mathbf{E}) = 0$$ Using $\mathbf{E} = -\nabla V$, where $V(\mathbf{r})$ is the electrostatic potential we have: $$\varepsilon(\mathbf{r}) \nabla^2 V + \nabla \varepsilon(\mathbf{r})
\cdot \mathbf{\nabla}V = 0$$ If $\varepsilon(\mathbf{r})$ has only a radial dependence (as in our case $\sim 1/r^2$), then $\nabla \varepsilon(\mathbf{r}) =
\hat{r} (\partial \varepsilon / \partial r)$ and we can separate the solution as $V(\mathbf{r}) = \frac{U(r)}{r} Y_{lm}(\theta,\phi)$ where the $Y_{lm}$ is the spherical harmonic and the radial part $U(r)$ satisfies: $$\varepsilon(r) \frac{ d^2 U}{d r^2} - \frac{l(l+1)}{r^2} U +\frac{d
\varepsilon}{dr} \left[ \frac{d U}{d r} - \frac{U}{r} \right] = 0$$ To have a solution as a single power of $r$, the only choices possible for the dielectric constant are either $\varepsilon = C$, a constant – the usual case , or $\varepsilon = C/r^2$. In the latter case the solution is $U(r) \sim r^{l+2}$ or $U(r) \sim r^{-(l-1)}$. The full solution can then be written as $$V(\mathbf{r}) = \sum_{l=0}^{\infty} \left[ A_{lm} r^{l+1} + B_{lm}r^{-l}
\right] Y_{lm}(\theta,\phi).$$
[99]{} J.B. Pendry, Phys. Rev. Lett. **85**, 3966 (2000). V.G. Veselago, Usp. Fiz. Nauk [**92**]{}, 517 (1964) \[Sov. Phys. Uspekhi, **10**, 509 (1968).\] D.R. Smith, W.J. Padilla, D.C. Vier, S.C. Nemat-Nasser and S. Schultz, Phys. Rev. Lett., [**84**]{}, 4184 (2000). R.A. Shelby, D.R. Smith and S. Schultz, [*Science,*]{} [**292**]{}, 77 (2001). A. Grbic and G.V. Eleftheriades, J. Appl. Phys. [**92**]{}, 5930 (2002). J.B. Pendry, A.J. Holden, W.J. Stewart and I. Youngs, Phys. Rev. Lett., [**76**]{}, 4773 (1996); J.B. Pendry, A.J. Holden, D.J. Robbins and W.J. Steward, J. Phys.: Condens. Matter, [**10**]{}, 4785 (1998). J.B. Pendry, A.J. Holden, D.J. Robbins and W.J. Steward, IEEE trans. MTT, [**47**]{}, 2075 (1999). N. Garcia and M. Nieto-Vesperinas, Opt. Lett. [**27**]{}, 885 (2002). C.G. Parazzoli, R.B. Greegor, K. Li, B.E.C. Kotenbah and M. Tanielan, Phys. Rev. Lett. [**90**]{}, 107401 (2003). A.A. Houck, J.B. Brock, I.L. Chuang, Phys. Rev. Lett. [**90**]{}, 137401 (2003). A. Grbic and G.V. Eleftheriades, Appl. Phys. Lett. [**82**]{}, 1815 (2003). P.V. Parimi, W.T. Lu, P. Vodo, J. Sokoloff and S. Sridhar, cond-mat/0306109. H. Raether, [*Surface Plasmons*]{}, (Springer-Verlag, Berlin, 1988) D.R. Smith, D. Schurig, M. Rosenbluth, S. Schultz, S.A. Ramakrishna and J.B. Pendry, Appl. Phys. Lett. [**82**]{}, 1506 (2003). F.D.M. Haldane, cond-mat/0206420. G. Gomez-Santos, Phys. Rev. Lett., [**90**]{}, 077401 (2003). S.A. Ramakrishna, J.B. Pendry, D.R. Smith, D. Schurig and S. Schultz, J. Mod. Optics, [**49**]{},1747 (2002) X.S. Rao and C.N. Ong, Phys. Rev. B [**68**]{}, 113103 (2003). S. Foteinopolou, E.N. Economou and C.M. Soukoulis, Phys. Rev. Lett. [**90**]{} 107402 (2003). J.B. Pendry and S.A. Ramakrishna, J. Phys.: Condens. Matter [**14**]{}, 8463 (2002). J.T. Shen and P. Platzmann, Appl. Phys. Lett. [**90**]{}, 3286 (2002) Z. Ye, Phys. Rev. B [**67**]{}, 193106 (2003). N. Fang and X. Zhang, Appl. Phys. Lett. [**82**]{}, 161 (2003). S.A. Ramakrishna, J.B. Pendry, M.C.K. Wiltshire and W.J. Stewart, J. Mod. Optics [**50**]{} 1419 (2003). S. A. Ramakrishna and J.B. Pendry, Phys. Rev. B. [**67**]{}, 201101(R) (2003). V.V. Klimov, Opt. Commun. [**211**]{}, 183 (2002). J.B. Pendry, Opt. Express [**11**]{}, 755 (2003). J.B. Pendry and S.A. Ramakrishna, J. Phys.: Condens. Matter [**15**]{}, 6345 (2003) M.C.K. Wiltshire, J.B. Pendry, I.R. Young, D.J. Larkman, D.J. Gilderdale and J.V. Hajnal, Science [**291**]{}, 848 (2001); Opt. Express, [**11**]{}, 709 (2003).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'According to the mathematical classification of topological band structures, there exist a number of fascinating topological states in dimensions larger than three with exotic boundary phenomena and interesting topological responses. While these topological states are not accessible in condensed matter systems, recent works have shown that synthetic systems, such as photonic crystals or electric circuits, can realize higher-dimensional band structures. Here, we argue that the 4D spinless topological insulator, due to its symmetry properties, is particularly well suited to be implemented in these synthetic systems. We explicitly construct a 2D electric circuit lattice, whose resonance frequency spectrum simulate the 4D spinless topological insulator. We perform detailed numerical calculations of the circuit lattice and show that the resonance frequency spectrum exhibit pairs of 3D Weyl boundary states, a hallmark of the nontrivial topology. These pairs of 3D Weyl states with the same chirality are protected by classical time-reversal symmetry that squares to $+1$, which is inherent in the proposed circuit lattice. We also discuss how the simulated 4D topological band structure can be observed in experiments.'
author:
- Rui Yu
- 'Y. X. Zhao'
- 'Andreas P. Schnyder'
bibliography:
- 'refs.bib'
title: 4D spinless topological insulator in a periodic electric circuit
---
Introduction {#introduction .unnumbered}
============
With the great success of topological band theory in condensed matter physics [@hasan_RMP_2010; @BOOK_Bernevig_2013; @QXL_RMP_2011; @chiu_Classification_RMP2016; @Weyl_review_2017; @Burkov_Weyl_rev_2018], recent research has branched out to the study of topological bands in synthetic lattices, such as, photonic crystals [@lu_topological_2014; @Photonic_rechtsman_photonic_2013; @LuLing_yan_experimental_2018; @ozawa_TopologicalPhotonics_RMP2019], ultracold atomic gases [@ColdAtom_goldman_realistic_2010; @ColdAtom_sun_topological_2012; @ColdAtom_jotzu_HaldaneModel_2014; @ColdAtom_aidelsburger_ChernNumber_Hofstadter_2015; @goldman_TopologicalQuantumMatter_N2016; @ZhuSL_Rev], and electric circuit networks [@Topological_Circuit_PRX_2015; @Topological_Circuit_PRL_2015; @Topolectrical_Circuits_Weyl_Ronny_CoomP_2018; @luo_topological_2018; @circuit_Weyl_probing_2018; @Topolectrical_Circuits_ssh_PRB_2018; @Thomale_Chern_2019; @Topolectrical_Circuits_Ronny_NP_cornermodes_2018; @hadad_self_inducedTopo_ne_2018; @ezawa_higher_order_PRB2018; @Garcia_LC_PRB2019; @luo_nodal_nodate; @li_4Dboundary_2019; @Franz_PRB2019; @Prodan_PRApp_2019; @JieRen_PRB_2019; @liu_topologically_2019; @Ronny_circuitband_PRB2019]. These synthetic lattices have several advantages compared to their condensed matter counterparts. One is the ability to precisely control and manipulate the band structure, another is the possibility to create lattices in dimensions greater than three. The celebrated ten-fold classification of topological materials [@kitaev_PeriodicTableTopological_2009; @Schnyder-PRB-classification; @ryu_Topological_Table_NJP2010; @ZYX-PRL-2013; @ZYX-PRB-2014] predicts a number of interesting higher-dimensional topological states, including four-dimensional (4D) topological insulators [@Schnyder-PRB-classification; @QXL_PRB2008], 4D topological superconductors, and a 4D generalization of the integer quantum Hall effect [@zhang_4D_Science2001]. These 4D topological states exhibit many interesting phenomena, e.g., quantized nonlinear responses [@zhang_4D_Science2001; @froehlich_pedrini_4D_hall_2000; @QXL_PRB2008; @ryu_moore_PRB2012; @lohse_Exploring4DQuantum_Nature2018; @price_FourDimensionalQuantumHall_PRL2015], topological charge pumping, and in-gap boundary modes with protected level crossings [@kraus_FourDimensionalQuantumHall_PRL2013]. Unfortunately, these 4D states cannot be realized in condensed matter systems, which are limited to three spatial dimensions. However, recent technological advances in photonics and cold atomic gases have allowed to synthetically engineer the 4D integer quantum Hall effect, using, e.g., internal degrees of freedom as additional effective dimensions [@price_FourDimensionalQuantumHall_PRL2015; @price_4D_AI_Topo_2018; @ozawa_SyntheticDimensionsIntegrated_PRA2016; @lohse_Exploring4DQuantum_Nature2018; @petrides_SixdimensionalQuantumHall_nature2018; @zilberberg_PhotonicTopologicalBoundary_Nature2018; @kraus_FourDimensionalQuantumHall_PRL2013]. These experiments have revealed signatures of charge pumping and topological transport [@lohse_Exploring4DQuantum_Nature2018; @zilberberg_PhotonicTopologicalBoundary_Nature2018]. Apart from these works, there has been no other experimental investigation of the 4D integer quantum Hall effect, and likewise no other 4D topological state has yet been realized experimentally. Among the five 4D topological states of the ten-fold classification [@kitaev_PeriodicTableTopological_2009; @Schnyder-PRB-classification; @ryu_Topological_Table_NJP2010], the spinless topological insulator, belonging to symmetry class AI, is particularly intriguing. Its energy bands are characterized by a 4D topological invariant, namely the second Chern number, which has the distinguishing property of taking on only *even* integer values [@ryu_Topological_Table_NJP2010]. This invariant leads to topological transport responses in the 4D bulk [@ryu_moore_PRB2012] and to pairs of Weyl fermions of same chirality on the 3D boundary [@ZYX-PRL-2013; @ZYX-PRB-2014]. Hence, an experimental realization of the 4D spinless topological insulator could allow to simulate chiral lattice gauge theory of high-energy physics [@Jackiw-Chiral-domain; @KAPLAN1992342; @BALL19891].
Besides these interesting properties, the 4D spinless topological insulator has the advantage that it can be realized easily and in a robust manner in bosonic synthetic or classical systems, such as photonic lattices or periodic electric circuits. This is because such systems naturally exhibit a time-reversal symmetry that squares to $+1$, which is the protecting symmetry of the 4D spinless topological insulator in class AI. Hence, there is no need to introduce artificial gauge fields or to engineer fine-tuned intra-unit-cell degrees of freedom for the simulation of additional symmetries. The time-reversal symmetry also guarantees that the first Chern numbers vanish, such that the topological responses originate purely from the second Chern number.
Motivated by these considerations, we propose in this paper an experimental realization of the 4D spinless topological insulator in a periodic electric circuit composed of inductors (L), capacitors (C), and operational amplifiers. By using a mapping between circuit Laplacians and single-particle Hamiltonians, we explicitly construct an circuit lattice, whose resonance frequency spectrum is identical to a 4D spinless topological insulator in class AI. We perform detailed numerical simulations of the resonance frequency spectrum for various boundary conditions. For open boundary conditions we observe pairs of 3D Weyl cones that traverse a gap in the resonance frequency spectrum. Since the LC circuit lattice is non-dissipative, it has a built-in time-reversal symmetry of class AI, which leads to a strong and robust protection of the 3D Weyl boundary states. Even though the proposed circuit lattice realizes a 4D state, it can readily be implemented on a 2D circuit board or integrated-circuit wafer by projecting the 4D hyperlattice onto the 2D plane. The crossings of the projected lattice links can be avoided by using a bridge structure for the wiring. The predicted pairs of Weyl modes can be experimentally observed using frequency-dependent measurements.
4D spinless topological insulator {#d-spinless-topological-insulator .unnumbered}
=================================
We start by discussing a minimal model for the 4D topological insulator in class AI and its boundary Weyl modes. A minimal model can be constructed from a four-band Hamiltonian of the form, $$\H(\boldsymbol{k})=\sum_{a=0}^{5}f_{a}(\boldsymbol{k})\gamma_{a}.\label{eq:Hk_4D}$$ Here $f_a(\k)$ are real functions of the 4D quasi-momentum $\boldsymbol{k}=(k_{1},k_{2},k_{3},k_{4})$, [ $\gamma_0 = \bm{1}_{4 \times 4}$, and $\gamma_i$ (with $i=1,2,\cdots,5$) are five $4\times 4$ gamma matrices, which satisfy the Clifford algebra $\{\gamma_i,\gamma_j\}=2\delta_{ij}$ and act on the spinors $\Psi^\dagger=(\psi_a^\dagger,\psi_b^\dagger,\psi_c^\dagger,\psi_d^\dagger)$. For concreteness we choose the following representation for the gamma matrices: $\gamma_{1,2,3}=\tau_{1,2,3}\otimes\rho_{1}$, $\gamma_4=\tau_0\otimes\rho_2$, and $\gamma_5=\tau_0\otimes\rho_3$, with $\tau_\alpha$ and $\rho_\alpha$ two sets of the Pauli matrices.]{} Time-reversal symmetry acts on $\H(\k)$ as $\H^{*}(\k)=\H(-\k)$, which implies that $f_{0,1,3,5 }$ ($f_{2,4}$) are even (odd) functions of $\k$. With this condition, one possible choice for $f_i$ that yields a finite second Chern number is: $f_{0}(\k)=\epsilon-t\cos(k_{2}+k_{3})$, $f_{1}(\boldsymbol{k})=-t(1+\cos k_{1}+\cos k_{2})$, $f_{2}(\boldsymbol{k})= t(\sin k_{1}+\sin k_{2})$, $f_{3}(\boldsymbol{k})=-t(1+\cos k_{3}+\cos k_{4})$, $f_{4}(\boldsymbol{k})= t(\sin k_{3}+\sin k_{4})$, and $f_{5}(\boldsymbol{k})=m-t\cos(k_{2}+k_{3})$, similar to a previous model introduced in a general context [@price_4D_AI_Topo_2018]. Since the term $f_0(\k)$ only affects the global energy at each $\k$, rather than the topological property as indicated by $E(\boldsymbol{k})=f_{0}(\mathbf{k})\pm(\sum_{a=1}^{5}f_{a}^{2}(\boldsymbol{k}))^{1/2}$, we choose its form only for the convenience of the realization of the 4D topological electric circuit. The topology of the gapped 4D class AI system can be characterized by the second Chern number. For the Dirac model, the second Chern number can be nicely simplified as the winding number of $\hat{\boldsymbol{f}}=\boldsymbol{f}/|\boldsymbol{f}|$ from the 4D Brillouin zone (BZ) to the 4D unit sphere $S^{4}$ [@QXL_PRB2008] $$C_{2}=\frac{3}{8\pi^{3}}\int d^{4}k~\epsilon^{\mu\nu\lambda\rho\sigma}\hat{f}_{\mu}\partial_{k_{1}}\hat{f}_{\nu}\partial_{k_{2}}\hat{f}_{\lambda}\partial_{k_{3}}\hat{f}_{\rho}\partial_{k_{4}}\hat{f}_{\sigma},\label{eq:Ch2}$$ where $\epsilon^{\mu\nu\lambda\rho\sigma}$ is the rank-5 Levi-Civita symbol with $\mu,\nu,\lambda,\rho,\sigma=1,2,\cdots,5$, and repeated indices are summed over. Straightforward calculation gives that $C_{2}=-2$ if $-t/2<m<t$, and otherwise $C_{2}=0$ as shown in Fig.\[4D-model\] **a**, for which a detailed derivation can be found in the supplemental materials.
![[\[fig:1\] Topological properties of the 4D model Hamiltonian]{}. **a**, The second Chern number $C_{2}$ as a function of $m$ (in unit of $t$). For $-t/2<m<t$, $C_{2}=-2$. **b**, The band structures for a slab geometry confined in the $\boldsymbol{r}_{1}$ direction, with $m=0$. The $k$-line in the $(k_{2},k_{3},k_{4})$ space is chosen to cross the point $w_{1}=2\pi( 5/12, 1/3,-1/3)$ along the $k_{2}$ direction. The bulk part of the band structure (gray) are obtained by projecting the eigenvalues of the Hamiltonian (\[eq:Hk\_4D\]) with $k_{1}\in [0,2\pi]$. The boundary states (red lines) are the eigenstates of the boundary effective Hamiltonian (\[eq:H\_surface\_3D\]) in the range of $k_{2}\in2\pi(1/3,2/3)$, namely between the two vertical lines (blue dashed lines). **c** and **d**, The bulk band structures and the boundary Weyl states along the $k_{3}$ and $k_{4}$ directions, respectively. The boundary states appear in the whole range of $k_{3,4}\in[0,2\pi]$. The local band structure around the point $w_2$ is related to that around $w_1$ by time-reversal symmetry. \[4D-model\] ](fig_1){width="1.1\columnwidth"}
According to the general theory of bulk-boundary correspondence of topological insulators, a nontrivial second Chern number leads to boundary Weyl fermions. We consider a 3D boundary perpendicular to the $\boldsymbol{r}_{1}$-axis, putting the semi-infinite system in the region with $r_1>0$. For simplicity we set $m=0$ and the system is in the topologically nontrivial phase with $C_{2}=-2$. For the Dirac model the boundary effective Hamiltonian can be derived analytically as [@mong_edge_2011] $$\H_{s}(\tilde{\k})=f_{0}(\tilde{\k})\sigma_{0}-f_{3}(\tilde{\k})\sigma_{1}+f_{4}(\tilde{\k})\sigma_{2}+f_{5}(\tilde{\k})\sigma_{3}.\label{eq:H_surface_3D}$$ Here $\sigma_\alpha$ acts in the sub-lattices $c$ and $d$, $f_{0,3,4,5}$ are functions defined in , depending only on $\tilde{\k}\equiv(k_2,k_3,k_4)$, where $k_{2}\in(2\pi/3,4\pi/3)$ and $k_{3,4}\in[0,2\pi]$. From the boundary spectrum $E(\tilde{\k})=f_{0}\pm(\sum_{a=3}^{5}f_{a}^{2})^{1/2}$, it is easy to obtain that there are two Weyl points located at $w_{1,2}=\pm 2\pi( 5/12, 1/3,-1/3)$ in the 3D boundary BZ as shown in Fig.\[fig:1\] **c**-**d**. [As the Weyl points are located at generic momenta, they have anisotropic dispersion relations due to the lack of rotational symmetry. The $\gamma_0$ term in Eq. leads to the unwanted $\sigma_0$ term in Eq. , which tilts the boundary Weyl points. But, we have to make a trade off between the simplicity of the model and the magnitude of the term. Since the two Weyl points are related by time-reversal symmetry, they have the same chirality, right-handedness as shown in the supplementary materials. More details are given in the supplementary materials. Generically the boundary states decay exponentially towards the bulk, which will be confirmed by our simulation results in Fig.\[fig:Simu\_result\].]{}
{width="1\linewidth"}
Tight-binding model and circuit lattice {#tight-binding-model-and-circuit-lattice .unnumbered}
=======================================
We now proceed to address the realization of the above 4D topological states by constructing a realistic electric-circuit in a practical way. For this purpose, it is more convenient to write the model Hamiltonian in real space, that reads $H=\sum_{\alpha,\beta,i,s}t_{\alpha\beta}(R^s_{\alpha\beta})c_{\alpha}^{+}(i+R^s_{\alpha\beta})c_{\beta}(i)$, where $\alpha, \beta$ label nodes in each unit cell and $i$ labels the unit cells. $R_{\alpha\beta}^s$ are hopping vectors, which can be obtained by the inverse Fourier transform of , and are listed as: $R_{ad}^s= R_{bc}^s=(0,0,0,0),(1,0,0,0),(0,1,0,0)$ and $R^s_{ab}=-R^s_{cd}=(0,0,0,0),(0,0,1,0),(0,0,0,1)$ with $s=1,2,3$, respectively, and $R^s_{aa}= R^s_{cc}=(0,1,1,0),(0,-1,-1,0)$ with $s=1,2$, respectively. Here, each number in the parentheses is in the unit of the corresponding lattice constant for the 4D hypercube lattice. The hopping amplitudes are assumed to be $t_{ab}=t_{bc}=t_{ad}=t_{aa}=t_{cc}=-t_{cd}=-t$, where $t$ is a real constant so that time-reversal symmetry is preserved. Exchanging the order of the subscripts, the amplitudes $t_{\alpha\beta}$ are unchanged while the vectors $R_{\alpha\beta}$ are reversed. The above 4D tight-binding model can be mapped to a 4D circuit lattice as detailed in the supplementary materials. As the property of a circuit lattice depends only on the connection relations among its nodes, regardless of the shape of circuit lattice, one can project the 4D circuit lattice onto a 2D plane to obtain an equivalent 2D circuit lattice as shown in Fig.\[fig\_circuit\], preserving the property of the circuit. In more detail, the circuit in Fig.\[fig\_circuit\] is constructed by the following two steps. First, the sub-circuits in the $\boldsymbol{r}_{3}$-$\boldsymbol{r}_{4}$ planes, with the Born-von Karman periodic boundary conditions, are constructed as shown in Fig.\[fig\_circuit\] **a**. The nodes and lines in Fig.\[fig\_circuit\] **a** are detailed in Fig.\[fig\_circuit\] **b**-**d**. Then, arrange the sub-circuits on the $\boldsymbol{r}_{1}$ and $\boldsymbol{r}_{2}$ plane, and connect the nodes between sub-circuits with capacitors to realize the connections in the $\boldsymbol{r}_{1}$-$\boldsymbol{r}_{2}$ planes as shown in Fig.\[fig\_circuit\] **e**. By the two steps, we have constructed a 2D circuit that is genuinely equivalent to the 4D circuit, since the connections of nodes in the two circuits have a one-to-one correspondence. According to the Kirchhoff current law, it is easy to check that the current equations for the circuit in Fig.\[fig\_circuit\] **e** lead to a Hamiltonian with exactly the same form as Eq. . Now the functions $f_a(\boldsymbol{k})$ have the parameters concretely specified in terms of capacitance values as $t=C$, $m=(C_{a0}-C_{b0}+2C)/2$ and $\epsilon=(C_{a0}+C_{b0})/2+7C$ as detailed in the supplementary materials. If the capacitance values satisfy $C_{a0}+2C=C_{b0}$, namely $m=0$, the circuit is in a topologically nontrivial phase with the second Chern number $C_{2}=-2$.
{width="1\linewidth"}
Simulation results {#simulation-results .unnumbered}
==================
In order to extract the resonance frequency spectrum of the circuit lattice, we perform the time-domain transient simulations to obtain the voltage $v(t,\boldsymbol{R},\alpha)$ on each node as a function of time. Here, $\boldsymbol{R}$ is the unit cell label, $\alpha=a,b,c,d$ is the index for the nodes in each unit cell, and $t$ is the time. Taking periodical boundary conditions in $\boldsymbol{r}_{1,2,3,4}$ directions, respectively, and performing the Fourier transform, the voltage $v(\omega,\boldsymbol{k},\alpha)$ can be obtained in the momentum $\boldsymbol{k}$ and frequency $\omega$ space. The band-structure-like dispersions are obtained by plotting $|v(\omega,\boldsymbol{k},\alpha)|$. As introduced in the supplementary materials, the eigenvalue $\varepsilon$ of the tight-binding model corresponds to the resonance frequency $\omega$ of the circuit lattice, with the relation $\varepsilon=1/(\omega^2L)$. Therefore, the vertical axis in Fig.\[fig:Simu\_result\] are plotted as $1/(\omega^{2}L)$, so as to compare with the eigenvalues of the Hamiltonian (\[eq:Hk\_4D\]). In Fig.\[fig:Simu\_result\] **a** and **b**, it is easy to see that the simulation results (gray points) are in good agreement with the bulk band dispersions (red dashed line) obtained from the model Hamiltonian (\[eq:Hk\_4D\]).
Next we study the surface states by assuming open boundary conditions in the $\boldsymbol{r}_{1}$ direction and periodic boundary conditions in the$\boldsymbol{r}_{2,3,4}$ directions. The technical details are provided in the supplementary materials. The pulse voltage source is connected to the (1,1,1,1) cell on the bottom edge, and thereby the voltage $v(t,\boldsymbol{R},\alpha)$ are obtained for a slab geometry with nine layers in the $\boldsymbol{r}_{1}$ direction. We then carry out the Fourier transforms for $\boldsymbol{r}_{2,3,4}$ and $t$, which gives $v(\omega,R_{1},\tilde{\k},\alpha)$, where $R_{1}$ is the lattice index in the $\boldsymbol{r}_{1}$ direction and $\tilde{\k}=(k_2,k_3,k_4)$. The corresponding boundary band structures for the voltage intensity along a number of selected $k$ lines crossing one of two Weyl points, $w_{1}$, are listed in Fig.\[fig:Simu\_result\] **c**-**h**. We now briefly introduce these figures, while more information can be found in the figure caption. For all of them, the data from simulation and analytic solutions from the model Hamiltonian are plotted by gray dots and dashed red lines, respectively, for comparison. In Fig.\[fig:Simu\_result\] **c**, the data from simulation shows that the surface Weyl states appear in the gap of the band structure, in good agreement with the results obtained from the model Hamiltonian. To reveal the local nature of the topological boundary states, the intensity of voltage for the fifth layer to the top layers is depicted in Fig.\[fig:Simu\_result\] **d**, where the surface Weyl states disappear, because the pulse source on the bottom boundary cannot excite the Weyl surface states located on the top boundary. Furthermore, according to the model Hamiltonian, the boundary states are contributed only by the node-$c$ and $d$ components, and have vanishing $a$ and $b$ components, which is confirmed by comparing the simulation result in Fig.\[fig:Simu\_result\] **f** with that in Fig.\[fig:Simu\_result\] **e**. The intensity of voltage for the bottom four layers is clearly visible only for the component of the $c$ and $d$ nodes plotted in Fig.\[fig:Simu\_result\] **f**, while the component of the $a$ and $b$ nodes plotted in Fig.\[fig:Simu\_result\] **e** is too weak to be seen. Finally, the band structures of the Weyl states along the $k_{3}$ and $k_{4}$ directions are depicted in Fig.\[fig:Simu\_result\] **g** and **h**, respectively. The linear dispersion relations in the vicinity of the point $w_1$ w.r.t. all boundary momentum components $\tilde{\boldsymbol{k}}$ show that the point $w_1$ is indeed a Weyl point.
Conclusions {#conclusions .unnumbered}
===========
In summary, 4D topological states exhibit many interesting phenomena that are markedly different from lower-dimensional topological phases. Unfortunately, they cannot be realized in condensed-matter materials, which are limited to three spatial dimensions. In this article we have shown that periodic electric circuits, composed of inductors, capacitors, and operational amplifiers, provide a realistic and ideal platform to create higher-dimensional topological states in the laboratory. We have explicitly constructed an electric circuit lattice that realizes the 4D spinless topological insulator. By projecting onto two dimensions, this 4D circuit lattice can readily be implemented on a printed circuit board or an integrated-circuit wafer. In this way, the higher dimensions of the 4D circuit lattice are faithfully realized through long-ranged lattice connectivity, rather than by internal degrees of freedom. Furthermore, the circuit implementation of higher-dimensional topological states has the advantage of being highly controllable and easily reconfigurable. This allows, for example, to investigate topological phase transitions, non-Hermitian phenomena, and the effects of nonlinear couplings [@Garcia_LC_PRB2019]. Using detailed numerical simulations, we have shown that the resonance frequency spectrum of our circuit lattice exhibits a pair of 3D Weyl boundary states, which is the hallmark of nontrivial topology.
[ Our work opens up the possibility of realizing topological phases in arbitrarily high dimensions, for example the 5D topological Weyl state [@lian_5D_PRB2016], or the 6D chiral topological superconductors [@kitaev_PeriodicTableTopological_2009; @Schnyder-PRB-classification; @ryu_Topological_Table_NJP2010]. Even topological states on non-orientable surfaces of any dimension could be realized, such as, topological phases on Möbius strips [@Topological_Circuit_PRX_2015], Klein bottles, or real projective planes. Other interesting directions for future research concern the study of quantum effects and interactions in higher-dimensional topological states. The former could be simulated by use of periodic Josephson junction arrays [@tsomokos_nori_PRA_10; @cosmic_nakamura_PRB_18]. Moreover, the nonlinearity effect could be achieved by bringing the electronic device into a nonlinear region. We hope that our work will stimulate further investigations along these lines.]{}
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Sheng Chang and Dong Zhang for fruitful discussion. This work was supported by the National Key Research and Development Program of China (No.2017YFA0303402, No.2017YFA0304700), the National Natural Science Foundation of China (No.11874048, No.11674077), and the GRF (HKU 173057/17P) of Hong Kong. The numerical calculations in this work have been done on the supercomputing system in the Supercomputing Center of Wuhan University.
REFERENCES {#references .unnumbered}
==========
|
{
"pile_set_name": "ArXiv"
}
|
-1cm 16.5cm 22.5cm
[SHEP-98/06]{}\
\
\
Physics Department,\
University of Southampton,\
Southampton SO17 1BJ, United Kingdom\
\
We discuss the significance of the next-to-leading order term in the BFKL equation on the energy dependence of diffractive processes controlled by the perturbative QCD pomeron. It is shown that whereas the large negative corrections do indeed reduce the rate of growth of diffractive amplitudes with increasing energy, this reduction is considerably less than previously expected.
SHEP-98/06\
April 1998
In a recent paper Fadin and Lipatov [@FL] have presented, in compact form, the results of the higher order corrections to the BFKL [@bfkl; @BL] equation. The solution to this equation is the amplitude, $f(s,t,{\mathbf k, \, k^\prime})$, for the scattering of two gluons with transverse momenta ${\mathbf k}$ and ${\mathbf k^\prime}$ respectively, centre-of-mass energy $\sqrt{s}$ and square momentum transfer $t$, in the ‘diffractive region’ where $ s \gg |t|$. It may be written $$\frac{\partial}{\partial \ln s} f(s,t,{\mathbf k, \, k^\prime})
= \int d^2{\mathbf l} \, {\cal K}(t,{\mathbf k, \, l})
f(s,t,\mathbf{l, \, k^\prime}). \label{eq1}$$ We are interested here in the forward scattering amplitude required for total cross-sections or deep-inelastic structure functions and so we henceforth set $t=0$ and suppress it.
The solutions to Eq.(\[eq1\]) are found through the identification of a complete set of eigenfunctions, $\phi_\nu(\mathbf{k})$, of the kernel ${\cal K}$ and their corresponding eigenvalues, $\omega(\nu)$ $$\int d^2{\mathbf l} \, {\cal K}({\mathbf k,l})
\phi_\nu({\mathbf l})
= \omega(\nu) \phi_\nu({\mathbf l})
, \label{eq2}$$ where, up to order $\alpha_s^2$, $$\omega(\nu)=\bar{\alpha} \chi^{(1)}(\nu) + \bar{\alpha}^2 \chi^{(2)}(\nu),
\label{eq3}$$ with $\bar{\alpha}=3\alpha_s/\pi$. The eigenfunctions $\phi_\nu({\mathbf k})$ are given by $$\phi_\nu({\mathbf k})= \frac{1}{\sqrt{ {\mathbf k}^2}}
\left(\frac{{\mathbf k}^2}{\sqrt{\alpha_s({\mathbf k}^2)}}
\right)^{i \nu}
\label{eq4}$$ where we have taken the advice in ref. [@FL] and included in $\phi_\nu$ a factor which breaks the conformal invariance in such a way that it has no effect on the leading order eigenvalue but guarantees an eigenvalue to next order which is a (real) even function of $\nu$.
It was pointed out in ref. [@FL] that the next order correction is large and of opposite sign, leading to a considerable reduction in the high energy (low-$x$) dependence of processes mediated by this amplitude ( the ‘perturbative pomeron’). Indeed if we look at $\nu=0$ we find [^1] $$\chi(0)=2 .77 \bar{\alpha} - 18.34 \bar{\alpha}^2, \label{eq5}$$ which changes sign for the small value of 0.16 for $\alpha$. The immediate conclusion is that even at HERA, where the running coupling is sufficiently small for a perturbative expansion to be reliable, the first two terms in the calculation of $\chi$ are insufficient. This may well be the case. However, in the absence (presumably indefinitely) of further terms in the expansion of $\chi$, one has no option but to assume that the expansion calculated so far represents a valid approximation and it is interesting to examine in more detail the consequences of the next order corrections for the $s$ dependence of diffractive amplitudes.
In Fig. 1 we have plotted the functions $\chi^{(1)}$ (solid line) and $\chi^{(2)}$ (dotted line) against $\nu$ in the region $0 < \nu < 2$. We observe that the higher order correction is largest for $\nu=0$ but that its magnitude diminishes rapidly as $\nu$ increases. Moreover, unlike $\chi^{(1)}$, $\chi^{(2)}$ has a turning point at $\nu \approx 0.6$. This is consistent with the asymptotic values, which are of the same sign, namely $$\lim_{\nu \to \infty} \chi^{(1)}(\nu) = -2 \ln \nu$$ $$\lim_{\nu \to \infty} \chi^{(2)}(\nu) = - \left( \frac{22}{3}-
4\frac{n_f}{9} \right) \ln^2(\nu)$$
Fig.1 shows us that it is [*not*]{} the $\nu=0$ component of the function $f(s,{\mathbf k, \, k^\prime})$ that dominates (except for exceedingly small values of $\alpha_s$). Indeed we see that the second order term has a positive maximum for $\nu \approx 0.6$, so that as $\alpha_s$ increases the intercept of the QCD pomeron also increases. There is no maximum value as there is for the $\nu=0$ component.
The maximum value of $\omega(\nu)$, which determines the high energy behaviour of the pomeron amplitude occurs at a value of $\nu$ which depends on the coupling constant $\alpha_s$. In Fig. 2 we have plotted both this maximum value, $\omega_{\max}$ and the value $\nu_{max}$ at which this occurs for values of $\alpha_s$ up to 0.5, which is the largest value for which one might reasonably expect to be able to rely on a perturbative calculation. The increasing value of $\omega_{\max}$ with $\alpha_s$ is unfortunate from the point of view of those who might have hoped that the maximum value of 0.1 for $\omega(0)$, which is very little greater than the intercept of the ‘soft’ pomeron [@DL], might have contained a clue to the reconciliation of the soft and hard pomerons from the higher order corrections. On the other hand it must be noted that since the larger values of $\omega_{max}$ refer to larger values of $\nu$, the components which dominate at large $s$ will have an oscillatory behaviour in transverse momenta and will therefore be somewhat suppressed when the pomeron amplitude is convoluted with the impact factors that determine the coupling of the pomeron to the scattering hadrons [@BL]. The upshot of this is that one may have to go to extremely large energies before these large $\omega$ components dominate. Such a scenario was proposed in ref.[@cdl].
More importantly, the large $s$ behaviour of the perturbative pomeron is not suppressed by the higher order corrections as much as has initially been anticipated. The steep change of $\chi^{(2)}(\nu)$ near $\nu=0$ shown in Fig. 1 means that for all but extremely small values of $\alpha_s$, an expansion of $\omega$ up to quadratic order in $\nu$ followed by a saddle point integration is not an appropriate technique for finding the solution to Eq.(\[eq1\]) up to ${\cal O}(1/\ln^2s)$. Instead we must expand up to fourth order in $\nu$ since it is this fourth order term that is negative. We find $$\chi^{(1)}(\nu)= 4 \ln 2 - 14 \zeta(3) \nu^2 \ + \ 62 \zeta(5) \nu^4 \
+ \ \cdots ,
\label{eq6}$$ where $$\zeta(n) \ = \ \sum_{i=1}^\infty \frac{1}{i^n},$$ is the Riemann zeta function, and $$\begin{aligned}
\chi^{(2)}(\nu)&=& -h_0-\frac{27}{128}\pi^3-\frac{11}{2} \zeta(3)
+\frac{67}{9} \ln(2) -\frac{1}{3}\pi^2 \ln(2) -\frac{22}{3} \ln^2(2) \nonumber
\\ & & \hspace*{1cm}
-n_f \left[ \frac{11}{3456}\pi^3 +\frac{10}{27} \ln (2) -\frac{4}{9} \ln^2 (2)
\right] \nonumber \\ & &
+\left( -h_2+\frac{199}{768} \pi^5 -\frac{1}{128} \pi^3 +186 \zeta(5)
-\frac{469}{18} \zeta(3) +\frac{7}{6} \pi^2 +\frac{154}{3} \zeta(3) \ln(2)
\right. \nonumber \\ & & \hspace*{1cm} \left.
+n_f \left[ \frac{55}{20736} \pi^5 -\frac{1}{3456} \pi^3
+\frac{35}{27} \zeta(3)-\frac{28}{9} \zeta(3) \ln (2) \right] \right)
\nu^2 \nonumber \\ & & \hspace*{-70pt}
+\left(-h_4- \frac{933}{5120} \pi^7 +\frac{5}{768} \pi^5 +\frac{1}{128} \pi^3
-1905 \zeta(7) + \zeta(5) \left(
\frac{2077}{18} -\frac{31}{6} \pi^2
-\frac{682}{3} \ln(2) \right) -\frac{539}{6} \zeta(3)^2
\right.
\nonumber \\ & & \hspace*{-2cm}\left.
+n_f \left[ -\frac{671}{414720} \pi^7 + \frac{5}{20736}\pi^5+\frac{1}{3456}
\pi^3- \zeta(5) \left( \frac{155}{27} -\frac{124}{9} \ln (2) \right)
+\frac{49}{9}\zeta(3)^2 \right] \right) \nu^4 \ + \ \cdots ,
\label{eq7} \end{aligned}$$ where $$h(\nu)\ =\ 2 \int_0^1 \frac{dx}{\sqrt{x}(1+x)} \cos(\nu\ln x)
{\mathrm Li}_2(1-x)
\ = \ h_0 +h_2 \nu^2 + h_4 \nu^4 \ + \ \cdots ,$$ ${\mathrm Li}_2$ being the dilogarithm function.
Numerical integration gives $$h_0=4.115, \ \ h_2=-24.92, \ \ h_4= 87.47,$$ so that if we write $$\omega(\nu) \ = \omega_0 \ + \ a \nu^2 \ - \ b \nu^4 \ + \ \cdots,$$ then for $\alpha_s=0.15$ and three flavours we obtain $$\omega_0=0.021, \ \ a=4.19, \ \ b=47.4$$ The solution to Eq.(\[eq1\]) up to ${\cal O}(1/\ln^2 s)$ is then obtained by expanding around the saddle point at $\nu^2=a/2b$ and gives $$f(s,{\mathbf k, k^\prime }) \sim
\frac{1}{\sqrt{{\mathbf k}^2 {\mathbf k^\prime}^2}}
s^{(\omega_0+a^2/4b)} \frac{1}{\sqrt{ a \ln s}}
\exp(\frac{3b}{4a^2\ln s})
\cos\left( \sqrt{\frac{a}{2b}}\left(1-\frac{3b}{4a^2\ln s}\right)\ln r\right),
\label{eq10}$$ where $$r \ = \ \frac{\sqrt{\alpha_s({\mathbf k^\prime}^2)} {\mathbf k}^2}
{\sqrt{\alpha_s({\mathbf k}^2)} {\mathbf k^\prime}^2}.$$ We can see immediately from Eq.(\[eq10\]) that the power behaviour of $s$ has an exponent of $0.12$ for $\alpha_s=0.15$ rather than $0.02$ which one obtains from a consideration of $\omega(0)$. These two behaviours are plotted for against $s$ for the case ${\mathbf k}^2={\mathbf k^\prime}^2$. $s$ is normalized by a typical square transverse momentum ${\mathbf k}^2$ ( in deep inelastic scattering $1/x$ plays the role of $s/\mathbf{k}^2$). The solid line is the plot of Eq.(\[eq10\]), whereas the dotted line is the $\omega_0$ only behaviour. These have been normalized so that they are equal at $s/{\mathbf k}^2=10^3$. We note that at $s/{\mathbf k}^2=10^5$, which represents the smallest $x$ values that one might currently expect at HERA, the difference between the two are not very significant. However the two curves are significantly different if we go up to much larger energies corresponding to rapidity gaps which one might hope to achieve in future hadron colliders such as LHC.
The exponent of $0.1$ for this value of $\alpha_s$ is slightly below the value of $\omega_{max}$ as can be seen from Fig. 2, and the value of $\nu$ at which this maximum value occurs is a little higher than the position of the saddle point, $\nu_{saddle}=0.21$. For larger values of the coupling one does not expect this saddle point approximation to be reliable. Here one expects to sample larger values of $\nu$ for which the quartic approximation to $\omega$ begins to break down. Rather we use Fig. 2 to read off the the corresponding $\omega_{max}$ and allow for the fact that the dominant value of $\omega$ is expected to be slightly below this value. Thus for $\alpha_s \approx 0.22$ we expect an exponent of 0.2. This is indeed what is found by a direct numerical solution of Eq.(\[eq1\]) with $\alpha_s$ taken to be 0.22, shown in Fig. 4. This corresponds to a pomeron intercept of 1.2, which is consistent with the value quoted by the H1 collaboration [@H1] from studies of the diffractive proton structure function [^2] at $Q^2=18 \ \mathrm{GeV}^2$ (for which $\alpha_s \approx 0.22$).
To summarize we have shown that although the order $\alpha_s^2$ correction to the eigenvalues of the BFKL kernel do indeed significantly suppress the growth of the diffractive amplitude with increasing $s$, this suppression is much less than that which is obtain from a consideration of the $\nu=0$ components. For $\alpha_s \approx 0.22$ we predict a pomeron intercept of 1.2, which is consistent with experimental data on diffractive deep-inelastic scattering.
[9]{}
V.S. Fadin and L.N. Lipatov, “BFKL Pomeron in the Next-to-Leading Approximation” [**hep-ph/9802290**]{} V.S. Fadin, E.A. Kurayev and L.N. Lipatov, [*Sov. Phys. JETP*]{} [**44**]{} 443 (1976); [*ibid.*]{} [**45**]{} 199 (1977) A. Donnachie and P.V. Landshoff, [*Nucl. Phys.*]{} [**B244**]{} 322 (1984); [*Phys. Lett.*]{} [**B296**]{} 227 (1992) Y.Y. Balitski and L.N. Lipatov, [*Sov. J. Nucl.Phys.*]{} [**28**]{} 822 (1978) J.R. Cudell, A.Donnachie and P.V. Landshoff, [*Nucl. Phys.*]{} [**B482**]{} 241 (1996) C. Adloff et. al., [*Z. Phys.*]{} [**C76**]{} 613 (1997) J. Breitweg et. al., [*Eur. Phys. Journal*]{} [**C1**]{} 81 (1998)
[^1]: The second order term in $\chi$ has a mild dependence on the number of active flavours. We set this number to three.
[^2]: On the other hand the Zeus collaboration [@zeus] favours a smaller value of the intercept consistent with the soft pomeron.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Yongheng Zhao [^1]\
University of Padova Tolga Birdal\
Stanford University Jan Eric Lenssen\
TU Dortmund Emanuele Menegatti\
University of Padova Leonidas Guibas\
Stanford University Federico Tombari\
Google
title: Quaternion Equivariant Capsule Networks for 3D Point Clouds
---
[^1]: Equal contribution.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the second fundamental form of the Prym map lifts the second gaussian map $\mu_A$ of the Prym-canonical bundle. We prove, by degeneration to binary curves, that $\mu_A$ is surjective for the general point $[C,A]$ of ${\mathcal R}_g$ for $g\geq 20$.'
address:
- 'Dipartimento di Matematica, Università di Milano, via Saldini 50, I-20133, Milano, Italy '
- ' Dipartimento di Matematica, Università di Pavia, via Ferrata 1, I-27100 Pavia, Italy '
author:
- Elisabetta Colombo
- Paola Frediani
title: 'Prym map and second gaussian map for Prym-canonical line bundles'
---
[^1]
Introduction
============
Similarly to the period map $P_g:{\mathcal M}_g \rightarrow
{\mathcal A}_g$, the Prym map $Pr_g: {\mathcal R}_g \rightarrow
{\mathcal A}_{g-1}$ provides a way to link the geometry of moduli spaces of curves to the geometry of moduli spaces of principally polarized abelian varieties. Recall that ${\mathcal R}_g$ denotes the moduli space which parametrizes isomorphism classes of pairs $[C, A]$, where $C$ is a smooth curve of genus $g$ and $A \in
Pic^0(C)[2] - \{\mathcal O_C\}$ is a torsion point of order 2, or equivalently isomorphism classes of unramified double coverings $\pi: \tilde{C}
\rightarrow C$. The Prym map associates to a point $[(C,A)] \in
{\mathcal R}_g$ the isomorphism class of the connected component of zero, $P(C,A)$ of the kernel of the norm map $Nm_{\pi}:J{\tilde
C} \rightarrow JC$, with its principal polarization. Both the period map and the Prym map have been extensively studied since a long time, but also recently there have been important developements on the birational geometry of ${\mathcal R}_g$ ([@fl]).
In this paper we focus on the study of the second fundamental form of the Prym map analogously to what it has been done for the second fundamental form of the period map and its link with the second gaussian map. In fact in [@cpt] it is shown that the second fundamental form of the period map lifts the second gaussian map of the canonical line bundle, as stated in an unpublished paper by Green and Griffiths (cf. [@green]). With this geometrical motivation, in [@cf2] we investigated curvature properties of ${\mathcal M}_g$ endowed with the Siegel metric. In fact, we computed the holomorphic sectional curvature of ${\mathcal M}_g$ along the tangent directions given by the Schiffer variations in terms of the second gaussian map. This also suggested that the second gaussian map itself could give interesting information on the geometry of the curves, hence its rank properties have been investigated in a series of papers (see [@ccm], [@cf1], [@cf3], [@cfp]).
Here we first generalize the lifting result of [@cpt] to the Prym map, namely at the point $[C,A] \in{\mathcal R}_g $ we have the following commutative diagram
$$\xymatrix{
& {\mathcal N^{*}_{{\mathcal R}^0_g/{\mathcal A}_{g-1},[C,A]}} \cong I_2(K_C \otimes A)\ar[d]^{-\frac{1}{2 \pi i} \mu_A} \ar[r]^{II } & S^2 \Omega^{1}_{{\mathcal R}_g^0, [C,A]} \cong S^2H^0(K_C^{\otimes 2}) \ar[dl]^{\bf{m}}\\
&H^0(K_C^{\otimes 4})
& }$$ where $I_2(K_C \otimes A)$ is the kernel of the multiplication map $S^2H^0(K_C \otimes A) \rightarrow
H^0(K_C^{\otimes 2})$, $II$ is the second fundamental form of the Prym map, $m$ is the multiplication map and $\mu_A$ is the second gaussian map associated to the Prym canonical bundle $K_C \otimes A$. This also allows us to generalize the results of [@cf2] on the holomorphic sectional curvature of ${\mathcal R}_g$ with the Siegel metric induced by ${\mathcal
A}_{g-1}$ via the Prym map.
In the second part of the paper we concentrate on the study of the second gaussian map $\mu_A$, $A \in Pic^0(C)[2]$ non trivial. The main result is the proof of the surjectivity of $\mu_A$ for the general curve $[C,A] \in {\mathcal R}_g$ of genus $g \geq 20$, generalizing analogous results on the surjectivity of the second gaussian map of the canonical line bundle for the general curve in ${\mathcal
M}_g$ for $g \geq 18$. For the canonical line bundle this surjectivity for general curves of high genus was proved in [@cf3] using curves on K3 surfaces, then the sharp result for genus $\geq 18$ has been shown in [@ccm] using degeneration to binary curves, i.e. stable curves which are the union of two rational curves meeting transversally at $g+1$ points. Here we generalize these degeneration techniques to prove the surjectivity of second gaussian maps $\mu_A$, for the Prym-canonical bundles $K_C \otimes A$.
In particular, this shows that the locus of curves $[C,A] \in
{\mathcal R}_g$ ($g \geq 20$) for which the map $\mu_A$ is not surjective is a proper subscheme of ${\mathcal R}_g$ and one observes that for $g =20$ it is an effective divisor in ${\mathcal R}_{20}$ of which we compute the cohomology class both in ${\mathcal R}_{20}$ and in a partial compactification $\tilde{{\mathcal R}}_{20}$ following computations developed in [@fl].
The paper is organized as follows: in Section 2 we describe the second fundamental form, we prove that it is a lifting of the second gaussian map $\mu_A$ and we compute the holomorphic sectional curvature along the Shiffer variations. In Section 3 we construct the Prym-canonical binary curves that we use for the degeneration. In Section 4 we explicitly describe the ideal of the quadrics containing the Prym-canonical binary curve. In Section 5 we prove, by induction on the genus, the surjectivity of $\mu_A$. In Section 6 we compute the cohomology class in $\tilde{{\mathcal
R}}_{20}$ of the degeneracy locus of the second gaussian map. Finally in the Appendix we list the Maple scripts used in the computations.
Finally we observe that the results and the techniques of section 2 are of different nature from the rest of the paper, which can be read separately, once one has looked at the definition of the gaussian maps in subsection 2.2.
We also notice that in the proof of the surjectivity of $\mu_A$, in particular in sections 4 and 5, we follow the lines of the proof of the surjectivity of the 2nd gaussian map of the canonical line bundle given in [@ccm].
The 2nd fundamental form and the 2nd gaussian map
=================================================
The second fundamental form of the Prym map
-------------------------------------------
We start by recalling the definition of the Prym map $$Pr: {\mathcal R}_g \rightarrow {\mathcal A}_{g-1},$$ which associates to a point $[(C,A)] \in {\mathcal R}_g$ its Prym variety $P(C,A)$ with its principal polarization. If $\pi: \tilde{C} \rightarrow C$ is the unramified double covering associated to the pair $(C,A)$, the Prym variety of the double covering is the principally polarized abelian variety of dimension $g-1$ defined as the connected component of zero of the kernel of the norm map $Nm_{\pi}:J{\tilde C} \rightarrow JC$, $$P(C,A) = Ker(Nm_{\pi})^0 \subset J{\tilde C}.$$ We recall that the Prym map is generically an embedding for $g \geq 7$ ([@fs], [@ka]). Hence there exists an open set ${\mathcal R}^0_g \subset {\mathcal R_g}$ where $Pr$ is an embedding and such that there exists the universal family $f: {\mathcal X} \rightarrow {\mathcal R}^0_g$. If $b \in {\mathcal R}^0_g$, we have $f^{-1}(b) = (C_b, A_b)$ where $C_b$ is a smooth irreducible curve of genus $g$ and $A_b \in Pic^0(C_b)[2]$ is a line bundle of order $2$ on $C_b$. Denote by ${\mathcal P} \in Pic({\mathcal X})$ the corresponding Prym bundle and by ${\mathcal F}^{Pr}:= f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g} \otimes {\mathcal P})$. Observe that ${\mathcal F}^{Pr}$ is the pullback of the Hodge bundle on ${\mathcal A}_{g-1}$ to ${\mathcal R}^0_g$. More precisely, if $\psi: {\mathcal Pr}({\mathcal X}) \to {\mathcal R}^0_g$ is the universal family of Prym varieties, so $\psi^{-1}(b)=P(C_b,A_b)$ is the Prym variety associated to the pair $(C_b, A_b)$, then $ {\mathcal F}^{Pr}$ is the Hodge bundle ${\mathcal H}^{1,0} \subset R^1 \psi_*{{\mathbb C}}$ of the family $\psi$.
On the local system $R^1 \psi_*{{\mathbb C}}$ we have the flat Gauss-Manin connection $\nabla^{GM}$, and a non degenerate bilinear form defined as follows. At the point $P(C,A)$, the fiber of $R^1 \psi_*{{\mathbb C}}$ is isomorphic to the vector space $H^1(\tilde{C}, {{\mathbb C}})^{-}$, where $\tilde{C} \stackrel{\pi} \rightarrow C$ is the double covering associated to $(C,A)$ and $H^1(\tilde{C}, {{\mathbb C}})^{-}$ is the anti-invariant part of the cohomology under the covering involution on $\tilde{C}$. If $[\omega_1], [\omega_2] \in H^1(\tilde{C}, {{\mathbb C}})^{-}$, we have the following non degenerate bilinear form $\langle [\omega_1],[\omega_2] \rangle = i \int_{\tilde C} \omega_1 \wedge \overline{\omega_2}$ and the Gauss-Manin connection $\nabla^{GM}$ is compatible with it, so it induces a metric connection ${\nabla}^{1,0}$ on ${\mathcal H}^{1,0}$, hence a connection on ${\mathcal F}^{Pr}$ (still denoted by ${\nabla}^{1,0}$) and on its second symmetric power, $S^2 {\mathcal
F}^{Pr}$. Observe that this metric on $S^2 {\mathcal
F}^{Pr}$ is the pullback via the Prym map of the metric on ${\mathcal A}_{g-1}$ induced by the unique (up to scalar) $Sp(2g-2,{\mathbb R})$-invariant metric on the Siegel space $H_{g-1}$. So we will call this metric the Siegel metric.
Consider the tangent bundle exact sequence of the Prym map $$\label{tangent}
0 \rightarrow T_{{\mathcal R}^0_g} \rightarrow T_{{{\mathcal A}_{g-1}}_{|{\mathcal R}^0_g}} \rightarrow {\mathcal N_{{\mathcal R}^0_g/{\mathcal A}_{g-1}}} \rightarrow 0$$ Its dual becomes $$\label{I2}
0 \rightarrow {\mathcal I}_2 \stackrel{i} \rightarrow S^2f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g} \otimes {\mathcal P}) \stackrel {m}\rightarrow f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g}^{\otimes 2}) \rightarrow 0$$ where $m$ is fibrewise the multiplication map and we denote by ${\mathcal I}_2$ the conormal bundle $ {\mathcal N^{*}_{{\mathcal R}^0_g/{\mathcal A}_{g-1}}}$. Recall that the second fundamental form of the exact sequence is defined as follows $$II: {\mathcal I}_2 \rightarrow f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g}^{\otimes 2}) \otimes \Omega^1_{{\mathcal R}^0_g}, \ \
II(s) = m(\nabla(i(s))),$$ where $\nabla$ is the metric connection on $S^2f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g} \otimes {\mathcal P})=S^2 {\mathcal F}^{Pr}$ defined above. At the point $(C,A) \in {\mathcal R}^0_g$ the exact sequence becomes $$0 \rightarrow I_2(K_C\otimes A) \rightarrow S^2H^0(K_C \otimes A) \stackrel {m}\rightarrow H^0(K_C^{\otimes 2})
\rightarrow 0.$$ Hence, if we identify $T_{{\mathcal R}^0_g,b_0}
\stackrel{\cong}\rightarrow H^{1}(T_C)$ via the Kodaira-Spencer map of the family ${\mathcal X} \stackrel{f} \rightarrow {\mathcal
R}^0_g$, the second fundamental form $II$ at $[C]$ can be seen as a map $II:I_2(K_C\otimes A)\rightarrow H^0(2K_C)\otimes
H^0(2K_C)$.
Gaussian maps
-------------
Let $Y$ be a smooth complex projective variety and let $\Delta_Y\subset
Y\times Y$ be the diagonal. Let $L$ and $M$ be line bundles on $Y$. For a non-negative integer $k$, the *k-th Gaussian map* associated to these data is given by restriction to the diagonal $$\label{gaussian1}\gamma^k_{L,M}:H^0(Y\times Y,I^k_{\Delta_Y}\otimes
L\boxtimes M )\rightarrow
H^0(Y\times Y,I^k_{\Delta_Y}\otimes
L\boxtimes M \otimes \OO_{\Delta_Y})\cong
H^0(Y,S^k\Omega_Y^1\otimes L\otimes M).$$ The exact sequence $$\label{Ik} 0 \rightarrow I^{k+1}_{\Delta_Y} \rightarrow
I^k_{\Delta_Y} \rightarrow S^k\Omega^1_Y \rightarrow 0,$$ twisted by $L\boxtimes M$, shows that the domain of the $k$-th gaussian map is the kernel of the previous one: $$\gamma^k_{L,M}:
ker \gamma^{k-1}_{L,M}\rightarrow H^0(Y,S^k\Omega_Y^1\otimes L\otimes
M).$$
In this paper, we will exclusively deal with the second gaussian map for curves $C$, assuming also that $L =M$.
The map $\gamma^0_L$ is the multiplication map of global sections $$H^0(C,L)\otimes
H^0(C,L)\rightarrow H^0(C,L^{\otimes 2})$$ which obviously vanishes identically on $\wedge^2 H^0(L)$. Consequently, $H^0(C
\times C, I_{\Delta_C}\otimes L\boxtimes L)$ decomposes as $\wedge^2 H^0(L)\oplus I_2(L)$, where $I_2(L)$ is the kernel of $S^2H^0(C,L)\rightarrow H^0(C,L^{\otimes 2})$. Since $\gamma^1_L$ vanishes on symmetric tensors, one writes $$\gamma^1_L:\wedge^2H^0(L)\rightarrow H^0(K_C\otimes
L^{\otimes 2}).$$ Again, $H^0(C\times C, I^2_{\Delta_C}\otimes L\boxtimes L)$ decomposes as the sum of $I_2(L)$ and the kernel of $\gamma^1_L$. Since $\gamma_L^2$ vanishes identically on skew-symmetric tensors, one usually writes $$\gamma^2_L:I_2(L)\rightarrow H^0(K_C^{\otimes 2}\otimes L^{\otimes 2})$$ Assume now that the line bundle $L$ is $K_C \otimes A$, with $A \in Pic^0(C)[2]$, and denote by $$\label{muA}
\mu_A:= \gamma^2_{K_C \otimes A}: I_2(K_C \otimes A) \rightarrow H^0(K_C^{\otimes 4})$$ the second gaussian map. It is useful to provide also a local description of it. Fix a basis $\{\omega_i\}$ of $H^0(K_C\otimes A)$ and write it in a local coordinate $z$ as $\omega_i = f_i(z)dz \otimes l$, where $l$ is a local generator of the line bundle $A$. For a quadric $Q \in I_2(K_C \otimes A)$ we have $Q = \sum_{i,j} a_{ij} \omega_i \otimes \omega_j $, where $a_{ij} = a_{ji}$ and $\sum_{i,j} a_{ij} f_i f_j \equiv 0$, hence we have $\sum_{i,j} a_{ij} f'_i f_j \equiv 0$. The local expression of $\mu_A(Q)$ is $$\mu_A(Q) =\sum_{i,j} a_{ij} f''_i f_j (dz)^4 = - \sum_{i,j} a_{ij} f'_i f'_j (dz)^4.$$
The maps $\mu_A$ glue together to give a map of vector bundles on ${\mathcal R}^0_g,$ $$\label{mu}
\mu: {\mathcal I}_2 \rightarrow f_*( (\omega_{{\mathcal X}/{\mathcal R}^0_g}\otimes {\mathcal P})^{\otimes 2}\otimes {\omega_{{\mathcal X}/{\mathcal R}^0_g}}^{\otimes 2})\cong f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g}^{\otimes 4}),$$ where ${\mathcal I}_2 $ is as in .
The theorem
-----------
In this subsection we show that the second fundamental form $II$ of the Prym map is a lifting of the second Gaussian map $\mu_A$ as it happens for the second fundamental form of the period map and the second Gaussian map of the canonical line bundle (see Theorems 2.1 and 4.5 of [@cpt]).
To this purpose, recall that given a holomorphic line bundle $A$ of degree zero on a curve $C$, there exists a unique (up to constant) hermitian metric $H$ on $A$ and a unique connection $D_H$ on $A$ which is compatible both with the holomorphic structure and with the metric and which is flat (see e.g. [@gh]). If moreover $A^{\otimes 2} = \OO_C$ and we denote by $\pi: \tilde{C} \rightarrow C$ the associated unramified double covering, we can take an atlas $\{(U_{\alpha}, s_{\alpha})\}$ of $A$ such that the sections $s_{\alpha}$ have values in $\tilde{C}$, hence the cocycle $g_{\alpha,\beta} = s_{\alpha}/s_{\beta}$ has values in $\{\pm1\}$, so it induces a flat structure on $A$ and a compatible flat hermitian metric on $A$, which is then equal to $H$ up to scalar (see [@ko]).
So we can write $D_H = D'_H +
\overline{\partial}$, where $D'_H$ is the $(1,0)$ component. Such a pair $(A,H)$ is also called a harmonic line bundle and we have the following properties (see [@sim]):
- The Kähler identities.
- The associated harmonic decomposition $${\mathcal A}^{\bullet}(A) = {\mathcal H}^{\bullet}(A) \oplus im(D_H) \oplus im(D^*_H)= {\mathcal H}^{\bullet}(A) \oplus
im(\overline{\partial}) \oplus im(\overline{\partial}^*),$$ where ${\mathcal H}(A)$ is the kernel of the laplacian operator $\Delta
= D_H D^*_H + D^*_H D_H =
2(\overline{\partial}\overline{\partial}^*+
\overline{\partial}^*\overline{\partial})$.
- The principle of two types $$ker(D'_H) \cap ker(\overline{\partial}) \cap (im(D'_H) + im(\overline{\partial})) = im(D'_H \overline{\partial}).$$
\[II\] The diagram $$\xymatrix{
& I_2(K_C \otimes A)\ar[d]^{-\frac{1}{2 \pi i}\mu_A} \ar[r]^{II \ \ } & \ \ S^2 H^0(K_C^{\otimes 2}) \ar[dl]^{\bf{m}}\\
&H^0(K_C^{\otimes 4})
& }$$ is commutative.
The proof follows the lines of the proof of Th.2.1 of [@cpt]. First of all we take $v\in H^1(C,T_C)$ and we compute $II(Q)(v)$ for every $Q\in I_2(K_C \otimes A)$. Using the Kodaira Spencer map $k$ we can assume that $v=k(\frac{\partial}{\partial t})$, where $t$ is the local coordinate of the unit disc $\Delta=
\{|t|<1\}$ parametrizing a one dimensional deformation ${\mathcal X} \stackrel{f}
\rightarrow \Delta$ where $(C,A) = f^{-1}(0)$. Take $Y$ a ${\mathcal C}^{\infty}$ lifting of the holomorphic vector field $\frac{\partial}{\partial t}$ on $\Delta$, so we have a ${\mathcal
C}^{\infty}$ trivialization $\tau: \Delta \times (C,A) \rightarrow
{\mathcal X}$, $\tau(t,x) := \Phi_{tY}(1)$, where $\Phi_Y(t)$ is the flow of the vector field $Y$. Then $\theta:=
\overline{\partial} Y_{|(C,A)}$ is a closed form in $A^{0,1}(T_C)$ such that $ [\theta]=v\in H^1(C,T_C)$. Denote by $(C_t, A_t)$ the fibre of $f$ over $t$, where $A_t$ is a holomorphic line bundle in $Pic^0(C_t)[2]$ endowed with the flat structure induced by the double covering $\pi_t: \tilde{C_t} \rightarrow C_t$. We denote by $H_t$ the flat hermitian metric and by ${D_{H_t}} = D'_{H_t}+ \overline{\partial}_t$ the flat Chern connection.
Let $\omega(t)$ be a section of ${\mathcal F}^{Pr}$, hence $\forall t \in \Delta$, $\omega(t) \in H^0(K_{C_t} \otimes A_t)
\cong H^{1,0}(A_t)$. Denote by $\tau_t: C \rightarrow C_t$ and by by $\sigma_t: \tilde{C} \rightarrow \tilde{C_t}$ the diffeomorphisms induced by $\tau$, where $\tilde{C}$ and $\tilde{C_t}$ are the unramified double coverings induced by $A$ and by $A_t$. We have the following commutative diagram: $$\xymatrix{
&\tilde{C}\ar[d]^{\pi } \ar[r]^{\sigma_t} & \tilde{C_t} \ar[d]^{\pi_t}\\
&C \ar[r]^{\tau_t}& C_t
& }$$ so we have an induced map by pullback $\tau_t^{*}: A^1(A_t)
\rightarrow A^1(A)$, and since $\omega_t \in A^{1,0}(A_t)$ is $
D_{H_t}$-closed, then also $\tau_t^{*}(\omega_t)$ is $D_H$-closed because we have $\tau_t^{*}(D_{H_t}) = D_H$. In fact by the commutativity of the diagram one immediately sees that the flat structure on $A_t$ given by the covering $\pi_t: \tilde{C_t} \rightarrow C_t$ induces by pullback the flat structure on $A$ given by the covering $\pi: \tilde{C} \rightarrow C$.
So $\tau_t^{*}(\omega_t)$ is $D_H$-closed, hence it has a power series expansion at $t=0$ $$\tau_t^{*}(\omega_t)= \omega + (\alpha + D_H h)t +o(t),$$ where $\omega := \omega(0)$, $\alpha \in A^1(A)$ is harmonic and $h$ is a ${\mathcal C}^{\infty}$ section of $A$ (by the harmonic decomposition for $D_H$). So we have $\nabla^{GM}_{\frac{\partial}{\partial t}}[\omega(t)]_{t=0} =
[\alpha]$, $\theta \cdot \omega = \alpha^{0,1} +
\overline{\partial}h$, so $k(\frac{\partial}{\partial t}) \cdot
[\omega] = [ \alpha^{0,1}]$, where $ \alpha^{0,1}$ is the $(0,1)$ component of $\alpha$.
Now assume that $\{\omega_i\}_{i=1,...,g-1}$ is a basis of $H^0(K_C \otimes A) $. Take a quadric $Q \in I_2(K_C \otimes A)$, $Q = \sum_{i,j} a_{i,j} \omega_i \otimes \omega_j$, with $a_{i,j}
= a_{j,i}$, then $\forall i$ we have $\nabla^{1,0}_{\frac{\partial}{\partial t}}[\omega_i(t)]_{t=0} =
[{\alpha_i}^{1,0}]$, so if $\tilde{Q}(t) = \sum_{i,j} a_{i,j}(t)
\omega_i(t) \otimes \omega_j(t)$ is a section of ${\mathcal I}_2$ such that $\tilde{Q}(0) = Q$, we have $II(Q)(v) = m (
\nabla_{\frac{\partial}{\partial t}} \tilde{Q}_{|t=0}) =
\sum_{i,j} a'_{i,j}(0) \omega_i \omega_j + 2 \sum_{i,j} a_{i,j}
\alpha^{1,0}_i \omega_j$. Since $\sum_{i,j} a_{i,j}(t) \omega_i(t)
\omega_j(t) \equiv 0$, also its derivative with respect to $t$ at $t=0$ must be zero, i.e. $2\sum_{i,j} a_{i,j}(\alpha_i + D_H h_i)
\omega_j + \sum_{i,j} a'_{i,j}(0) \omega_i \omega_j \equiv 0$, and if we take the $(1,0)$ part we have $2\sum_{i,j}
a_{i,j}(\alpha^{1,0}_i + D'_H h_i) \omega_j + \sum_{i,j} a'_{i,j}
\omega_i \omega_j \equiv 0,$ so $II(Q)(v) = -2 \sum_{i,j} a_{i,j}
\omega_j D'_H h_i. $
Now we observe that $ \sum_{i,j} a_{i,j} \omega_j D'_H
h_i=\rho(Q)(v)$ where $\rho$ is the map defined in theorem 4.4. of [@cpt]. So we conclude by theorem 4.5 of [@cpt] that asserts that $\rho$ is a lifting of $\mu_A$.
Let $\xi_P \in H^1(T_C)$ be a Schiffer variation at a point $P \in C$. Then we have $$\mu_A(Q)(P) = -\frac{1}{2\pi i} II(Q)(\xi_p \odot \xi_P) = -\frac{1}{2\pi i}\xi_P(II(Q)(\xi_P)).$$
Recall that given a point $P \in C$, a Schiffer variation $\xi_P \in H^1(T_C)$ is a generator of the image of the coboundary map $H^0(T_C(P)_{|P}) \rightarrow H^1(T_C)$. Given a local coordinate in a neighborhood of $P$, under the Dolbeault isomorphism $H^1(T_X) \cong H^{0,1}(T_X)$, $\xi_P$ is represented by the form $\theta_P = \frac{1}{z-z(P)} \overline{\partial}b_P \otimes
\frac{\partial}{\partial z},$ where $b_P$ is a bump function around $P$. Notice that if we choose $b_P$ to be one in a neighborhood of $P$, $\xi_P$ depends only on the choice of $z$. The choice of the local coordinate also allows us to see the evaluation $val_P$ in $P$ as an element of $H^0(K_C^{\otimes 4})^{*}$ and it holds: $m^*(val_P) = \frac{1}{(2 \pi i)^2} \xi_P \odot \xi_P$ (see [@cpt] p. 139). In fact, if $\{\lambda_i \}$ is a basis of $H^0(K_C^{\otimes 2})$, such that locally $\lambda_i = g_i(z) (dz)^2$, we have $(\xi_P \odot \xi_P)(\lambda_i \odot \lambda_j) = \xi_P(\lambda_i) \xi_P(\lambda_j)$ and by Serre duality $$\label{conto}
\xi_P(\lambda_i) = \int_{C} g_i(z) dz \wedge \frac{\overline{\partial} b_P}{z-z(P)} = -\int_C d \left(\frac{b_Pg_i(z) }{z-z(P)}dz \right) = \int_{\Gamma} \frac{g_i(z)}{z - z(P)} dz=(2\pi i) g_i(P),$$ where $\Gamma$ is a small circle around $P$. Hence by Theorem we have $\mu_A(Q)(P) = -\frac{1}{2\pi i} II(Q)(\xi_p \odot \xi_P) = -\frac{1}{2\pi i}\xi_P(II(Q)(\xi_P))$.
Curvature
---------
In this subsection we would like to give an explicit formula for the holomorphic sectional curvature of the Siegel metric on ${\mathcal
R}_g$ along the tangent directions given by the Schiffer variations. This formula is analogous to the formula of the holomorphic sectional curvature of the Siegel metric on ${\mathcal
M}_g$ induced by the Period map, given in Cor.3.8 [@cf2].[^2]
Assume that $\{Q_i\}$ is an orthonormal basis of $I_2(K_C\otimes
A)$, $\{\omega_i\}$ an orthonormal basis of of $H^0(K_C\otimes A)$ and choose a local coordinate $z$ at $P$ and a local generator $a$ of $A$ such that locally $\omega_i=f_i(z)dz\otimes a$.
\[curv\] The holomorphic sectional curvature $H$ of $T_{{\mathcal R}^0_g}$ at $[C,A]\in {\mathcal R}^0_g$ computed at the tangent vector $\xi_P$ given by a Schiffer variation in $P$ is given by: $$H(\xi_P)=-1-\frac{1}{16\alpha_{P}^4 \pi^2}\sum_i|\mu_A(Q_i)(P)|^2$$ where $\alpha_{P}=\sum_i|f_i(P)|^2$.
Also the proof follows the lines of [@cf2]. We start recalling that $$H(\xi_P) =\frac{\langle R(\xi_P), \xi_P\rangle(\xi_P, \overline{\xi_P}) }{{\langle \xi_P, \xi_P \rangle}^2},$$ and by the Gauss formula we have $\langle R(\xi_P),
\xi_{P}\rangle(\xi_P, \overline{\xi_P}) = \langle
\tilde{R}(\xi_P), \xi_{P}\rangle (\xi_P,
\overline{\xi_P}) - \langle\sigma(\xi_P),
\sigma(\xi_{P})\rangle(\xi_P, \overline{\xi_P}),$ where $R$ is the curvature of the Siegel metric, $\tilde{R}$ is the curvature of ${\mathcal A}_{g-1}$ and $\sigma$ is the second fundamental form of ${\mathcal R}_g^0$ in ${\mathcal A}_{g-1}$ (see the tangent bundle exact sequence ).
First of all observe that the holomorphic sectional curvature of ${\mathcal A}_{g-1}$ along the Schiffer variations is equal to $-1$ (see the argument below corollary 3.8 of [@cf2]). In fact a Schiffer variation $\xi_P$, seen as a symmetric homomorphism $H^0(K_C \otimes A) \rightarrow H^0(K_C \otimes A)^* \cong H^1(A)$ has rank 1, since its kernel is $H^0(K_C \otimes A(-P))$. To compute $\langle\sigma(\xi_P), \sigma(\xi_{P})\rangle(\xi_P, \overline{\xi_P})$, recall that $\sigma(\xi_P)(Q) = II(Q)(\xi_P)$, for all $Q \in I_2(K_C\otimes A)$, hence $\sigma(\xi_P) = \sum_i II(Q_i)(\xi_P) \otimes Q^*_i$, where $\{Q_i\}$ is an orthonormal basis of $I_2(K_C\otimes A)$. Hence $\langle\sigma(\xi_P),
\sigma(\xi_{P})\rangle = \sum_i \langle II(Q_i)(\xi_P) , II(Q_i)(\xi_P) \rangle$. Now recall that the Schiffer variations at $3g-3$ general points of $C$ give a basis of $H^1(T_C)$, so we can write $II(Q_i)(\xi_P) = \sum_S \xi_S(II(Q_i)(\xi_P)) \xi_S^*$, therefore $ \langle\sigma(\xi_P), \sigma(\xi_{P})\rangle(\xi_P, \overline{\xi_P}) = \sum_i (\xi_P(II(Q_i)(\xi_P))\overline{( \xi_P(II(Q_i)(\xi_P))} .$ By \[schiffer\] we have $\sum_i (\xi_P(II(Q_i)(\xi_P))\overline{( \xi_P(II(Q_i)(\xi_P) )}= 4 \pi^2 \sum_i|\mu_A(Q_i)(P)|^2$. Now it remains to show that $\langle \xi_P, \xi_P \rangle = 8 \pi^2 \alpha_P^2$ as in Lemma 2.2 of [@cf2]. To do this we write $\xi_P$ as an element of $S^2(H^0(K_C \otimes A)^*)$ as $\xi_P = \sum_{i,j} \xi_P(\omega_i)(\omega_j)(\omega_j^* \odot \omega_i^*)$, where $\{\omega_i^*\}$ is the dual basis of the orthonormal basis $\{\omega_i\}$. One computes $\xi_P(\omega_i)(\omega_j) = \xi_P(\omega_i \omega_j) = 2 \pi i f_i(P) f_j(P)$ by , then the proof follows exactly as in lemma 2.2 of [@cf2].
Prym-canonical binary curves
============================
Strategy of the proof of surjectivity
-------------------------------------
The rest of the paper is devoted to the proof of the surjectivity of the 2nd gaussian map $\mu_A$ for the general point $[C,A] \in
{\mathcal R}_g$. We will do it by degeneration to binary curves following the method used in [@ccm] for the second gaussian map of the canonical line bundle. We recall that ${\mathcal R}_g$ admits a suitable compactification $\overline{{\mathcal R}}_g$, which is isomorphic to the coarse moduli space of the stack ${\bf{R}}_g$ of Beauville admissible double covers ([@b], [@acv]) and to the coarse moduli space of the stack of Prym curves ([@bcf]).
Consider the partial compactification $\tilde{\mathcal R}_g$ of ${\mathcal R}_g$ introduced in [@fl]. Denote by $f: {\mathcal X} \rightarrow \tilde{{\bf{R}}}_g$ the universal family and by ${\mathcal P} \in Pic({\mathcal X})$ the corresponding Prym bundle as in [@fl] 1.1. The map of vector bundles over ${\mathcal R}^0_g$, $\mu: {\mathcal I}_2 \rightarrow f_*( (\omega_{{\mathcal X}/{\mathcal R}^0_g}\otimes {\mathcal P})^{\otimes 2}\otimes {\omega_{{\mathcal X}/{\mathcal R}^0_g}}^{\otimes 2})\cong f_*(\omega_{{\mathcal X}/{\mathcal R}^0_g}^{\otimes 4})$ defined in , extends to a map $$\label{mutilde}
\tilde{\mu}: \tilde{{\mathcal I}}_2 \rightarrow f_*( (\omega_f\otimes {\mathcal P})^{\otimes 2}\otimes S^2({\Omega^1_f}))\cong f_*(\omega_f^{\otimes 4}\otimes {\mathcal P}^{\otimes 2}\otimes {\mathcal I}_Z^{\otimes 2}),$$ where $\tilde{{\mathcal I}}_2 $ is the kernel of the multiplication map $ S^2f_*( \omega_f \otimes {\mathcal P})
\rightarrow f_{*}(\omega_f^{\otimes 2} \otimes {\mathcal
P}^{\otimes 2}),$ and $Z$ is the locus of nodes of fibres of $f$, so $\Omega^1_f \cong \omega_f
\otimes {\mathcal I}_Z$.
If $[C,A] \in $ is a point in $\tilde{\mathcal R}_g$, the local expression of $$\label{gauss}
\mu_A: I_2(\omega_C \otimes A) \rightarrow H^0((\omega_C \otimes A)^{\otimes 2} \otimes S^2({\Omega^1_C}))$$ is as follows. Let $\{\omega_i\}$ be a basis of $H^0(\omega_C\otimes A)$ and write it in a local coordinate as $\omega_i = f_i(z)\xi \otimes l$, where $\xi$ and $l$ are local generators of the line bundles $\omega_C$, respectively $A$. For a quadric $Q = \sum_{i,j} a_{ij} \omega_i \otimes \omega_j \in I_2(\omega_C \otimes A)$, $\mu_A(Q)$ is locally defined as $$\label{mulocal}
\mu_A(Q) = - \sum_{i,j} a_{ij} (df_i) (df_j) \xi^{\otimes2} \otimes l^{\otimes 2}.$$ To prove by semicontinuity the surjectivity of $\mu_A$ for the general point in ${\mathcal R}_g$ in the following we will exhibit a Prym-canonical binary curve $(C,A)$ for which $\mu_A$ is surjective.
Construction of Prym-canonical binary curves
--------------------------------------------
Recall that a binary curve of genus $g$ is a stable curve consisting of two rational components $C_j$, $j=1,2$ meeting transversally at $g+1$ points. Moreover one can check that if $A\in Pic^0(C)$ then $H^0(C,\omega_C \otimes A)$ has dimension $g-1$ and the restriction of $\omega_C \otimes A$ to the component $C_j$ is $K_{C_j}(D_j)$ where $D_j$ is the divisor of nodes on $C_j$. Since $K_{C_j}(D_j)\cong \OO_{{{\mathbb P}}^{1}}(g-1)$ we observe that the components are embedded by a linear subsystem of $\OO_{{{\mathbb P}}^{1}}(g-1)$, hence they are projections from a point of rational normal curves in ${{\mathbb P}}^{g-1}$. Viceversa, let us take 2 rational curves embedded in ${{\mathbb P}}^{g-2}$ by non complete linear systems of degree $g-1$ intersecting transversally at $g+1$ points. Then their union $C$ is a binary curve of genus $g$ embedded either by a linear subsystem of $\omega_C$ or by a complete linear system $|\omega_C \otimes A|$, where $A\in Pic^0(C)$ is nontrivial (see e.g. [@capo], Lemma 10). In this section we will construct a binary curve $C$ embedded in ${{\mathbb P}}^{g-2}$ by a linear system $|\omega_C \otimes A|$ with $A^{\otimes 2}\cong \OO_C$, and $A$ is non trivial.
Assuming that the first $g-1$ nodes, $P_1,...,P_{g-1}$ are in general position, up to projective transformations we will take $P_i=(0,...,0,1,0,...0)$ with 1 at the $i$-th place. Then we can assume that $C_j$ is the image of the map $$\label{phi}
\begin{gathered}\phi_j:{{\mathbb P}}^1 \rightarrow {{\mathbb P}}^{g-2}, \
j=1,2\\
\phi_j(t,u):= [M_j(t,u)\frac{({\delta}_{1,j}t-c_{1,j}u)}{(t-a_{1,j}u)},
..., M_j(t,u)\frac{({\delta}_{g-1,j}t-c_{g-1,j}u)}{(t-a_{g-1,j}u)}]
\end{gathered}$$ with $M_j(t,u):= \prod_{r=1}^{g-1} (t-a_{r,j}u)$, $j=1,2$ and $\phi_j([a_{l,j},1]) = P_l$, $l=1,...,g-1$.
We will also impose that the remaining two nodes $P_g:=[t_1,...,t_{g-1}]$ and $P_{g+1}:=[s_1,...,s_{g-1}]$ are the images of $[0,1]$ and $[1,0]$ through the maps $\phi_j$, $j=1,2$. This is equivalent to $$\ c_{i,j}= \frac{d_j t_ia_{i,j}}{A_j}, \
{\delta}_{i,j}= \mu_js_i,$$ where $\mu_j, d_j$ are non zero scalars and $A_j= \prod_{k=1}^{g-1} a_{k,j}$, $j=1,2$.
\[lemmadeltac\] Let us choose $s_i=1$ for $i
=1,...,[\frac{g}{2}]$, $s_i =0$, for $i =
[\frac{g}{2}]+1,...,g-1$, while $t_i=0$ for $i
=1,...,[\frac{g}{2}]$, $t_i =1$, for $i =
[\frac{g}{2}]+1,...,g-1$, $\mu_1 = \mu_2 =: \mu$, $d_1=-\frac{d_2
A_1}{A_2}$. Then, for a general choice of $a_{i,j}$’s, $C=C_1\cup
C_2$ is a binary curve embedded in ${{\mathbb P}}^{g-2}$ by a linear system $|\omega_C \otimes A|$ with $A^{\otimes 2}\cong \OO_C$ and $A$ nontrivial.
One can easily check that if we choose the elements $a_{k,j}$ general ($j=1,2$, $k=1,...,g-1$), $C_j$ are smooth rational curves and $C$ has exactly $g+1$ nodes at the points $P_k$, $k
=1,...,g+1$, and no other singularity. Then, by the above discussion we know that $C$ is a binary curve embedded in ${{\mathbb P}}^{g-2}$ by a linear system of $\omega_C \otimes A$, with $deg(A)=0$. We will now show that $A$ is a 2-torsion non trivial element in $Pic^0(C)$. In fact, recall that $Pic^0(C) \cong {{{\mathbb C}}^*}^g$ and if we denote by $\alpha:N \rightarrow C$ the normalization map, we have an exact sequence $$\label{KA} 0 \rightarrow (\omega_C\otimes A) \rightarrow
\alpha_*(\alpha^*(\omega_C\otimes A)) \rightarrow
\oplus_{i=1}^{g+1}{{\mathbb C}}_{P_i} \rightarrow 0.$$ If we set $\{q_i,r_i \} = \alpha^{-1}(P_i)$, with $q_i \in C_1$, $r_i \in C_2$, $i=1,...,g+1$, $D_1:=\sum_{i=1}^{g+1}q_i$, $D_2:=\sum_{i=1}^{g+1}r_i$ we have $\alpha^*(\omega_C\otimes A) =
K_N(D_1+D_2)$. So if we take the long exact sequence in cohomology associated to (\[KA\]), we have $$0 \rightarrow H^0(\omega_C\otimes A) \rightarrow H^0(K_{C_1}(D_1))
\oplus H^0(K_{C_2}(D_2)) \stackrel{e}\rightarrow {{\mathbb C}}^{g+1}
\rightarrow 0.$$ Clearly $ H^0(K_{C_1}(D_1)) \cong H^0({\mathcal
O}_{{{\mathbb P}}^1}(g-1)) \cong {{\mathbb C}}^g \cong H^0(K_{C_2}(D_2))$. Recall that the line bundle $A$ corresponds to an element in $
{{{\mathbb C}}^*}^g$ as follows. Consider the natural isomorphisms $f_i:
({\alpha^*(A)_{|C_1}})_{q_i} \rightarrow
({\alpha^*(A)_{|C_2}})_{r_i}$, and choose local trivializations $
({\alpha^*(A)_{|C_1}})_{q_i} \cong {{\mathbb C}}$, $
({\alpha^*(A)_{|C_2}})_{r_i} \cong {{\mathbb C}}$, $\forall i =1,...,g+1$. Then $f_i$ is given by multiplication by an element $h_i \in
{{\mathbb C}}^*$, $\forall i = 1,...,g+1$, and we associate to $A$ the element $(h_1,...,h_{g+1})$ modulo the diagonal action of ${{\mathbb C}}^*$.
Notice that if $\sigma \in H^0(\omega_C\otimes A)$ and $\alpha^*(\sigma) = (\sigma_1, \sigma_2) \in H^0(K_{C_1}(D_1))
\oplus H^0(K_{C_2}(D_2))$, then we have $$Res_{q_i}(\sigma_1) - h_i Res_{r_i}(\sigma_2) =0, \ \forall i=1,...,g+1.$$ We claim that with our assumptions the line bundle $A$ corresponds to the element $[(h_1,...,h_{g+1})] \in {{{\mathbb C}}^*}^{g+1}/{{\mathbb C}}^*$, where $h_i=1$, for $i< [\frac{g}{2}]+1$, $h_i = -1$, for $i
=[\frac{g}{2}]+1,...,g-1$, $h_g=-1$, $h_{g+1} = 1$, so $A$ is of 2-torsion. In fact, consider the hyperplane $x_i =0$, $i=1,...,g-1$ in ${{\mathbb P}}^{g-2}$, and set $\sigma_{i,1} :=
\phi_{1}^*(x_i) \in H^0({\mathcal O}_{{{\mathbb P}}^1}(g-1)) \cong
H^0(K_{C_1}(D_1))$, $\sigma_{i,2} := \phi_{2}^*(x_i) \in
H^0({\mathcal O}_{{{\mathbb P}}^1}(g-1)) \cong H^0(K_{C_2}(D_2))$. We have $$\sigma_{i,j} = \frac{({\delta}_{i,j}t -c_{i,j})}{(t - a_{i,j})t} dt, \ j=1,2.$$ Notice that, with our assumptions, we have $$\label{deltac}
\begin{gathered}
{\delta}_{i,1} = {\delta}_{i,2} = \mu, \ i < [\frac{g}{2}]+1,\ {\delta}_{i,1} = {\delta}_{i,2} = 0, \ i \geq [\frac{g}{2}]+1,\\
c_{i,1} = c_{i,2} =0, \ i< [\frac{g}{2}]+1,\\
c_{i,1} = \frac{{d_1} a_{i,1}}{A_1}= -\frac{{d_2}a_{i,1}}{A_2}, \
c_{i,2} = \frac{{d_2} a_{i,2}}{A_2}, \ i \geq [\frac{g}{2}]+1,
\end{gathered}$$ and for simplicity we shall choose $d_2 =\mu = 1$, so the only parameters are the $a_{i,j}$’s. Hence, for $j =1,2$, we have $Res_{q_i}(\sigma_{i,1}) = {\delta}_{i,1}
- c_{i,1}/a_{i,1} = \mu = {\delta}_{i,2} - c_{i,2}/a_{i,2}=
Res_{r_i}(\sigma_{i,2})$ for $i < [\frac{g}{2}]+1$, so $$h_i = \frac{Res_{q_i}(\sigma_{i,1})}{Res_{r_i}(\sigma_{i,2}) }= 1, \ for \ i < [\frac{g}{2}]+1,$$ $$h_i = \frac{Res_{q_i}(\sigma_{i,1})}{Res_{r_i}(\sigma_{i,2}) }= \frac{\frac{c_{i,1}}{a_{i,1}}}{ \frac{c_{i,2}}{a_{i,2}}}= -1, \ for \ i = [\frac{g}{2}]+1,..., g-1,$$ $$h_{g}= \frac{Res_{q_{g}}(\sigma_{g-1,1})}{Res_{r_{g}}(\sigma_{g-1,2}) }= \frac{\frac{c_{g-1,1}}{a_{g-1,1}}}{\frac{c_{g-1,2}}{a_{g-1,2}}}= -1,$$ $$h_{g+1} = \frac{Res_{q_{g+1}}(\sigma_{1,1})}{Res_{r_{g+1}}(\sigma_{1,2}) } = \frac{Res_0( \frac{({\delta}_{1,1}-u c_{1,1})} {(1 - {\delta}_{1,1}u)} (-\frac{1}{u}) du)}{ Res_0( \frac{({\delta}_{1,2}-u c_{1,2})} {(1 - {\delta}_{1,2}u)} (-\frac{1}{u}) du)}= \frac{{\delta}_{1,1}}{{\delta}_{1,2}} = 1.$$
Quadrics
========
In this section we explicitly describe the ideal $I_2(C) := I_2(\omega_C \otimes A)$ of the quadrics containing the Prym-canonical binary curve $C$ embedded in ${{\mathbb P}}^{g-2}$ by $\omega_C \otimes A$ as in the previous section for a general choice of the $a_{i,j}$’s. Similarly as in Proposition 7 of [@ccm], the ideal $I_2(C)$ is described as the space of solutions of the linear system given in Proposition which has maximal rank $2g-2$, so the curve is quadratically normal.
Observe that, since the curves $C_1$ and $C_2$ pass through the coordinate points, the equation of a quadric $Q \subset {{\mathbb P}}^{g-2}$ containing $C_k$ has the form $$\label{Qua} \sum_{1\leq i<j\leq g-1} s_{ij} x_i x_j =0.$$
In the next lemma we give a set of generators of $I_2(C_k)$ of the above form.
Set $$Q_{n,k}:= \sum_{1 \leq i<j\leq g-1} \tilde{q}_{g-1-n,k;i,j} \cdot
s_{ij}, \ n=0,...,g-1,\ k=1,2$$ with $$\begin{gathered}
\tilde{q}_{0,k;i,j} := q_{0,k;i,j} {\delta}_{i,k} {\delta}_{j,k} \\
\tilde{q}_{1,k;i,j} := q_{1,k;i,j} {\delta}_{i,k} {\delta}_{j,k}- q_{0,k;i,j}({\delta}_{i,k}c_{j,k} + c_{i,k} {\delta}_{j,k}) \\
\tilde{q}_{r,k;i,j} := q_{r,k;i,j} {\delta}_{i,k} {\delta}_{j,k}-
q_{r-1,k;i,j}({\delta}_{i,k}c_{j,k} + c_{i,k} {\delta}_{j,k}) +
q_{r-2,k;i,j} c_{i,k} c_{j,k},\ (r\geq 2),
\end{gathered}$$ where $$q_{h,k;i,j}:= \sum_{m=0}^h a_{i,k}^m a_{j,k}^{h-m}$$ and ${\delta}_{i,k}$, $c_{i,k}$, $a_{i,k}$ are as in .
Then the quadrics in $I_2(C_k)$ ($k=1,2$) are the solutions of the linear system: $$\label{Qsistem} Q_{n,k}(s_{ij}) = 0, \ \
n=0,...,g-1.$$
The quadrics of the form containing $C_k$ are the quadrics which satisfy the equations: $$\label{quadric} P_k(t)= \sum_{1\leq i<j\leq g-1}
M_k(t,1)\frac{({\delta}_{i,k}t - c_{i,k})({\delta}_{j,k}t -
c_{j,k})}{(t-a_{i,k})(t-a_{j,k})} s_{ij} = \sum_{n=0}^{g-1}
P_{n,k}(s_{ij}) t^n \equiv 0,$$ $k=1,2$, where the coefficients $P_{n,k}(s_{ij})$ of the polynomial $P_k(t)$ are linear in the $s_{ij}$’s. We will show that the linear system $P_{n,k}(s_{ij})=0$ is equivalent to the system .
By expanding the product $M_k(t,1)$ one sees that the coefficients $p_{h,k;i,j}$ of $s_{ij}$ in $P_{g-1-h, k}$ are $$p_{0,k;i,j}= {\delta}_{i,k}{\delta}_{j,k}, \ p_{1,k;i,j}= - \sum_{i_1 \neq i,j} a_{i_1,k} {\delta}_{i,k}{\delta}_{j,k} -({\delta}_{i,k}c_{j,k} + c_{i,k} {\delta}_{j,k})$$
$$\label{pij}
\begin{gathered}
p_{h,k;i,j}= (-1)^h (\sum_{\substack{1\leq i_1 < i_2 < \cdots < i_h \leq g-1\\ \text{all} \ne i,j}} a_{i_1,k}\cdot \cdot \cdot a_{i_h,k} ){\delta}_{i,k}{\delta}_{j,k} \\
+ (-1)^h (\sum_{\substack{1\leq i_1 < i_2 < \cdots < i_{h-1} \leq g-1\\ \text{all} \ne i,j}} a_{i_1,k}\cdot \cdot \cdot a_{i_{h-1},k} )({\delta}_{i,k}c_{j,k} + c_{i,k} {\delta}_{j,k}) \\
+ (-1)^h (\sum_{\substack{1\leq i_1 < i_2 < \cdots < i_{h-2} \leq g-1\\
\text{all} \ne i,j}}a_{i_1,k}\cdot \cdot \cdot a_{i_{h-2},k}
)c_{i,k} c_{j,k},
\end{gathered}$$
for $ 2 \leq h \leq g-1$.
Set $$\gamma_{0,k} = 1, \ \gamma_{h,k} = (-1)^h \sum_{\substack{1\leq i_1 < i_2 < \cdots < i_h\leq g-1\\ \text{all} \ne i,j}}a_{i_1,k}\cdot \cdot \cdot a_{i_h,k} ,$$
Then we have (cf. [@ccm] (12)) $$\label{formula12} (-1)^h \sum_{\substack{1\leq i_1 < i_2 < \cdots
< i_h \leq g-1\\ \text{all} \ne i,j}} a_{i_1,k}\cdot \cdot \cdot
a_{i_h,k} = \sum_{l=0}^h \gamma_{l,k} q_{h-l,k;i,j}$$ for $h = 0,...,g -1$.
So, by (\[formula12\]), formula (\[pij\]) becomes
$$\label{pijneu}
\begin{gathered}
p_{h,k;i,j}=\sum_{l=0}^{h-2}\gamma_{l,k}(q_{h-l,k;i,j} {\delta}_{i,k}{\delta}_{j,k} -q_{h-1-l,k;i,j} ({\delta}_{i,k}c_{j,k}+c_{i,k}{\delta}_{j,k})+q_{h-2-l,k;i,j} c_{i,k}c_{j,k}) +\\
\gamma_{h-1,k}(q_{1,k;i,j} {\delta}_{i,k}{\delta}_{j,k} -q_{0,k;i,j}({\delta}_{i,k}c_{j,k}+c_{i,k}{\delta}_{j,k})) +
\gamma_{h,k} {\delta}_{i,k}{\delta}_{j,k}= \sum_{l=0}^h \gamma_{l,k} \tilde{q}_{h-l,k;i,j} \end{gathered}$$
So, we have $$P_{n,k}= \sum_{m=0}^{g-1-n} \gamma_{m,k} Q_{n+m,k}$$ and one immediately checks that the linear systems $P_{n,k}(s_{ij}) =0$ and are equivalent.
\[quad1\] Let $g \geq 6$. For a general choice of $a_{i,k}$, $k=1,2$, $i =
1,...,g-1$ and with conditions on $c_{i,k},{\delta}_{i,k}$, the linear system (\[Qsistem\]) has maximal rank $g$.
Consider the matrix $$M(a_{1,k},...,a_{g-1,k}):= (\tilde{q}_{h,k;i,j})_{0\leq h\leq g-1, 1\leq i<j\leq g-1}$$ of size $g \times \frac{(g-1)(g-2)}{2}$. We will show that the minor $B_g$ determined by the columns with indexes $(i,j)=
(1,2),...,(1,g-1),(2,[\frac{g}{2}] + 1),(g-2,g-1)$ is non zero.
Notice that $$\begin{gathered}
\tilde{q}_{0,k;1,j} = \mu^2, \ \ \tilde{q}_{1,k;1,j} = \mu^2 q_{1,k;1,j}, \ \tilde{q}_{h,k;1,j} = \mu^2 q_{h,k;1,j}, \ \ j < [\frac{g}{2}] + 1, \ h \geq 2\\
\tilde{q}_{0,k;1,j} = 0, \ \ \tilde{q}_{1,k;1,j} = -\mu c_{j,k}, \ \tilde{q}_{h,k;1,j} = -\mu c_{j,k} q_{h-1,k;i,j}, \ \ j \geq [\frac{g}{2}] + 1, \ h \geq 2\\
\tilde{q}_{0,k;2,[\frac{g}{2}] + 1} = \tilde{q}_{0,k;g-2,g-1}= 0, \ \tilde{q}_{1,k;2,[\frac{g}{2}] + 1} = -\mu c_{[\frac{g}{2}] + 1,k} , \ \tilde{q}_{1,k;g-2,g-1} = 0 ,\\
\tilde{q}_{h,k;2,[\frac{g}{2}] + 1} = -\mu c_{[\frac{g}{2}] + 1,k}q_{h-1,k;2,[\frac{g}{2}] + 1}, \ h \geq 2\\
\tilde{q}_{h,k;g-2,g-1} = q_{h-2,k;g-2,g-1} c_{g-2,k} c_{g-1,k}, \ h \geq 2 \\
\end{gathered}$$ So, dividing the first $[\frac{g}{2}] $ columns by $\mu^2$, the column indexed by $(2,[\frac{g}{2}] + 1)$ by $-\mu
c_{[\frac{g}{2}] + 1,k}$,the last column by $c_{g-2,k} c_{g-1,k}$ and all the other columns indexed by $(1,j)$ with $[\frac{g}{2}]
+ 1 \leq j \leq g-1$ by $- \mu c_{j,k}$, we see that $B_{g}$ is a non zero multiple of the determinant $d$ of the following matrix:
$$\label{Mtilde} {\small \left( \begin{array}{cccccccc}
1&..&1&0&..&0&0&0 \\
q_{1,k;1,2} &..&q_{1,k;1,[\frac{g}{2}]}&1&..&1&1&0\\
q_{2,k;1,2} &..&q_{2,k;1,[\frac{g}{2}]}&q_{1,k;1,[\frac{g}{2}]+ 1}&..&q_{1,k;1,g-1}&q_{1,k;2,[\frac{g}{2}] + 1}&1\\
q_{3,k;1,2} &..&q_{3,k;1,[\frac{g}{2}]}&q_{2,k;1,[\frac{g}{2}]+ 1}&..&q_{2,k;1,g-1}&q_{2,k;2,[\frac{g}{2}] + 1}&
q_{1,k;g-2,g-1}\\
.&..&.&.&..&.&.&. \\
.&..&.&.&..&.&.&. \\
.&..&.&.&..&.&.&. \\
q_{g-1,k;1,2} &..&q_{g-1,k;1,[\frac{g}{2}]}&q_{g-2,k;1,[\frac{g}{2}]+ 1}&..&q_{g-2,k;1,g-1}&q_{g-2,k;2,[\frac{g}{2}]+ 1}&q_{g-3,k;g-2,g-1}\\
\end{array} \right)}$$
One can inductively compute the determinant $d$ up to sign, $$\label{det1} d= V(a_{3,k},...,a_{g-1,k})\cdot
\prod_{\substack{r=1\\ r \ne 2,[\frac{g}{2}]+1}}^{g-1}
(a_{r,k}-a_{2,k})\cdot \prod_{s=3}^{[\frac{g}{2}]} a_{s,k} \cdot
\prod_{\substack{j=1}}^{g-3} a_{j,k},$$ where $V(a_{3,k},...,a_{g-1,k})$ is the Vandermonde determinant in the variables $a_{3,k},...,a_{g-1,k}$. To do this one can perform column and row operations. [^3]
So we reduce to a $(g-3)\times (g-3)$ matrix whose columns except the last one are the columns of the Vandermonde matrix in the variables $a_{3,k},...,a_{[\frac{g}{2}],k},$ $a_{[\frac{g}{2}]+2,k},...,a_{g-1,k}$. Hence repeating recursively standard row and columns operations,[^4] we obtain formula (\[det1\]).
In the following proposition we give an explicit description of the ideal $I_2(C)$ of quadrics containing $C=C_1\cup C_2$ and we prove that $C$ is quadratically normal.
\[Z\] Let $g \geq 6$. For a general choice of $a_{i,k}$, $k=1,2$, $i =
1,...,g-1$ and with conditions on $c_{i,k},{\delta}_{i,k}$, the linear system $$\label{qqij} Q_{0,1}(s_{ij})=...=Q_{g-1,1}(s_{ij})=Q_{1,2}(s_{ij})=...=Q_{g-2,2}(s_{ij})=0,$$ has maximal rank $2g-2$.
Since we want to prove the statement for generic $a_{i,j}$, it suffices to show it for the following choice of $a_{i,j}$, $j=1,2$, $i=1,...,g-1$: $$\label{scelta}
a_{i,1} := i\cdot a, i=1,...,g-1; \ a_{1,2}:= 1, \ a_{r,2}:=r+1, \ r=1,...,g-1,$$ where $a\neq 1$ is a non zero constant. Consider the matrix $Z:=Z(a_{i,j})$ of size $(2g-2) \times \binom{g-1}{2}$ obtained by concatenating vertically $M(a_{1,1},...,a_{g-1,1})=
(\tilde{q}_{h,1;i,j})_{0\leq h\leq g-1, 1\leq i<j\leq g-1}$, $N(a_{1,2},...,a_{g-1,2})$ $= (\tilde{q}_{h,2;i,j})_{1\leq h\leq g-2,
1\leq i<j\leq g-1}$. Let us set $k:=[\frac{g}{2}]$ and consider the submatrix $Z_1$ of $Z$ formed by the columns of $Z$ indexed by $(1,2),...,(1,g-1),(2,k+1),(g-2,g-1),(2,3),...,(2,k),(2,
k+2),...,(2,g-1),(k, k+1), (k, g-1)$. We will prove that $Z_1$ has maximal rank $2g-2$. Note that the submatrix given by the first $g$ rows and columns is the matrix of Proposition \[quad1\] which is proved to be non singular. So doing operations on the columns we can assume that $Z_1$ is a matrix whose submatrix given by the last $g-2$ columns ad the first $g$ rows is zero. Hence we just need to prove that the submatrix $A$ given by the last $g-2$ rows and columns has maximal rank. If we denote by $v_i$ the column indexed by $(1,i)$, $i
=1,...,g-1$, by $w_i$ the column indexed by $(2,i)$, $i
=1,...,g-1$, by $w$ the column indexed by $(k,k+1)$, by $\zeta$ the column indexed by $(k,g-1)$, the operations that we do on the columns of $Z_1$ are the following:
- for $i=3,...,k$, substitute the column $w_i$ with the vector $w_i+\frac{1-i}{i-2}v_i + \frac{1}{i-2}v_1$.
- for $i=k+2,...,g-1$, substitute the column $w_i$ with the vector
$w_i+\frac{k\cdot c_{i,1}}{c_{k+1,1}(i-2)}v_{k+1} -\frac{c_{i,1}(k-1)}{c_{k+1,1}(i-2)}w_{k+1}- \frac{i-1}{i-2} v_{i}.$
- substitute the column $w$ with the vector
$w+\frac{(k-1)\cdot c_{k+1,1}}{ka}v_{1} -\frac{c_{k+1,1}(k-1)}{ka}v_{k}- (2-k)v_{k+1}-\frac{2(1-k)^2}{k}w_{k+1}$.
- substitute the column $\zeta$ with the vector
$\zeta+\frac{(k-1)\cdot c_{g-1,1}}{ka(g-k-1)}v_{1} -\frac{c_{g-1,1}(k-1)}{ka(g-k-1)}v_{k}+ \frac{2(k-1)c_{g-1,1}}{(g-k-1)c_{k+1,1}}v_{k+1}-\frac{2(1-k)^2c_{g-1,1}}{(g-k-1)kc_{k+1,1}}w_{k+1}- \frac{g-2}{g-k-1}v_{g-1}$.
To prove that the matrix $A$ is of maximal rank $g-2$, we argue as follows. First of all one can easily check (with the same procedure as in \[quad1\]) that the submatrix $C$ of $Z_1$ formed by the columns indexed by $(1,2), (1,4), ...,
(1,g-1),(2,k+1)$ and by the last $g-2$ rows has rank $g-2$. Denote by $(\lambda_1,...,\lambda_{g-2})$ the coordinates of the vector given by the column of $Z_1$ indexed by $(1,3)$ and the last $g-2$ rows, with respect to the basis of ${{\mathbb C}}^{g-2}$ given by the columns of $C$. Then the coordinates of the columns of $A$ with respect to the basis given by the columns of $C$ are given by the following matrix which we will show to have maximal rank: $${\small \left( \begin{array}{ccccccccccc}
\lambda_1-1&-\frac{1}{2}&-\frac{1}{3}&...&-\frac{1}{k-2}&0&...&0&\alpha_1&\beta_1\\
\lambda_2&\frac{1}{2}&0&...&0&0&...&0&0&0 \\
\lambda_3&0&\frac{1}{3}&...&0&0&...&0&0&0 \\
.&...&.&.&...&.&.&.&.&.& \\
.&...&.&.&...&.&.&.&.&.& \\
.&0&.&.&...&.&.&0&0&0& \\
\lambda_{k-2}&0&0&...&\frac{1}{k-2}&0&...&0&-\alpha_1&-\beta_1 \\
\lambda_{k-1}&0&0&...&0&\mu_{1,k+2}&...&\mu_{1,g-1}&\alpha_3&\beta_3 \\
\lambda_{k}&0&0&...&0&\mu_{2,k+2}&...&0&0&0 \\
&...&.&.&...&0&.&.&.&.& \\
.&...&.&.&...&.&.&0&.&0& \\
\lambda_{g-3}&0&0&0&...&0&0&\mu_{2,g-1}&0&\beta_4& \\
\lambda_{g-2}&0&0&0&...&\mu_{3,k+2}&..&\mu_{3,g-1}&\alpha_4&\beta_5& \\
\end{array} \right),}$$ where, for $j=k+2,...,g-1$, $$\mu_{1,j} = \frac{1}{j-2} (\frac{kc_{j,1}}{c_{k+1,1}}-\frac{(k+1)c_{j,2}}{c_{k+1,2}}),
\ \mu_{2,j} = \frac{1}{j-2},$$ $$\mu_{3,j} = (\frac{k-1}{j-2}) (-\frac{c_{j,1}}{c_{k+1,1}}+\frac{c_{j,2}}{c_{k+1,2}}), \ \alpha_3= \frac{k-2}{2},$$ $$\alpha_4= \frac{(k-1)(4-k^2)}{2k(k+1)}, \ \beta_3 = \frac{1}{g-k-1} (\frac{-3kc_{g-1,2}}{2c_{k+1,2}}+\frac{2(k-1)c_{g-1,1}}{c_{k+1,1}}),$$ $$\beta_4= \frac{1}{g-k-1}, \ \beta_5 = \frac{1}{g-k-1} (\frac{3k(k-1)c_{g-1,2}}{2(k+1)c_{k+1,2}}-\frac{2(k-1)^2c_{g-1,1}}{kc_{k+1,1}}).$$
Substracting from each of the last two columns a suitable multiple of the $(k-2)$’s column, we can assume that $\alpha_1=\beta_1 =0$, hence the submatrix formed by the last $g-k$ columns and the first $k-2$ rows is zero. The determinant of the submatrix given by the first $(k-2)$ rows and columns is $\frac{1}{(k-2)!} (\sum_{i=1}^{k-2} \lambda_i -1)$, and the determinant of the submatrix given by the last $g-k$ rows and columns is a non zero multiple of $$det\left( \begin{array}{ccc}
\mu_{1,g-1}&\alpha_3&\beta_3 \\
\mu_{2,g-1}&0&\beta_4 \\
\mu_{3,g-1}&\alpha_4&\beta_5 \\
\end{array} \right)= \frac{-(k-1)(k-2)^2(g-k-2)}{2k(k+1)^2(k+2)(g-3)(g-k-1)} \neq 0.$$ So it remains to show that $\sum_{i=1}^{k-2} \lambda_i \neq 1$. To do this, it suffices to show that the matrix obtained by adding the row $(1,...,1)$ to the submatrix of $Z_1$ formed by the columns indexed by $(1,2),(1,3),...,(1,g-1),(2,k+1)$ and the last $g-2$ rows has rank $g-1$. This can be easily seen with a procedure similar to the one used in Proposition \[quad1\].
Surjectivity
============
In this section we will prove by induction on the genus the surjectivity of $\mu_A$ for a general Prym-canonical binary curve $(C,A)$ of genus $\geq 20$.
The 2nd Gaussian map
--------------------
Let us first of all analyze in detail the map $\mu_A$ of when $C= C_1 \cup C_2$ is a Prym-canonical binary curve embedded in ${{\mathbb P}}^{g-2}$ by $\omega_C \otimes A$, where $A \in Pic^0(C)$ is nontrivial of order 2.
Since ${\omega_C}_{|C_i} = K_{C_i}(D_i)$ where $D_i$ is the divisor of nodes in $C_i$, we have $$H^0(S^2(\Omega^1_C) \otimes \omega_C^{\otimes 2}) \cong T \oplus (\oplus_{i=1,2} H^0(C_i, K^{\otimes 4}_{C_i}(2D_i))),$$ where $T$ is the torsion of $S^2(\Omega^1_C)$, which is supported at the nodes (see lemma 2 of [@ccm]). In fact we have an exact sequence $$0 \rightarrow T \rightarrow S^2(\Omega^1_C) \rightarrow {\mathcal F}_C \rightarrow 0,$$ where $ {\mathcal F}_C$ is a non-locally free, rank 1, torsion free sheaf on $C$.
To prove the surjectivity of $\mu_A$ we will show the surjectivity of the components of $\mu_A$ on both non torsion and torsion parts of $H^0(S^2(\Omega^1_C) \otimes \omega_C^{\otimes 2})$.
Consider first the non torsion component $\nu = \nu_1 \oplus \nu_2$, where $$\nu_k: I_2(C)\rightarrow H^0(C_k, K^{\otimes 4}_{C_k}(2D_k)) \cong H^0({{\mathbb P}}^1, {\mathcal O}_{{{\mathbb P}}^1}(2g-6)), \ k =1,2.$$
Recall that the curves $C_l$, $l=1,2$ are the images of the maps $\phi_l$ defined in , $\phi_l:{{\mathbb P}}^1 \rightarrow {{\mathbb P}}^{g-2}, \ l=1,2$ $$\phi_l(t,u):= (f_{1,l}(t),...,f_{g-1,l}(t)), \ f_{i,l}(t)= M_l(t)\frac{({\delta}_{i,l}t-c_{i,l})}{(t-a_{i,l})}.$$
Assume $Q \in I_2(C)$ is of the form (\[Qua\]) where $s_{ij}$ are solutions of (\[Qsistem\]). Then using the local expression given in , we have $$\label{nuk}
\nu_k(Q) = \sum_{1\leq i<j\leq g-1} M^2_k(t)(\frac{{\delta}_{i,k}t-c_{i,k}}{t-a_{i,k}})^{'}( \frac{{\delta}_{j,k}t-c_{j,k}}{t-a_{j,k}})^{'}s_{ij}(dt)^4, \ k=1,2$$
As an element of $H^0({{\mathbb P}}^1, {\mathcal O}_{{{\mathbb P}}^1}(2g-6))$, $\nu_k(Q)$ can be identified with the polynomial of degree $2g-6$ in $t$, $$\label{Rk} R_k(t) = \sum_{1\leq i<j\leq g-1} M^2_k(t)\frac{(c_{i,k}-{\delta}_{i,k}a_{i,k})}{(t-a_{i,k})^2}
\frac{(c_{j,k}-{\delta}_{j,k}a_{j,k})}{(t-a_{j,k})^2}s_{ij}.$$
To study the torsion component, we consider as in [@ccm] the restriction $\tau$ of $\mu_A$ to $ker(\nu)$, which lands in the torsion part $T$ of $H^0(S^2(\Omega^1_C) \otimes \omega_C^{\otimes 2})$.Then using Lemma 2 of [@ccm] one sees that the composition of $\tau$ with the projection on the torsion part $T_{P_h}$ at the nodes $P_1,...,P_{g-1},P_g$ is as follows: a quadric $Q \in ker(\nu)$ as in (\[Qua\]) is mapped to $$\label{torsion}
\begin{gathered}
dx dy \sum_{i \neq j}s_{ij}
f_{i,1}^{'}(a_{h,1})f_{j,2}^{'}(a_{h,2}) + 2xdx dy \sum_{i \neq
j}s_{ij}
f_{i,1}^{''}(a_{h,1})f_{j,2}^{'}(a_{h,2}) + \\
2y dx dy \sum_{i \neq j}s_{ij}
f_{i,1}^{'}(a_{h,1})f_{j,2}^{''}(a_{h,2}),
\end{gathered}$$ where $h=1,...,g$, $s_{ij} = s_{ji}$ and $x$, $y$ are local coordinates around $P_h$ such that $C_1$ is given locally by $x=0$ and $C_2$ by $y=0$ and since $P_g$ is the image of $[0,1]$, we set $a_{g,1} = a_{g,2} =0$. The description of the torsion at the point $P_{g+1}$ is similar: $$\label{torsionPg+1}
\begin{gathered}
dx dy \sum_{i \neq j}s_{ij} ({\delta}_{i,1}a_{i,1} - c_{i,1})({\delta}_{j,2}a_{j,2}-c_{j,2}) +\\ 2xdx dy \sum_{i \neq j}s_{ij} a_{i,1}({\delta}_{i,1}a_{i,1}-c_{i,1})({\delta}_{j,2}a_{j,2} - c_{j,2}) +\\
2y dx dy \sum_{i \neq j}s_{ij} a_{j,2}({\delta}_{i,1}a_{i,1} - c_{i,1})({\delta}_{j,2}a_{j,2} - c_{j,2}),
\end{gathered}$$ where $s_{ij} = s_{ji}$ and $x$, $y$ are local coordinates around $P_{g+1}$ such that $C_1$ is given locally by $x=0$ and $C_2$ by $y=0$.
Proof of surjectivity
---------------------
Let $C \subset {{\mathbb P}}^{g-2}$ be a Prym-canonical binary curve embedded by $\omega_C \otimes A$, with $A^{\otimes 2} \cong \OO_C$, as in (\[deltac\]) and set $k :=[\frac{g}{2}]$. Denote by $\tilde{C}$ the partial normalization of $C$ at the node $P$, where $P=P_k$ if $g =2k$, $P= P_{k+1}$ if $g = 2k+1$ and by $p_1,p_2$ the preimages of $P$ in $\tilde{C}$. Observe that for a general choice of the $a_{i,j}$’s, the projection $\pi$ from $P$ sends the curve $C$ to the Prym-canonical model of $\tilde{C}$ in ${{\mathbb P}}^{g-3}$ given by the line bundle $K_{\tilde{C}} \otimes A'$ where $A'$ corresponds to the point $(h'_1,...,h'_{g-1}, 1) \in {{{\mathbb C}}^*}^{g}/{{\mathbb C}}^*$, with $h'_i = 1$ for $i\leq [\frac{g-1}{2}]$, $h'_i = -1$ for $i=[\frac{g-1}{2}]+1,...,g-1$, as in section 2. In fact if $g =2k$, we have $k-1 = [\frac{g-1}{2}]=:k'$ and $(\tilde{C},A')$ is as in , with $a'_{i,j} = a_{i,j}$ for $i \leq k'$, $j =1,2$, $a'_{i,j} =
a_{i+1,j}$ for $i \geq k'+1$, $j =1,2$. If $g = 2k+1$ we have $[\frac{g-1}{2}] = k$, so $(\tilde{C},A')$ is parametrized by $a'_{i,j} = a_{i,j}$ for $i
\leq k$, $j =1,2$, $a'_{i,j} = a_{i+1,j}$ for $i \geq k+1$, $j=1,2$.
Consider the following commutative diagrams with horizontal exact sequences $$\label{chi} \xymatrix{
&0\ar[r] & \oplus_{i=1,2} H^0(C_i, K^{\otimes 4}_{C_i}(2\tilde{D_i}))\ar[r] & \oplus_{i=1,2} H^0(C_i, K^{\otimes 4}_{C_i}(2D_i)) \ar[r] & \oplus_{i=1,2}\OO_{2p_i}\\
&0 \ar[r] &I_2({\tilde C}) \ar[u]^{\tilde\nu} \ar[r]& I_2(C)
\ar[u]^{\nu}\ar[ur]_{\chi} & }$$
$$\label{tau} \xymatrix{
&0\ar[r] & \tilde T \ar[r] & T \ar[r] & T_P \\
&0 \ar[r] &\ker(\tilde\nu) \ar[u]^{\tilde\tau} \ar[r]& ker(\nu)
\ar[u]^{\tau} \ar[ur]_{\tau_P} & }$$
where $D_i$ is the divisor of nodes of $C$ on $C_i$ and $\tilde{D_i} = D_i + p_i$ and $\nu, \tau$ and $\tilde\nu, \tilde\tau$ are the maps defined in the previous section for $C$ and $\tilde C$. Hence, if $\tilde\nu$ and $\chi$ ($\tilde\tau$ and $\tau_P$, resp.) are surjective, then $\nu$ ($\tau$, resp.) is also surjective.
\[thm20\] If $(C=C_1\cup C_2, \omega_C\otimes A)$ is a Prym-canonical general binary curve of genus $g \geq 20$, then $\mu_A$ is surjective for $C$.
The case $g=20$ is done by a direct computation with Maple (see Appendix A). We then proceed by induction on $g$: the commutativity of the diagrams shows that it is enough to prove the surjectivity of $\chi$ and $\tau_P$, where $P=P_k$ for $g=2k$ and $P=P_{k+1}$ for $g=2k+1$, as above. Recall that the map $\nu$ is $\nu_1\oplus\nu_2$ where $\nu_1$ and $\nu_2$ are defined in (\[nuk\]), so we can write $\chi=\chi_1\oplus\chi_2$, where $\chi_l$ is the composition of $\nu_l$ with the restriction to $\OO_{2p_l}$, $l=1,2$. We want to compute $\chi(Q)$, where $Q\in I_2(C)$. Notice that if $Q\in
I_2(C)$ is of the form , with the $s_{ij}$’s satisfying , then $\chi_l(Q)$ is the pair $(R_l(a_{r,l}),R'_l(a_{r,l}))$ (where $r=k$ for $g$ even and $r=k+1$ for $g$ odd) corresponding to the evaluation of the polynomial $R_l(t)$ ($l=1,2$) of and of its derivative at $P$. Recall that $R_l(t)$ is linear in the $s_{ij}$’s and denote by $R^l_{i,j}(t)$ the coefficient of $s_{ij}$ in $R_l(t)$.
To prove the surjectivity of $\chi$ we have to show that the matrix $Y$ of size $(2g+2)\times \binom{g-1}{2}$ obtained by concatenating vertically the matrix $Z$ in the proof of Proposition \[Z\], and the matrix of size $4\times
\binom{g-1}{2}$ whose rows are the evaluations in $P$ of $R^1_{i,j},(R^1_{i,j})',R^2_{i,j},(R^2_{i,j})'$ is of maximal rank.
By formula we have: $$R^l_{i,j}(t)=(c_{i,l}-a_{i,l}\delta_{i,l})(c_{j,l}-a_{j,l}\delta_{j,l})\cdot\prod_{r\neq
i,j}(t-a_{r,l})^2.$$ Therefore, if $i,j\neq n$, $R^l_{i,j}(a_{n,l})=0$ and $(R^l_{i,j})'(a_{n,l})=0$. So it remains to compute $R^l_{i,k}(a_{k,l}),R^l_{k,j}(a_{k,l}) $ and $(R^l_{i,k})'(a_{k,l}),(R^l_{k,j})'(a_{k,l})$ , for $g=2k$, and $R^l_{i,k+1}(a_{k+1,l}),$ $R^l_{k+1,j}(a_{k+1,l}) $ and $(R^l_{i,k+1})'(a_{k+1,l}),$ $(R^l_{k+1,j})'(a_{k+1,l})$, for $g=2k+1$. If $g=2k$ and we denote by $D_{k,l}:=\prod_{r\neq
k}(a_{k,l}-a_{r,l})^2$, we have $$R^l_{i,k}(a_{k,l})= \frac{D_{k,l}\cdot
a_{i,l}a_{k,l}}{(a_{k,l}-a_{i,l})^2} ,\ R^l_{k,j}(a_{k,l})=-
\frac{D_{k,l}\cdot c_{j,l}a_{k,l}}{(a_{k,l}-a_{j,l})^2}$$
$$(R^l_{i,k})'(a_{k,l})= \frac{2D_{k,l}\cdot
a_{i,l}a_{k,l}}{(a_{k,l}-a_{i,l})^2}\cdot \sum_{r\neq
i,k}\frac{1}{(a_{k,l}-a_{r,l})},$$ $$(R^l_{k,j})'(a_{k,l})=- \frac{2D_{k,l}\cdot
c_{j,l}a_{k,l}}{(a_{k,l}-a_{j,l})^2}\cdot \sum_{r\neq
j,k}\frac{1}{(a_{k,l}-a_{r,l})}$$
If $g=2k+1$, and we denote by $D_{k+1,l}:=\prod_{r\neq
k+1}(a_{k+1,l}-a_{r,l})^2$ we have $$R^l_{i,k+1}(a_{k+1,l})=-
\frac{D_{k+1,l}\cdot a_{i,l}c_{k+1,l}}{(a_{k+1,l}-a_{i,l})^2},\
R^l_{k+1,j}(a_{k+1,l})= \frac{D_{k+1,l}\cdot
c_{j,l}c_{k+1,l}}{(a_{k+1,l}-a_{j,l})^2}$$
$$(R^l_{i,k+1})'(a_{k+1,l})=- \frac{2D_{k+1,l}\cdot
a_{i,l}c_{k+1,l}}{(a_{k+1,l}-a_{i,l})^2}\cdot \sum_{r\neq
i,k+1}\frac{1}{(a_{k+1,l}-a_{r,l})}$$ $$(R^l_{k+1,j})'(a_{k+1,l})= \frac{2D_{k+1,l}\cdot
c_{j,l}c_{k+1,l}}{(a_{k+1,l}-a_{j,l})^2}\cdot \sum_{r\neq
j,k+1}\frac{1}{(a_{k+1,l}-a_{r,l})}$$
To show that the matrix $Y$ has maximal rank $2g+2$ we will show that the minor $det N$ is non zero, where $N$ is determined by the columns indexed by $(1,i)$, $(2,j)$, with $2\leq i\leq g-1$, $3\leq j\leq g-1$, $(k, k+1), (k, g-1),
(g-2,g-1)$ and we choose the columns $(3,k),(4,k),(k+1,g-2),(k+1,g-1)$, in the case $g=2k$, and the columns $(3,k),(4,k+1),(k+1,g-2),(k+1,g-1)$, in the case $g=2k+1$. Notice that the square submatrix of $N$ given by the first $2g-2$ rows and columns is the submatrix $Z_1$ of $Z$ introduced in Prop.\[Z\], which is shown to be non singular for a general choice of the $a_{i,l}$’s. The columns of the submatrix $G$ of $N$ given by its last four columns and its first $2g-2$ rows are clearly also columns of $Z$ hence linearly dependent on the columns of $Z_1$. Therefore we perform operations on the last four columns of $N$ to bring $G$ to the zero matrix. So it suffices to prove that the submatrix $A$ of order 4, given by the last 4 rows and columns is nonsingular for general $a_{i,l}$. To do this we choose the set of the $a_{i,l}$’s as in , we compute with Maple the determinant of $A$ and we see that as a function of $k$ it does not vanish for any integer $k \geq 10$ (see Appendix B). This proves that $\chi$ is surjective.
It remains to show that $\tau_P$ is surjective. Recall that $\ker(\nu)$ is defined in $ I_2(C)$ by the vanishing of the polynomials $R_l(t)$, $l=1,2$. By the description of the torsion at the point $P$ given in , we need to show the rank maximality of the matrix $X$ of size $(2g+5)\times
\binom{g-1}{2}$ obtained by concatenating vertically the above matrix $Y$ and the matrix of size $3\times \binom{g-1}{2}$ whose rows are, for $g=2k$ (hence $P=P_k$) $$(T_1)_{ij}=f_{i,1}^{'}(a_{k,1})f_{j,2}^{'}(a_{k,2})+f_{j,1}^{'}(a_{k,1})f_{i,2}^{'}(a_{k,2}),$$ $$(T_2)_{ij}=f_{i,1}^{''}(a_{k,1})f_{j,2}^{'}(a_{k,2})+ f_{j,1}^{''}(a_{k,1})f_{i,2}^{'}(a_{k,2}),$$ $$(T_3)_{ij}=f_{i,1}^{'}(a_{k,1})f_{j,2}^{''}(a_{k,2}) + f_{j,1}^{'}(a_{k,1})f_{i,2}^{''}(a_{k,2})$$ and for $g=2k+1$, hence $P=P_{k+1}$, $$(T_1)_{ij}=f_{i,1}^{'}(a_{k+1,1})f_{j,2}^{'}(a_{k+1,2})+ f_{j,1}^{'}(a_{k+1,1})f_{i,2}^{'}(a_{k+1,2}),$$ $$(T_2)_{ij}=f_{i,1}^{''}(a_{k+1,1})f_{j,2}^{'}(a_{k+1,2}) + f_{j,1}^{''}(a_{k+1,1})f_{i,2}^{'}(a_{k+1,2}),$$ $$(T_3)_{ij}=f_{i,1}^{'}(a_{k+1,1})f_{j,2}^{''}(a_{k+1,2}) + f_{j,1}^{'}(a_{k+1,1})f_{i,2}^{''}(a_{k+1,2}).$$
We claim that the minor $det M$ of the submatrix $M$ of $X$ determined by the $2g+5$ columns, indexed as the columns of $N$ plus $(5,k+1), (k,g-4),(k+1,g-3)$ if $g = 2k$, and $(5,k+1), (k+1,g-4),(k+1,g-3)$ if $g = 2k+1$ is nonzero. This will conclude the proof that $\tau_P$ is surjective, hence the proof of the theorem.
As above the square submatrix of $M$ given by the first $2g-2$ rows and columns is the submatrix $Z_1$ of $Z$ introduced in Prop.\[Z\], which is non singular for a general choice of the $a_{i,l}$’s. The columns of the submatrix $H$ of $M$ given by its last seven columns and its first $2g-2$ rows are clearly also columns of $Z$ hence linearly dependent on the columns of $Z_1$. Therefore we perform operations on the last seven columns of $M$ to bring $H$ to the zero matrix. So it suffices to prove that the submatrix $B$ of order 7, given by the last 7 rows and columns is nonsingular for general $a_{i,l}$. To this purpose we choose the set of the $a_{i,l}$’s as in , we compute again with Maple the determinant of $B$ and we see that for any integer $k \geq 10$ it does not vanish (see Appendix B). This proves that $\tau_P$ is surjective, hence by induction $\mu_A$ is surjective.
The class
=========
In the previous section we have proved by semicontinuity that the 2nd Gaussian map $\mu_A: I_2(C) \rightarrow H^0(S^2(\Omega^1_C)
\otimes K_C^{\otimes 2})$ has maximal rank for the general pair $[C,A]$ in ${\mathcal R}_{20}$. Notice that for $g =20$, $dim(I_2(C)) = dim(H^0(S^2(\Omega^1_C) \otimes K_C^{\otimes 2})) =
133. $ Consider the locus ${\mathcal D} = \{[C,A] \in {\mathcal
R}_{20} \ | \ rk(\mu_A) <133 \}$. We have proved that ${\mathcal
D} \neq {\mathcal R}_{20}$, hence, if it is not empty, it is an effective divisor in ${\mathcal R}_{20}.$ Let $\pi:
\overline{\mathcal R}_{g} \rightarrow \overline{{\mathcal M}}_g$ be the finite map which extends the forgetful map ${\mathcal R}_g
\rightarrow {\mathcal M}_g$ (see [@fl] Section 1). The partial compactification $\tilde{\mathcal R}_g$ of ${\mathcal R}_g$ introduced in [@fl] Section 1 is the inverse image $\pi^{-1}(\tilde{\mathcal M}_g)$, where $\tilde{\mathcal
M}_g:={\mathcal M}_g\cup \tilde{\Delta }_0$ and $\tilde{\Delta
}_0$ is the locus of one-nodal irreducible curves. Denote by $f:
{\mathcal X} \rightarrow \tilde{{\bf{R}}}_g$ the universal family and by ${\mathcal P} \in Pic({\mathcal X})$ the corresponding Prym bundle as in [@fl] 1.1. Assume $g =20$, then if $\tilde{{\mathcal D}}$ is the closure of ${\mathcal D}$ in ${\tilde{\mathcal R}_{20}}$, $\tilde{{\mathcal D}}$ is the degeneracy locus of the map $$\tilde{\mu}: \tilde{{\mathcal I}}_2 \rightarrow f_*( (\omega_f\otimes {\mathcal P})^{\otimes 2}\otimes
{S^2(\Omega^1_f)})\cong f_*(\omega_f^{\otimes 4}\otimes {\mathcal
P}^{\otimes 2}\otimes {\mathcal I}_Z^{\otimes 2}),$$ of . Denote by ${\mathcal F}_i :=
f_{*}(\omega_f^{\otimes i} \otimes {\mathcal P}^{\otimes i})$. Using Grothendieck-Riemann-Roch and Proposition 1.6 of [@fl] one computes as in Proposition 1.7 of [@fl] $$c_1({\mathcal F}_i) = \frac{i(i-1)}{2} (12 \lambda - \delta'_0- \delta''_0 - 2 \delta_0^{ram}) +
\lambda - \frac{i^2}{4} \delta_0^{ram},$$ where $\lambda$ is the pullback of the Hodge class $\lambda \in {\overline{\mathcal M}}_g$ and $ \delta'_0$, $\delta''_0$, and $\delta_0^{ram}$ are the boundary classes defined in [@fl] section 1. So we have $$c_1({\mathcal F}_1) = \lambda - \frac{\delta_0^{ram}}{4}, \ \ c_1({\mathcal F}_2) = 13 \lambda -
\delta'_0- \delta''_0- 3\delta_0^{ram}, \ \
c_1(S^2({\mathcal F}_1)) = 20 \cdot c_1({\mathcal F}_1)= 20 \lambda - 5 \delta_0^{ram},$$ therefore $$c_1({\mathcal I}_2) = c_1(S^2{\mathcal F}_1) - c_1({\mathcal F}_2) = 7 \lambda + \delta'_0 +
\delta''_0 - 2 \delta_0^{ram}.$$
Notice that by Grothendieck-Riemann-Roch we have $$c_1( f_*( \omega_f^{\otimes 4} \otimes {\mathcal P}^{\otimes 2} \otimes {\mathcal I}_Z^{\otimes 2})) =
f_*[(1 + c_1( \omega_f^{\otimes 4} \otimes {\mathcal P}^{\otimes 2})
+ \frac{1}{2} c_1^2( \omega_f^{\otimes 4} \otimes {\mathcal P}^{\otimes 2})-2 [Z])
\cdot (1 - \frac{c_1(\omega_f)}{2} + \frac{ c_1(\omega_f)^2+ [Z]}{12}) ]_2$$ $$= 73 \lambda -8 (\delta'_0 + \delta''_0) -17 \delta_0^{ram},$$ since $f_*(c_1(\omega_f) \cdot {\mathcal P}) =0$, $f_*(c_1({\mathcal P})^2) = - \delta_0^{ram}/2$, by Proposition 1.6 of [@fl] and by Mumford’s formula, $f_*(c_1(\omega_f)^2) = 12 \lambda + f_*([Z])$ and $f_*[Z] = \delta'_0 + \delta''_0+ 2 \delta_0^{ram}$ ([@fl], 1.1). So, finally we have $$c_1(\tilde{{\mathcal D}}) = c_1( f_*( \omega_f^{\otimes 4} \otimes {\mathcal P}^{\otimes 2}
\otimes {\mathcal I}_Z^{\otimes 2}))\cdot rk({\mathcal I}_2 )- c_1({\mathcal I}_2)\cdot
rk (f_*( \omega_f^{\otimes 4} \otimes {\mathcal P}^{\otimes 2}\otimes {\mathcal I}_Z^{\otimes 2}))=$$ $$=133 (66 \lambda -9 (\delta'_0 + \delta''_0) -15 \delta_0^{ram}),$$ and $c_1({\mathcal D}) = 8778 \lambda$, hence ${\mathcal D}$ is an effective divisor in ${\mathcal R}_{20}$, $\tilde{{\mathcal D}}$ is an effective divisor in $\tilde{{\mathcal R}}_{20}$ and if we denote by $\overline{{\mathcal D}}$ the closure of ${\mathcal D}$ in $\overline{\mathcal R}_{20}$, we have computed $$\label{Dbar}
c_1(\overline{{\mathcal D}}) = 133 (66 \lambda -9 (\delta'_0 + \delta''_0) -15 \delta_0^{ram}- ...)$$ In fact, since the partial compactification $\tilde{{\mathcal
R}}_{g} \subset \overline{\mathcal R}_{g}$ has the property that $\pi^{-1}({\mathcal M}_g \cup \Delta_0)-\tilde{{\mathcal R}}_{g} $ has codimension $\geq 2$, the expression (\[Dbar\]) computes the coefficients of $\lambda$, $\delta'_0$, $\delta''_0$, $\delta_0^{ram}$ in $c_1(\overline{{\mathcal D}})$.
- Using proposition 1.9 of [@fl] one can find lower bounds on some of the other boundary coefficients of $\overline{{\mathcal D}}$.
- Pushing forward $\overline{{\mathcal D}}$, one gets $$c_1(\pi_*(\overline{{\mathcal D}})) =
133 (66(2^{40}-1) \lambda -(33\cdot2^{38}-9) \delta_0-...),$$ hence its slope is $\geq 8+\frac{2}{3023656976381}$.
Appendix: Maple scripts for computations {#appendix-maple-scripts-for-computations .unnumbered}
========================================
[**[A. Surjectivity for $g=20$]{}**]{} We list here the Maple script we run. We will explain it afterwards: for this purpose, we added line numbers.
[`a[1]:=[25,35,54,47,67,97,73,81,22,33,76,27,38,44,58,69,63,80,99]:`\
`a[2]:=[1,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20]:`\
`listsij:=[seq(seq(s[i,j],j=i+1..19),i=1..19)]:`\
`t[1]:=[0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1]:`\
`5``A1:=mul(a[1][i],i=1..19): `\
`A2:=mul(a[2][i],i=1..19): `\
`d2:=1:`\
`d1:= -d2*A1/A2:`\
`delta:=[1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0]:`\
`10``c[1]:=[seq(t[1][i]*a[1][i]*d1/A1,i=1..19)]:`\
`c[2]:=[seq(t[1][i]*a[2][i]*d2/A2,i=1..19)]:`\
`Z:=Matrix([seq([seq(seq((delta[i]*delta[j])*add(a[1][i]^m*a[1][j]^(h-m),m=0..h)-`\
`(delta[j]*c[1][i]+delta[i]*c[1][j])*add(a[1][i]^m*a[1][j]^(h-1-m),m=0..h-1)+ `\
`(c[1][i]*c[1][j])*add(a[1][i]^m*a[1][j]^(h-2-m),m=0..h-2), j=i+1..19),i=1..19)],`\
` h=0..19), seq([seq(seq((delta[i]*delta[j])*add(a[2][i]^m*a[2][j]^(h-m),m=0..h)-`\
`(delta[j]*c[2][i]+delta[i]*c[2][j])*add(a[2][i]^m*a[2][j]^(h-1-m),m=0..h-1)+`\
` (c[2][i]*c[2][j])*add(a[2][i]^m*a[2][j]^(h-2-m),m=0..h-2), j=i+1..19),i=1..19)],`\
`h=0..19)]):`\
`19``Zref:=Gausselim(Z,'r0') mod 131:`\
` r0;`\
` 38`\
`M[1] := mul(t-a[1][i], i = 1 .. 19):`\
`M[2] := mul(t-a[2][i], i = 1 .. 19):`\
`for i from 1 to 19 do phi1[1, i] := diff(M[1]*(delta[i]*t-c[1][i])/(t-a[1][i]), t):`\
`25``phi1[2, i] := diff(M[2]*(delta[i]*t-c[2][i])/(t-a[2][i]), t) end do:`\
`R[1] := add(add(s[i, j]*phi1[1, i]*phi1[1, j], j = i+1 .. 19), i = 1 .. 19):`\
`R[2] := add(add(s[i, j]*phi1[2, i]*phi1[2, j], j = i+1 .. 19), i = 1 .. 19):`\
`Eqskernu := [seq(seq(primpart(coeff(R[k], t, n)), n = 0 .. 34), k = 1 .. 2)]:`\
`K:= Gausselim(linalg[stackmatrix](Zref,linalg[genmatrix](Eqskernu,listsij)),'r1')`\
`mod 131):`\
` r1;`\
` 108`\
`32``for i from 1 to 19 do phi2[1,i]:= diff(phi1[1,i],t): phi2[2,i] := diff(phi1[2,i],t): `\
`phi1e0[1,i]:= eval(phi1[1,i], t = 0): phi2e0[1,i] := eval(phi2[1,i], t = 0): `\
`phi1e0[2,i] := eval(phi1[2,i], t = 0): phi2e0[2,i]:= eval(phi2[2,i], t = 0): `\
`for h to 19 do phi1e[1, i, h]:= eval(phi1[1,i], t = a[1][h]):`\
`phi2e[1, i, h]:= eval(phi2[1,i], t = a[1][h]):`\
`phi1e[2, i, h]:= eval(phi1[2,i], t = a[2][h]):`\
`phi2e[2, i, h] := eval(phi2[2,i], t = a[2][h]) end do end do:`\
`39``for h from 1 to 19 do`\
`tors[h,1]:= add(add(s[i,j]*(phi1e[1,i,h]*phi1e[2, j, h]+`\
`phi1e[1,j,h]*phi1e[2,i,h]), j = i+1 .. 19), i = 1 .. 19):`\
`tors[h,2]:= add(add(s[i,j]*(phi2e[1,i,h]*phi1e[2, j, h]+`\
`phi2e[1,j,h]*phi1e[2,i,h]), j = i+1 .. 19), i = 1 .. 19):`\
`tors[h,3]:= add(add(s[i,j]*(phi1e[1,i,h]*phi2e[2, j, h]+`\
`45``phi1e[1,j,h]*phi2e[2,i,h]), j = i+1 .. 19), i = 1 .. 19) end do:`\
`tors[20,1]:= add(add(s[i j]*(phi1e0[1,i]*phi1e0[2, j]+phi1e0[1, j]*phi1e0[2, i]), `\
`j = i+1 .. 19), i = 1 .. 19): `\
`tors[20,2]:= add(add(s[i,j]*(phi2e0[1,i]*phi1e0[2, j]+phi2e0[1, j]*phi1e0[2, i]), `\
`j = i+1 .. 19), i = 1 .. 19):`\
`tors[20,3]:= add(add(s[i,j]*(phi1e0[1,i]*phi2e0[2, j]+phi1e0[1, j]*phi2e0[2, i]),`\
`j = i+1 .. 19), i = 1 .. 19):`\
`52``tors[21,1]:= add(add(s[i,j]*((delta[i]*a[1][i]-c[1][i])*`\
`(delta[j]*a[2][j]-c[2][j])+`\
`(delta[j]*a[1][j]-c[1][j])*(delta[i]*a[2][i]-c[2][i])), j = 1 .. 19), i = 1 .. 19):`\
`tors[21,2]:= add(add(s[i, j]*((delta[i]*a[1][i]-c[1][i])*a[1][i]*`\
`(delta[j]*a[2][j]-c[2][j])+`\
`(delta[j]*a[1][j]-c[1][j])*(delta[i]*a[2][i]-c[2][i])*`\
`a[1][j]), j = 1 .. 19), i = 1 .. 19):`\
`tors[21,3]:= add(add(s[i, j]*((delta[i]*a[1][i]-c[1][i])*a[2][j]*`\
`(delta[j]*a[2][j]-c[2][j])+`\
`(delta[j]*a[1][j]-c[1][j])*(delta[i]*a[2][i]-c[2][i])*`\
`a[2][i]), j = 1 .. 19), i = 1 .. 19):`\
`63``Eqskertau:= [seq(seq(primpart(tors[h,l]), l = 1 .. 3), h = 1 .. 21)]:`\
`Gausselim(linalg[stackmatrix](K, linalg[genmatrix](Eqskertau, listsij)),'r2')`\
`mod 131):`\
`r2;`\
` 171`\
`r2-r1;`\
` 63` ]{}
In lines 1–2, we define the $a_{i,j}$’s which will be used. We chose them randomly. In line 3, we collect the unknowns $\{s_{i,j}\}_{1\le i < j \le g}$ in the list `listsij`: there are $\binom{g-1}{2}$ of them. In line 4,5,6 we define $\texttt{t[1]}$ which is the vector $P_g$ as in (\[lemmadeltac\]), $A_i = \prod_{i=1,...,19} a_{r,i}$, $i=1,2$. In line 7,8 we define $d_2=1$, $d_1 = -\frac{d_2 A_1}{A_2}$, as in (\[deltac\]). In line 9 we define the vector $\delta$ whose components are $\delta_{i,1} = \delta_{i,2}$, as in (\[deltac\]). In lines 10,11 we collect the $c_{i,1}$, $c_{i,2}$ $i=1,...,19$ as in (\[deltac\]), and we call them $ \texttt{c[1][i]}$, $ \texttt{c[2][i]}$. These data give the curve $C$ and the line bundle $A$ as in (\[lemmadeltac\]). In lines 13–19, we define the matrix `Z` associated to the linear system , whose solutions give us the quadrics in $I_2(C)$, cf. Proposition \[Z\]. In line 20, Maple computes the rank `r0` of `Z` via Gaussian elimination, by calculating modulo 131 to speed up computations. The resulting matrix is called `Zref`. As expected by Proposition \[Z\], Maple finds $\texttt{r0}=38=2g-2$ and it prints it in line 21. In lines 22, 23 we define `M[j]` as $\prod_{i=1,...,19}(t-a_{i,j})$, $j=1,2$. In lines 24, 25 we define `phi1[j, i]` as the i-th component of $\frac{d}{dt}(\phi_j(t,1))$, $j=1,2$, where $\phi_j$ is defined in . In lines 26, 27 we define `R[k]` as the polynomial $R_k(t)$ of , $k=1,2$. In line 28, we collect in `EqsKerNu` the list of equations which determine $\ker(\nu)$, cf. the definition of $\nu$ in . In lines 29–31, Maple computes the rank `r1` of the linear system $\texttt{EqsKerNu} \cap\ker(\texttt{Zref})$, again via Gaussian elimination modulo 131, and the resulting matrix is called `K`. Maple finds that $\texttt{r1}=108$ and it prints it in line 31. Since $rank(\nu) = \texttt{r1}- \texttt{r0} = 70 = 2 h^0(\OO_{{\mathbb P}^1}(2g-6))$, for $g=20$, we have shown that $\nu$ has maximal rank. In line 32, we define the 2nd derivative `phi2` of the $\phi_j(t,1)$’s of . In lines 33-34 we define the evaluations `phi1e0`, `phi2e0`, of the first and the second derivatives of the $\phi_j(t,1)$’s at $t=0$, i.e. at the point $P_{20}$ and in lines 35-38 we define their evaluations at the points $P_i$, $i =1,...,19$. Using them, in lines 39–51 we compute the torsion at $P_i$, $i=1,\ldots,20$, cf. , and, in lines 52–62, the torsion at the point $P_{21}$, cf. . In lines 63 we collect in `EqsKerTau` the equations which determine $\ker(\tau)$ and Maple computes the rank `r2` of $\texttt{EqsKerTau} \cap\ker(\text{\texttt{K}})$, via Gaussian elimination modulo 131 as before. Maple finds that $\texttt{r2}=171$, therefore the rank of $\tau$ is $\texttt{r2}-\texttt{r1}=171-108=63= 3(g+1)= dim(T)$, $(g=20)$, hence also $\tau$ has maximal rank. So we have shown that $\mu_A$ is surjective for $g=20$.
[**[B. Results of computations in Thm.\[thm20\]]{}**]{}
Here we give the formulas of the determinants of the matrices $A$ and $B$ in the proof of Thm.\[thm20\]. The Maple files of these computations are available under request to the authors.
If $g=2k$, $det A=\frac{ -4(4k^5+14k^4+15k^3+k^2-7k+1)}{p_1}$ where $$p_1=
k^2 (k - 4)^3 (k + 1) (2 k - 1) (k - 1)^3 (k^2 - 4) (2 k^3 - 9 k^2
+ 12 k - 4) (k - 3)^2 ((2 k-1)! )^4 .$$
$det(B)=\frac{147456}{ 5 }(k-5)(k^4-9k^3+16k^2+3k-8)\cdot p_2\cdot p_3/(q_1 \cdot q_2),$ where $$p_2=16k^9-14k^8-87k^7+121k^6+75k^5-138k^4-52k^3+54k^2+18k-12,$$ $$p_3=6k^5-k^4-12k^3-4k^2+12k-4$$ $$q_1 = k^3(k-3)^3(k^2-4k+4)(-3k+k^2+2)(2k-1)(k-1)^8a(k-4)^7(2k^3-9k^2+12k-4)$$ $$q_2= (k^2-4)(2k^2-7k+6)((2k)!)^8$$
and one can check all the functions appearing in these expressions do not have any integral zero $k \geq 10$.
If $g=2k+1$, $ detA= \frac{-16}{15\cdot p_4}\cdot (k-4)p_5$ where $$p_4=(-2+k)^4(2k-1)k^2(k-3)^3(k-1)^6(k+2)(k+1)((2+2k)!^2)),$$ $$p_5=(1168k^{14}+2216k^{13}-22360k^{12}-41218k^{11}+17145k^{10}+47730k^9+46525k^8+38736k^7$$ $$-70488k^6-58080k^5+35288k^4+14726k^3-6093k^2+66k-465).$$
$ det(B)=-\frac{3072}{25}(k-5) \cdot p_6 \cdot p_7/p_8,$ where $$p_6=(270336+1257472k+25884500k^4-5217504k^2-15573704k^3-6492143k^{17}+68438542k^5$$ $$-28031103k^6-108784825k^7-49730235k^8-30298961k^9 +50987804k^{10}+197670424k^{11}+60883960k^{12}$$ $$-162484142k^{13}-9462204k^{18} -79976k^{19}+945456k^{20}+45632k^{21}-44288k^{22}-1216k^{23}+768k^{24}$$ $$-92612465k^{14}+54292657k^{15}+44402735k^{16}),$$ $$p_7=2+4k-k^2+(2k+1)!(k+1)(k^4-2k^3-k^2+12k+4),$$ $$p_8=ak^6(k-4)^5(k-1)(k^3-4k^2+5k-2)^3(2k^2-7k+3)(k^2-3k+2)^3(k+1)^3(k-3)^5(k+2)(2k-3)(4k^2-1)((2k+1)!)^7$$ and again one can check all these functions do not have any integral zero $k \geq 10$.
[99]{} D. Abramovich, A. Corti and A. Vistoli, [*[Twisted bundles and admissible coverings]{}*]{}, math.AG/0106211, Comm. Algebra **31** (2003), 3547-3618. E. Ballico, C. Casagrande and C. Fontanari, [*[Moduli of Prym curves]{}*]{}, Documenta Mathematica **9** (2004), 265-281. A. Beauville, [*[Prym varieties and the Schottky problem]{}*]{}, Inventiones Math. **41** (1977), 149-96. Calabri, A., Ciliberto, C., Miranda, R., The rank of the 2nd Gaussian map for general curves. Michigan Math. J., Vol. 60, no.3 (2011), 545-559. Caporaso, L., Brill Noether theory of binary curves, Math. Res. Letters, Vol. 17 no.2 (2010), 243-262.
Colombo, E., Frediani, P., Some results on the second Gaussian map for curves. Michigan Math. J. Vol. 58, 3 (2009), 745-758. Colombo, E., Frediani, P.,Siegel metric and curvature of the moduli space of curves. Transactions of the Amer. Math. Soc. 362 (2010), no. 3, 1231-1246. Colombo, E., Frediani, P.,On the second Gaussian map for curves on a K3 surface. Nagoya Math. J. Volume 199 (2010), 123-136. Colombo, E., Frediani, P., Pareschi, G., Hyperplane sections of abelian surfaces. J. Algebraic Geometry 21 (2012) 183-200. Colombo, E., Pirola, G.P., Tortora, A., Hodge-Gaussian maps, Ann. Scuola Normale Sup. Pisa Cl. Sci. (4) [**30**]{} (2001), no. 1, 125-146. Farkas, Gavril; Ludwig, Katharina, The Kodaira dimension of the moduli space of Prym varieties. J. Eur. Math. Soc. (JEMS) 12 (2010), no. 3, 755-795. Friedman, Robert; Smith, Roy, The generic Torelli theorem for the Prym map. Invent. Math. 67 (1982), no. 3, 473Ð490. Green, M. L., Infinitesimal methods in Hodge theory, in [*Algebraic Cycles and Hodge Theory*]{}, Torino 1993, Lecture Notes in Mathematics, 1594. Springer, Berlin, (1994), 1-92. Griffiths, Phillip; Harris, Joseph Principles of algebraic geometry. Pure and Applied Mathematics. Wiley-Interscience \[John Wiley & Sons\], New York, 1978. xii+813 pp. Kanev, V. I., A global Torelli theorem for Prym varieties at a general point. (Russian) Izv. Akad. Nauk SSSR Ser. Mat. 46 (1982), no. 2, 244-268, 431. Kobayashi, Shoshichi, Differential geometry of complex vector bundles. Publications of the Mathematical Society of Japan, 15. Kano Memorial Lectures, 5. Princeton University Press, Princeton, NJ; Iwanami Shoten, Tokyo, 1987. xii+305 pp. Simpson, Carlos T., Higgs bundles and local systems. Inst. Hautes ƒtudes Sci. Publ. Math. No. 75 (1992), 5-95.
[^1]: Partially supported by FAR 2010 (PV) “Varietà algebriche, calcolo algebrico, grafi orientati e topologici”, by INdAM (GNSAGA), and by MIUR of the Italian Government in the framework of the National Research Project “Geometria algebrica e aritmetica, teorie coomologiche e teoria dei motivi”(PRIN 2008). AMS Subject classification: 14H10, 14K12.
[^2]: Notice that between the formula of Cor 3.8 of [@cf2] and the formula given in Prop. \[curv\] there is a difference by a factor 4, due to a small mistake in Cor 3.8 of [@cf2] where we identified $\rho$ with $II$, while $II = -2 \rho$ ([@cpt] p.136).
[^3]: Substitute column $(2,[\frac{g}{2}])$ with $(2,[\frac{g}{2}])-(1,[\frac{g}{2}])$ then divide it by $a_{2,k}-a_{1,k}$; substitute any row by itself minus $a_{1,k}$ times the preceding row; substitute each column from $(1,3)$ to $(1,[\frac{g}{2}])$ by itself minus the first column, eliminate the first row and column and divide the column $(1,i)$ ($j=3,...,[\frac{g}{2}]$) by $a_{i,k}-a_{2,k}$; substitute each row by itself minus $a_{[\frac{g}{2}]+1,k}$ times the preceding row, eliminate the first row and the $(1,[\frac{g}{2}]+1)$-column and divide the column $(1,j)$ with $j=[\frac{g}{2}]+2,...,g-1$ by $a_{j,k}-a_{[\frac{g}{2}]+1,k}$; substitute any row by itself minus $a_{2,k}$ times the preceding row, eliminate the first row and the $(2,[\frac{g}{2}]+1)$-column and divide the column $(1,j)$ with $j=3,...,[\frac{g}{2}]$ by $a_{j,k}-a_{[\frac{g}{2}]+1,k}$.
[^4]: substitute each column except the first one and the last one by itself minus the first column, substitute the last one by itself minus the first column multiplied by the first coefficient of the last column, eliminate the first row and column and divide all the columns except the last one by $a_{j,k}-a_{3,k}$ and repeat.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Considering a planar gravitating thick domain wall of the $\lambda
\phi^4$ theory, we demonstrate how the Darmois junction conditions written on the boundaries of the thick wall with the embedding spacetimes reproduce the Israel junction condition across the wall when one takes its thin wall limit.\
\
KEY WORDS: Thick planar domain walls; Darmois junction condition; Thin planar domain walls; Israel junction condition.
author:
- |
S. Ghassemi$^{1}$[^1], S. Khakshournia$^{2}$[^2], and R. Mansouri$^{3}$[^3]\
\
$^{1,3}$Department of Physics, Sharif University of Technology,\
Tehran 11365-9161, Iran\
$^{2}$Nuclear Research Center, Atomic Energy Organization of Iran, Tehran, Iran
title: On the thin wall limit of thick planar domain walls
---
$$$$ $$$$
Introduction
============
Domain walls are solutions to the coupled Einstein-scalar field equations with a potential having a spontaneously broken discrete symmetry and a discrete set of degenerate minima. In the simplest case of two minima, a domain wall having a non-vanishing energy density appears in the separation layer, with the scalar field interpolating between these two values.\
Domain walls in the cosmological context have a long history [@vilen]. It was realized very early that the formation of domain walls with a typical energy scale of $\geq 1 MeV$ must be ruled out [@zel], because a network of such objects would dominate the energy of the universe, violating the observed isotropy and homogeneity. Domain walls were reconsidered in a possible late time phase transition scenario at the scale of $\leq 100 MeV$. Such walls were supposed to be thick because of the low temperature of the phase transition [@hill]. The suggestion that Planck size topological defects could be regarded as triggers of inflation, revived the discussion of thin and thick domain walls [@vil2]. The realization of our universe as a $(3+1)$-dimensional domain wall immersed in a higher dimensional spacetime has led to the recent numerous studies [@rand].\
The first attempts to investigate the gravitational properties of domain walls were based on the so called thin wall approximation. In this approach one forgets about the underlying field theory and simply treats the domain wall as a zero thickness (2+1)-dimensional timelike hypersurface embedded in a four-dimensional spacetime. The Israel thin wall formalism [@israel] is then used to continue the solutions of the Einstein equations on both sides of the wall in the embedding spacetime across the thin wall. However, such spacetimes have delta function-like distributional curvature and energy-momentum tensor supported on the hypersurface. Using this procedure, the first vacuum solutions for a spacetime containing a infinitely thin planar domain wall was found by Vilenkin [@vil3], and Ipser and Sikivie [@ipser]. The very interesting feature of such domain walls is that they are not static, but have a de Sitter-like expansion in the wall’s plane. External observers experience a repulsion from the wall, and there is an event horizon at finite proper distance from the wall’s core. These results were initially obtained within the framework of the Israel thin wall formalism which has been shown to be an approximative description of a real thick wall by using an expansion scheme in powers of the wall thickness [@widrow; @garfin]. Typically, a self gravitating domain wall has two length scale, its thickness $w$ and the distance to event horizon which can be compared to $w$. Since these lengths are expressed in terms of the coupling constants of the theory, thin walls turn out to be an artificial construction in terms of these underlying parameters as mentioned in [@gre99].\
The first exact dynamical solution to thick planar domain walls was obtained by Goetz [@goetz] and later a static solution was recovered with the price of sacrificing reflection symmetry [@mukh]. Within the context of a fully nonlinear treatment of a scalar field coupled to gravity, Bonjour, Charmousis and Gregory (BCG) found an approximate but analytic description of the spacetime of a thick planar domain wall of the $\lambda \phi^4$ model by examining the field equations perturbed in a parameter characterizing the gravitation interaction of the scalar field [@gre99]. Recently, the thin wall limit of Goetz’s solution was studied in [@rom] and it has been shown that this solution has a well-defined limit. But to date the thin wall limit of BCG’s solution has never been investigated.\
In this paper, we study the thin wall limit of the thick planar domain wall described by BCG spacetime. To do so, we use the formalism developed by Mansouri and Khakshournia (MK) [@khak02] to treat dust thick shells.\
The organization of this paper is as follows. In section 2 we give a brief introduction to BCG thick wall solution and summarize all the useful equations we will need in the present work . In Section 3 we describe the thick wall formalism of reference [@khak02] and apply it to the thick planar domain wall solved by BCG. Section 4 considers the thin wall limit of the thick domain wall solution followed by the conclusion.
The thick planar domain wall solution
=====================================
Domain walls as the regions of varying scalar field separating two vacua with different values of field are usually described by the matter Lagrangian: $$\label{Lagrangian}
L=\nabla_{\mu}\phi\nabla^{\mu}\phi-V(\phi),$$ where $\phi$ is a real scalar field and $V(\phi)$ is a symmetry breaking potential which we take to be $V(\phi)=\lambda(\phi^2-\eta^2)^2$, where $\lambda$ is a coupling constant and $\eta$ the symmetry breaking scale. Looking for a static solution of the equation of motion derived from this lagrangian in flat space-time, we get $$\label{phi}
X=\tanh(\frac z w),$$ where $X=\frac\phi\eta$, $w=\frac{1}{\sqrt{\lambda}\eta}$, and z is the coordinate normal to the wall. This particular solution represents an infinite planar domain wall centered at $z=0$. From the stress-energy tensor of the wall, one can easily observe that the wall energy density peaked around $z=0$ falls down effectively at $z=w$. So $w$ a length scale in the system is called the effective thickness of the wall within the theory.\
We now look at the planar gravitating domain wall solutions. The line element of a plane symmetric spacetime may be written in the general form $$\label{metric}
ds^2=A^2(z)dt^2-B^2(z,t)(dx^2+dy^2)-dz^2,$$ which displays reflection symmetry around the wall’s core located at z=0, where z is the proper length along the geodesics orthogonal to the wall. In order to obtain a thick domain wall solution one should solve the coupled system of the Einstein and scalar matter field equations as follows $$\label{Richi}
R_{\mu\nu}=8\pi
G\eta^2\left(2X_{,\mu}X_{,\nu}-\frac{1}{w^2}g_{\mu\nu}(X^2-1)^2\right),$$ $$\label{field}
\square X+\frac{2}{w^2}X(X^2-1)=0,$$ where $R_{\mu\nu}$ is the spacetime Ricci tensor. For a static field $X(z)$, Einstein equations (\[Richi\]) constraint $B(z,t)$ as $B(z,t)=A(z)\exp(kt)$.\
BCG investigated the spacetime of a thick gravitating planar domain wall for a $\lambda \phi^4$ potential [@gre99]. In the context of their work a dimensionless parameter $\epsilon$ arisen from equation (\[Richi\]) is singled out to characterize the coupling of gravity to the the scalar field namely $$\label{epsilon}
\epsilon=8\pi G\eta^2.$$ Supposing that gravity is weakly coupled to the scalar field, $A_0(z)$ and $X(z)$ may be expanded in the powers of $\epsilon$: $$\begin{aligned}
A(z)=A_0(z)+\epsilon A_1(z)+O(\epsilon^2),\\\
X(z)=X_0(z)+\epsilon X_1(z)+O(\epsilon^2).\end{aligned}$$ In the $\epsilon\rightarrow 0$ limit, the results should be the same as non-gravitating planar wall’s which are $A(z)=1$ and $X(z)=\tanh(\frac z w)$. Using these expansions, BCG solved the coupled Einstein and scalar matter field equations to first order in $\epsilon$ and obtained the following results: $$\label{inside metric}
A_i(z)=1-\frac{\epsilon}{3}\left[2\ln[\cosh(\frac z w)]+\frac 1 2
\tanh^2(\frac z w)\right]+O(\epsilon^2),$$ $$\label{inside metric2}
k_i=\frac 2 3\frac{\epsilon}{w}+O(\epsilon^2),$$ $$\label{inside metric3}
X_i(z)=\tanh(\frac z w)-\frac{\epsilon}{2}{\rm sech}^2(\frac z w
)\left[\frac z w+\frac 1 3 \tanh(\frac z w )\right]+ O(\epsilon^2).$$ Thus, (\[inside metric\]) is a perturbative solution to the spacetime of the thick wall (\[metric\]) obtained by BCG. In the following sections we will use this solution.
The thick wall formalism
========================
In this section we first make a short review of the thick wall formalism developed by MK in [@khak02]. Then we apply it to the thick planar domain wall described by the metric (\[metric\]) . In MK formalism a thick wall is modelled with two boundaries $\Sigma_1$ and $\Sigma_2$ dividing a spacetime $\cal M$ into three regions. Two regions $\cal M_{+}$ and $\cal M_{-}$ on either side of the wall and region $\cal M$$_{0}$ within the wall itself. Treating the two surface boundaries $\Sigma_1$ and $\Sigma_2$ separating the manifold $\cal M$$_{0}$ from two distinct manifolds $\cal M_{+}$ and $\cal
M_{-}$, respectively, as nonsingular timelike hypersurfaces, we do expect the intrinsic metric $h_{\mu\nu}$ and extrinsic curvature tensor $K_{\mu\nu}$ of $\Sigma_j\hspace{0.2cm}$ (j=1,2) to be continuous across the corresponding hypersurfaces. These requirements named the Darmois conditions are formulated as $$\label{hmn}
[h_{\mu\nu}]_{\Sigma_j}=0\hspace {1cm} j=1,2,$$ $$\label{thick israel}
[K_{\mu\nu}]_{\Sigma_j}=0\hspace {1cm} j=1,2,$$ where the square bracket denotes the jump of any quantity that is discontinuous across $\Sigma_j$. To apply the Darmois conditions on two surface boundaries of a given thick wall one needs to know the metrics in three distinct spacetimes $\cal M_{+}$, $\cal M_{-}$ and $\cal M$$_{0}$ being jointed at $\Sigma_j$. While the metrics in $\cal M_{+}$ and $\cal M_{-}$ are usually given in advance, knowing the metric in the wall spacetime $\cal M$$_{0}$ requires a nontrivial work.\
Let us now impose these junction conditions for a self gravitating thick planar domain wall described in the previous section. Recalling $w$ is the effective thickness of the wall, we first follow Ref. [@garfin] to introduce a parameter $\Delta \gg1$ to assure that the scalar field takes its vacuum values on the wall boundaries $\Sigma_1$ and $\Sigma_2$ being located at the proper distances $z=\pm \:\: \Delta w/2$ far from the wall’s core surface at $z=0$. We can then think of $\Delta w$ as the proper thickness of the planar domain wall. In the coordinate frame of the metric (\[metric\]) in which the wall is stationary, the nonvanishing components of the intrinsic metric $h_{\mu\nu}$ and extrinsic curvature $K_{\mu\nu}$ of $\Sigma_j$ take the following simple forms: $$\begin{aligned}
\label{hcompon}
h_{\mu\nu}=g_{\mu\nu},\hspace {1cm} \mu,\nu\neq z,\end{aligned}$$ $$\label{Kcompon}
K_{\mu\nu}=-\frac 1 2 g_{\mu\nu,z}.$$ In order to find the spacetime metric on both sides of $\Sigma_j$, we first note that within the vacuum region $\cal M_{+}$ ($\cal
M_{-}$) in which $\phi= \eta(-\eta)$, the spacetime metric can be easily determined by solving the Einstein equations (\[Richi\]) yielding $$\label{vacuum metric}
A_o(z)=-k_o |z|+ C.$$ Using (\[hcompon\]) and (\[Kcompon\]) we write down junction conditions (\[hmn\]) and (\[thick israel\]) as $$\label{k}
k_i=k_0,$$ $$\label{metric constraints1}
A_i(z)|_{z=\Delta w/2}=A_o(z)|_{z=\Delta w/2},$$ $$\label{metric constraints2}
\frac{\partial A_i(z)}{\partial z}|_{z=\Delta w/2}=\frac{\partial
A_o(z)}{\partial z}|_{z=\Delta w/2}.$$ We now use the solutions (\[inside metric\]) and (\[vacuum metric\]) due to BCG for the wall metric in the region $\cal
M$$_{0}$, where the scalar field varies according to (\[inside metric3\]), and for the metric in the vacuum regions, respectively. Then the junction conditions (\[metric constraints1\]) and (\[metric constraints2\]) lead to the following constraints on the vacuum metric constants $C$ and $k_o$ $$\label{c}
C=1+k_o\frac{\Delta w
}{2}-\frac{\epsilon}{3}\left(2\ln(\cosh(\Delta/2))+\frac 1 2
\tanh^2(\Delta/2)\right),$$ $$\label{ko}
k_o=\frac {\epsilon}{w}\left(\tanh(\frac{\Delta}{2})-\frac 1 3
\tanh^3(\frac{\Delta}{2})\right).$$ Note that within the context of BCG work, it is supposed that the boundaries of the wall where the scalar field takes its vacuum values are at infinity. But here we have modelled the thick planar wall in such a way that the wall boundaries $\Sigma_j$ are situated at the finite proper distances $\pm\Delta\,w/2$ from the core of the wall. Hence, we choose $\Delta$ to be sufficiently large in order to simulate BCG solution within our wall model.
From the thick to thin domain walls
===================================
We now turn our attention to the thin wall limit of our thick wall model. First let us define the process of passing from a thick gravitating domain wall to a thin one by letting $\epsilon$ and $w$ go to zero while keeping their ratio $\frac {\epsilon}{w}$ fixed. This has the effect that the distance of the event horizon to the domain wall remains finite.\
We then rewrite the Darmois junction condition (\[thick israel\]) demanding the continuity of the extrinsic curvature tensor $K_{\mu\nu}$ across the thick wall boundary, say $\Sigma_1$, located at the proper distance $z=\Delta
w/2$ as $$\label{kmunu}
K^o_{\mu\nu}|_{z=\Delta w/2}=K^i_{\mu\nu}|_{z=\Delta w/2}.$$ From the formula (\[Kcompon\]) one can evaluate the right hand side of the $(tt)$ component of the equation (\[kmunu\]) for the metric (\[metric\]) using the BCG wall metric solution (\[inside metric\]). This yields $$\label{ktt}
K^o_{tt}|_{z=\Delta
w/2}=\frac{\epsilon}{w}\left(\tanh(\frac{\Delta}{2})-\frac 1 3
\tanh^3(\frac{\Delta}{2})\right)\left(1-\frac{\epsilon}{3}\left[(2\ln[\cosh(\frac
\Delta 2)]+\frac 1 2 \tanh^2(\frac \Delta 2)\right]\right).$$ Imposing the above thin wall limit prescription, the equation (\[ktt\]) reduces to $$\label{ktt2}
K^o_{tt}|_{z=0}=\frac
{\epsilon}{w}\left(\tanh(\frac{\Delta}{2})-\frac 1 3
\tanh^3(\frac{\Delta}{2})\right).$$ To identify the right hand side of the equation (\[ktt2\]) we recall the definition of the surface energy density $\sigma$ of an infinitely thin wall. Within our thick wall model it takes the form $$\begin{aligned}
\label{sigmap}
\sigma=\lim_{(w\rightarrow 0, \epsilon \rightarrow 0)}\int^{\Delta
w/2}_{-\Delta w/2}\rho dz,\end{aligned}$$ where $\rho=\rho(z)$ is the energy density of the scalar field which is computed for the BCG scalar field solution (\[inside metric3\]). Finally, we get the following expression for $\sigma$ $$\begin{aligned}
\label{sigmap2}
\sigma=\frac{\epsilon}{2\pi G w}\left(\tanh(\frac{\Delta}{2})-\frac
1 3 \tanh^3(\frac{\Delta}{2})\right),\end{aligned}$$ where we used the definition $\epsilon$ given by (\[epsilon\]). Comparing the results (\[ktt2\]) and (\[sigmap2\]) one immediately obtains $$\begin{aligned}
\label{thin}
K^o_{tt}|_{z=0}=2\pi G\sigma.\end{aligned}$$ Now, consider a planer thin domain wall placed at $z=0$. The Israel thin wall approximation treats the wall as a singular hypersurface with the surface energy $\sigma$ separating the two plane symmetric vacuum spacetimes $\cal M_{+}$ and $\cal M_{-}$ from each other. Then the Israel junction condition across the wall is written as $$\label{israelp}
K^o_{\mu\nu}=2\pi G\sigma h_{\mu\nu}|_{z=0}.$$ Not surprisingly, we now see that the equation (\[thin\]) is just the same as the $(tt)$ component of the Israel’s equation (\[israelp\]), since from the equations (\[hcompon\]) and (\[vacuum metric\]) it follows that $h_{tt}|_{z=0}=A_o^2=C$, where in the thin wall limit, the equation (\[c\]) reduces to $C=1$. Using the condition (\[k\]) being held at $z=0$, one can easily show the same results for the $(xx)$ and $(yy)$ components of the equation (\[kmunu\]).
Conclusion
==========
We have studied the thin wall limit of the thick planar domain wall solution obtained by Bonjour, Charmousis and Gregory in Ref. [@gre99]. Treating the thick planar wall as a defect having two boundaries at the same proper distance from the wall’s core, as formulated by Mansouri-Khakshournia for the case of a dust shell in Ref. [@khak02], we have shown that the Darmois junction conditions for the extrinsic curvature tensor at the wall boundaries with the two embedding spacetimes generate the well-known Israel jump condition at the separating boundary of the corresponding thin wall with the same embedding spacetimes. We have realized that in the process of passing from a thick planar domain wall to the thin one all the information about the internal structure of the wall are squeezed in the parameter $\sigma$ characterizing the wall surface energy density as introduced in (\[sigmap2\]).\
\
[**Acknowledgement**]{}: We would like to thank S. Khosravi for useful discussions.
[99]{} A. Vilenkin, Phys.Rev. D23, 852 (1981). Ya, B. Zel’dovich,I. Yu. Kobzarev, and L. N. okun, Sov. Phys. JETP 40,1 (1974). C. T. Hill, D. N. Schramm, J. N. Fray, Comm. Nucl. Part. Sci. 19, 25 (1989). A. Vilenkin, Phys. Rev. Lett. 72, 3137 (1994).
L. Randall, R Sundrum, Phys. Rev. Lett. 83, 3370 (1999); 83, 4690 (1999). W. Israel, Nuovo Cimento B 44, 1 (1966). A. Vilenkin, Phys. Lett. 133B, 177 (1983). J. Ipser and P. Sikivie, Phys. Rev. D 30, 712 (1984). L. M. Widrow, Phys. Rev. D 39, 3571 (1989); 39, 3576 (1989). D. Garfinkle and R. Gregory, Phys. Rev. D 41, 1989 (1990). F. Bonjour, C. Charmousis and R. Gregory, Class. Quantum Grav. 16, 2427 (1999),\
\[gr-qc/9902081\]. G. Goetz, J. Math. Phys. 31, 2683 (1990). M. Mukherjee, Class. Quantum Grav. 10, 131 (1993). **16**, 2427 (1999). R. Guerrero, A. Melfo and N. Pantoja, Phys. Rev. D 65, 125010 (2002), \[gr-qc/0202011\]. S. Khakshournia and R. Mansouri, Gen. Rel. Grav. **34**, 1847 (2002), \[gr-qc/0308025\]
[^1]: e-mail: sima\[email protected]
[^2]: e-mail: [email protected]
[^3]: e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'During software front-end development, the work to convert *Graphical User Interface(GUI)* image to the corresponding front-end code is an inevitable tedious work. There have been some attempts to make this work to be automatic. However, the GUI code generated by these models is not accurate due to the lack of attention mechanism guidance. To solve this problem, we propose PixCoder based on an artificially supervised attention mechanism. The approach is to train a neural network to predict the style sheets in the input GUI image and then output a vector. PixCoder generate the GUI code targeting specific platform according to the output vector. The experimental results have shown the accuracy of the GUI code generated by PixCoder is over 95%.'
author:
- |
Xiaoling Huang$^a$, and Feng Liao$^a$\
$^a$ Dept. of Computer Science, Sun Yat-sen University, Guangzhou, China\
[email protected],
bibliography:
- 'ijcai18.bib'
title: Automatically Generating Codes from Graphical Screenshots Based on Deep Autocoder
---
Introduction
============
Using machine learning techniques to build an automatic programming system is a relatively new field. Automatic programming can be applied to many fields. For example, Automatic programming can help developers to implement some simple functions. Balog et al. and Riedel et al. have done related work[@DBLP:journals/corr/BalogGBNT16; @DBLP:journals/corr/RiedelBR16]. Balog et al. proposed DeepCoder, which take five input-output pairs as input, and output a program that satisfy these five input-output pairs. However, the program generated by DeepCoder is very simple, so Deepcoder is still far away from the actual application.
Moreover, automatic programming can prevent developers from time-consuming software front-end development. Beltramelli propose pix2code [@DBLP:journals/corr/Beltramelli17], which is a system that takes a GUI image as input and outputs GUI code. Pix2code consists encoding part and decoding part. The encoding part contains a *Convolutional Neural Network(CNN)* model and a RNN model. Pix2code accepts the GUI image and the corresponding DSLs as input. At this time, the CNN model is responsible for encoding the GUI image into a vector, and the RNN model is responsible for encoding a segment of the corresponding DSLs. And then input them to the decoding part. The decoding part is another RNN model. It generates the next character in the DSLs. By comparing the generated character with the character in the actual DSLs, a gradient can be obtained to optimize the parameters in pix2code. However, the result of pix2code is not satisfying since the code generation of pix2code lacks attention mechanism [@DBLP:journals/corr/LuongPM15]. As a result, the generation of every token is based on the whole image as shown in Figure \[contrast\]. In fact, each piece of the frond-end code corresponds to a specific block in the image. This led to low accuracy of the front-end code generated by pix2code.
{width="50.00000%"}
\[contrast\]
In this work, we focus on the problem in pix2code and propose PixCoder, which is a precise descriptive programming system using deep learning techniques. We apply the method named PixCoding, which applies an artificially supervised attention mechanism, to fix the problems in pix2code. In more detail, we first artificially encode the GUI image to a vector. As shown in Figure \[contrast\], Each bit in the vector is associated with some blocks in the image. In this way, we artificially establish the attention mechanism. As a result, the model generates code based on specific block of GUI image. And then We feed the images and vectors into a vision model based on CNN for supervised learning. During the training process, CNN learns the mapping between the style sheet in specific block of the GUI image and the specific bits in the vector. Finally, based on the predicted vector output by the vision model, we can directly generate the precise front-end code of GUI.
In summary, our concrete contributions are:
1. proposing an impressive automatic descriptive programming system. In some domains, the capacity of this system is comparable to humans.
2. defining a method in image recognition and classification. This method decomposes the original classification tasks into simple sub-classification tasks, which reduces the difficulty of the whole task and improves the classification accuracy.
We organize the paper as follows. We first review related work. After that, we present the formal details of our framework, and then give a detailed description of our algorithm. Finally, we evaluate PixCoder in three different datasets and conclude our work with a discussion on future work.
Related Work
============
As we mentioned, there has been prior work on automatic programming. In automatic functional programming field, Bunel et al. propose an adaptive neural-compilation framework to address the problem of efficient program learning [@DBLP:conf/nips/BunelDMKT16]. In the work of Riedel et al., through a neural implementation of the dual stack machine that underlies Forth, programmers can write program sketches with slots that can be filled with behaviour trained from program input-output data [@DBLP:journals/corr/RiedelBR16]. Ling et al. explored the generation of source code from a mixed natural language and structured program specification [@DBLP:conf/acl/LingBGHKWS16]. In the work of Gaunt et al., the source code is generated through a distinguishable interpreter that learns the relationship between input-output examples [@DBLP:journals/corr/GauntBSKKTT16]. Balog et al. propose DeepCoder, which leverages statistical predictions to generate computer programs [@DBLP:journals/corr/BalogGBNT16].
However, code generation with visual inputs is still an unexplored research area until Beltramelli propose pix2code [@DBLP:journals/corr/Beltramelli17]. The architecture of pix2code is similar to some models applied to other areas (c.f. [@DBLP:conf/cvpr/KarpathyL15; @DBLP:conf/cvpr/VinyalsTBE15; @DBLP:conf/icml/XuBKCCSZB15; @DBLP:conf/cvpr/DonahueHGRVDS15]). Due to the code generation of pix2code lacks attention mechanism [@DBLP:journals/corr/LuongPM15], the result of pix2code is far from expectation. For this problem, we apply an artificially supervised attention mechanism and adopt an approach similar to DeepCoder.
Most of these works rely on *Domain Specific Languages(DSLs)*. DSLs are programming languages that are designed for a specialized domain. Compare with full-featured programming languages, DSLs are more restrictive. As a result, DSLs limit the complexity of programming language, which makes automatic programming easier and makes special-purpose search algorithm efficient [@DBLP:conf/oopsla/PolozovG15]. In this work, we are only interested in the GUI layout, the different graphical components, and their relationships. Thus we choice a simple DSLs designed by Beltramelli [@DBLP:journals/corr/Beltramelli17].
Problem Definition
==================
The problem we focus on is designing an automatic descriptive programming system, which input is a GUI image ${\mathcal I}$ and the output is the front-end code ${\mathcal C}$ of the GUI. We denote $\alpha({\mathcal I})={\mathcal C}$ as the automatic descriptive programming process. So the problem can can be considered to find mapping $\alpha$ between ${\mathcal I}$ and ${\mathcal C}$.
According to different target platforms, ${\mathcal I}$ can be a web-based UI image, an iOS UI or an Android UI image. Figure \[imgdslcode\](a) is an input example of web-based UI image. Different UI images vary in size, so we resize the input image to 256$\times$256 pixels and the pixel values are normalized. After this process, we get standardized image ${\widetilde{\mathcal{I}}}$.
The output of the problem is the front-end code ${\mathcal C}$ of the GUI. However, as shown in Figure \[imgdslcode\](c), the front-end code is complicated. And it is difficult for the model to directly generate the front-end code. So our approach is designing DSLs ${\mathcal D}$ for target front-end code ${\mathcal C}$ and let model output ${\mathcal D}$ instead of ${\mathcal C}$. Figure \[imgdslcode\](b) is the DSLs of web-based UI image. The syntax of DSLs is relatively simple here, since we only need to use ${\mathcal D}$ to describe the outline of the GUI image.
Since the output of our model is ${\mathcal D}$ and the problem need ${\mathcal C}$. So we need to compile ${\mathcal D}$ to get ${\mathcal C}$. This process is denoted as $\beta({\mathcal D})={\mathcal C}$ and we denote $\delta({\widetilde{\mathcal{I}}})={\mathcal D}$ as the automatic descriptive programming process of model. The compile process is static so the key work is to find $\delta$, the mapping between ${\widetilde{\mathcal{I}}}$ and ${\mathcal D}$.
{width="42.50000%"}
\[imgdslcode\]
PixCoder
========
To implement such an automatic descriptive programming system, we have tried to add attentional mechanism to pix2code just like the work of Luong et al. [@DBLP:journals/corr/LuongPM15]. However, this did not improve the performance of the model. Therefore, we use a method named PixCoding to add a supervised attentional mechanism artificially. And in combination with the ideas of DeepCoder [@DBLP:journals/corr/BalogGBNT16], we propose PixCoder. The automatic descriptive programming process of PixCoder can be divided into the following phases: The first phase is a vision model based on CNN, which is used to understanding the given GUI image ${\mathcal I}$ and inferring the objects present. The vision model outputs a predicted vector ${\mathcal V}$ and each bit in ${\mathcal V}$ corresponds to a kind of style sheet in different block in ${\mathcal I}$. This can be seen as a vision model identifies each block in ${\mathcal I}$ and then record the result into ${\mathcal V}$. In the second phase, we exploit a parser in solving code generation problem. The parser leverages ${\mathcal V}$ generated by vision model to generate DSLs ${\mathcal D}$ describing ${\mathcal I}$. The last phase is a compiler. In this phase, with traditional compiler design techniques, ${\mathcal D}$ is compiled to the front-end code ${\mathcal C}$ targeting specific platform.
{width="40.00000%"}
\[architecture\]
PixCoding
---------
PixCoding is a universal image recognition method: we analyze the possible changes in different areas of the image and then devise a vector. Each bit in the vector corresponds to a different change in different areas of the image. And then we can make some rules based on our experience, according to which we are able to know that the changes in various combinations of different areas corresponds to which label. And then the task of image recognition can be completed. This image recognition method decomposes the image classification task into sub-classification tasks in different areas of the image, and then combines the results of the sub-classification tasks to obtain the final classification result. At this point, the accuracy of the classification depends on the accuracy of the sub-classification. Because sub-tasks are relatively simple, this method can improve the accuracy of the whole task.
In this work, we apply this method to automatic descriptive programming field. We design the corresponding vectors for a particular type of GUI image. Each bit in the vector corresponds to the changes of style sheet in different blocks in the GUI image. By combining each of the vectors we know all the style sheets in the GUI image. As a result, we can directly generate DSLs corresponding to the input GUI image. This process is no different from the process of combining the results of a sub-classification task to get the final classification result. In addition, PixCoding mentioned above can be applied to image text recognition, such as recognition of long strings. We only need to identify the different areas of the characters belong to which number, and then combined them to complete the string recognition task. This method can also be applied to motion recognition. We divide the hand, foot, torso and other parts of human into different areas and then encode the vectors so that the vectors imply all possible situations in these areas.
However, PixCoding has some limitations: it can only be used when the data in the dataset is very normative and has a clearly defined area.
**Input:** a GUI image ${\mathcal I}$.\
**Output:** the GUI’s front-end code ${\mathcal C}$.
resize ${\mathcal I}$ to 256$\times$256 pixels and normalized the pixel values. And get standardized image ${\widetilde{\mathcal{I}}}$: *Image\_Standardization*(${\mathcal I}$)=${\widetilde{\mathcal{I}}}$; input ${\widetilde{\mathcal{I}}}$ to vision model and get predicted vector ${\mathcal V}$: $\zeta({\widetilde{\mathcal{I}}})={\mathcal V}$; standardize ${\mathcal V}$ and get standardized vector ${\widetilde{\mathcal{V}}}$: *Vector\_Standardization*(${\mathcal V}$)=${\widetilde{\mathcal{V}}}$; generate DSLs ${\mathcal D}$ according to ${\widetilde{\mathcal{V}}}$: $\theta({\widetilde{\mathcal{V}}})={\mathcal D}$; compile ${\mathcal D}$ to get ${\mathcal C}$: $\beta({\mathcal D})={\mathcal C}$; ${\mathcal C}$;
Image To Vector
---------------
### Vector Design
First, we need to analyze and design a specific type of GUI image. An GUI image example from iOS UI dataset is shown in Figure \[blockinimg\]. After analyzing the iOS UI dataset, we learned that the GUI image here can be divided into two blocks: a stack block and a footer block (the area in the Figure \[blockinimg\] is surrounded by the black frame). There can be a maximum of eight row blocks in the stack block, and at least one row block (the area in the Figure \[blockinimg\] that is surrounded by the red frame). There are at most four controls in each row block, and there are five types of controls, and up to four controls in the footer block and four types of footer controls. If all possibilities are under consideration, we shall design a vector of 176 (4$\times$5$\times$8+4$\times$4) bits using one-hot encoding. In this case, the vector contains all the blocks and the changes of the style sheet of them and in the GUI image.
But if we do further analysis, we can prune the bits of the vector. After further analysis, we know that the number of row blocks is from 1 to 8; the number of controls in the row block is from 2 to 4. The number of controls in the footer block ranges from 2 to 4. The appearance of controls in the row block is somewhat regular: such that the slider does not appear alone, it always appears as a group like label, slider, label. According to the analysis results, we can prune the median of the vector, and finally get a 72-bit vector shown in Figure \[vector\].
Using the vector before and after pruning separately, we found that the use of pruned vectors can reduce training time and improve the accuracy of vision recognition. In addition, the pruned vector can also avoid some unusual circumstances. Such as the error of beyond the length: Due to the length of each control is different, so a row block cannot contain four longest slider controls. In the pruned vector, the situation which four sliders appear in a row block cannot happen. As a result, we can avoid this error of beyond the specified length, which in turn enhances our automatically descriptively programming system.
{width="40.00000%"}
\[blockinimg\]
### Vision Model
After designing the vector, we need a vision model to achieve the image-to-vector mapping $\zeta$. we employ a CNN to model and learn $\zeta$ due to it can learn rich latent representations from the image [@DBLP:conf/nips/KrizhevskySH12; @DBLP:journals/corr/SermanetEZMFL13]. CNN is one of the most widely used methods in vision problems. Compared with the traditional image processing algorithms, CNN avoids the complex pre-processing of images, which improves the efficiency of the experiment greatly.
Firstly, ${\mathcal I}$ are initially re-sized to 256$\times$256 pixels and normalized the pixel values. After this process, we get standardized image ${\widetilde{\mathcal{I}}}$ and fed it in CNN. To recognize different objects in input image, we exclusively used 3$\times$3 receptive fields which are convolved with stride 1 as VGGNet [@DBLP:journals/corr/SimonyanZ14a]. These operations are applied twice before to down-sample with max-pooling. The width of the first convolutional layer is 32, followed by a convolutional layer of width 64, and the width of the third convolutional layeris 128. Our CNN ends with two fully connected layers. The size of the first fully connected layer is 1024 and the size of the final connection layer is based on the complexity of the style sheets in given GUI images. Except for the final fully connected layer applying the softmax activation function, each layer of our model applying the rectified linear unit activation function.
The Training process of vision model is supervised. Since only ${\mathcal I}$ and ${\mathcal D}$ are provided in the dataset, after designing the vector, we need to generate ${\mathcal V}$ corresponding to ${\mathcal I}$ and use it as the label for the supervised training. After completing this pre-process, we can start to train the vision model.
-------------------- ----------------------- ---------------
Before pruning After pruning
web-based UI Image 1.08$\times$$10^{8}$ 3528
iOS UI Image 4.66$\times$$10^{25}$ 9072
Android UI Image 1.16$\times$$10^{28}$ 42768
-------------------- ----------------------- ---------------
: Patterns of GUI image statistics.[]{data-label="Number of all possible patterns"}
{width="40.50000%"}
\[varietyofvectorsize\]
Vector To DSLs
--------------
The central idea of our work is to use a vision model to guide code generation. The parser is responsible for generating DSLs based on vectors. The parser has two main functions: to standardize ${\mathcal V}$ output by the vision model, and to generate ${\mathcal D}$ based on the ${\widetilde{\mathcal{V}}}$($\theta({\widetilde{\mathcal{V}}})={\mathcal D}$).
As mentioned above, the vector we design contains all the stylistic changes in the GUI image. Therefore, we can know from the vector that the presence or absence of individual style sheets in individual block of the given input GUI image. In accordance with the previously designed DSLs syntax rules, we can generate the corresponding DSLs directly from the output vector.
However, ${\mathcal V}$ output by the vision model does not meet our expectations. First, the value of each bit in the output vector is not a standard value like 0 or 1. In addition, it is possible to have ambiguities in the output vectors, such as specifying conflicting patterns for a block, and so on. Therefore, before generating the corresponding DSLs based on vectors, we need to standardize ${\mathcal V}$ and get ${\widetilde{\mathcal{V}}}$. We can only generate ${\mathcal D}$ based on ${\widetilde{\mathcal{V}}}$.
{width="40.50000%"}
\[vector\]
### Vector Standardization
First, we need to use thresholds to process the value of each bit in ${\mathcal V}$ into a standard value like 0 or 1. Here we cannot use the method of setting the maximum value in each region of ${\mathcal V}$ directly to 1, and the rest of the value all to 0. Because this method does not consider the situation that ${\mathcal V}$ can have some regions of all 0. Using the iOS UI dataset as an example, with our method, for any input GUI image, the final output is a DSLs containing eight row blocks. Therefore, we need to use thresholds to deal with this problem.
In this work, we automatically get the threshold during the training of vision model. After each epoch after training halfway, we take a part of the samples from the training set and input them into the vision model to get the output vector. By comparing the output vector with the target vector, we can know which bit in the output vector correspond to 1 and which bit correspond to 0. The smallest one of the output values corresponding to 1 is recorded as a candidate for the threshold value. At the same time, we will also record the largest of the output values corresponding to 0. After training, We will apply a complex algorithm to get the threshold. For example, if the maximum value corresponding to 0 is smaller than the minimum one corresponding to 1, then we take the average of this interval as the threshold. Otherwise, We use other methods to get a reasonable threshold. According to Figure \[accthreshold\], The choice of threshold is important to the result of model. And the threshold generated by our algorithm is close to the best threshold.
For all possible ambiguities in ${\mathcal V}$, we have corresponding solution. For example, for the situation of specifying conflict patterns for a block, our approach is to take the larger one of the two models corresponding to the prediction and set it to 1, and the other set to 0. And for the situation that an empty row block in the middle of two row blocks, our approach is to swap the empty row block with next row block contains objects.
### DSLs Generation
After completing the process of standardizing ${\mathcal V}$, we can easily generate the corresponding DSLs according to ${\widetilde{\mathcal{V}}}$ based on the previously designed DSLs grammar rules($\theta({\widetilde{\mathcal{V}}})={\mathcal D}$). The normalized vector is error-free, which guarantees there will be no syntax error exists in the generated DSLs.
DSLs to Front-end code
----------------------
When we get ${\mathcal D}$ from parser, our work have not finished. We need a compiler to compile the ${\mathcal D}$ into ${\mathcal C}$ for various platforms(i.e. Android and iOS native mobile interfaces, and multi-platform web-based HTML/CSS interfaces). As mentioned, This process denotes as $\beta({\mathcal D})={\mathcal C}$. The compile process is static. Thanks to the parser guarantees ${\mathcal D}$ without syntax error, ${\mathcal C}$ generated by compiler is legal.
Experiments
===========
Dataset
-------
The dataset provided by Tony Beltramelli [@DBLP:journals/corr/Beltramelli17] has three types, including iOS GUI images and corresponding DSLs, Android GUI images and corresponding DSLs and web-based GUI images and corresponding DSLs. Each type of dataset is divided into two parts, including 1500 training set and 250 testing set. Each example contains an original image and the corresponding DSLs.
According to Figure \[varietyofvectorsize\], the complexity of these three datasets is increasing. The number of the style sheets of the GUI image are least in the Web-based UI dataset, the iOS UI dataset is the second, and the Android UI datasets is the most.
-------------- ------- ------ ------- ------
Image DSLs Image DSLs
web-based UI 1500 1500 250 250
iOS UI 1500 1500 250 250
Android UI 1500 1500 250 250
-------------- ------- ------ ------- ------
: Dataset statistics.[]{data-label="dataset"}
Criterion
---------
There are many ways to evaluate the quality of codes generated by the model. For example, screenshot the image showed by the generated GUI code, and then compare the similarities between this and the input image. Or, compare the similarities between the generated code and the target code. After experimental comparison, we evaluated the PixCoder with the second way: compare the similarities between the generated DSLs and the target DSLs.
In our experiment, every DSLs that in line with grammatical rules is in tree structure. We first convert the generated DSLs and the target DSLs to a tree, then using the similarities between tree structures to represent the similarities between DSLs. The classic algorithm used to match the similarity of traditional tree models is based on the similarity of edit distance [@DBLP:journals/jacm/Tai79], which allows cross-layer matching of tree nodes. Cross-layer matching and replacement may be useful when comparing the general tree similarity, but not suited to the tree model used in our experiment. The tree model of our experiment is similar to the HTML tree model. HTML tree model document label nodes will be read by the browser rendering to the screen, different root node corresponds to a different set of child nodes. Therefore, even if we replace the root node, the children will not be matched. So cross-layer matching and replacement cannot obfuscate the relationship between nodes in the HTML tree, which is rendered in the browser effect.
Therefore, we need another way to measure the similarity of the tree model in our experiment. Here we use the Simple Tree Matching(STM) algorithm [@DBLP:journals/spe/Yang91; @DBLP:journals/jcrd/He07]. STM bases on the principle of maximum matching, dynamic programming is used to calculate the maximum number of matching nodes of two trees, and then the similarity between two trees is obtained. This process does not allow cross-layer matching and node replacement, and requires that child nodes be completely ordered. If the root nodes of two subtrees do not match, the other nodes of the two subtrees are not considered, so as to achieve the effect of pruning. The time complexity of this algorithm is O($n^{2}$).
$$Similarity(T_1, T_2) = \frac{SimpleTreeMatching(T_1, T_2)}{(|T_1|+|T_2|)/2}$$
Above is the formula that calculates the similarity between the generated DSLs and the target DSLs. $T_{1}$ and $T_{2}$ represent the tree obtained by converting the generated DSLs and the target DSLs.
SimpleTreeMatching($T_{1}$, $T_{2}$) represents the maximum number of matched nodes for two trees. $|T_{1}|$ and $|T_{2}|$ represent the number of nodes in two trees. If and only if the maximum number of matching nodes of two trees is larger, the Similarity($T_{1}$, $T_{2}$) is larger, too. That is, SimpleTreeMatching($T_{1}$, $T_{2}$) is closer to 1. At this point the trees represent the generated DSLs and the target DSLs are more similar.
------------------ -------------- --------- ------------
web-based UI iOS UI Android UI
Baseline 62.882% 70.303% 65.825%
pix2code(beam 3) 76.905% 68.640% 54.644%
pix2code(greedy) 88.591% 87.621% 85.073%
PixCoder 98.699% 95.562% 98.177%
------------------ -------------- --------- ------------
: Experimental results in different dataset.[]{data-label="result"}
{width="40.50000%"}
\[accthreshold\]
Experimental Results
--------------------
We evaluate PixCoder in following aspects. We first compare the similarities between the target DSLs and the DSLs generated by PixCoder and Baseline. We then compare the similarities between the target DSLs and the DSLs generated by PixCoder and pix2code, specially, pix2code has two types, one in use of beam 3 method, and the other in use of greedy method. We use a randomly generated grammar-compliant DSLs with a similarity to the target DSLs of the test set as Baseline.
As shown in Figure \[accthreshold\], we can get the conclusion that the choice of threshold is important to the result of PixCoder. Threshold can not be too large or too small. In this experiment, we set threshold to 0.01, which is close to the best threshold according to the experimental results. As shown in Figure \[accthreshold\], the size of the dataset affects the result. As the dataset increases, the model can learn more knowledge about the style sheets of GUI image. As a result, the model is more capable. In this work, we feed all 1500 examples from dataset to the model while training.
The experimental results in the Table \[result\] are the average similarities of 250 samples in the test set obtained using baseline, pix2code and PixCoder respectively. As described above, the complexity of these three datasets is increasing. The average similarities of Baseline on these datasets do not reflect the complexity of the style sheet changes in the GUI image. We believe this is caused by the unevenly distributed style sheets of the GUI image in the test datasets. The main contrast model of PixCoder in this experiment is pix2code. Constrained by the experimental results, the DSLs generated by pix2code in use of beam 3 method has a lot of syntax errors. In the calculation of average similarities, our method is to write 0 for the similarity of DSLs that have syntax errors. This results in the similarities of pix2code in use of beam 3 method being particularly low. Pix2code in the use of the greedy method almost has no syntax errors, so its average similarities are relatively high. There is no syntax error in DSLs generated by PixCoder, and its accuracy is far better than pix2code. The difference in average similarities between Pix2code and PixCoder does not reflect the difference in the quality of the code they generate. In fact, the quality of the code generated by Pix2code is far from PixCoder. As shown in Figure \[result\], There is only one error(the area surrounded by the red frame) in the output of PixCoder and there are eight error in pix2code. But their difference in similarity is only 12.821%. In a word, to the best of our knowledge, PixCoder is the best model in automatic descriptive programming field.
{width="45.00000%"}
\[result\]
Conclusion
==========
In this paper, we present a method called PixCoding for image recognition and classification. PixCoding decomposes the image classification tasks into sub-classification tasks in different regions of the image. This process is an artificially supervised attention mechanism. We artificially design a vector to encode the image. The vector-image pairs guide CNN model to identify specific regions in the image and classify style sheets in these regions. In the end, following some rules, we can get the final classification results by integrating the results of sub-classification tasks. Our experimental results also show that PixCoding has an excellent effect on image recognition and classification tasks.
We apply the PixCoding to automatic programming field and propose PixCoder. PixCoder is an impressive automatic descriptive programming system, which takes an GUI image as input, and then generates the corresponding front-end code. Compared to other automatic descriptive programming systems, PixCoder produces extremely accurate GUI code that is close to human levels.
However, the GUI image in the dataset we used are simpler than the actual GUI image. Therefore, how to make PixCoder still perform excellent in more complex datasets is our future work. Moreover, we also interested in how to apply PixCoding in other image classification tasks.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Time-division SQUID multiplexers are used in many applications that require exquisite control of systematic error. One potential source of systematic error is the pickup of external magnetic fields in the multiplexer. We present measurements of the field sensitivity figure of merit, effective area, for both the first stage and second stage SQUID amplifiers in three NIST SQUID multiplexer designs. These designs include a new variety with improved gradiometry that significantly reduces the effective area of both the first and second stage SQUID amplifiers.'
author:
- 'G.M. Stiehl, [^1] H.M. Cho, [^2] G.C. Hilton, K.D. Irwin, J.A.B. Mates, C.D. Reintsema, [^3] and B.L. Zink [^4]'
bibliography:
- 'Stiehl\_IEEE.bib'
title: 'Time-division SQUID multiplexers with reduced sensitivity to external magnetic fields'
---
=1
Gradiometry, Multiplexer, SQUID, Transition Edge Sensor.
Introduction
============
recent years, large arrays of superconducting Transition Edge Sensors (TESs) have been implemented in various microcalorimetric and bolometric detector schemes for millimeter and sub-millimeter wavelengths, as well as single-photon detection of optical [@Cab00], x-ray [@Ull05] and gamma-ray [@Dor07; @Zin06] wavelengths. Superconducting QUantum Interference Devices (SQUIDs) are the readout amplifier of choice for TES detectors due to their low noise, low impedance and low power dissipation. Practical readout of TES arrays requires multiplexing at the cold stage to reduce the power load and minimize wiring complexity. At the National Institute of Standards and Technology (NIST) we have developed various time-domain SQUID multiplexers [@Irw04] that have been implemented in bolometric, kilopixel cameras such as the Atacama Cosmology Telescope (ACT) [@Fowler10] and the second iteration of the Background Imaging of Cosmic Extragalactic Polarization telescope (BICEP-2) [@orlando:77410H]. Experiments such as ACT and BICEP-2 use the SQUID multiplexer designs termed MUX06a and MUX07a. Experiments such as BICEP-2, SPIDER [@montroy:62670R] and ACTPol [@niemack:77411S] require exquisite control of systematic error in order to resolve the polarization of the Cosmic Microwave Background (CMB). A potential source of systematic error is SQUID pickup from external magnetic fields, which can be synchronous with a telescope’s rotation through Earth’s magnetic field. We have therefore designed a new SQUID multiplexer, the MUX09a, with the goal of reducing sensitivity to external magnetic fields through improved gradiometry in the SQUID input transformers.
An important figure of merit for SQUID sensitivity to extraneous magnetic fields is the effective area, $A_{eff}=\Phi/B$, where $\Phi$ is the flux coupled to the SQUID by uniform DC magnetic field $B$. The effective area of the first stage SQUID amplifier in the NIST multiplexer is of particular interest, as scan synchronous magnetic pickup from the second stage amplifier can be servoed out [@Niemack08]. However, the combined $V$-$\Phi$ response of the first and second stage SQUIDs in the multiplexer designs makes effective area measurements of the first stage SQUID difficult. In this paper, we describe a measurement technique for separating the first and second stage effective areas for SQUID multiplexers. We report effective areas of first and second stage SQUIDs for NIST SQUID multiplexer designs MUX06a, MUX07a and MUX09a. These values show that the improved gradiometry in the MUX09a input transformers has significantly reduced the sensitivity to external magnetic fields.
NIST Time-Division SQUID Multiplexer Designs
============================================
![A schematic of a 2x2 subset of the NIST SQUID multiplexer circuit. Each column has two first stage SQUIDs (SQ1) with inputs inductively coupled to a TES. The output of the SQ1s are transformer-coupled to a second stage SQUID (SQ2), which is then read out by a SQUID series array and room-temperature electronics. The four shaded areas represent distinct TES pixels.[]{data-label="fig_sim"}](SMUX){width="3.5in"}
In this section, we briefly describe the NIST time-division SQUID multiplexer circuit and discuss the features that determine effective area contributions in the MUX06a, MUX07a and MUX09a designs.
General NIST Time-Division Multiplexer Design
---------------------------------------------
The NIST time-division SQUID multiplexer circuit consists of 33 first stage SQUIDs (SQ1s), each inductively coupled to a unique TES detector. These SQUIDs are biased sequentially so that only one SQ1 is on at any time. The response from SQ1 couples to the second stage SQUID (SQ2) through a superconducting inductive summing coil that runs the length of the multiplexer chip. This combines the $V$-$\Phi$ response curve of SQ1 and SQ2. The combined SQUID response is further amplified by a SQUID Series Array (SSA) and room-temperature amplifier electronics for readout. A schematic of the multiplexer is shown in Figure \[fig\_sim\]. The NIST time-division SQUID multiplexer circuit is described in greater detail elsewhere [@Dor04].
Design Differences: Gradiometry and Input Transformers
------------------------------------------------------
![A picture of the MUX09a SQUID Gradiometer. For size reference, the inner coil of one of the lobes is 40 $\mu m$ across.[]{data-label="MUX09a_SQ1_top"}](MUX09a_SQ1_top "fig:"){width="3.25in"}\
Since the development of gradiometric SQUIDs at NIST in 1971 [@Zimm71], it has become common practice to use gradiometry in order to reduce a SQUID amplifier’s coupling to the external environment. A SQUID gradiometer consists of pickup coils wound in such a way that external magnetic fields do not couple flux into the SQUID. Clover-leaf SQUID gradiometers similar to those developed at PTB [@Drung07] are used throughout the NIST multiplexer designs. Figure \[MUX09a\_SQ1\_top\] shows the clover-leaf gradiometer used in the MUX09a SQ1.
![A diagram of the half-loop input coil path around the SQUID gradiometer. The arrows show the direction of current in the input coil. The inner circles represent the SQUID gradiometer lobes, and the plus and minus signs give the magnetic coupling polarity. The cross-hatching shows the area susceptible to coupling from external magnetic fields.[]{data-label="HalfLoop"}](halfloopv6 "fig:"){width="3.45in"}\
![A diagram of the whole-loop input coil path around the SQUID gradiometer. The arrows show the direction of current in the input coil. The wider trace is the incoming current path and the thin trace is the return line. The difference in trace size is only for ease of viewing. The inner circles represent the SQUID gradiometer lobes, and the plus and minus signs give the magnetic coupling polarity. Notice how in this configuration flux coupled through the area outlined by the wider trace is canceled by the flux coupled through the area outlined by the thin trace.[]{data-label="WholeLoop"}](WholeLoopv3 "fig:"){width="3.45in"}\
In order to couple a signal to the gradiometric SQUID pickup coils, an input coil is wound such that the polarity of the flux coupled into the SQUID matches the polarity of each gradiometer lobe. One such layout is shown in Figure \[HalfLoop\]. The input coil windings in this figure make half-loop turns of alternating polarity around the SQUID gradiometer lobes in order to couple flux into the SQUID. The MUX06a and MUX07a SQUID multiplexer circuits use half-loop input coil geometries similar to that shown in Figure \[HalfLoop\]. The MUX06a uses a $4\frac{1}{2}$ turn variation of Figure \[HalfLoop\] on SQ1, and the MUX07a uses a $1\frac{1}{2}$ turn variation on SQ1. Both designs use a $4\frac{1}{2}$ turn input coil geometry on SQ2. If input coil geometries like that shown in Figure \[HalfLoop\] are part of a superconducting transformer, screening currents induced in the input coil by an extraneous uniform DC field will couple flux into the SQUID. The new design, the MUX09a, utilizes whole-loop input coil geometries like that shown in Figure \[WholeLoop\]. The SQ1 has a 2 turn input coil and the SQ2 has a 4 turn input coil. This configuration minimizes coupling to uniform external fields.
The poor gradiometry of the half-loop input coil contributes to the effective area only if the coil is part of a closed superconducting loop, such as a transformer. There is no such input transformer on the first stage SQUID in the MUX06a. Therefore, we do not expect the input coil to couple flux from a uniform DC magnetic field into the SQUID. The MUX06a does however have a superconducting input transformer on the second stage SQUID (the summing coil). Thus, we expect a significant contribution to the effective area of the second stage. The MUX07a design utilizes both the half-loop input coil geometry and input transformers for SQ1 and SQ2. Thus, we expect to see a large effective area for each amplification stage. The MUX09a also has input transformers on SQ1 and SQ2. However, with the whole-loop input coil geometry we do not expect a significant effective area contribution from either amplification stage.
Measurement Scheme
==================
A liquid helium immersion probe used for SQUID characterization was adapted to make effective area measurements. A solenoid is mounted axially to the cold end of the probe and applies a uniform DC magnetic field perpendicular to the SQUID multiplexer. The solenoid consists of 2100 turns and is 0.2667 meters in length. A high-permeability magnetic shield mounts around the solenoid. The cold end of the probe is submersed in liquid helium.
As the $V$-$\Phi$ response curves of SQ1 and SQ2 are combined upon readout, measurements of the separate contributions to effective area by the first and second stage SQUID circuits can be cumbersome. We therefore devised a method of separating the effective area contributions through the use of the NIST digital feedback electronics [@Rei03] and the multiplexed nested feedback loop scheme shown in Figure \[ClosedLoopScheme\].
![A timing diagram of the scheme used in the closed-loop methodology for deconvolving SQ1 and SQ2 effective area contributions. The row period is about 10 microseconds and the arrows depict where the error signal is sampled for the corresponding feedback system.[]{data-label="ClosedLoopScheme"}](NestedFBloop3 "fig:"){width="3.45in"}\
In this two-row scheme, both SQ1 and SQ2 are closed in flux-locked loops using the same error signal sampled at different times. In the first row of the scheme (row 0), the error signal is sampled and feedback flux is applied to SQ2 prior to SQ1 being turned on. SQ1 is then turned on, the error signal is sampled again and feedback flux is applied to SQ1.
In this configuration, we first measure the flux from the applied external magnetic field coupling into the summing coil and SQ2. The feedback electronics null out the extraneous flux by applying a signal to the SQ2 feedback. The amount of feedback flux required to accomplish this is exactly the amount that is coupled into the SQUID by the magnet. This is the SQ2 contribution to the effective area.
Once the SQ1 is turned on, the flux from the applied field couples into the SQ1. The error signal is sampled again, and feedback is applied to the SQ1 to null the applied flux. This is the SQ1 contribution to the effective area. We thus separate out the SQ1 and SQ2 effective areas.
The solenoid is stepped through 100 different DC bias values. At each solenoid bias many thousands of flux data points are taken and then averaged. The average values for SQ2 feedback and SQ1 feedback are recorded. A linear fit is applied to the feedback flux as a function of applied magnetic field for both SQ1 and SQ2. The absolute value of the slope from each linear fit determines the effective area.
The magnet current bias values is converted to field values using an equation for the field inside a solenoid: $$B = \frac{V}{R_{\rm{bias}}}\frac{2100}{0.2667\rm{m}}\mu_{o}
\label{BfieldEQ}$$ The magnet calibration was checked with a Gauss meter at higher values of magnetic field. There was no measurable deviation from Equation \[BfieldEQ\]. The magnet response is thus assumed to be linear at smaller current values.
Results
=======
![Separated effective area contribution of SQ1 for the MUX06a, MUX07a and MUX09a.[]{data-label="SQUID1EffA"}](SQUID1EffA.pdf "fig:"){width="3.45in"}\
![Separated effective area contribution of SQ2 for the MUX06a, MUX07a and MUX09a.[]{data-label="SQUID2EffA"}](SQUID2EffA.pdf "fig:"){width="3.45in"}\
We performed effective area measurements for SQ1 and SQ2 in the MUX06a, MUX07a and MUX09a. The voltages across the SQ1 feedback and SQ2 feedback lines are recorded and converted into flux as follows: $$\Phi_{\rm{SQ Fb}} = (V/R_{\rm{FB}})M_{\rm{SQ FB}}
\label{SQ1Feedback}$$ where $M_{\rm{SQ Fb}}$ is the feedback mutual inductance and $R_{\rm{FB}}$ is the bias resistor for the SQUID feedback line.
Figures \[SQUID1EffA\] and \[SQUID2EffA\] show the amount of flux plotted as a function of applied magnetic field for SQ1 and SQ2 (respectively) for all multiplexer designs. These plots also show a linear fit to the data. The absolute value of the slopes for the linear fits are the effective areas and are tabulated in Table \[EffaTable\].
-------- -- ------------- -- -------------
SQ1 SQ2
($\mu m^2$) ($\mu m^2$)
MUX06a 5.0 468.5
MUX07a 882.6 482.7
MUX09a 0.6 21.5
-------- -- ------------- -- -------------
: Effective Areas for SQUID multiplexers
\[EffaTable\]
A variation in effective area values of a few square micrometers from chip to chip and row to row has been observed. An exhaustive study of this variation and possible row-position dependence is planned but has not yet been conducted.
Conclusion
==========
We have successfully measured SQ1 and SQ2 effective areas for three NIST SQUID multiplexer designs: MUX06a, MUX07a and MUX09a. From the results, it is clear that the largest contributions to effective area are due to poor gradiometry in the superconducting transformers used in the SQUID multiplexer circuit. Changing the input coil from half-loop to whole-loop gradiometry in the MUX09a multiplexer has made a significant impact by reducing the effective area of SQ2 by factor of 20, and more importantly, by reducing SQ1 by approximately three orders of magnitude when compared to the MUX07a. As a result, the MUX09a is far less sensitive to extraneous magnetic fields than earlier designs. This should significantly reduce the systematic error introduced by pickup from external magnetic fields in instruments that utilize NIST SQUID multiplexers.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors thank D.A. Bennett, W.B. Doriese, R.D. Horansky, G.C. O$'$Neil, M.D. Niemack, and D.R. Schmidt for their insight into SQUID electronics and measurement techniques. We also thank L. Ferreira and T. Sundby for their continued technical support. G.M. Stiehl thanks those involved with the PREP program both at NIST and the University of Denver for their support.
[^1]: Manuscript received 3 August 2010
[^2]: G.M. Stiehl, H.M. Cho, G.C. Hilton, K.D. Irwin, J.A.B. Mates, and C.D. Reintsema are with the National Institute of Standards and Technology, Boulder, CO 80305 USA (contact G.M. Stiehl at 303-497-5215 or [email protected])
[^3]: B.L. Zink is with the University of Denver, Denver, CO 80210 USA
[^4]: Contribution of NIST, not subject to copyright
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Gunnar Bali,\
Institut für Theoretische Physik, Universität Regensburg,\
93040 Regensburg, Germany\
E-mail: ,\
- '(QCDSF Collaboration)'
title: 'Potentials between pairs of static-light mesons'
---
Introduction
============
Potentials between static-light mesons ($\mathcal{B}=Q\bar{q}$) are of interest since they give insights in the nature of strong interactions from first principles for multiquark systems. For large heavy quark masses, e.g., the spectra of heavy-light mesons are determined by excitations of the light quark and gluonic degrees of freedom. In particular, the vector-pseudoscalar splitting vanishes and the static-light meson $\mathcal{B}$ can be interpreted as either a $\overline{B}$, a $\overline{B}^*$, a $D$ or a $D^*$ heavy-light meson. Calculating potentials between two $\mathcal{B}$ mesons then will also enable investigations of possible bound tetraquark states or for particles that are close to the meson-antimeson thershold, such as the $X(3872)$ or the $Z^{+}(4430)$. Therefore, $\mathcal{BB}$ and $\mathcal{B\overline{B}}$ potentials have been studied by many groups. First calculations were performed by Michael and Pennanen [@Michael:1999nq; @Pennanen:1999xi]. A more detailed quenched study can be found in ref. [@Detmold:2007wk]. For the computation of static-light meson-antimeson systems we refer to [@Sesam]. Recent dynamical simulations with twisted mass fermions were carried out by Wagner in refs. [@Wagner:2010ad; @Wagner:2011ev] and with Sheikholeslami-Wohlert fermions in our Lattice 2010 proceedings [@Bali:2010xa].
Computation
===========
In these proceedings we present the latest results from our investigations of potentials between two static-light mesons. Of special interest is the question of attraction and repulsion and their dependence on the separation between the static quarks. Therefore, we numerically determine ground and excited states of ${\mathcal B}$ mesons as well as intermeson potentials between pairs of static-light mesons, ${\mathcal B}({\mathbf r}){\mathcal B}({\mathbf 0})$ and ${\mathcal B}({\mathbf r})\overline{\mathcal B}({\mathbf 0})$. The static quark-quark (or quark-antiquark) separation is given by $r=|{\mathbf r}|=Ra, R\in \mathbb{N}_0$. $a$ denotes the lattice spacing, $Q$ a static colour source and the positions of the mass-degenerate light quarks $q \in\{u,d\}$ are not fixed. Thus, ${\mathcal B}$ mesons carry isospin $I=1/2$ and for $\mathcal{BB}$ and $\mathcal{B\overline{B}}$ states $I \in \{0,1\}$ with $I_z\in \{-1,0,1\}$. For ${\mathcal B}$ mesons the isosinglet corresponds to the representation $q_1 q_2 = ud + du$ and the isotriplet to $q_1 q_2 \in \{uu, dd, ud-du\}$ respectively. The corresponding representations for $\mathcal{B\overline{B}}$ states are given by $\bar{q}_1 q_2 = \bar{u}u + \bar{d}d$ for $I=0$ and $\bar{q}_1 q_2 \in \{\bar{d}u, \bar{u}d, \bar{u}u-\bar{d}d\}$ for $I=1$. Graphically, the quark line diagrams with respect to isospin that we evaluate can be depicted as, $$\begin{aligned}
\mbox{Isospin $I=0$:} & \quad C_{\mathcal{BB}}(t) = \;\;
\parbox{.06\textwidth}{
\begin{fmffile}{diag_para}
\begin{fmfgraph*}(20,27)
\fmfstraight
\fmfkeep{static-light}
\fmfleft{lu,lo}
\fmfright{ru,ro}
\fmfbottom{lu,um,ru}
\fmf{fermion}{lu,lo}
\fmf{fermion}{ru,ro}
% \fmflabel{$\leftarrow$ $R$ $\rightarrow$}{um}
\fmf{wiggly,tension=0,left=0.5}{lo,lu}
% \fmf{phantom_arrow,tension=0,left=0.5}{lo,lu}
\fmf{wiggly,tension=0,right=0.5}{ro,ru}
% \fmf{phantom_arrow,tension=0,right=0.5}{ro,ru}
\end{fmfgraph*}
\end{fmffile}},
&&
C_{\mathcal{B\overline{B}}}(t) = \; \;
\parbox{.06\textwidth}{
\begin{fmffile}{diag_antipara}
\begin{fmfgraph*}(20,27)
\fmfstraight
\fmfkeep{static-light}
\fmfleft{lu,lo}
\fmfright{ru,ro}
\fmfbottom{lu,um,ru}
\fmf{fermion}{lu,lo}
\fmf{fermion}{ro,ru}
% \fmflabel{$\leftarrow$ $R$ $\rightarrow$}{um}
\fmf{wiggly,tension=0,left=0.5}{lo,lu}
% \fmf{phantom_arrow,tension=0,left=0.5}{lo,lu}
\fmf{wiggly,tension=0,left=0.5}{ru,ro}
% \fmf{phantom_arrow,tension=0,left=0.5}{ru,ro}
\end{fmfgraph*}
\end{fmffile}}
-\;2\;\;
\parbox{.06\textwidth}{
\begin{fmffile}{diag_box}
\begin{fmfgraph*}(20,27)
\fmfstraight
\fmfkeep{static-light}
\fmfleft{lu,lo}
\fmfright{ru,ro}
\fmf{fermion}{lu,lo}
\fmf{fermion}{ro,ru}
\fmf{wiggly}{lu,ru}
\fmf{wiggly}{lo,ro}
\end{fmfgraph*}
\end{fmffile}}, \\
\mbox{Isospin $I=1$:} & \quad C_{\mathcal{BB}}(t) = \;\;
\parbox{.06\textwidth}{
\begin{fmffile}{diag_para}
\begin{fmfgraph*}(20,27)
\fmfstraight
\fmfkeep{static-light}
\fmfleft{lu,lo}
\fmfright{ru,ro}
\fmfbottom{lu,um,ru}
\fmf{fermion}{lu,lo}
\fmf{fermion}{ru,ro}
% \fmflabel{$\leftarrow$ $R$ $\rightarrow$}{um}
\fmf{wiggly,tension=0,left=0.5}{lo,lu}
% \fmf{phantom_arrow,tension=0,left=0.5}{lo,lu}
\fmf{wiggly,tension=0,right=0.5}{ro,ru}
% \fmf{phantom_arrow,tension=0,right=0.5}{ro,ru}
\end{fmfgraph*}
\end{fmffile}}
-\;\;
\parbox{.06\textwidth}{
\begin{fmffile}{diag_cross}
\begin{fmfgraph*}(20,27)
\fmfstraight
\fmfkeep{static-light}
\fmfleft{lu,lo}
\fmfright{ru,ro}
\fmf{fermion}{lu,lo}
\fmf{fermion}{ru,ro}
\fmf{wiggly}{lo,ru}
\fmf{wiggly}{lu,ro}
\end{fmfgraph*}
\end{fmffile}},
&&
C_{\mathcal{B\overline{B}}}(t) = \; \;
\parbox{.06\textwidth}{
\begin{fmffile}{diag_antipara}
\begin{fmfgraph*}(20,27)
\fmfstraight
\fmfkeep{static-light}
\fmfleft{lu,lo}
\fmfright{ru,ro}
\fmfbottom{lu,um,ru}
\fmf{fermion}{lu,lo}
\fmf{fermion}{ro,ru}
% \fmflabel{$P_{+}\;\;P_{-}$}{um}
\fmf{wiggly,tension=0,left=0.5}{lo,lu}
% \fmf{phantom_arrow,tension=0,left=0.5}{lo,lu}
\fmf{wiggly,tension=0,left=0.5}{ru,ro}
% \fmf{phantom_arrow,tension=0,left=0.5}{ru,ro}
\end{fmfgraph*}
\end{fmffile}} .\end{aligned}$$ Straight lines represent static quark propagators and wiggly lines light quark propagators.
volume $L_{\sigma}^3\times L_{\tau}$ $\beta$ $\kappa_{\mathrm{val}} = \kappa_{\mathrm{sea}}$ $c_{\mathrm{SW}}$ $a/\mathrm{fm}$ $La/\mathrm{fm}$ $m_{\mathrm{PS}}/\mathrm{MeV}]$ $N_{\mathrm{conf}}$
-------------------------------------- --------- ------------------------------------------------- ------------------- ----------------- ------------------ --------------------------------- ---------------------
$16^3\times32$ $5.29$ $0.13550$ $1.9192$ $0.084$ $1.34$ $770(9)$ $180$
$24^3\times48$ $5.29$ $0.13620$ $1.9192$ $0.077$ $1.85$ $400(4)$ $200$
: Lattice parameters.[]{data-label="tab:lattice"}
Masses are calculated from the asymptotic behavior of Euclidean-time correlation functions. We employ $N_{\mathrm f}=2$ Sheikholeslami-Wohlert configurations generated by the QCDSF Collaboration [@AliKhan:2003br]. The parameter values are listed in table \[tab:lattice\], where the scale is set using $r_0(\beta,\kappa)=0.5$ fm. The pseudoscalar mass corresponds to its infinite volume value. We use the Chroma software system [@Edwards:2004sx]. For the techniques and improvement methods we use, we refer to [@Bali:2010xa] and the references therein. To analyze our data and to extract also excited states we apply the variational method [@Vari], solving a generalized eigenvalue problem for a $3\times 3$ cross correlation matrix generated by different amounts of Wuppertal smearing [@Wuppertalsmear] applied to the source and sink operators. Errors are calculated using the jackknife method.
Representations and classification of states
============================================
To create $\mathcal{B}$ meson states as well as $\mathcal{BB}$ and $\mathcal{B\overline{B}}$ systems of different $J^{P(C)}$ we use interpolators $\mathcal{B}=Q\mathcal{O}\bar{q}$, where the operators $\mathcal{O}$ contain combinations of Dirac $\gamma$-matrices and covariant lattice derivatives. This has been discussed in [@Bali:2010xa] and we give a summary for the representation and classification of our states.\
In the continuum limit, the static-light states can be classified according to fermionic representations $J^P$ of the rotation group ${\mathrm O(3)}$. At vanishing distance ${\mathbf r}={\mathbf 0}$ the ${\mathcal B}{\mathcal B}$ and ${\mathcal B}\overline{\mathcal B}$ states can be characterized by integer $J^P$ and $J^{PC}$ quantum numbers, respectively. However at $r=|{\mathbf r}|>0$ the $\mathrm{O(3)}$ (or $\mathrm{O(3)}\otimes{\mathcal C}$) symmetry is broken down to its cylindrical $\mathrm{D_{\infty h}}$ subgroup. The irreducible representations of this are conventionally labeled by the spin along the axis $\Lambda$, where $\Sigma,\Pi,\Delta$ refer to $\Lambda=0,1,2$, respectively, with a subscript $\eta=g$ for gerade (even) $PC=+$ or $\eta=u$ for ungerade (odd) $PC=-$ transformation properties with respect to the midpoint. All $\Lambda\geq 1$ representations are two-dimensional. The one-dimensional $\Sigma$ representations carry an additional $\sigma_v=\pm$ superscript for their reflection symmetry with respect to a plane that includes the two endpoints.
**$\mathcal{O}$** wave [@Michael:1998sg] $\mathrm{O_h}'$ rep. continuum $J^{P}$ $J^{P}$ (heavy-light)
----------------------------------------------------------------------- ------------------------ ---------------------- ------------------- -----------------------
$\gamma_5$ $S$ $G_1^{+}$ $\frac{1}{2}^{+}$ $ 0^{-}, 1^{-}$
$\Eins$ $P_-$ $G_1^{-}$ $\frac{1}{2}^{-}$ $ 0^{+}, 1^{+}$
$\gamma_i\nabla_i$ $P_-$ $G_1^{-}$ $\frac{1}{2}^{-}$ $0^{+}, 1^{+}$
$\left( \gamma_1\nabla_1 - \gamma_2\nabla_2 \right) + \mathrm{cycl.}$ $P_+$ $H^{-}$ $\frac{3}{2}^{-}$ $ 1^{+}, 2^{+}$
: Operators and representations for static-light mesons. In the last column we display the $J^P$ for a heavy-light meson, obtained by substituting the (spinless) static source by a heavy fermion.[]{data-label="tab:operators_sl"}
The operators that we used to create the static-light mesons are displayed in table \[tab:operators\_sl\]. The intermeson potentials were obtained by combining two static-light mesons of different (or the same) quantum numbers. This can be projected into an irreducible $\mathrm{D_{\infty h}}$ representation, either by coupling the light quarks together in spinor space [@Wagner:2010ad] or by projecting the static-light meson spins into the direction $\hat{\mathbf r}$ of the static source distance, by applying $\frac12(\Eins\pm i\gamma_5\pmb{\gamma} \cdot \hat{\mathbf{r}})$, and taking appropriate symmetric ($\Lambda_z=1$) or antisymmetric ($\Lambda_z=0$) spin combinations. These two approaches can be related to each other via a Fierz transformation. For our coarse lattice and the operator combinations that couple to total angular momentum $J=0$ we have performed this projection. For the other combinations and our fine lattice different representations will mix. The analyzed operators and the corresponding representations are listed in table \[tab:operators\_potential\]. We note that the operator combinations $\gamma_5 \times \gamma_5$ and $\Eins \times \Eins$ carry the same quantum numbers as well as the combinations $\gamma_5 \times \Eins$ and $\gamma_5 \times \gamma_i\nabla_i$.
Results
=======
The eigenvalues $\lambda^{(k)}(t,t_0)$ of the generalized eigenvalue problem [@Vari], are fitted to one- and two-exponential ansätze, to obtain the $k$th mass. The appropriate values of $t_0$ and the fit ranges in $t$ are determined from monitoring the effective masses as described in [@Bali:2010xa]. Let us first discuss $\mathcal{BB}$ meson systems. We define intermeson potentials as the differences between the meson-meson energy levels and the $r\rightarrow\infty$ two static-light meson limiting cases: $$\begin{aligned}
V_{\!\mathcal{B}_1\mathcal{B}_2}(r)=E_{\!\mathcal{B}_1\mathcal{B}_2}(r) -
\left(m_{\!\mathcal{B}_1}+ m_{\!\mathcal{B}_2}\right)\quad\stackrel{r\rightarrow\infty}{\longrightarrow}\quad 0\,.\end{aligned}$$ In the left panel of figure \[fig:mp\_diff\_spin\_isospin\], we display the ground state ($\Sigma_g^+$) and the first excited state ($\Sigma_g^{+\prime}$) of the $\gamma_5 \otimes \gamma_5$ operator as well as the $\Sigma_u^-$ ground state and the $\Sigma_u^{-\prime}$ first excited state of the $\gamma_5 \otimes \Eins$ operator, both in the $I=1,\Lambda_z=0$ channel. In the last case the lowest lying $r\rightarrow \infty$ combination of states would be a radially excited $\frac12^{+\prime}$ state ($\gamma_5)$ plus a $\frac12^-$ ($\Eins$) ground state. The next level would be the sum of $\frac12^+$ and $\frac12^{-\prime}$. It is not clear to which one of these states our creation operator has best overlap. In the figure we display both possibilities. The latter assignment would mean that in the excited state channel (like for the ground state) there is repulsion at intermediate distances.
In the right panel of figure \[fig:mp\_diff\_spin\_isospin\] we show the $\gamma_5 \otimes \gamma_5$ operator in the different spin $\Lambda_z=0,1$ and isospin $I=0,1$ channels for the ground state. For short distances we observe attraction in all spin and isospin channels. In fact at very short distances we find attraction in all analyzed channels, see table \[tab:operators\_potential\], for ground and excited states. This may not be too surprising as this is expected from gluon exchange in the $3^*$ channel between the two static sources. When comparing the ground state of the $\gamma_5 \otimes \gamma_5$ operator for different spin and isospin channels we figure out that the $\Lambda_z=0$ channel is more attractive than the $\Lambda_z=1$ channel for isospin $I=0$. For isospin $I=1$ this pattern is reversed. In both cases the difference is of the order of $150$ MeV at a distance of $0.11$ fm. For the other spin-projected operator combination ($\Eins\otimes\Eins$, $\gamma_5\otimes\Eins$ and $\gamma_5\otimes\gamma_i\nabla_i$) we also find attractive forces of similar sizes for the $\Lambda_z=0,1$ and isospin $I=0$ ground states while for isospin $I=1$ the $\Lambda_z=1$ channel is more attractive than the $\Lambda_z=0$ channel.
On our coarse lattice we observe repulsive potentials at distances between $0.2$ fm and $0.45$ fm for the ground state of the $\gamma_5 \otimes \Eins$ operator in all spin and isospin channels. In addition we find repulsion in the $I=1,\Lambda_z=1$ channel for the $\Sigma_g^+$ ground state of the $\gamma_5 \otimes \gamma_5$ operator at distances between $0.3$ fm and $0.45$ fm. In the case of the fine lattice we did not perform the $\Lambda_z$ projection so that here we cannot distinguish between $\Sigma$ and $\Pi$ states. In agreement with the coarse lattice results we obtain repulsion of $O(50\,\,\mathrm{MeV})$ in the ground states of the operator combinations $\gamma_5 \otimes \Eins, \gamma_5 \otimes \nabla_i\gamma_i\, (\Sigma_u^-)$ and $\gamma_5 \otimes \gamma_5\, (\Sigma_g^+)$ for isospin $I=1$ at intermediate distances $r>0.35$ fm.
In figure \[fig:mp\_diff\_quark\_masses\] we compare coarse and fine lattice results in the $\gamma_5 \otimes \gamma_5$ ($\Sigma_g^+$ ground state) and $\gamma_5 \otimes \Eins$ ($\Sigma_u^-$ ground state) channels. In the first channel we observe reasonable scaling while in the latter channel the fine lattice potentials appear to be more attractive. This may be related to the lighter pion mass resulting in a different Yukawa interaction.
In figure \[fig:mp\_antipara\_box\] we display the $I=0$, $\Lambda_z=0$ ground state for the $\mathcal{B}\overline{\mathcal{B}}$ meson-antimeson case in the $\gamma_5 \otimes \gamma_5 \, (\Sigma_g^+$) channel. On the left hand side we see the effective ground state energy levels $E_{\mathrm{eff}}$ for different $t_0$ of the $\gamma_5 \otimes \gamma_5$ operator combination at a distance of $0.11$ fm. One finds a short plateau of poor quality in a range $t/a\in\{4,\ldots,8\}$ for $t_0>2a$. Then the effective energy level decreases again and forms another plateau from $t/a=8$ onwards. This can be explained by the observation that this state has the same quantum numbers as the $Q\overline{Q}$ static potential (and multiparticle states of the static potential plus a $P$ wave $\eta$ meson, the static potential plus 2 pions etc.). Our interpolator basis however, has very little overlap with these states. Therefore, we cannot easily disentangle the $Q\overline{Q}$ static potential and this background of multiparticle excitations from the lowest lying ${\mathcal B}\overline{\mathcal B}$ state that we are interested in. The ansatz to rewrite the correlator as $C_{ij}(t) = C^{\mathcal{B\overline{B}}}_{ij}(t) + d_{ij}\cdot \exp(-V_{Q\overline{Q}}(t)\cdot t)$ and extract $C^{\mathcal{B\overline{B}}}_{ij}(t)$ failed as well. Thus, we tried to fit the ground state $\mathcal{B\overline{B}}$ mass without using the variational approach, but from our single correlation function that has the largest overlap with the $\mathcal{B\overline{B}}$ ground state. Masses could be extracted for separations $r<0.26$ fm. For larger distances the coupling to the $\mathcal{B\overline{B}}$ state could not be resolved. The result is displayed in the right panel of figure \[fig:mp\_antipara\_box\] (blue circles). The two horizontal lines correspond to twice the ground state mass of the $\frac12^+$ static-light meson, the expected $r\rightarrow\infty$ limit. At first sight there appear to be very substantial short distance attractive forces in this channel. Also the static potential $V_{Q\overline{Q}}$ is lying much lower and can be disentangled. However, states consisting of a static potential and a scalar $I=0$ particle will have the same quantum numbers. For our lattice parameters the $P$-wave pseudoscalar $\eta$ meson (that at our quark mass will have a similar mass to the pion) is the lowest such state, with masses of a $f_0$ meson as well as two pseudoscalars lying higher. We include these sums in the figure where we approximate $m_{\eta}$ by $m_{\pi}$ and $m_{f_0}$ by $m_{a_0}$. The ground state ${\mathcal B}\overline{\mathcal B}$ lies between these states and the static quark potential. So it is hard to decide whether we see a substantial attraction between the static-light meson-antimeson pair in this channel or bound states between the static quark potential and additional light mesons.
Conclusions
===========
We investigated interactions between pairs of static-light mesons and found attraction for short distances in all spin and isospin channels. For distances of the order of $0.4$ fm some operator combinations yield repulsion, in particular the combination $\gamma_5 \otimes \Eins$. The interaction ranges are larger on the fine lattice with a smaller pion mass than on the coarse lattice. Meson-antimeson potentials are also very interesting with respect to charmonium threshold states [@Brambilla:2010cs] ($D\overline{D}$ molecules or tetraquarks) but difficult to disentangle from mesons that are bound to the static potential (hadro-quarkonium [@Dubynskiy:2008mq]). Analyzing $I=0, \Lambda_z=0$ ${\mathcal B}\overline{\mathcal B}$ states is a very challenging task since they couple directly to the $Q\overline{Q}$ static potential, e.g. in the case of the $\gamma_5 \otimes \gamma_5$ combination, or to the vacuum state, e.g. for the $\gamma_5 \otimes \Eins$ channel.
We thank Sara Collins, Christian Ehmann, Christian Hagen and Johannes Najjar for their help. We also thank our collaborators from the QCDSF Collaboration for generating the gauge ensemble. The computations were mainly performed on Regensburg’s Athene HPC cluster. We thank Michael Hartung and other support staff. We acknowledge support from the GSI Hochschulprogramm (RSCHAE), the Deutsche Forschungsgemeinschaft (Sonderforschungsbereich/Transregio 55) and the European Union grant 238353, ITN STRONGnet.
[99]{} C. Michael and P. Pennanen \[UKQCD Collaboration\], *Two heavy-light mesons on a lattice*, *Phys. Rev. D* **60** (1999) 054012 \[[arXiv:hep-lat/9901007](http://arxiv.org/abs/hep-lat/9901007)\]. P. Pennanen, C. Michael and A. M. Green \[UKQCD Collaboration\], *Interactions of heavy-light mesons*, *Nucl. Phys. Proc. Suppl.* **83** (2000) 200 \[[arXiv:hep-lat/9908032](http://arxiv.org/abs/hep-lat/9908032)\]. W. Detmold, K. Orginos and M. J. Savage, *$BB$ potentials in quenched lattice QCD*, *Phys. Rev. D* **76** (2007) 114503 \[[arXiv:hep-lat/0703009](http://arxiv.org/abs/hep-lat/0703009)\]. G. S. Bali, H. Neff, T. Düssel, T. Lippert and K. Schilling, *Observation of string breaking in QCD*, *Phys. Rev. D* **71** (2005) 114513 \[[[arXiv:hep-lat/0505012]{}](http://arxiv.org/abs/hep-lat/0505012)\]. M. Wagner \[ETM Collaboration\], *Forces between static-light mesons*, [arXiv:1008.1538 \[hep-lat\]](http://arxiv.org/abs/1008.1538). M. Wagner \[ETM Collaboration\], *Static-static-light-light tetraquarks in lattice QCD*, [arXiv:1103.5147 \[hep-lat\]](http://arxiv.org/abs/1103.5147). G. S. Bali and M. Hetzenegger \[QCDSF Collaboration\], *Static-light meson-meson potentials*, *PoS* **LAT2010** 142 [arXiv:1011.0571 \[hep-lat\]](http://arxiv.org/abs/1011.0571). A. Ali Khan *et al.* \[QCDSF Collaboration\], *Accelerating the hybrid Monte Carlo algorithm*, *Phys. Lett. B* **564** (2003) 235 \[[arXiv:hep-lat/0303026](http://arxiv.org/abs/hep-lat/0303026)\]. R. G. Edwards and B. Joó, *The Chroma software system for Lattice QCD*, *Nucl. Phys. Proc. Suppl.* **140** (2005) 832 \[[[hep-lat/0409003]{}](http://arxiv.org/abs/hep-lat/0409003)\]; C. Michael, *Adjoint sources in Lattice Gauge Theory*, *Nucl. Phys. B* **259** (1985) 58. S. Güsken *et al.*, *Non-singlet axial vector couplings of the baryon octet in lattice QCD*, *Phys. Lett. B* **227** (1989) 266. C. Michael and J. Peisa, *Maximal variance reduction for stochastic propagators with applications to the static quark spectrum*, *Phys. Rev. D* **58** (1998) 034506 \[[arXiv:hep-lat/9802015](http://arxiv.org/abs/hep-lat/9802015)\]. N. Brambilla *et al.*, *Heavy quarkonium: progress, puzzles, and opportunities*, [arXiv:1010.5827 \[hep-ph\]](http://arxiv.org/abs/arXiv:1010.5827). S. Dubynskiy and M. B. Voloshin, *Hadro-charmonium*, *Phys. Lett. B* **666** (2008) 344 \[[arXiv:0803.2224 \[hep-ph\]](http://arxiv.org/abs/0803.2224)\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The transition metal-oxygen bond appears prominently throughout chemistry and solid-state physics. Many materials, from biomolecules to ferroelectrics to the components of supernova remnants contain this bond in some form. Many of these materials’ properties strongly depend on fine details of the TM-O bond and intricate correlation effects, which make accurate calculations of their properties very challenging. We present quantum Monte Carlo, an explicitly correlated class of methods, to improve the accuracy of electronic structure calculations over more traditional methods like density functional theory. We find that unlike s-p type bonding, the amount of hybridization of the d-p bond in TM-O materials is strongly dependant on electronic correlation.'
address: |
Center for High Performance Simulation and Department of Physics\
North Carolina State University, Raleigh, NC 27695
author:
- 'Lucas K. Wagner [^1]'
bibliography:
- 'review.bib'
title: Transition metal oxides using quantum Monte Carlo
---
Introduction
============
Transition metal chemistry is a particularly exciting area of research, with applications from astrophysics to biology to potential inexpensive high-efficiency solar cells and high-temperature superconductivity. Because of the partially filled d-shell, transition metals can form many types of bonds and can also exhibit ferroelectric and ferromagnetic ordering. Transition metal oxides are particularly interesting because they are one of the most common transition metal complexes, and exhibit most of the above effects. This rich physics is quite difficult to describe theoretically, however, since electronic correlation is very strong in these materials. Current approximate density functional theories tend to perform quite poorly on transition metals, particularly in comparison to its quite good accuracy on elements with s and p type bonding. Problematic quantities are not hard to find; they include the dipole moment in molecules, binding (or cohesive energies), the lattice constants of perovskites, high pressure behavior, and band gaps/excitation energies.
Rather than attempting to improve the approximate density functional, quantum Monte Carlo (QMC) approaches take a different direction–explicitly treating the electronic correlation in a wave function based approach, while maintaining reasonable scaling with system size. It can be made to scale from O(1) to O(N$^3$) in the number of electrons[@qmc_ordern], depending on the quantity of interest. QMC attains very low upper-bound energies on medium-sized electronic problems (up to thousands of electrons at the time of writing), and has been used as a benchmark method on s-p systems[@jeff_benchmark]. Since it treats the electronic correlation explicitly in the many-body wave function, it is a promising method for strongly correlated TMO systems.
The goal of this review is to summarize the current state of the art of QMC as applied to TMO’s. This is a fairly new field, with few calculations. Most of these calculations have benchmarked the method to determine the accuracy that one should expect. This accuracy has generally been quite high on most of the quantities studied, particularly for energetics. In the course of this work, it has also been determined what trial function (starting guess, as explained in the methods section) is necessary to obtain this accuracy. The upper-bound property of diffusion Monte Carlo has been critical in this success. By this, we have also learned that the electronic correlation in transition metal oxides is entangled with the d-p orbital hybridization in these materials.
Quantum Monte Carlo
===================
The most common flavors of Quantum Monte Carlo that have been used on TMO’s are variational, diffusion, and reptation Monte Carlo (VMC, DMC, and RMC, respectively). We will summarize them here; one can find a more complete review in Ref [@Foulkes_review]. Another flavor, auxillary field Monte Carlo[@zhang_afqmc], has been used in a few calculations, but will not be discussed here.
VMC is a direct application of the variational theorem. We write the many-body wave function as a function of many-body coordinates ${{\bf R}}=[ {{\bf r}}_1,{{\bf r}}_2,\ldots,{{\bf r}}_{N_e}]$ and a set of variational parameters ${{\bf P}}$. One then approximates the ground state wave function by minimizing the energy expectation value $$E({{\bf P}})=\int \Psi^*({{\bf R}},{{\bf P}}) H \Psi({{\bf R}},{{\bf P}}) d{{\bf R}},$$ assuming that the wave function is normalized. For a complicated variational ansatz such as we will introduce later, this integral cannot be evaluated analytically. One can, however, evaluate it using Monte Carlo by rearranging the integral to read $$E({{\bf P}})=\int |\Psi({{\bf R}},{{\bf P}})|^2 \frac{H \Psi({{\bf R}},{{\bf P}})}{\Psi({{\bf R}},{{\bf P}})} d{{\bf R}}.$$ Since $|\Psi({{\bf R}},{{\bf P}})|^2$ is a probability distribution function, one can sample it using Markov chain Monte Carlo and evaluate the energy expectation value as an average over the local energy $E_L({{\bf R}})=\frac{H\Psi({{\bf R}})}{\Psi({{\bf R}})}$. The lowest-energy approximate wave function is then found by minimizing the energy. In practice, a combination of energy and the variance of the local energy[@umrigar_optimization2] or variance only[@umrigar_varopt] is optimized.
Many wave functions can be used with VMC, since the only requirement is that one can evaluate the wave function and its derivatives quickly. For the work covered in this article, we start with a Slater determinant of one-particle orbitals, $D$, or a linear combination of Slater determinants. We then multiply $D$ by the explicitly correlated inhomogeneous Jastrow correlation factor $e^U$ to obtain the Slater-Jastrow variational wave function $De^U$. We write $$U= \sum_{ijI} u(r_{iI},r_{jI},r_{ij})$$ where the lower case indices stand for electronic coordinates, and the upper case indices are ionic coordinates. There is considerable choice on how to expand $u$; for concreteness, we show one expansion that performs well enough and has been applied to TMO’s. The correlation factor is expanded in the Schmidt-Moskowitz form[@schmidt:4172]: $$\begin{aligned}
u(r_{iI},r_{jI},r_{ij})=\sum_k c_k^{ei}a_k(r_{iI}) + \sum_m c_m^{ee} b_k(r_{ij}) \\
+ \sum_{klm} c_{klm}^{eei} (a_k(r_{iI})a_l(r_{jI})+a_k(r_{jI})a_l(r_{iI}))b_k(r_{ij}),\end{aligned}$$ where the $a_k$ and $b_k$ functions are written as $$\frac{1-z(r/r_{cut})}{1+\beta_k z(r/r_{cut})},$$ with different $\beta_k$ for the different types of functions. The polynomial $z(x)=x^2(6-8x+3x^2)$ is chosen so the functions go smoothly to zero at $r_{cut}=$7.5 bohr. The $\beta_k$’s and all the expansion coefficients $c^{ei}$,$c^{ee}$, and $c^{eei}$ are optimized. If there are multiple determinants, their coefficients can also be optimized. We then use the VMC wave function as a trial function for RMC or DMC.
DMC and RMC are based on the so-called imaginary time Schrödinger equation $$-\frac{d\Psi({{\bf R}},\tau)}{d\tau}=(H-E_0)\Psi({{\bf R}},\tau),
\label{eqn:imagtime_diff}$$ which has a steady-state solution $\Phi_0$, the lowest energy eigenfunction with eigenvalue $E_0$ as long as $\Psi({{\bf R}},0)$ has a non-zero overlap with $\Phi_0$. All non-steady-state solutions converge exponentially to the eigenstate $\Phi_0$ as $\tau$ goes to infinity. Transforming to an integral equation, we have $$\Phi_0({{\bf R}}_1)=\lim_{\tau \to \infty} \int G({{\bf R}}_1,{{\bf R}}_0,\tau) \Psi_T({{\bf R}}_0) d{{\bf R}}_0,
\label{eqn:imagtime}$$ where $G$ is the Green’s function of the imaginary time Schrödinger equation and $\Psi_T({{\bf R}}_0)$ is the trial wave function that we obtain from VMC. Solving for the exact $G$ for large $\tau$ is as difficult as solving for $\Phi_0$, so we choose some constant small value of $\tau$ for which we know $G$ accurately (for example, see Refs [@Foulkes_review; @unr]), and compound the operations (suppressing the $\tau$ dependence of $G$): $$\Phi_0({{\bf R}})=\lim_{n \to \infty} \int G({{\bf R}},{{\bf R}}_n)\ldots G({{\bf R}}_1,{{\bf R}}_0)\Psi_T({{\bf R}}_0) d{{\bf R}}_0 d{{\bf R}}_1 \ldots d{{\bf R}}_n.$$ Each application of $G$ is interpreted as a stochastic process, in the same way that the diffusion equation can be mapped onto Brownian particles and vice versa (in fact, for a free particle, the Hamiltonian is $-\frac{1}{2}\nabla^2$ and the simulation is a diffusion process).
DMC performs a simulation of these random particles for large $n$. All implementations of DMC use a particularly clever importance sampling transformation by multiplying the imaginary time Schrödinger equation (Eqn \[eqn:imagtime\_diff\]) by the trial function $\Psi_T({{\bf R}})$ and working with the time-dependent function $\Psi_T({{\bf R}})\Psi({{\bf R}},\tau)$. Since the time dependence is the same, it eventually obtains samples distributed according to the probability distribution function $P_{R_\infty}({{\bf R}})=\Phi_0({{\bf R}})\Psi_T({{\bf R}})$. This transformation improves the efficiency of the calculation by several orders of magnitude[@Foulkes_review] by using information that we already have about the ground state in the form of a trial function. The final probability distribution function can be used to evaluate the ground-state energy as follows: $$\langle E_0 \rangle = \int d{{\bf R}}\Psi_T({{\bf R}})\Phi_0({{\bf R}}) \frac{H\Psi_T({{\bf R}})}{\Psi_T({{\bf R}})},$$ since $\Phi_0$ is an eigenstate of $H$ and $H$ can operate forwards or backwards. Any operators that do not commute with the Hamiltonian will have expectation values that are biased, only becoming unbiased in the limit of $\Psi_T=\Phi_0$.
We can remove the error in these operators by using reptation Monte Carlo[@Baroni_RMC; @pierleoni_rmc], where the random walk is performed in the space of paths: $s=[{{\bf R}}_0, {{\bf R}}_1, \ldots,{{\bf R}}_{n-1}, {{\bf R}}_n]$. We sample the path probability distribution $$\Pi(s)=\Psi_T({{\bf R}}_0) G({{\bf R}}_0,{{\bf R}}_1)\ldots G({{\bf R}}_{n-1},{{\bf R}}_n) \Psi_T({{\bf R}}_n)
$$ This can be interpreted in several different ways. If we examine the distribution at ${{\bf R}}_0$, we can view the samples of Green’s functions as acting on $\Psi_T({{\bf R}}_n)$, and therefore $P_{R_0}({{\bf R}}_0)=\Psi_T({{\bf R}}_0)\Phi_0({{\bf R}}_0)$. This is the same distribution as we obtain in DMC as the path length goes to infinity. Alternatively, since $G$ is symmetric on exchange of the two ${{\bf R}}$ coordinates, the probability distribution of ${{\bf R}}_n$ is the same. Finally, we can split the path in two, one projecting on $\Psi_T({{\bf R}}_0)$, and the other projecting on $\Psi_T({{\bf R}}_n)$. We then have $$\begin{aligned}
P_{R_{n/2}}({{\bf R}}_{n/2})=( G({{\bf R}}_{n/2},{{\bf R}}_{n/2-1})\ldots G({{\bf R}}_1,{{\bf R}}_0) \Psi_T({{\bf R}}_0) )\\
\times ( G({{\bf R}}_{n/2},{{\bf R}}_{n/2+1})\ldots G({{\bf R}}_{n-1},{{\bf R}}_n) \Psi_T({{\bf R}}_n) ) \\
= \Phi_0^2({{\bf R}}_{n/2})\end{aligned}$$ for $n \rightarrow \infty$, which allows us to obtain correct expectation values of operators that do not commute with the Hamiltonian.
Geometry optimization
=====================
In TMO materials, it is particularly useful to be able to optimize the geometry of the system within QMC. The usual way of doing this in mean-field calculations is to calculate the forces on the atoms and use one of many minimization routines. Unfortunately, there are not yet any reliable methods to calculate the force within diffusion Monte Carlo, despite much work in that direction[@pierleoni_rmc; @filippi_force; @assaraf_force; @chiesa_force]. These methods all require high-accuracy trial wave functions, which we usually do not have for transition metals. Thus, with the current state of the art, we are only able to optimize a few key degrees of freedom using the total energies from DMC calculations and line minimization. Even this must be done carefully because of the statistical uncertainty in the DMC energy. What follows is the scheme used in the work presented here, which has been found to be quite robust.
According to Bayes’ theorem, given a model $M$ and a set of data $D$, the probability of the model given the set of data is $$P(M|D)=\frac{P(D|M)P(M)}{P(D)}.$$ $P(D)$ is an unimportant normalization constant and $P(M)$ is called the prior distribution, which we are free to set to reflect the [*a priori*]{} probability distribution on the set of models. One usually sets $P(M)=1$, the unbiased maximum entropy/least knowledge condition. In the case of normally distributed data on a set of points $\lbrace x_1,x_2,...,x_N \rbrace $, $$P(D|M) \propto \exp[-\sum_i (M(x_i)-D(x_i))^2/2\sigma^2(x_i)],$$ where $\sigma(x)$ is the statistical uncertainty of $D(x)$.
For example, in the case of bond lengths, we can limit our space of models to $M(x)=c_1+c_2x+c_3x^2$, for $x$ close to the minimum bond length. This is equivalent to setting the prior distribution equal to one for all quadratic functions and to zero for non-quadratic functions. One then calculates several data points $D(x)$ with statistical uncertainties $\sigma(x)$. The probability distribution function of the bond length $b$ is then obtained by calculating the marginal distribution $$p(b)=\frac{\int \delta(-c_2/2c_3-b) P(D|M)P(M) dc_1dc_2dc_3}{\int P(D|M)P(M) dc_1dc_2dc_3}.$$ This integral is only three-dimensional, and as such could be calculated by a grid method, but it is convenient to calculate it by Monte Carlo, by sampling $P(D|M)P(M)$ and binning the bond length. The probability distribution function for the bond length is typically a Gaussian function to high accuracy, so it can be described as a mean value with a statistical uncertainty.
To make this scheme more efficient, we would like to calculate QMC energies as far away from the minimum as possible while still maintaining accuracy. This is because the energy changes much more quickly far from the minimum, which mitigates the stochastic uncertainties. That is, the energy scale is larger far from the minimum, so less precision is necessary. Thus, we should use a fitting function that is valid as far from the minimum as possible, while containing as few parameters as possible. For minimum energy geometries, it has been found [@ryo_vinet; @lucas_thesis] that the Vignet or modified Morse potentials are quite good for this purpose.
Approximations
==============
Pseudopotentials
----------------
In QMC, we can increase the efficiency significantly by using pseudopotentials to replace the core electrons with an effective potential. This has the effect of removing the large fluctuations near the core, which do not contribute much to the valence electrons’ correlation, which is the important for chemical properties. This introduces two approximations in the technique: first, the pseudopotential itself, and second, the small localization error[@lubos_psp] in diffusion Monte Carlo.
It has been found that small-core pseudopotentials are necessary for high accuracy on transition metals[@lee_mno; @dolg_psp_tm]. On the 3d metals, which are the primary focus in this paper, this means a Ne-core pseudopotential. The reason for this is that the 3d electrons occupy much the same space as the semicore 3p and, to a lesser extent, the 3s electrons. Since the 3d electrons are strongly affected by bonding, they in turn interact with the semicore. This interaction will change with correlation and chemical environment, so we must include the semicore electrons in accurate electronic structure calculations. This is not unique to QMC and is generally done in density functional theory where high accuracy is needed[@ferroelectric_with_small_core].
Finite size errors
------------------
When performing calculations for extended systems such as crystals, it is necessary to introduce periodic boundary conditions. This is an approximation on two levels. The first is the standard one-body level that is corrected by using reciprocal space sampling (i.e., k-points). The second level is inherent in a many-body correlated method, where the periodic boundary conditions force the electron to interact unphysically with its periodic image. This is similar to the finite simulation cell error in classical molecular dynamics simulation. This is typically corrected by either modifying the Coulomb interaction to remove the spurious interaction[@finite_size99] or by an correction[@chiesa_sk; @lucas_thesis]. Neither of these methods has clearly been demonstrated to be superior, and both methods or similar ones have been used successfully. Even with these corrections, a QMC calculation of an extended system usually involves on the order of 40 to 100 atoms, regardless of the size of the primitive cell, followed by extrapolation to infinite size.
Fixed node
----------
The algorithms described above are exact when the wave function can be written as a positive function, since then $\Psi_T\Phi_0$ is a probability distribution function. For fermions, it is not usually the case that $\Psi_T$ has the same zeros as the exact ground state, so we make the fixed-node approximation, where the nodal surface of the exact wave function are assumed to be the same as the trial wave function. This approximation typically results in recovering 90-95% of the correlation energy, and can be relaxed, but at the cost of exponential scaling of the system size[@Foulkes_review].
Given that the pseudopotential localization approximation is usually quite small for energy differences[@casula_lrdmc], we are mostly concerned with the fixed-node error. The Jastrow factor does not change the nodes of the wave function, so in the method outlined above, the nodes (and thus the final accuracy) are fixed to be the nodes of the Slater determinant of orbitals from the mean-field method. It is currently not feasible to vary the orbital expansion directly for a large system, since the number of parameters grows to the thousands for even moderately sized systems. However, partial optimizations can be done, and, as we shall see, are very effective for transition metal-oxygen systems.
TM-O molecules
==============
Simple molecular systems are excellent starting points for the study of transition metal oxides, since they are small enough to study carefully in a reasonable amount of time, and are also treatable by accurate but expensive quantum chemistry techniques like Coupled Cluster. This provides an additional much-needed data point to compare accuracy of the various electronic structure methods.
Near-optimal one-particle orbitals
----------------------------------
Wagner and Mitas[@CPL_lucas] performed the first calculations using DMC on simple two-atom transition metal oxides (TiO and MnO), and found a strong dependence of the calculated binding energy on the orbitals used in the Slater determinant. They used the B3LYP hybrid DFT/Hartree-Fock functional, and varied the percentage of Hartree-Fock mixing. They found the optimal percentage to be very close to the semi-empirical value fitted by Becke for his B3PW potential[@becke_3parm]. We have plotted the energy gain of B3LYP orbitals versus Hartree-Fock for the first five transition metal monoxide molecules in Fig \[fig:energy\_gain\_b3lyp\]. Upon examining the orbitals, they found a large difference in the d-p hybridization for both TiO (Fig \[fig:tio\_hf\_vs\_b3lyp\]) and MnO. This is a direct consequence of the importance of electronic correlation in transition metals.
To understand the importance of the one-particle orbitals, one can conceptually divide the total energy in three parts, each described by a different part of the wave function:
- One-body and antisymmetry: the Slater determinant
- Two-body electron interaction: Jastrow factor
- Higher orders : implicit diffusion Monte Carlo wave function.
The first part, the Slater determinant, determines the nodes of the wave function and therefore the ultimate accuracy of the calculation. Empirically, in materials containing only s and p-type elements, these three parts are almost independent of each other–the Hartree-Fock orbitals are close to optimal for a Slater-Jastrow wave function. In transition metal oxides, however, this situation changes, and the two-body and higher interactions are strong enough to change the one-body part significantly. In TMO’s, this effect seems to be largely in the d-p hybridization between oxygen and the transition metal.
![The energy gain in DMC from using B3LYP orbitals as a function of the metal monoxide. The line is a guide to the eye. Taken from Ref [@wagner_jcp].[]{data-label="fig:energy_gain_b3lyp"}](figure1){width="\columnwidth"}
![The d-p hybridization orbital (doubly occupied) isosurface for TiO in Hartree-Fock (left) and B3LYP (right). B3LYP enhances the hybridization significantly, which leads to lower energy in QMC. Figure generated using VMD and POV-Ray[@vmd; @povray][]{data-label="fig:tio_hf_vs_b3lyp"}](figure2a "fig:"){width="5cm"} ![The d-p hybridization orbital (doubly occupied) isosurface for TiO in Hartree-Fock (left) and B3LYP (right). B3LYP enhances the hybridization significantly, which leads to lower energy in QMC. Figure generated using VMD and POV-Ray[@vmd; @povray][]{data-label="fig:tio_hf_vs_b3lyp"}](figure2b "fig:"){width="5cm"}
By using the reptation Monte Carlo algorithm, we can obtain the unbiased one-particle density within the fixed-node approximation (Fig \[fig:tio\_dens\_profile\]), which gives further insight into the importance of correlation in the one-particle density. QMC tends to enhance the density in the bonding region (the hybridization) over both Hartree-Fock and B3LYP, but is not able to completely repair the erroneous Hartree-Fock density because of the fixed-node approximation. This is the reason for the large energy gain from using B3LYP orbitals to fix the nodal surface.
![The density of the Ti-O molecule projected onto the Ti-O axis in the bonding region for various methods. []{data-label="fig:tio_dens_profile"}](figure3){width="\columnwidth"}
Energetic Performance
---------------------
--------------------------- ---------- ---------- ---------- ---------- ---------- ------
Method ScO TiO VO CrO MnO RMS
LDA[@furche:044103] 9.09 9.13 8.48 6.26 6.51 2.19
CCSD(T)[@baushlicher:189] 6.71 6.64 6.13 4.20 3.43 0.31
TPSSh[@furche:044103] 7.11 7.18 6.44 4.45 4.62 0.38
DMC[@wagner_jcp] 7.06(3) 6.81(3) 6.54(3) 3.98(2) 3.66(3) 0.21
AFQMC[@al-saidi_tio] - 7.02(21) - - 3.79(34) -
Exp[@merer_review] 7.01(12) 6.92(10) 6.44(20) 4.41(30) 3.83(8) 0
--------------------------- ---------- ---------- ---------- ---------- ---------- ------
: Binding energies of the first five transition metal monoxides by different theoretical methods, along with RMS deviations from the experiment(all in eV). Statistical uncertainties in units of $10^{-2}$ eV are shown in parentheses for Monte Carlo and experimental results. Zero point energy corrections are estimated to be much less than the uncertainty in experiment. There are too few AFQMC data to calculate meaningful RMS values. []{data-label="table:binding"}
The total energy of a system is quite important for determination of lowest-energy spin states, competing phases, reactions, etc, and is a place where traditional density functional theory has encountered difficulties on transition metal oxides. In Table \[table:binding\], we compare the binding energy obtained by DMC using B3LYP orbitals and several other methods. We find excellent accuracy, with the RMS deviations of DMC within the experimental uncertainty for most materials. CrO is the only molecule with a large deviation from experiment; however, it is not very far outside the experimental uncertainty. DMC is also able to consistently obtain a minimum energy bond length with errors below 0.01 Å (Table \[table:bond\_lengths\]), better than any other published result.
--------------------------- ---------- ---------- ---------- ---------- ---------- -------
Method ScO TiO VO CrO MnO RMS
LDA[@furche:044103] 1.644 1.597 1.564 1.584 1.602 0.033
CCSD(T)[@baushlicher:189] 1.680 1.628 1.602 1.634 1.66 0.011
TPSSh[@furche:044103] 1.659 1.613 1.582 1.612 1.628 0.012
DMC[@wagner_jcp] 1.679(2) 1.612(3) 1.587(3) 1.617(4) 1.652(4) 0.008
Exp[@merer_review] 1.668 1.623 1.591 1.621 1.648 0
--------------------------- ---------- ---------- ---------- ---------- ---------- -------
: Bond lengths in Åfor the first five transition metal monoxide molecules. []{data-label="table:bond_lengths"}
Dipole moments
--------------
While energetics are very important for electronic structure calculations, one is also often interested in non-energetic properties, such as dipole moments. There has been little work done on such things within QMC, even in the context of simpler s and p systems. To our knowledge, the only study of dipole moments other than on TMO’s is of the CO molecule[@co_dipole]. A primary reason for this lack of calculations is that until the development of RMC, there has not been an easy to implement method to obtain expectation values without the mixed-estimator bias. The commonly used methods, pure diffusion Monte Carlo and forward-walking[@caffarel_pdmc1; @caffarel_pdmc2; @forward_walking] do not scale well with the system size[@assaraf_fixed_num], since they suffer from increased fluctuations of weights as the number of particles increases. One can also use extrapolated estimation, where the expectation value of an operator is estimated as $\langle{\cal O}\rangle=2 \langle{\cal O}\rangle_{DMC} - \langle{\cal O}\rangle_{VMC}$, but that method introduces an additional approximation that one would like to avoid if possible.
RMC, on the other hand, scales quite well, and is easily applicable to medium-sized systems such as TMO molecules. As we have noticed above, the electronic correlation and hybridization are very intertwined, and therefore, the electronic correlation and dipole moment are also closely related. In Table \[table:dipole\_moments\], we report the dipole moments for the first five transition metal monoxides using RMC with B3LYP orbitals. RMC obtains dipole moments much higher than that found in experiment, which is somewhat surprising given the high accuracy seen in energetic properties. We will explore the fixed node approximation and its effect on the dipole moment in the next section.
--------------------------- --------- -------------------------- --------- --------- -------- --
Method ScO TiO VO CrO MnO
LDA[@furche:044103] 3.57 3.23 3.10 3.41 –
CCSD(T)[@baushlicher:189] 3.91 3.52 3.60 3.89 4.99
TPSSh[@furche:044103] 3.48 3.43 3.58 3.97 –
RMC[@wagner_jcp] 4.61(5) 4.11(5) 4.64(5) 4.76(4) 5.3(1)
Exp[@steimle_review] 4.55 3.34(1)[@steimle_tio_03] 3.355 3.88 –
--------------------------- --------- -------------------------- --------- --------- -------- --
: Dipole moments in Debye. The fixed-node RMC results have been obtained with a single determinant of B3LYP orbitals. See text for an analysis of the errors involved for the case of TiO.[]{data-label="table:dipole_moments"}
Beyond the Slater-Jastrow form
------------------------------
In this section, we explore one of the biggest advantages of the QMC method-the ability to go beyond a Slater-Jastrow trial function if needed. As we saw in the previous section, RMC with the Slater-Jastrow trial function does not obtain dipole moments in agreement with experiment. The dipole moment is very sensitive to electronic correlation, and we wish to perform as accurate a calculation as possible to approach the exact value. We can do this in QMC by expanding the wave function in determinants. We write the trial wave function as $$\Psi_T({{\bf R}})= \left( \sum_i c_i D_i \right) e^U,$$ where the $D_i$’s are determinants of one-particle orbitals, $e^U$ is the Jastrow factor, and $c_i$’s are variational parameters. These determinants and the initial coefficients are taken from a Configuration Interaction calculation, and the coefficients are reoptimized using Variational Monte Carlo in the presence of the Jastrow factor.
This last reoptimization step is crucial, since the DMC energy increases if the CI coefficients are kept constant. This is a result of the strong correlation of these systems–the first order correlations are taken care of by the Jastrow factor, which the CI calculation tries to describe (inefficiently) with determinants.
In Fig \[fig:tio\_multidet\], we see the convergence of this expansion for TiO. The energy has a smooth convergence in the number of determinants, but the dipole moment oscillates significantly, with smaller oscillations as the number of determinants increases. The final result is approximately 3.8(1) Debye, a significant change from the Slater-Jastrow trial wave function, but still quite far from the experimental value of 3.34(1). While this calculation is probably not at the exact limit, the dipole moment does not appear to change enough to reconcile with experiment. Somewhat reassuringly, though, the Coupled Cluster value also predicts a larger value for the dipole moment, so it is possible that the experiment may be in error. More studies of non-energy properties using Quantum Monte Carlo are sorely needed, however, to obtain an estimate of the expected accuracy.
![The number of determinants versus the energy and dipole moment for TiO.The dipole moments are shifted downwards by 0.1 Debye to correct for the pseudopotential error. []{data-label="fig:tio_multidet"}](figure4){width="\columnwidth"}
Solids
======
Calculations on extended TMO systems using QMC are particularly challenging, since QMC suffers not only from one body finite size effects (i.e., that described by k-point sampling), but also from many-body finite size effects, which require large simulation cells. For this reason, complete studies as those reported above for molecules are not usually attainable, and most work is still in progress. We will discuss a few preliminary studies and a few private communications of work that remains unpublished at the time of this writing. Clearly, the details of the calculations may change, so this section is meant more as a comment on the current state of the art.
Using QMC, there have been studies of the antiferromagnet NiO[@tanaka_nio; @towler_nio] and MnO[@lee_mno]. Except for Tanaka[@tanaka_nio], who performed a very rough optimization of the lattice constant within Variational Monte Carlo, all the published studies calculated only the cohesive energy, which comes quite close to experiment (Table \[table:solid\_cohesive\]) for the materials available. In the very recent work of Kolorenc and Mitas[@jindra_unpublished], they obtain similar accuracy for the cohesive energy of FeO and also obtain the correct ordering of phases for that material, which DFT mispredicts. In most of these materials, researchers have found a large dependence on the mean-field orbitals used, with the optimal orbitals ranging from Hartree-Fock to LDA. Apparently, there is no universal optimal mean-field method.
------------------------------- ------------------------- -------------- ---------------------
Material DMC binding energy (eV) Experimental mean-field orbitals
NiO[@towler_nio] 9.442(2) 9.5 Hartree-Fock
MnO[@lee_mno] 9.40(5) 9.5
BaTiO$_3$[@lucas_unpublished] 31.2(3) 31.57 LDA
FeO[@jindra_unpublished] 9.47(4) 9.7 PBE0[@pbe0]
------------------------------- ------------------------- -------------- ---------------------
: Cohesive energies for several materials using QMC, all calculated per formula cell. Also listed are the optimal mean-field orbitals if reported. LDA is the local density approximation of DFT, and PBE0 is a hybrid functional. []{data-label="table:solid_cohesive"}
Wagner and Mitas[@lucas_unpublished] have also reported using the Bayesian optimization scheme to find the minimum energy lattice constant of BaTiO$_3$, which is well-known to be underestimated by over 1% in the local density approximation to density functional theory, and overestimated by a similar margin in the gradient corrections. This 1% error in the lattice constant can affect the calculated spontaneous polarization up to 50%, so even this small error is not acceptable for a truely first-principles description of this material. DMC obtains a cubic lattice constant in error only by 0.015 $\pm$ 0.005 Å, which is somewhat less than half a percent, a significant improvement over the density functional results. Also, in BaTiO$_3$, there is an energy gain in DMC of $\sim$1 eV/formula cell by using LDA orbitals instead of Hartree-Fock orbitals, and they report that it is due to a similar change in d-p hybridization that is seen in the transition metal monoxide molecules.
Conclusions
===========
On the systems that have been tested thus far, QMC offers unprecedented accuracy in a completely first-principles and scalable method, particularly in the energetics of the systems. The d-p hybridization of transition metal oxides is strongly affected by electronic correlation. Using QMC methods, we can clearly see this, both by investigating the minimum-energy orbitals and by examining the one-particle density and dipole moment within QMC. The dipole moment in particular is strongly affected by the level of correlation present in the quantum mechanics approximation.
On TMO molecules, we have a significant gain in the total energy on expansion into determinants, of about 0.5 eV. This means that we are relying on cancellation of errors for the high accuracy of QMC, although to a much lesser degree than post-Hartree-Fock approaches and DFT. We see this error in the dipole moment, which does not benefit from cancellation of errors. On the molecules, however, we can use a brute-force approach by expanding in determinants and come quite close to the true ground state. However, this kind of expansion will ultimately fail for large systems, since the number of determinants grows very quickly with system size. In order to reliably check the QMC results, it is vital to develop new reasonably scaling wave functions that go beyond the Slater-Jastrow form. Some work has been done in this direction with the RVB[@casula_bcs], Pfaffian[@michal_prl], and backflow[@backflow1; @backflow2] wave functions in QMC. These wave functions’ accuracy should be tested on TMO systems in the future. Equally important are optimization schemes within VMC that can systematically minimize the energy with respect to the wave functions’ parameters despite the stochastic nature of VMC, which is under serious investigation[@umrigar_optimization2; @umrigar_opt07]. Finally, we need to be able to calculate forces within QMC accurately and efficiently. The current state of the art is not sufficient to treat transition metal oxides[@lucas_thesis], and the Bayesian method of geometry optimization is only efficient for a few dimensions.
The future looks promising for QMC calculations of TMO solids, with the only drawback that the calculations are very expensive on today’s computers, since one must use a large supercell. However, the scaling with system size is quite favorable, and QMC is very easy to operate in parallel, so it can take advantage of low-cost processors. It has already been shown for a few important transition metal oxide solids that QMC can obtain binding energies and other energetic properties with excellent accuracy, well worth the additional cost when high accuracy is needed. It remains to be seen how well the method performs for non-energetic properties, and what sort of trial wavefunctions are necessary to obtain even higher accuracy.
I would like to acknowledge Lubos Mitas, Jindrich Kolerenc, and Michal Bajdich for their support and discussions in much of the work discussed in this article, as well as E. Ertekin and V. Srinivasan for their comments on the article itself. I would also like to thank a NSF Graduate Research Fellowship and NSF grant EAR-0530110 for funding during the course of this work.
[^1]: Present address: 366 Le Conte Hall \#3700; Berkeley, CA 94720; [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Durga P. Choudhury${}^{a,b}$, Balam A. Willemsen${}^{a,b}$, John S. Derov${}^b$ and S. Sridhar${}^a$\
[${}^a$Physics Department, Northeastern University, Boston, MA 02115]{}\
[${}^b$Rome Laboratory, Hanscom AFB., Bedford, MA 01730]{}
bibliography:
- 'strings.bib'
- 'big.bib'
title: Nonlinear Response of HTSC Thin Film Microwave Resonators in an Applied DC Magnetic Field
---
11.5 pt =.75truecm
[^1]
Introduction
============
The microwave response of high-T${}_c$ superconductors (HTSC) is important both from the point of view of microwave applications of HTSC[@ZYShen94a] and fundamental physics[@TCLGSollner96a]. An understanding of the loss mechanisms, field and current profiles and nature of non-linearities can lead to improvement in fabricated devices that use them. While numerous experimental studies of non-linear microwave response of HTSC have been done [@CWilker95a], none of them to our knowledge have probed the non-linear response in the presence of DC magnetic fields, where the DC and microwave fields are of comparable magnitude. In this situation the effect of the microwave field cannot be considered as a small perturbative Lorentz force on the vortex lattice generated by the DC field, as is often done at high fields[@MWCoffey92a]. Such a situation can also act to test various models that have been proposed for the non-linear response.
Experimental Techniques
=======================
We used a patterned suspended resonant thin film of YBa${}_2$Cu${}_3$O$
{}_{7-\delta }$ housed in a rectangular copper package to carry out this series of experiments. Similar methods have been used before[@BAWillemsen94b; @BAWillemsen95a] with great success to investigate vortex dynamics and non-linearities in HTSC. The film, procured from Neocera Inc, was deposited on a 0.6 in $\times$ 0.22 in $\times$ 0.010 in LaAlO${}_3$ substrate by laser ablation, and was subsequently patterned in-house to a straight line of dimension 0.56 in $\times$ 0.004 in using methods described elsewhere [@BAWillemsen95t]. In order to obtain the highest possible $Q^{\prime }$s, the package was mechanically polished and thoroughly cleaned before each set of experiments. The resonator was made symmetric by placing a blank substrate of the same material and dimensions on top of the film before it was loaded into the package.
The assembly, complete with a controlling heater and temperature sensor, was inserted into the sample chamber of a Cryo Industries Variable Temperature cryostat. Two independent carbon glass sensors and temperature controllers were used to stabilize the temperature of the cavity to the degree required for these experiments. A LakeShore DR91C was used for gross control of the vaporizer temperature, which was set slightly below the desired sample temperature. The desired sample temperature was obtained and finely controlled with a LakeShore DR93CA. Temperature stability of the order of 1 mK were typical for the experiments presented here, where the data took up to two hours to obtain for each run.
DC magnetic fields up to $~$1000 Oe was applied parallel to the $c$-axis of the sample using a custom-built Walker Scientific copper solenoid and a LakeShore 622 superconducting power supply. Unlike typical Helmholtz coil configurations, the solenoid has no polecaps, thus ensuring that there is no remanent field, save for possibly the geomagnetic fields. It is worth pointing out that our experiment does not use any superconducting ground planes unlike parallel plate or microstrip resonators, thus avoiding complications due to demagnetization effects from such plates.
Microwaves were inductively coupled to and from the resonator by means of loops at the ends of stainless steel coaxial lines. The microwave transmission amplitude $S_{21}$ was then measured using an HP 8510C Automatic Network Analyzer. The coupling strength was adjusted by varying the distance between the loops and the resonator as well as their relative orientation. Coupling could thus be reduced to the point that the loaded and unloaded $Q^{\prime }$s are indistinguishable, simplifying the data extraction process. At the low input power levels that were used to carry out these experiments, the trace was very noisy. This coupled with the fact that we require extreme sensitivity to very small changes mean that we could not simply determine the $Q$ from the maximum frequency and $-3$dB bandwidth as is often done. In order to reduce the noise and obtain the required sensitivity:
- The network analyzer was used in the “Step” mode, in which every frequency point is individually synthesized
- The signal was heavily averaged to get rid of the random noise,
- The trace was fitted using the method of least squares to Lorentzian shape, and
- The frequency span was kept as narrow as possible, usually only about 20% larger than the $-3$ dB bandwidth.
The center frequency and the $-3$dB bandwidth obtained from the fit agreed very well with those directly read off the trace, especially at low power levels where the trace is closest to Lorentzian shape, but provided significantly enhanced sensitivity to small changes.
Results and Discussion
======================
Ubiquitous intrinsic non-linearities have been observed in thin film specimens of High-T${}_c$ materials[@PPNguyen93a; @BAWillemsen95a], and some aspects of these non-linearities appear to be explained by a current-induced critical state model[@SSridhar94a]. The present experiments, which involve both microwave and DC fields of comparable magnitude so as to study the interplay of these two effects, were designed to further test these critical state and other ideas. Our experiments show that the presence of even relatively low microwave powers can induce vortices in the film, emulating the response of a DC field. The signature of this fact come from the observation that low DC field hysteresis does not show the virgin state response.
In a typical sample, the signature of the virgin state (i.e. absence of trapped flux tubes) in the low DC field hysteresis experiments manifests itself as a sharp rise in the $-3$dB bandwidth as field is slowly increased from zero corresponding to initial penetration of flux. As the field is further increased to a value $H_{max}$ and then cycled between $H_{max}$ and $-H_{max}$, where $H_{max}$ is a field of the order of a few hundred Gauss, this initial behavior is never reproduced; instead, it goes through a butterfly-shaped hysteresis loop[@BAWillemsen96a].
The same experiment, performed on the films under discussion, yields two new observations :
- The initial “virgin” response vanishes at higher microwave powers. This seems to indicate that microwave fields can create enough vortices in in the sample to wash away the virgin state response, mimicking the effect of an applied DC field.
- A sharp dip in $R_s$ is observed at a field scale $H_{DC}\sim 5$G in the virgin response, indicating that a small applied DC field serves to [*lower*]{} $R_s$.
The second result was verified when we did a measurement of $R_s$ against applied microwave power in a fixed $H_{DC}$. The decrease in $R_s$ reproduces itself, as is evident from fig.3. Another observation from the microwave power ramp experiments is that the non-linearities in the sample also get substantially suppressed at these low field for low microwave powers. As $H_{DC}$ is increased, $R_s$ gradually rises and finally goes above it’s zero field value. To further ensure that this observation is genuine and is not an artifact of some experimental inconsistency, we repeated the measurements on a film patterned out of a different albeit similar wafer. Although results obtained from this film were not quantitatively identical with those of the other, which would be expected because of the differences on growth, deposition and patterning of the two films, the two characteristic features described above was observed to a comparable extent in the second film. Also to rule out the effect of any stray remanent DC field, we carried out the microwave power ramp measurements with the DC field reversed and no such effects were found.
We have examined the present data in the framework of two proposed explanations for non-linear response in HTSC, viz. weak links and dynamics of a current-driven critical state. The weak link picture can be viewed in terms of a resistively shunted junction (RSJ) model, taking the effect of the DC field to be a DC current flowing on the surface in addition to the rf current. The equation of motion for the relative phase between the coupled grain then become $\dot{\phi}+\sin \phi =i_{dc}+i_\omega \cos \omega
t$. The dynamic impedance can be calculated from $Z_\omega =\dot{\phi}
_\omega /i_\omega $. While numerical calculations of this response yield very interesting effects as $i_{dc}$ is varied, this approach does not seem to yield the results that are observed in this experiment.
The critical state model should also lead to ac + dc effects, so that the non-linearity should be modified by the DC magnetic field. However a calculation of this effect is not straight-forward, since it requires a prescription for the present case where the loss needs to be calculated when $i_{rf}$ is varied over on rf cycle for finite $H_{DC}$. The available prescription in the literature [@EHBrandt93a] does not consider this case, but instead considers a different method of varying $i_{rf}$ and $H_{DC}$. Hence, at the present, it is not possible to determine if the critical state model can explain these unusual results.
It is worth noting that the unusual decrease in $R_s$ observed here can occur due to non-equilibrium effects and in fact have been seen in low $T_c$ superconductors [@SSridhar83t]. There it was shown that when the microwave frequency $\omega >\tau ^{-1}$, where $\tau $ is the quasiparticle relaxation time, a non-equilibrium quasiparticle distribution can occur which leads to a decrease of $R_s$ in the presence of an $i_{dc}$. Another related phenomenon which occurs is an enhancement of the superconducting gap. While this condition is met in pure metals at low temperatures, it is not clear if this happens in the high $T_c$ materials.
However it is interesting to note that in YBa${}_2$Cu${}_3$O${}_{7-\delta }$ crystals, $R_s(T)$ is non-monotonic and there are regions of temperature where $%
(\partial R_s/\partial T)<0$. This unusual, apparently non-thermodynamic result, may imply that $(\partial R_s/\partial i_{dc})<0$ need not be surprising.
Conclusion
==========
We have described a novel effect in which both the microwave losses and non-linear response decrease in the presence of small magnetic fields. Although a clear explanation of this effect is lacking, and it could arise from non-equilibrium quasiparticle effects, the present observation implies that losses can be reduced by as much as $30\%$ and could have interesting implications for device performance.
[^1]: Manuscript received August 27, 1996. D. P. Choudhury, 617-373-2948, fax 617-373-2943, [email protected], http://sagar.physics.neu.edu/ Work at Northeastern University was supported by the AFOSR through Rome Labs, Hanscom AFB.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Android apps must be able to deal with both *stop event*s, which require immediately stopping the execution of the app without losing state information, and *start event*s, which require resuming the execution of the app at the same point it was stopped. Support to these kinds of events must be explicitly implemented by developers who unfortunately often fail to implement the proper logic for saving and restoring the state of an app. As a consequence apps can lose data when moved to background and then back to foreground (e.g., to answer a call) or when the screen is simply rotated. These faults can be the cause of annoying usability issues and unexpected crashes.
This paper presents a public benchmark of 110 data loss faults in Android apps that we systematically collected to facilitate research and experimentation with these problems. The benchmark is available on GitLab and includes the faulty apps, the fixed apps (when available), the test cases to automatically reproduce the problems, and additional information that may help researchers in their tasks.
author:
-
bibliography:
- 'bibliography.bib'
title: A Benchmark of Data Loss Bugs for Android Apps
---
Data loss, Android, benchmark, bug detection
### Acknowledgements {#acknowledgements .unnumbered}
This work has been partially supported by the EU H2020 Learn project, funded under the ERC Consolidator Grant 2014 program (Grant Agreement n. 646867) and the GAUSS national research project, funded by the MIUR under the PRIN 2015 program (Contract 2015KWREMX).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Random intersection graphs have received much attention recently and been used in a wide range of applications ranging from key predistribution in wireless sensor networks to modeling social networks. For these graphs, each node is equipped with a set of objects in a random manner, and two nodes have an undirected edge in between if they have at least one object in common. In this paper, we investigate connectivity and robustness in a [*general random intersection graph*]{} model. Specifically, we establish sharp asymptotic zero–one laws for $k$-connectivity and $k$-robustness, as well as the asymptotically exact probability of $k$-connectivity, for any positive integer $k$. The $k$-connectivity property quantifies how resilient is the connectivity of a graph against node or edge failures, while $k$-robustness measures the effectiveness of [*local-information-based*]{} consensus algorithms (that do not use global graph topology information) in the presence of adversarial nodes. In addition to presenting the results under the general random intersection graph model, we consider two special cases of the general model, a [*binomial*]{} random intersection graph and a [*uniform*]{} random intersection graph, which both have numerous applications as well. For these two specialized graphs, our results on asymptotically exact probabilities of $k$-connectivity and asymptotic zero–one laws for $k$-robustness are also novel in the literature.
as the results for these
which both have been numerous applications as well. For these two specialized graphs, the results on zero–one laws for $k$-robustness and asymptotically exact probabilities of $k$-connectivity we present novel results including .
author:
- 'Jun Zhao, Osman Yağan, and Virgil Gligor, [^1]'
title: On Connectivity and Robustness in Random Intersection Graphs
---
Complex networks, connectivity, consensus, random intersection graphs, robustness.
Introduction
============
Background {#sec:GraphModel}
----------
Since random intersection graphs were introduced by Singer-Cohen [@RIGThesis], different classes of these graphs have received considerable attention [@Rybarczyk; @yagan; @virgil; @2013arXiv1301.0466R; @zz; @r1; @ryb3; @bloznelis2013] recently. In these graphs, each node is assigned a set of objects selected by some random mechanism. An undirected edge exists between any two nodes that have at least one object in common. Random intersection graphs have been used in modeling and analyzing real-world networks in a wide variety of applications. Examples include secure wireless sensor networks [@Rybarczyk; @yagan; @virgil], social networks [@ryb3; @zz; @2013arXiv1301.0466R], classification analysis [@Models], and cryptanalysis [@herdingRKG]. Several properties such as clustering [@bloznelis2013], component evolution [@Rybarczyk] and degree distribution [@Models] have been analyzed for different classes of random intersection graphs.
The graph model in this paper, hereafter referred to as a [*general random intersection graph*]{}, represents a generalization of the random intersection graphs studied by Bloznelis [*et al.*]{} [@bloznelis2013; @Rybarczyk], and is defined on a node set $\mathcal {V}_n = \{v_1,
v_2, \ldots, v_n \}$ as follows. Each node $v_i$ ($i=1,2,\ldots,n$) is assigned an object set $S_i$ from an object pool $\mathcal {P}_n$ \[poolref\] consisting of $P_n$ distinct objects, where $P_n$ is a function of $n$. Each object set $S_i$ is constructed from the following two steps: First, the size of $S_i$, $|S_i|$, is determined according to some probability distribution $\mathcal {D}_n:\{1,
2,\ldots, P_n\} \to [0,1]$. Of course, we have $\sum_{x = 1}^{P_n}
\mathbb{P}[|S_i| = x] = 1$, with $\mathbb{P}[A]$ denoting the probability that event $A$ occurs. Next, $S_i$ is formed by selecting $|S_i|$ distinct objects uniformly at random from the object pool $\mathcal {P}_n$. In other words, conditioning on $|S_i| =
s_i$, set $S_i$ is chosen uniformly among all $s_i$-size subsets of $\mathcal {P}_n$. This process is repeated independently for all object sets $S_1, \ldots, S_n$. Finally, an undirected edge is assigned between two nodes if and only if their corresponding object sets have at least one object in common; namely, distinct nodes $v_i$ and $v_j$ have an undirected edge in between if and only if $S_i {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}S_j \neq \emptyset$. The graph defined through this adjacency notion is denoted by $G(n,P_n,\mathcal {D}_n)$.
A specific case of the general model $G(n,P_n,\mathcal {D}_n)$, known as the *binomial* random intersection graph, has been widely explored to date \[9\]–\[14\]. Under this model, each object set $S_i$ is constructed by a Bernoulli-like mechanism; i.e., by adding each object to $S_i$ independently with probability $p_n$. Like integer $P_n$, probability $p_n$ is also a function of $n$. The term binomial" accounts for the fact that $|S_i|$ now follows a binomial distribution with $P_n$ as the number of trials and $p_n$ as the success probability in each trial. We denote the binomial random intersection graph by $G_b(n,P_n,p_n)$, where the subscript “b” stands for “binomial”.
Another well-known special case of the general model $G(n,P_n,\mathcal
{D}_n)$ is the *uniform* random intersection graph [@r1; @yagan; @ryb3; @virgil]. Under the uniform model, the probability distribution $\mathcal {D}_n$ concentrates on a single integer $K_n$, where $1\leq K_n \leq P_n$; i.e., for each node $v_i$, the object set size $|S_i|$ equals $K_n$ with probability $1$. Note that $P_n$ and $K_n$ are both integer functions of $n$. We denote by $G_u(n,P_n,K_n)$ the uniform random intersection graph, with the subscript “u” meaning “uniform”.
Applications of Random Intersection Graphs
------------------------------------------
A concrete example for the application of random intersection graphs can be given in the context of secure wireless sensor networks. As explained in detail in numerous other places [@Rybarczyk; @ryb3; @yagan; @yagan_onoff; @yagan2015zero; @ZhaoYaganGligor], the uniform random intersection graph model $G_u(n,P_n,K_n)$ is induced naturally by the Eschenauer–Gligor random key predistribution scheme [@virgil], which is a typical solution to ensure secure communications in wireless sensor networks. In particular, let the set of $n$ nodes in graph $G_u(n,P_n,K_n)$ stand for the $n$ sensors in the wireless network. Also, let the object pool $\mathcal{P}_n$ (with size $P_n$) represent the set of cryptographic keys available to the network and let $K_n$ be the number of keys assigned to each sensor (selected uniformly at random from the key pool $\mathcal{P}_n$). Then, the edges in $G_u(n,P_n,K_n)$ represent pairs of sensors that share at least one cryptographic key and thus that can [securely]{} communicate over existing wireless links in the Eschenauer–Gligor scheme. In the above application, objects that nodes have are cryptographic keys, so uniform random intersection graphs are also referred to as random key graphs [@yagan; @yagan2015zero; @yagan_onoff; @ZhaoYaganGligor].
In the secure sensor network area, the general random intersection graph model captures the differences that may exist among the number of keys assigned to each sensor. These differences appear for a variety of reasons including (a) the number may vary from sensor to sensor in a heterogeneous sensor network due to differences in the sizes of sensor memories [@Rybarczyk]; (b) the number may decrease due to the revocation of compromised nodes and keys [@ChenGligorPerrigMuralidharanTDSC2005]; and (c) the number may increase due to the establishment of path keys, where new keys are generated and distributed to participating sensors after deployment [@virgil].
Random intersection graphs can also be used to model social networks, where a node represents an individual, and an object could be an hobby of individuals, a book being read, or a movie being watched, etc. [@bloznelis2013; @2013arXiv1301.0466R; @Bloznelis201494; @ZhaoYaganGligor]. Then a link between two individuals characterizes a common-interest relation; e.g., two individuals have a connection if they have a common hobby, read the same book, or watch the same movie. In this setting, binomial/uniform/general random intersection graphs represent common-interest networks where the sets of interests that individuals have are constructed in different ways. Specifically, in binomial random intersection graphs, each interest is attached to each person independently with the same probability; in uniform random intersection graphs, all individuals have the same number of interests; and general random intersection graphs provide general possibilities for assigning individuals’ interest sets; e.g., without probability or number-of-interest restrictions.
Problem Formulation {#sec:ConnectivityandRobustness}
-------------------
We now introduce the graph properties that we are interested in. First, $k$-connectivity is formally defined as follows.
A graph is said to be $k$-connected if each pair of nodes has at least $k$ internally node-disjoint path(s) in between, where two paths are internally node-disjoint if except the source and destination, the intermediate nodes are different. Equivalently, by Menger’s theorem, a graph is $k$-connected if it cannot be disconnected by deleting at most $(k-1)$ nodes or edges, where a graph is connected if there exists at least a path of edges between any two nodes.
Clearly, $k$-connectivity quantifies well-established measures of strength. For instance, it captures the resiliency of graphs against node or edge failures. It also captures the resiliency of consensus protocols in the presence of $h$ adversarial nodes in a graph with node size greater than $3h$; i.e., a necessary and sufficient condition is that the graph is $(2h+1)$-connected [@Dolev:1981:BGS:891722].
Many graph algorithms rely on sufficient connectivity; e.g, algorithms to achieve consensus [@6425841; @add-ref-2; @zhang2012robustness]. However, these algorithms typically assume that nodes have full knowledge of the graph topology, which is often impractical [@6425841]. To account for the lack of full topology knowledge in the general case, Zhang and Sundaram introduce the notion of [*graph robustness*]{} [@6425841], which has received much attention recently [@6481629; @leblanc2013resilient; @zhang2012robustness; @zhang2012simple; @7061412]. Formally, $k$-robustness is defined as follows.
A graph with a node set $\mathcal {V}$ is $k$-robust if at least one of (a) and (b) below hold for every pair of non-empty, disjoint subsets $A$ and $B$ of $\mathcal {V}$: (a) there exists at least a node $v_a \in A$ such that $v_a$ has no less than $k$ neighbors inside $\mathcal {V}\setminus
A$; and (b) there exists at least a node $v_b \in B$ such that $v_b$ has no less than $k$ neighbors inside $\mathcal {V}\setminus B$.
Zhang and Sundaram [@6425841] show that when nodes have local topology knowledge, consensus can still be reached in a sufficiently robust graph in the presence of adversarial nodes, but not in a sufficiently connected and insufficiently robust graph.
Graph robustness provides a different notion of strength than $k$-connectivity. That is, it quantifies the effectiveness and resiliency of local-information-based consensus algorithms in the presence of adversarial nodes. We detail the application of robustness to consensus in the next subsection. Robustness also has broad relevance in graph processes beyond consensus; e.g., robustness plays a key role in information cascades [@6425841].
Application of Robustness to Consensus
--------------------------------------
To study consensus in a graph, we consider that all nodes are synchronous and the time is divided into different time slots. From one time slot to the next time slot, each node updates its value. Let $x_i[t]$ denote the value of node $v_i$ at time slot $t$ for $t=0,1,\ldots$. We first suppose all nodes are benign. Then consensus means $\lim_{t \to \infty} |x_i[t] - x_j[t]| = 0 $ for each pair of nodes $v_i$ and $v_j$. The updating process of each node’s value is as follows. With $V_i$ denoting the neighborhood set of each node $v_i$, from time slot $t$ to $t+1$, $v_i$ updates its value $x_i [t]$ to $x_i[t+1]$ by incorporating node $v_j$’s value $x_j [t]$ that $v_j$ sends to $v_i$, for $v_j \in V_i$; i.e., there is a function $f_i(\cdot)$ such that $$\begin{aligned}
x_i[t+1] & = f_i\big( \big\{ x_j[t] \ \big| \ v_j \in V_i {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}\{ v_i \} \big\}\big) . \nonumber\end{aligned}$$ Now we consider the case where there might exist adversarial nodes, i.e., nodes that are not benign. A node $v_i$ is said to be benign if it sends $x_i [t]$ to all of its neighbors and applies $f_i(\cdot)$ at every time slot $t$, and is called adversarial otherwise. In the presence of adversarial nodes, consensus means $\lim_{t \to \infty} |x_i[t] - x_j[t]| = 0 $ for each pair of *benign* nodes $v_i$ and $v_j$.
Under the adversary model that the total number of adversarial node(s) in the graph is upper bounded by $h$, then consensus can be achieved if and only if the graph is $(2h+1)$-connected, given the graph has more than $3h$ nodes [@Dolev:1981:BGS:891722]. However, the algorithms often assume that all nodes know the entire network topology [@6425841]. Suppose each node does not know the entire network topology and only knows the number of adversarial nodes in its neighborhood, Zhang and Sundaram [@6425841] show the usefulness of robustness in studying consensus. Specifically, under the adversary model that each benign node has at most $h$ adversarial node(s) as neighbors, then consensus can be achieved if the graph is $(2h+1)$-robust [@6425841]. With the above, we can use consensus dynamics to motivate the study of both connectivity and robustness, where connectivity (resp., robustness) is applicable to the case where each node knows the global (resp., local) network topology.
, where a specific algorithm explained below assumes that each benign node knows $h$ as an upper bound on the number of adversarial nodes in its neighborhood, and does not need to know the entire network topology. In this algorithm, the function $f_i(\cdot)$ for a benign node $v_i$ is given by a weighted linear combination of $v_i$’s own value and the values that $v_i$’s some neighbors send to $v_i$. In particular, the algorithm for a benign node $v_i$ has the following steps:
- At each time slot $t$, $v_i$ receives values from all of its neighbors, and ranks them (either from largest to smallest, or from smallest to largest).
- In the ranked values, if there are $h$ or more values larger than $x_i [t]$, $v_i$ removes the $h$ largest values; if there are fewer than $h$ values larger than $x_i [t]$, $v_i$ removes all of these values. Similarly, if there are $h$ or more values smaller than $x_i [t]$, $v_i$ removes the $h$ smallest values; if there are fewer than $h$ values smaller than $x_i [t]$, $v_i$ removes all of these values. Let $R_i[t]$ denote the set of nodes whose values sent to $v_i$ are removed by $v_i$.
- From time slot $t$ to $t +1$, $v_i$ updates its value $x_i [t]$ to $x_i[t+1]$ by $$\begin{aligned}
x_i[t+1] & = w_{ii}[t] x_i[t] + \sum_{v_j \in V_i \setminus R_i[t]} w_{ij}[t] x_j[t], \nonumber\end{aligned}$$ where the weights $w_{\cdot \cdot }[\cdot]$ satisfy $w_{ij}[t] \geq \alpha$ for each $v_j \in V_i {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}\{ v_i \}$ ($\alpha \in (0,1)$ is a constant), $w_{ij}[t] = 0$ for each $v_j \not\in V_i {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}\{ v_i \}$ and $\sum_{j=1}^n w_{ij}[t] = 1$.
Related Work {#related}
------------
For connectivity (i.e., $k$-connectivity with $k=1$) in binomial random intersection graph $G_b(n,P_n,p_n)$, Rybarczyk establishes the exact probability [@2013arXiv1301.0466R] and a zero–one law [@zz; @2013arXiv1301.0466R]. She further shows a zero–one law for $k$-connectivity [@zz; @2013arXiv1301.0466R]. Our Theorem \[thm:rig\] provides not only a zero–one law, but also the exact probability to understand $k$-connectivity precisely.
For connectivity in uniform random intersection graph $G_u(n,P_n,K_n)$, Rybarczyk [@ryb3] derives the exact probability and a zero–one law, while Blackburn and Gerke [@r1], and Yağan and Makowski [@yagan] also obtain zero–one laws. Rybarczyk [@zz] implicitly shows a zero–one law for $k$-connectivity in $G_u(n,P_n,K_n)$. Our Theorem \[thm:urig\] also gives a zero–one law. In addition, it gives the exact probability to provide an accurate understanding of $k$-connectivity.
For general random intersection graph $G(n,P_n,\mathcal {D}_n)$, Godehardt and Jaworski [@Models] investigate its degree distribution and Bloznelis *et al.* [@Rybarczyk] explore its component evolution. Recently, Yağan [@yagan2015zero] obtains a zero–one law for connectivity.
Since asymptotic probability results of $k$-connectivity in random graphs are often established by first showing the corresponding results for the property of having minimum degree at least $k$, and then proving the probability of having minimum degree at least $k$ yet not being $k$-connected converges to zero asymptotically, all the above references on $k$-connectivity (resp., connectivity) also establish the corresponding results for the property of having minimum degree at least $k$ (resp., $1$).
To date, there have not been results on ($k$-)robustness of random intersection graphs reported by others. As noted in Lemma \[er\_robust\], Zhang and Sundaram [@6425841] present a zero–one law for $k$-robustness in an Erdős–Rényi graph. Specifically, their result is that if $\hat{p}_n =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}$, then Erdős–Rényi graph $G(n,\hat{p}_n)$ is almost surely $k$-robust (resp., not $k$-robust) if $\lim\limits_{n \to
\infty}{\alpha_n} =\infty$ (resp., $\lim\limits_{n \to
\infty}{\alpha_n} =-\infty$).
For random intersection graphs in this paper, two nodes have an edge in between if their object sets share at least one object. A natural variant is to define graphs with edges only between nodes which have at least $s$ objects in common (instead of just $1$) for some positive integer $s$. Recent researches [@Bloznelis201494; @ANALCO] investigate $k$-connectivity in graphs under this definition.
Contributions and Organization
------------------------------
With the above notions of $k$-connectivity and $k$-robustness in mind, a natural questions to ask is when will random intersection graphs become $k$-connected or $k$-robust? We answer this question and summarize the key contributions as follows:
1. We derive sharp zero–one laws and asymptotically exact probabilities for $k$-connectivity in general random intersection graphs.
2. We establish sharp zero–one laws for $k$-robustness in general random intersection graphs.
3. For the two specific instances of the general graph model, a binomial random intersection graph and a uniform random intersection graph, we provide the first results on asymptotically exact probabilities of $k$-connectivity and zero–one laws for $k$-robustness.
This paper extends the conference version [@ZhaoCDC] in several ways:
1. We strengthen the known results on binomial/uniform/general random intersection graphs. Specifically, Theorems 1–6 in this paper eliminate the condition $|\alpha_n| = o(\ln n)$ in [@ZhaoCDC Theorems 1–6].
2. For $k$-connectivity of a uniform random intersection graph, we provide a complete proof in Section \[prf:thm:urig\]. Note that this result serves as the building block for all other results.
3. We enhance numerical experiments to better confirm the theoretical results; see Section \[sec:expe\].
4. We discuss the parameter conditions of the theorems in detail; see Section \[ParameterConditions\].
5. We compare our results of binomial/uniform/general random intersection graphs with those of Erdős–Rényi graphs; see the last paragraph of Section \[sec:main:res\].
The rest of the paper is organized as follows. Section \[sec:main:res\] presents the main results as Theorems \[thm:rig\]–\[thm:grig:rb\]. To improve the readability of the paper, we defer the proofs of the theorems to the end of the paper. We provide numerical experiments in Section \[sec:expe\]. Afterwards, we introduce some auxiliary lemmas in Section \[sec:factlem\], before establishing the main results in Sections \[sec:thmprf:kcon\], \[prf:thm:urig\] and \[sec:thmprf:krb\]. Section \[sec:prf:fact:lem\] details the proofs of the lemmas. Finally, we conclude the paper Section \[sec:Conclusion\].
The Results {#sec:main:res}
===========
Our main results are presented in Theorems \[thm:rig\]–\[thm:grig:rb\] below. We defer the proofs of all theorems to Sections \[sec:thmprf:kcon\]–\[sec:thmprf:krb\]. Throughout the paper, $k$ is a positive integer and does not scale with $n$; and $e$ is the base of the natural logarithm function, $\ln$. All limits are understood with $n\to \infty$. We use the standard Landau asymptotic notation $o(\cdot), O(\cdot),
\omega(\cdot), \Omega(\cdot),\Theta(\cdot)$ and $ \sim$; see [@ZhaoYaganGligor Page 2-Footnote 1]. In particular, for two positive sequences $f_n$ and $g_n$, the relation $f_n \sim
g_n$ signifies $\lim_{n \to
\infty} ( {f_n}/{g_n})=1$. For a random variable $X$, the terms $\mathbb{E}[X]$ and $\text{Var}[X]$ stand for its expected value and variance, respectively.
As noted in Section \[sec:GraphModel\], we denote a binomial (resp., uniform) random intersection graph by $G_b(n,P_n,p_n)$ (resp., $G_u(n,P_n,K_n)$). Clearly, $G_b(n,P_n,0)$ (resp., $G_u(n,P_n,0)$) is an empty graph, while $G_b(n,P_n,1)$ (resp., $G_u(n,P_n,P_n)$) being a complete graph is $k$-connected for $n\geq k+1$ and is $k$-robust for $n\geq 2k$. Then for each $n\geq 2k$, with $P_n$ fixed and $p_n$ increasing from $0$ to $1$ (resp., $K_n$ increasing from $0$ to $P_n$), the probabilities of $k$-connectivity and $k$-robustness of $G_b(n,P_n,p_n)$ (resp., $G_u(n,P_n,K_n)$) increase from $0$ to $1$. In addition, for random graphs, results are often obtained in the asymptotic sense since the analysis becomes intractable in the finite regime [@citeulike:4012374; @erdos61conn; @RIGThesis; @YaganThesis; @yagan_onoff].
Given the above, it is natural to anticipate that our results are presented in the form of zero–one laws, where a zero–one law means that the probability of a graph having a certain property asymptotically converges to $0$ under some conditions and to $1$ under some other conditions. Moreover, it is useful to have a complete picture by obtaining the asymptotically exact probability result [@YaganThesis]. For binomial/uniform/general random intersection graphs, we derive asymptotically exact probabilities for $k$-connectivity in Theorems \[thm:rig\]–\[thm:grig\], and zero–one laws for $k$-robustness in Theorems \[thm:rig:rb\]–\[thm:grig:rb\]. A future work is to establish asymptotically exact probabilities for $k$-robustness.
Noting that for any graph/network, $k$-connectivity implies that the minimum node degree is at least $k$ [@citeulike:4012374], we often present results for the property of minimum node degree being at least $k$ together with $k$-connectivity results.
Asymptotically Exact Probabilities for $k$-Connectivity and the Property of Minimum Node Degree Being at Least $k$ {#sec:kconasy}
------------------------------------------------------------------------------------------------------------------
### $k$-Connectivity and Minimum Node Degree in Binomial Random Intersection Graphs
For a binomial random intersection graph, Theorem \[thm:rig\] below shows asymptotically exact probabilities for $k$-connectivity and the property of minimum node degree being at least $k$.
\[thm:rig\] For a binomial random intersection graph $G_b(n,P_n,p_n)$, with a sequence $\alpha_n$ for all $n $ defined through $$\begin{aligned}
{p_n}^2 P_n & =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}, \label{thm:rig:pe}\vspace{2.00000pt}\end{aligned}$$ if $P_n = \omega \big(n(\ln n)^5\big)$, $$\begin{aligned}
& \hspace{-25pt} \lim_{n \to \infty}\mathbb{P}
\big[\hspace{2pt}\text{Graph }G_b(n,P_n,p_n)\text{
is $k$-connected}.\hspace{2pt}\big] \nonumber
\end{aligned}$$
[ =]{} 0, &, \[bin-kon-0\]\
1, & \[bin-kon-1\]\
e\^[- ]{}, & \[bin-kon-e\]
and $$\begin{aligned}
& \hspace{-39pt} \lim_{n \to \infty}\mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_b(n,P_n,p_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \nonumber
\end{aligned}$$
[ =]{} 0, &, \[bin-kon-0-mnd\]\
1, & \[bin-kon-1-mnd\]\
e\^[- ]{}, & \[bin-kon-e-mnd\]
\[rm\]
As we will explain in Section \[pfrig\] within the proof of Theorem \[thm:rig\], for (\[bin-kon-0\]) (\[bin-kon-1\]) (\[bin-kon-0-mnd\]) (\[bin-kon-1-mnd\]) (i.e., the zero–one laws), the condition $P_n =
\omega \big(n(\ln n)^5\big)$ can be weakened as $P_n = \Omega
\big(n(\ln n)^5\big)$, while we enforce $P_n = \omega \big(n(\ln
n)^5\big)$ for (\[bin-kon-e\]) (\[bin-kon-e-mnd\]).
### $k$-Connectivity and Minimum Node Degree in Uniform Random Intersection Graphs
For a uniform random intersection graph, Theorem \[thm:urig\] below gives asymptotically exact probabilities for $k$-connectivity and the property of minimum node degree being at least $k$.
\[thm:urig\] For a uniform random intersection graph $G_u(n,P_n,K_n)$, with a sequence $\alpha_n$ for all $n $ defined through $$\begin{aligned}
\frac{{K_n}^2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n}, \label{thm:urig:pe}\end{aligned}$$ if $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$, then $$\begin{aligned}
& \hspace{-25pt} \lim_{n \to \infty}\mathbb{P}
\big[\hspace{2pt}\text{Graph }G_u(n,P_n,K_n)\text{
is $k$-connected}.\hspace{2pt}\big] \nonumber
\end{aligned}$$
[ =]{} 0, &, \[uni-kon-0\]\
1, & \[uni-kon-1\]\
e\^[- ]{}, & \[uni-kon-e\]
and $$\begin{aligned}
& \hspace{-39pt} \lim_{n \to \infty}\mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_u(n,P_n,K_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \nonumber
\end{aligned}$$
[ =]{} 0, &, \[uni-kon-0-mnd\]\
1, & \[uni-kon-1-mnd\]\
e\^[- ]{}, & \[uni-kon-e-mnd\]
### $k$-Connectivity and Minimum Node Degree in General Random Intersection Graphs
For a general random intersection graph, Theorem \[thm:grig\] below provides asymptotically exact probabilities for $k$-connectivity and the property of minimum node degree being at least $k$.
\[thm:grig\] Consider a general random intersection graph $G(n,P_n,\mathcal {D}_n)$. Let $X_n$ be a random variable following probability distribution $\mathcal {D}_n$. With a sequence $\alpha_n$ for all $n $ defined through $$\begin{aligned}
\frac{\big\{\mathbb{E}[X_n]\big\}^2}{P_n} & =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n},
\label{thm:grig:pe}\end{aligned}$$ if $\mathbb{E}[X_n] = \Omega\big(\sqrt{\ln n}\hspace{2pt}\big)$ and $\text{Var}[X_n] =
o\mathlarger{\mathlarger{\big(}}\frac{\{\mathbb{E}[X_n]\}^2}{ n(\ln
n)^2 }\mathlarger{\mathlarger{\big)}}$, then $$\begin{aligned}
& \hspace{-25pt} \lim_{n \to \infty}\mathbb{P}
\big[\hspace{2pt}\text{Graph }G(n,P_n,\mathcal {D}_n)\text{
is $k$-connected}.\hspace{2pt}\big] \nonumber
\end{aligned}$$
[ =]{} 0, &, \[grg-kon-0\]\
1, & \[grg-kon-1\]\
e\^[- ]{}, & \[grg-kon-e\]
and $$\begin{aligned}
& \hspace{-39pt} \lim_{n \to \infty}\mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G(n,P_n,\mathcal {D}_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \nonumber
\end{aligned}$$
[ =]{} 0, &, \[grg-kon-0-mnd\]\
1, & \[grg-kon-1-mnd\]\
e\^[- ]{}, & \[grg-kon-e-mnd\]
Asymptotic Zero–One Laws for $k$-Robustness {#sec:main:res:rb}
-------------------------------------------
### $k$-Robustness in Binomial Random Intersection Graphs
Theorem \[thm:rig:rb\] below gives an asymptotic zero–one law for $k$-robustness in a binomial random intersection graph.
\[thm:rig:rb\] For a binomial random intersection graph $G_b(n,P_n,p_n)$, with a sequence $\alpha_n$ for all $n $ defined through $$\begin{aligned}
{p_n}^2 P_n & =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}, \label{thm:rig:pe:rb}\end{aligned}$$ if $P_n = \Omega \big(n(\ln n)^5\big)$, then $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}\text{Graph }G_b(n,P_n,p_n)\text{
is $k$-robust}.\hspace{2pt}\big] \nonumber
\end{aligned}$$
[ =]{} 0, &,\[bin-krb-0\]\
1, & \[bin-krb-1\]
### $k$-Robustness in Uniform Random Intersection Graphs
Theorem \[thm:urig:rb\] below presents an asymptotic zero–one law for $k$-robustness in a uniform random intersection graph.
\[thm:urig:rb\] For a uniform random intersection graph $G_u(n,P_n,K_n)$, with a sequence $\alpha_n$ for all $n $ defined through $$\begin{aligned}
\frac{{K_n}^2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n}, \label{thm:urig:pe:rb}\end{aligned}$$ if $K_n = \Omega \big((\ln n)^3\big)$, then $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}\text{Graph }G_u(n,P_n,K_n)\text{
is $k$-robust}.\hspace{2pt}\big] \nonumber
\end{aligned}$$
[ =]{} 0, &,\[uni-krb-0\]\
1, & \[uni-krb-1\]
### $k$-Robustness in General Random Intersection Graphs
Theorem \[thm:grig:rb\] as follows provides an asymptotic zero–one law for $k$-robustness in a general random intersection graph.
\[thm:grig:rb\] Consider a general random intersection graph $G(n,P_n,\mathcal {D}_n)$. Let $X_n$ be a random variable following probability distribution $\mathcal {D}_n$. With a sequence $\alpha_n$ for all $n $ defined through $$\begin{aligned}
\frac{\big\{\mathbb{E}[X_n]\big\}^2}{P_n} & =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n},
\label{thm:grig:pe:rb}\end{aligned}$$ if $\mathbb{E}[X_n] = \Omega \big((\ln n)^3\big)$ and $\text{Var}[X_n] =
o\mathlarger{\mathlarger{\big(}}\frac{\{\mathbb{E}[X_n]\}^2}{ n(\ln
n)^2 }\mathlarger{\mathlarger{\big)}}$, then $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}\text{Graph }G(n,P_n,\mathcal {D}_n)\text{
is $k$-robust}.\hspace{2pt}\big] \nonumber \\
& \quad =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$.} \end{cases} \nonumber
\end{aligned}$$
In view of Theorems \[thm:rig\]–\[thm:grig:rb\], for each binomial/uniform/general random intersection graph, its $k$-connectivity, $k$-robustness and the property of minimum node degree being at least $k$ asymptotically obey the same zero–one laws. Moreover, these zero–one laws are all *sharp* since $|\alpha_n|$ can be much smaller compared to $\ln n$; e.g., even $\alpha_n = \pm \cdot \ln \ln \cdot
\cdot \cdot \ln n$ satisfies $\lim_{n \to
\infty}{\alpha_n} =\pm\infty$.
We compare our results of random intersection graphs with those of Erdős–Rényi graphs below. From [@PES:6114960 Section 1.1], ${p_n}^2 P_n$ in the scaling conditions (\[thm:rig:pe\]) and (\[thm:rig:pe:rb\]) of Theorems \[thm:rig\] and \[thm:rig:rb\] is an asymptotics of the edge probability in a binomial random intersection graph $G_b(n,P_n,p_n)$. Also, by [@bloznelis2013 Lemma 1], $\frac{{K_n}^2}{P_n}$ in the scaling conditions (\[thm:urig:pe\]) and (\[thm:urig:pe:rb\]) of Theorems \[thm:urig\] and \[thm:urig:rb\] (resp., $\frac{\{\mathbb{E}[X_n]\}^2}{P_n}$ in the scaling conditions (\[thm:grig:pe\]) and (\[thm:grig:pe:rb\]) of Theorems \[thm:grig\] and \[thm:grig:rb\] is an asymptotics of the edge probability in a uniform random intersection graph $G_u(n,P_n,K_n)$ (resp., a general random intersection graph $G(n,P_n,\mathcal {D}_n)$). Then comparing Theorems \[thm:rig\]–\[thm:grig\] with Lemma \[er\_kconmnd\], and comparing Theorems \[thm:rig:rb\]–\[thm:grig:rb\] with Lemma \[er\_robust\], we conclude binomial/uniform/general random intersection graphs under certain parameter conditions[^2] exhibit the same behavior with Erdős-Rényi graphs in the sense that for each of (i) $k$-connectivity, (ii) the property of minimum node degree being at least $k$, and (iii) $k$-robustness, a common point for the transition from a zero-law to a one-law occurs when the edge probability equals $\frac{\ln n + {(k-1)} \ln \ln n}{n}$. The term $\alpha_n$ in Equations (1) (resp., (4) and (7)), or Equations (10) (resp., (12) and (14)) measures how much ${p_n}^2 P_n$ (resp., $\frac{{K_n}^2}{P_n}$ and $\frac{\{\mathbb{E}[X_n]\}^2}{P_n}$) is away from the critical value $\frac{\ln n + {(k-1)} \ln \ln n}{n}$.
A Discussion of Parameter Conditions {#ParameterConditions}
------------------------------------
Note that we impose conditions on the parameters in the theorems; e.g., $P_n = \omega \big(n(\ln n)^5\big)$ in Theorem \[thm:rig\], and $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ in Theorem \[thm:urig\]. These conditions are enforced to have the proofs get through and are not that conservative as explained below. We take a binomial random intersection graph as an example and note that Theorem \[thm:rig\] for $k$-connectivity in a binomial random intersection graph does not hold if the condition $P_n = \omega \big(n(\ln n)^5\big)$ in Theorem \[thm:rig\] is replaced by $P_n = n^{\tau}$ for a positive constant $\tau<1$. Specifically, we use [@zz Theorem 4] and [@zz Conjecture 1] confirmed by later work [@2013arXiv1301.0466R] to have the following claim:
\[nPtaulem\] [Under $P_n = n^{\tau}$ for a positive constant $\tau<1$, with a sequence $\gamma_n$ for all $n $ defined through $$\begin{aligned}
p_n P_n =
\ln n + {\gamma_n},\vspace{-1pt} \label{pnPnscaling}
\end{aligned}$$ then $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P}
[G_b(n,P_n,p_n)\text{
is $k$-connected}.]
\nonumber \\ & \quad =\begin{cases} 0, &\text{ if\hspace{2pt}~$\lim_{n \to \infty}{\gamma_n}
=-\infty$}, \\ 1, &\text{ if\hspace{2pt}~$\lim_{n \to \infty}{\gamma_n}
=\infty$.} \end{cases}\vspace{-1pt} \nonumber\end{aligned}$$]{}
Note that different from (\[thm:rig:pe\]), the scaling condition (\[pnPnscaling\]) above does not depend on $k$.
Claim \[nPtaulem\] has $P_n = n^{\tau}$ for a positive constant $\tau<1$, while Theorem \[thm:rig\] has $P_n = \omega \big(n(\ln n)^5\big)$. We let $\delta$ denote an arbitrary constant with $\tau<\delta<\frac{\tau+1}{2}$ below. Claim \[nPtaulem\] shows that the probability of $G_b(n,n^{\tau},n^{-\delta})$ (i.e., $G_b(n,P_n,p_n)$ with $P_n = n^{\tau}$ and $p_n = n^{-\delta}$) being $k$-connected asymptotically converges to $0$ since $\gamma_n$ specified by (\[pnPnscaling\]) satisfies $$\begin{aligned}
\gamma_n & = p_n P_n - \ln n = n^{\tau-\delta} - \ln n \to -\infty, \text{ as $n \to \infty$}. \nonumber
\end{aligned}$$ In contrast, Theorem \[thm:rig\] with $P_n = \omega \big(n(\ln n)^5\big)$ replaced by $P_n = n^{\tau}$ for a positive constant $\tau<1$ presents that the probability of $G_b(n,n^{\tau},n^{-\delta})$ (i.e., $G_b(n,P_n,p_n)$ with $P_n = n^{\tau}$ and $p_n = n^{-\delta}$) being $k$-connected asymptotically approaches to $1$ because $\alpha_n$ defined by (\[thm:rig:pe\]) satisfies $$\begin{aligned}
\alpha_n & = n {p_n}^2 P_n - [\ln n + {(k-1)} \ln \ln n] \nonumber \\ & = n^{1+\tau-2\delta} - [\ln n + {(k-1)} \ln \ln n] \to \infty, \text{ as $n \to \infty$}. \nonumber\end{aligned}$$ Hence, Claim \[nPtaulem\] shows that if the condition $P_n = \omega \big(n(\ln n)^5\big)$ of Theorem \[thm:rig\] is replaced by $P_n = n^{\tau}$ for a positive constant $\tau<1$, we will not obtain the $k$-connectivity result of Theorem \[thm:rig\]. A future work is to investigate the intermediate range $\omega \big(n^{\tau}\big) = P_n = O\big(n(\ln n)^5\big)$. We have discussed the parameter conditions for binomial random intersection graphs. It is unclear whether $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ in Theorem \[thm:urig\] and $K_n = \Omega \big((\ln n)^3\big)$ in Theorem \[thm:urig:rb\] for uniform random intersection graphs can be weakened since these conditions are also often enforced in related work [@Rybarczyk; @ZhaoYaganGligor; @yagan]. Moreover, these conditions are applicable to secure sensor networks since it has been shown that $K_n$ is at least on the order of $\ln n$ to have reasonable connectivity and resiliency [@YaganThesis; @4198829; @DiPietroTissec]. For a general random intersection graph, Yağan [@yagan2015zero] recently obtains a zero–one law for connectivity and shows in [@yagan2015zero Section 3.3] that Theorem \[thm:grig\] with $\text{Var}[X_n] =
o\mathlarger{\mathlarger{\big(}}\frac{\{\mathbb{E}[X_n]\}^2}{ n(\ln
n)^2 }\mathlarger{\mathlarger{\big)}}$ replaced by a broader condition does not hold.
To conclude, the parameter conditions in our theorems are not that conservative.
Comparison with the Conference Version [@ZhaoCDC] {#Discussion:Comp}
-------------------------------------------------
In this paper, we extend the 8-page conference version [@ZhaoCDC] in the following aspects:
1. We improve results in [@ZhaoCDC] on binomial/uniform/general random intersection graphs. Specifically, Theorems 1–6 in this paper eliminate the condition $|\alpha_n| = o(\ln n)$ appearing in [@ZhaoCDC Theorems 1–6]. The approach is to use graph coupling as follows. First, as detailed in Section \[seca:conf:bin\], Lemma \[graph\_Hs\_cpln\] on graph coupling shows that we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in proving Theorems \[thm:rig\] and \[thm:rig:rb\]. Similarly, as shown in Section \[seca:conf:unig\], Lemma \[graph\_Gs\_cpl\] on graph coupling shows that we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in establishing Theorems \[thm:urig\] and \[thm:urig:rb\]. Theorems \[thm:grig\] and \[thm:grig:rb\] are proved using Theorems \[thm:urig\] and \[thm:urig:rb\] which remove $|\alpha_n| = o(\ln n)$ as mentioned above, so Theorems \[thm:grig\] and \[thm:grig:rb\] also eliminate $|\alpha_n| = o(\ln n)$ appearing in [@ZhaoCDC Theorems 3 and 6].
2. For a uniform random intersection graph’s $k$-connectivity result (i.e., Theorem \[thm:urig\]), we detail a complete proof in Section \[prf:thm:urig\]. This result serves as the building block for all other results. In the conference version [@ZhaoCDC], we establish this result from technical report [@mobihoca] on $k$-connectivity in the superposition of a uniform random intersection graph and an Erdős–Rényi graph. This paper presents a complete proof for clarity and completeness.
3. We extend numerical experiments in Section \[sec:expe\] to better confirm the theoretical results.
4. We add a discussion in Section \[ParameterConditions\] on the parameter conditions of the theorems.
5. We compare our results of random intersection graphs with those of Erdős–Rényi graphs below. From [@PES:6114960 Section 1.1], ${p_n}^2 P_n$ in the scaling conditions (\[thm:rig:pe\]) and (\[thm:rig:pe:rb\]) of Theorems \[thm:rig\] and \[thm:rig:rb\] is an asymptotics of the edge probability in a binomial random intersection graph $G_b(n,P_n,p_n)$. Also, by [@bloznelis2013 Lemma 1], $\frac{{K_n}^2}{P_n}$ in the scaling conditions (\[thm:urig:pe\]) and (\[thm:urig:pe:rb\]) of Theorems \[thm:urig\] and \[thm:urig:rb\] (resp., $\frac{\{\mathbb{E}[X_n]\}^2}{P_n}$ in the scaling conditions (\[thm:grig:pe\]) and (\[thm:grig:pe:rb\]) of Theorems \[thm:grig\] and \[thm:grig:rb\] is an asymptotics of the edge probability in a general random intersection graph $G(n,P_n,\mathcal {D}_n)$ (resp., a uniform random intersection graph $G_u(n,P_n,K_n)$). Then comparing Theorems \[thm:rig\]–\[thm:grig\] with Lemma \[er\_kconmnd\], and comparing Theorems \[thm:rig:rb\]–\[thm:grig:rb\] with Lemma \[er\_robust\], we conclude binomial/uniform/general random intersection graphs under certain parameter conditions[^3] exhibit the same behavior with Erdős-Rényi graphs in the sense that for each of (i) $k$-connectivity, (ii) the property of minimum node degree being at least $k$, and (iii) $k$-robustness, a common point for the transition from a zero-law to a one-law occurs when the edge probability equals $\frac{\ln n + {(k-1)} \ln \ln n}{n}$.
$p_n = n^{-\delta}$ for a constant $\delta$ satisfying $\tau<\delta<\frac{\tau+1}{2}$
We refer to ?? with as critical scalings. Then under $P_n = n^{\tau}$ with $0<\tau<1$, the critical $p_n$ for $k$-connectivity in binomial random intersection graph $G_b(n,P_n,p_n)$ is given by $\frac{\ln n}{P_n} = \frac{\ln n}{n^{\tau}} = n^{-\tau}\ln n$, while the critical $p_n$ for $k$-connectivity in Erdős–Rényi graph $G_b(n,P_n,p_n)$ is given by $\sqrt{\frac{\ln n + {(k-1)} \ln \ln n }{nP_n} }= \sqrt{\frac{\ln n + {(k-1)} \ln \ln n }{n\times n^{\tau}} } = n^{-\frac{\tau+1}{2}}[\ln n + {(k-1)} \ln \ln n]$. Given $n^{-\tau}\ln n > n^{-\frac{\tau+1}{2}}[\ln n + {(k-1)} \ln \ln n]$ for all $n$ sufficiently large with $0<\tau<1$,
Numerical Experiments {#sec:expe}
=====================
We present numerical experiments in the non-asymptotic regime to confirm our theoretical results. As we will see, the experimental observations are in accordance with our asymptotic findings.
Figure \[fig\] depicts the probability that a binomial random intersection graph $G_b(n,P,p)$ has $k$-connectivity or $k$-robustness, for $k = 1,2$. Similarly, Figure \[figa\] illustrates the probability of $k$-connectivity or $k$-robustness for $k = 2, 3$ in a uniform random intersection graph $G_u(n,P,K)$. In all set of experiments, we fix the number of nodes at $n=2000$ and the object pool size $P = 20000$. For each pair $(n,P,p)$ (resp., $(n,P,K)$), we generate $1000$ independent samples of $G_b(n,P,p)$ (resp., $G_u(n,P,K)$) and count the number of times that the obtained graphs are $k$-connected or $k$-robust. Then the counts divided by $1000$ become the corresponding empirical probabilities. As illustrated in Figures \[fig\] and \[figa\], there is an evident transition in the probabilities of $k$-connectivity and $k$-robustness. Also, for each $k$, the curves of $k$-connectivity and $k$-robustness are close to each other. Furthermore, the vertical lines in Figure \[fig\] specify $p$ such that ${p}^2 P$ equals $\frac{\ln n + {(k-1)} \ln \ln n}{n}$, while the vertical lines in Figure \[figa\] specify $K$ such that $\frac{{K}^2}{P}$ is closest to $\frac{\ln n + {(k-1)} \ln \ln n}{n}$ (since $K$ and $P$ are both integers, there might not exist $K$ satisfying $\frac{K^2}{P} = \frac{\ln n + {(k-1)} \ln \ln n}{n}$).
The vertical lines in Figure \[fig\] are at $4.4\times 10^{-4}$ and $4.9\times 10^{-4}$ because under $n=2000$ and $P=20000$, $\sqrt{\frac{\ln n + {(k-1)} \ln \ln n}{nP}}$ is $\sqrt{\frac{\ln 2000}{2000 \times 20000}}\approx 4.4\times 10^{-4}$ for $k=1$ and is $\sqrt{\frac{\ln 2000 + \ln \ln 2000}{2000 \times 20000}}\approx 4.9\times 10^{-4}$ for $k=2$. The vertical lines in Figure \[figa\] are at $10$ and $11$ because under $n=2000$ and $P=20000$, ${\operatornamewithlimits{argmin}}_{K} \big| \frac{K^2}{P} - \frac{\ln n + {(k-1)} \ln \ln n}{n}\big|$ equals $10$ for $k=2$ from $\frac{\ln 2000 + \ln \ln 2000}{2000}\approx 0.00481$, $\frac{9^2}{20000}\approx 0.00405$ and $\frac{10^2}{20000}\approx 0.005$, and equals $11$ for $k=3$ from $\frac{\ln 2000 + 2\ln \ln 2000}{2000}\approx 0.00583$, $\frac{10^2}{20000}\approx 0.005$ and $\frac{11^2}{20000}\approx 0.00605$.
These numerical results are in agreement with our analytical findings in the theorems.
We define $p^*$ such that $({p^*})^2 P=\frac{\ln n + {(k-1)} \ln \ln n}{n}$; i.e., $p^*:= \sqrt{\frac{\ln n + {(k-1)} \ln \ln n}{nP}}$. Under $n=2000$ and $P=20000$, we have i) for $k=1$, $p^* = \sqrt{\frac{\ln 2000}{2000 \times 20000}}\approx 4.4\times 10^{-4}$, and ii) for $k=2$, $p^* = \sqrt{\frac{\ln 2000 + \ln \ln 2000}{2000 \times 20000}}\approx 4.9\times 10^{-4}$.
Since $K$ and $P$ are both integers, there might not exist $K$ such that $\frac{K^2}{P} = \frac{\ln n + {(k-1)} \ln \ln n}{n}$. In view of this, we find $K^*$ such that $\big| \frac{K^2}{P} - \frac{\ln n + {(k-1)} \ln \ln n}{n}\big|$ is minimized; i.e., $K^* : = {\operatornamewithlimits{argmin}}_{K} \big| \frac{K^2}{P} - \frac{\ln n + {(k-1)} \ln \ln n}{n}\big|$. Under $n=2000$ and $P=20000$, we have i) $K^* = 10$ for $k=2$, given $\frac{\ln 2000 + \ln \ln 2000}{2000}\approx 0.00481$, $\frac{9^2}{20000}\approx 0.00405$ and $\frac{10^2}{20000}\approx 0.005$, and ii) $K^* = 11$ for $k=3$, given $\frac{\ln 2000 + 2\ln \ln 2000}{2000}\approx 0.00583$, $\frac{10^2}{20000}\approx 0.005$ and $\frac{11^2}{20000}\approx 0.00605$.
![A plot of the empirical probabilities that a binomial random intersection graph $G_b(n,P,p)$ has $k$-connectivity or $k$-robustness as a function of $p$, with $n=2000$, $P=20000$ and $k=2,6$. []{data-label="fig"}](TAC_bin_kconnkrb.eps){width="50.00000%"}
![A plot of the empirical probabilities that a uniform random intersection graph $G_u(n,P,K)$ has $k$-connectivity or $k$-robustness as a function of $K$, with $n=2000$, $P=20000$ and $k=3,4$. []{data-label="figa"}](TAC_uni_kconnkrbv5.eps){width="50.00000%"}
Auxiliary Lemmas {#sec:factlem}
================
We present lemmas that are used in proving the theorems.
Relationships between $k$-Robustness, $k$-Connectivity, and Minimum Node Degree
-------------------------------------------------------------------------------
Lemma \[lem-k-robu-mnd\] below, taken from [@6425841 Lemma 1], provides relationships between $k$-robustness, $k$-connectivity, and minimum node degree.
\[lem-k-robu-mnd\] For any graph/network, $k$-robustness implies $k$-connectivity, which further implies that the minimum node degree is at least $k$.
To prove that $k$-robustness implies $k$-connectivity, [@6425841 Lemma 1] shows that a graph $G$ that is not $k$-connected is also not $k$-robust. The idea is that for $G$ being not $k$-connected, there exists a set of $k-1$ nodes, whose deletion gives two disjoint subgraphs with node sets $V_a$ and $V_b$, respectively. Then in graph $G$, each node in $V_a$ has less than $k$ neighbors outside of $V_a$, and each node in $V_b$ has less than $k$ neighbors outside of $V_b$, so $G$ is not $k$-robust. Then it follows that $k$-robustness implies $k$-connectivity. In addition, it is clear that $k$-connectivity implies that the minimum node degree is at least $k$.
Lemma \[lem-k-robu-mnd\] is used in the proofs of Theorems \[thm:rig:rb\] and \[thm:urig:rb\].
Facts
-----
Recalling that $k$ does not scale with $n$, we obtain Facts \[fact1\] and \[fact\_ln\_n\_n\] below, whose proofs are straightforward and thus omitted here.
\[fact1\]
For $|\alpha_n| = o(\ln n)$, it holds that $$\begin{aligned}
& \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}
\sim \frac{\ln n}{n}. \nonumber \end{aligned}$$
\[scale=.9,auto=left\] (v1) at (1,3) [Theorem 3]{}; (v2) at (6,3) [Theorem 2]{}; (v3) at (6,3.5) [Lemma 1]{}; (v4) at (6,2.5) [Lemma 3]{}; (v5) at (10,3) [Theorem 5]{}; (l52) at (10,2.5) [Lemma 5]{}; (l5) at (10.7,2.48) [ ]{}; (v6) at (14,3) [Theorem 6]{}; (v56) at (12,3.038)[ ]{} ; (v567) at (11.991,2.962)[ ]{} ; (v25) at (8,3)[ ]{} ; (v251) at (8,2.962)[ ]{} ; (v252) at (8,3.038)[ ]{} ; (v7) at (17.5,2.1) [Theorem 4]{}; (v8) at (1,1.2) [Lemma 2]{}; (v82) at (1.65,1.13) [ ]{}; (v92) at (3.5,2.5) [Lemma 4]{}; (v9) at (3.4,2.55) [ ]{}; (v10) at (4.53,3.063) [ ]{}; (v11) at (2.655,1.827) [ ]{}; (v111) at (2.95,1.735) [ ]{}; (v112) at (2.948,1.873) [ ]{}; (v12) at (2.25,1.5) [ ]{}; (v13) at (4.427,1.827) [Theorem 1]{}; (v46) at (15.687,2.13)[ ]{} ; (v461) at (16,2.063)[ ]{} ; (v462) at (16,2.139)[ ]{} ; (v1) – (v2); (v2) – (v5); (v5) – (v6); (v6) – (v461); (v9) – (v10); (v1) – (v111); (v8) – (v112); (v11) – (v13); (v1) edge\[out=16,in=164\] (v567); (v3) – (v251); (v4) – (v252); (l52) – (v56); (v82) – (v462); (v46) – (v7);
\[fact\_ln\_n\_n\]
For $|\alpha_n| = o(\ln n)$, we have $$\begin{aligned}
& \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}
\cdot \left[1 \pm
o\left(\frac{1}{ \ln n}\right)\right] \nonumber \\
& \quad = \frac{\ln n +
{(k-1)} \ln \ln n + {\alpha_n} \pm o(1)}{n}, \nonumber \end{aligned}$$ and $$\begin{aligned}
& \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}
\cdot \left[1 \pm
O\left(\frac{1}{ \ln n}\right)\right] \nonumber \\
& \quad = \frac{\ln n +
{(k-1)} \ln \ln n + {\alpha_n} \pm O(1)}{n}. \nonumber \end{aligned}$$
Results of Erdős-Rényi Graphs
-----------------------------
Lemma \[er\_kconmnd\] below by Erdős and Rényi [@erdos61conn] investigates $k$-connectivity and minimum node degree in Erdős-Rényi graphs. An Erdős–Rényi graph $G(n,\hat{p}_n)$ [@citeulike:4012374] is defined on a set of $n$ nodes such that any two nodes have an edge in between independently with probability $\hat{p}_n$.
\[er\_kconmnd\] For an Erdős–Rényi graph $G(n,\hat{p}_n)$, with a sequence $\alpha_n$ for all $n$ through $$\begin{aligned}
\hat{p}_n = & \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n} , \nonumber
\end{aligned}$$ then it holds that $$\begin{aligned}
& \lim_{n \to \infty} \hspace{-1pt}\mathbb{P} \big[G(n,\hat{p}_n)\text{
is $k$-connected}.\big] \nonumber \\ & =
\lim_{n \to \infty} \hspace{-1pt}\mathbb{P} \big[G(n,\hat{p}_n)\text{
has a minimum node degree at least $k$}.\big]\nonumber \\ &= \begin{cases} 0, \text{ if $\lim_{n \to
\infty}{\alpha_n} \hspace{-2pt}=\hspace{-2pt}-\infty$}, \\ 1,
\text{ if $\lim_{n \to \infty}{\alpha_n}
\hspace{-2pt}=\hspace{-2pt}\infty$.}
\end{cases} \nonumber
\end{aligned}$$
Lemma \[er\_kconmnd\] is used in the comparison of random intersection graphs and Erdős–Rényi graphs.
Lemma \[er\_robust\] below presents the result on $k$-robustness in Erdős-Rényi graphs.
\[er\_robust\] For an Erdős–Rényi graph $G(n,\hat{p}_n)$, with a sequence $\alpha_n$ for all $n$ through $$\begin{aligned}
\hat{p}_n = & \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n},
\nonumber \end{aligned}$$ then it holds that $$\begin{aligned}
& \lim_{n \to \infty} \hspace{-1pt}\mathbb{P} \big[G(n,\hat{p}_n)\text{
is $k$-robust}.\big] = \begin{cases} 0, \text{ if $\lim_{n \to
\infty}{\alpha_n} \hspace{-2pt}=\hspace{-2pt}-\infty$}, \\ 1,
\text{ if $\lim_{n \to \infty}{\alpha_n}
\hspace{-2pt}=\hspace{-2pt}\infty$.}
\end{cases} \label{er_rb}
\end{aligned}$$
Lemma \[er\_robust\] is applied to Section \[prf:thm:rig:rb\] for proving Theorem \[thm:rig:rb\]. Lemma \[er\_robust\] is also used in the comparison of random intersection graphs and Erdős–Rényi graphs.
To prove Lemma \[er\_robust\], we note the following three facts. (a) The desired result (\[er\_rb\]) with $|\alpha_n| = o(\ln \ln n)$ is demonstrated in [@6425841 Theorem 3]. (b) By [@zz Facts 3 and 7], for any monotone increasing graph property $\mathcal {I}$, the probability that graph $G(n,\hat{p}_n)$ has property $\mathcal {I}$ is non-decreasing as $\hat{p}_n$ increases, where a graph property is called monotone increasing if it holds under the addition of edges. (c) $k$-Robustness is a monotone increasing graph property according to [@6481629 Lemma 3]. In view of (a) (b) and (c) above, we obtain Lemma \[er\_robust\].
Lemmas for Graph Coupling
-------------------------
We present several lemmas for graph coupling below. Formally, a coupling [@zz; @2013arXiv1301.0466R; @ZhaoCDC] of two random graphs $G_1$ and $G_2$ means a probability space on which random graphs $G_1'$ and $G_2'$ are defined such that $G_1'$ and $G_2'$ have the same distributions as $G_1$ and $G_2$, respectively. If $G_1'$ is a spanning subgraph (resp., spanning supergraph) of $G_2'$, we say that under the graph coupling, $G_1$ is a spanning subgraph (resp., spanning supergraph) of $G_2$, where a spanning subgraph (resp., spanning supergraph) is a subgraph (resp., supergraph) that has the same node set with the original graph.
Following Rybarczyk’s notation [@zz], we write $$\begin{aligned}
G_1 \preceq G_2 \quad (\text{resp.}, G_1 \preceq_{1-o(1)} G_2)
\label{g1g2coupling}\end{aligned}$$ if there exists a coupling under which $G_1$ is a [spanning subgraph]{} of $G_2$ with probability $1$ (resp., $1-o(1)$). We write $$\begin{aligned}
G_2 \succeq G_1 \quad (\text{resp.}, G_2 \succeq_{1-o(1)} G_1)
\label{g2g1coupling}\end{aligned}$$ if there exists a coupling under which $G_2$ is a [spanning supergraph]{} of $G_1$ with probability $1$ (resp., $1-o(1)$). According to the definitions above, $G_1 \preceq G_2$ and $G_2 \succeq G_1$ are equivalent, while $G_1 \preceq_{1-o(1)} G_2$ and $G_2 \succeq_{1-o(1)} G_1$ are equivalent.
In view that $k$-connectivity and $k$-robustness are monotone increasing graph properties [@erdos61conn; @6481629], it is natural to obtain that under $G_1 \preceq G_2$ or $G_1 \preceq_{1-o(1)} G_2$, if $G_1$ is $k$-connected (resp. $k$-robust) with a probability at least $1-o(1)$, then $G_2$ is also $k$-connected (resp. $k$-robust) with a probability at least $1-o(1)$. This result is formally presented in Lemma \[mono-gcp\] below given by Rybarczyk [@zz]. Lemma \[mono-gcp\] considers any monotone increasing graph property for generality.
\[mono-gcp\]
For two random graphs $G_1$ and $G_2$, the following results hold for any monotone increasing graph property $\mathcal {I}$.
- If $G_1 \preceq G_2$, then $$\mathbb{P} \big[ G_2\text{
has $\mathcal {I}$.}\big] \geq \mathbb{P} \big[ G_1\text{ has
$\mathcal {I}$.}\big].$$
- If $G_1 \preceq_{1-o(1)} G_2$, then $$\mathbb{P} \big[ G_2\text{
has $\mathcal {I}$.}\big] \geq \mathbb{P} \big[ G_1\text{ has
$\mathcal {I}$.}\big]- o(1).$$
### Couplings for Graphs Modeling the Studied Networks
$$\begin{aligned}
& \mathbb{P} \big[\hspace{2pt}G_u(n, P_n, K_n)\text{
has property $\mathcal {I}$}.\hspace{2pt}\big] \nonumber \\
& \leq \mathbb{P} \big[\hspace{2pt}G_u(n,P_n, \widetilde{K_n})\text{
has property $\mathcal {I}$}.\hspace{2pt}\big].\nonumber
\end{aligned}$$
Hence, we obtain that:
$$\begin{aligned}
\hspace{-10pt}\begin{array}{l}\text{if $\lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}G_u(n,P_n, K_n)\text{
has property $\mathcal {I}$}.\hspace{2pt}\big] = 0$,} \\ \text{then $\lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}G_u(n,P_n, \widetilde{K_n})\text{
has property $\mathcal {I}$}.\hspace{2pt}\big] = 0$.}\end{array}\end{aligned}$$
$$\begin{aligned}
& \mathbb{P} \big[\hspace{2pt}G_u(n, P_n, K_n)\text{
has property $\mathcal {I}$}.\hspace{2pt}\big] \nonumber \\
& \leq \mathbb{P} \big[\hspace{2pt}G_u(n,P_n, \widehat{K_n})\text{
has property $\mathcal {I}$}.\hspace{2pt}\big].\nonumber
\end{aligned}$$
Hence, we obtain that:
$$\begin{aligned}
\hspace{-10pt}\begin{array}{l}\text{if $\lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}G_u(n,P_n, K_n)\text{
has property $\mathcal {I}$}.\hspace{2pt}\big] = 0$,} \\ \text{then $\lim_{n \to \infty}\mathbb{P} \big[\hspace{2pt}G_u(n,P_n, \widehat{K_n})\text{
has property $\mathcal {I}$}.\hspace{2pt}\big] = 0$.}\end{array}\end{aligned}$$
Lemma \[mono-gcp\] is used in many places of this paper. We then present Lemmas \[lem:cp\]–\[cp\_urig\_rig\]. Except Lemma \[rkgikg\] which is from [@Rybarczyk Lemma 4], the proofs of other lemmas are deferred to Section \[sec:prf:fact:lem\].
### Coupling between general random intersection graphs and uniform random intersection graphs {#secffalem:cp}
\[lem:cp\] Let $X_n$ be a random variable with probability distribution $\mathcal {D}_n$. If $\text{Var}[X_n] =
o\mathlarger{\mathlarger{\big(}}\frac{\{\mathbb{E}[X_n]\}^2}{n (\ln
n)^2 }\mathlarger{\mathlarger{\big)}}$, then there exists $\epsilon_n = o\big(\frac{1}{\ln n}\big)$ such that $$\begin{aligned}
G_u\big(n, P_n,
(1 -& \epsilon_n) \mathbb{E}[X_n]\big) \preceq_{1-o(1)} G(n,P_n,\mathcal {D}_n) \nonumber \\
& \preceq_{1-o(1)} G_u\big(n, P_n,
(1 + \epsilon_n)\mathbb{E}[X_n]\big). \label{cptac15new}\end{aligned}$$
Lemma \[lem:cp\] is shown in Section \[Couplinggeneraluniform\], and is used to prove Theorems \[thm:grig\] and \[thm:grig:rb\].
### Coupling between binomial random intersection graphs and Erdős–Rényi graphs {#secfagcp_rig_er}
\[cp\_rig\_er\]
If $p_n = O\left( \frac{1}{n\ln n} \right)$ and ${p_n}^2
P_n =
O\left( \frac{1}{\ln n} \right)$, then there exists $\hat{p}_n =
{p_n}^2 P_n \cdot \left[1-
O\left(\frac{1}{ \ln n}\right)\right]$ such that $$\begin{aligned}
G(n,\hat{p}_n) \preceq_{1-o(1)} G_b(n,P_n,p_n). \label{cp_res_rig_er}\end{aligned}$$
Lemma \[cp\_rig\_er\] is shown in Section \[CouplingbinomialER\], and is used to prove Theorem \[thm:rig:rb\].
### Coupling between binomial random intersection graphs and uniform random intersection graphs {#secacp_rkgikg}
\[rkgikg\] If $p_n P_n = \omega\left( \ln n \right)$, and for all $n$ sufficiently large, $$\begin{aligned}
K_{n,-} & \leq p_n P_n - \sqrt{3(p_n P_n + \ln n) \ln n} ,
\nonumber \\
K_{n,+} & \geq p_n P_n + \sqrt{3(p_n P_n + \ln n) \ln n} ,
\nonumber\end{aligned}$$ then $$\begin{aligned}
G_u(n,P_n,K_{n,-})
& \preceq_{1-o(1)}
G_b(n,P_n,p_n) \nonumber \\
& \preceq_{1-o(1)}
G_u(n,P_n,K_{n,+}). \nonumber\end{aligned}$$
Lemma \[rkgikg\] is used in the proof of Theorem \[thm:rig\].
\[cp\_urig\_rig\]
If $K_n \hspace{-1pt}=\hspace{-1pt} \omega\left( \ln n \right)$ and $p_n \hspace{-1pt}=\hspace{-1pt} \frac{K_n}{P_n}
\left(1 - \sqrt{\frac{3\ln
n}{K_n }}\hspace{2pt}\right)$, then $$\begin{aligned}
G_u(n,P_n,K_n) \succeq_{1-o(1)} G_b(n,P_n,p_n). \nonumber\end{aligned}$$
Lemma \[cp\_urig\_rig\] is established in Section \[Couplingcp\_urig\_rig\], and is used to prove Theorem \[thm:urig:rb\].
We will use each of Lemmas \[lem:cp\]–\[cp\_urig\_rig\] along with Lemma \[mono-gcp\]. For simplicity, we just use Lemma \[lem:cp\] as an example to explain its implication with Lemma \[mono-gcp\]. From property (b) of Lemma \[mono-gcp\] and result (\[cptac15new\]) of Lemma \[lem:cp\], we obtain for any monotone increasing graph property $\mathcal {I}$ that $$\begin{aligned}
& \mathbb{P} \big[\hspace{2pt}\text{Graph }G_u(n, P_n,
(1 - \epsilon_n)\mathbb{E}[X_n])\text{ has $\mathcal {I}$}.
\hspace{2pt}\big] - o(1)
\nonumber \\
& \leq \mathbb{P} \big[\hspace{2pt}\text{Graph }G(n,P_n,\mathcal {D}_n)\text{
has $\mathcal {I}$}.\hspace{2pt}\big] \nonumber \\
& \leq \mathbb{P} \big[\hspace{2pt}\text{Graph }G_u(n, P_n,
(1 + \epsilon_n)\mathbb{E}[X_n])\text{ has $\mathcal {I}$}.
\hspace{2pt}\big] + o(1).\nonumber
\end{aligned}$$
Establishing Theorems \[thm:rig\] and \[thm:grig\] {#sec:thmprf:kcon}
==================================================
Theorems \[thm:rig\]–\[thm:grig\] describe results on $k$-connectivity for binomial/uniform/general random intersection graphs. We prove Theorems \[thm:rig\] and \[thm:grig\] in this section, and present the proof of Theorem \[thm:urig\] separately as Section \[prf:thm:urig\] next due to the length of the proof.
We briefly explain the idea of proving Theorems \[thm:rig\] and \[thm:grig\] from Theorem \[thm:urig\] below. First, we demonstrate Theorem \[thm:rig\] from Theorem \[thm:urig\] using the coupling between binomial random intersection graphs and uniform random intersection graphs given by Lemma \[rkgikg\] of Section \[secacp\_rkgikg\]. Second, we establish Theorem \[thm:grig\] from Theorem \[thm:urig\] using the coupling between general random intersection graphs and uniform random intersection graphs given by Lemma \[lem:cp\] of Section \[secffalem:cp\].
The Proof of Theorem \[thm:rig\] {#pfrig}
--------------------------------
As explained in Appendix \[seca:conf:bin\], we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in proving Theorem \[thm:rig\]. Then from Theorem \[thm:urig\], Lemmas \[mono-gcp\] and \[rkgikg\], and and the fact that both $k$-connectivity and the property of minimum node degree being at least $k$ are monotone increasing graph properties, the proof of Theorem \[thm:rig\] is completed once we show that with $K_{n,\pm}$ given by $$\begin{aligned}
K_{n,\pm} & = p_n P_n \pm \sqrt{3(p_n P_n + \ln n) \ln n} ,
\label{Kngeqsb}\end{aligned}$$ under conditions of Theorem \[thm:rig\] and $|\alpha_n| = o(\ln n)$, we have $K_{n,\pm} =
\Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and with $\alpha_{n,\pm}$ defined by $$\begin{aligned}
\frac{{K_{n,\pm}}^2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_{n,\pm}}}{n}, \label{thm:urig:peab}
\end{aligned}$$ then $$\begin{aligned}
\alpha_{n,\pm} & = \alpha_{n} \pm o(1). \label{thm:urig:peaaph}\end{aligned}$$
From conditions (\[thm:rig:pe\]) and $|\alpha_n| = o(\ln n)$, it is clear that $$\begin{aligned}
{p_n}^2 P_n & \sim
\frac{\ln n}{n} .\label{thm:rig:pe_sim}\end{aligned}$$ Substituting (\[thm:rig:pe\_sim\]) and condition $P_n = \omega
\big(n(\ln n)^5\big)$ into (\[Kngeqsb\]), we obtain $$\begin{gathered}
p_n P_n = \sqrt{{p_n}^2 P_n \cdot P_n} = \omega
\bigg( \frac{\ln n}{n} \cdot n(\ln n)^5\bigg) = \omega \big((\ln n)^3\big) , \nonumber
\\
K_{n,\pm} = \omega \big((\ln n)^3\big) = \Omega \big(\sqrt{\ln
n}\hspace{2pt}\big) , \label{zh1}\end{gathered}$$ and $$\begin{aligned}
\frac{{K_{n,\pm}}^2}{P_n} & = {p_n}^2 P_n \cdot \Bigg[1 \pm \sqrt{3\bigg(1 + \frac{\ln n}{p_n P_n}\bigg) \frac{\ln n}{p_n P_n}}\,\Bigg] \nonumber \\ & = {p_n}^2 P_n \cdot \bigg[1 \pm o\bigg(\frac{1}{\ln n}\bigg)\bigg]. \label{thm:urig:pea}\end{aligned}$$ Then from (\[thm:rig:pe\]) (\[thm:urig:peab\]) and (\[thm:urig:pea\]), we obtain (\[thm:urig:peaaph\]). As explained before, with (\[thm:urig:peab\]) (\[thm:urig:peaaph\]) and (\[zh1\]), Theorem \[thm:rig\] is proved from Theorem \[thm:urig\] and Lemmas \[mono-gcp\] and \[rkgikg\].
Finally, as noted in Remark \[rm\] after Theorem \[thm:rig\], to prove zero–one laws (\[bin-kon-0\]) (\[bin-kon-1\]) (\[bin-kon-0-mnd\]) (\[bin-kon-1-mnd\]) but not (\[bin-kon-e\]) (\[bin-kon-e-mnd\]) in Theorem \[thm:rig\], condition $P_n = \omega
\big(n(\ln n)^5\big)$ can be weakened as $P_n = \Omega \big(n(\ln
n)^5\big)$. This is seen by the argument that under $P_n =
\Omega \big(n(\ln n)^5\big)$, $K_{n,\pm} =
\Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ still holds and (\[thm:urig:peaaph\]) is weakened as $\alpha_{n,\pm} = \alpha_{n} \pm O(1)$, so we still have zero–one laws (\[bin-kon-0\]) (\[bin-kon-1\]) (\[bin-kon-0-mnd\]) (\[bin-kon-1-mnd\]).
To prove Theorem \[thm:rig\] using Theorem \[thm:grig\], we will show that given conditions in Theorem \[thm:rig\], conditions in Theorem \[thm:grig\] also hold.
For binomial random intersection graph $G_b(n,P_n,p_n)$, the corresponding $\mathcal {D}_n$ is a binomial distribution with parameters $P_n$ as the number of trials and $p_n$ as the success probability in each trial. Then for a random variable $X_n$ following probability distribution $\mathcal {D}_n$, we obtain $$\begin{aligned}
\mathbb{E}[X_n] & = p_n P_n , \label{thm:rig:ex}\end{aligned}$$ and $$\begin{aligned}
\text{Var}[X_n] & = p_n P_n(1-p_n). \label{thm:rig:varx}\end{aligned}$$ Given (\[thm:rig:ex\]) and condition (\[thm:rig:pe\]), we have $$\begin{aligned}
\frac{\big\{\mathbb{E}[X_n]\big\}^2}{P_n} & = {p_n}^2 P_n =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}. \label{expn}\end{aligned}$$
Then from conditions (\[thm:rig:pe\]) and $|\alpha_n| = o(\ln n)$, and Fact \[fact1\], it is clear that $$\begin{aligned}
{p_n}^2 P_n & \sim
\frac{\ln n}{n} .\label{thm:rig:pe_sim}\end{aligned}$$
From (\[thm:rig:pe\_sim\]) and condition $P_n = \omega \big(n(\ln
n)^3\big)$, it holds that $$\begin{aligned}
p_n & \sim
\sqrt{\frac{\ln n}{nP_n}} = o\Bigg(\sqrt{\frac{\ln n}{n^2(\ln n)^3}}\hspace{2pt}\Bigg)
= o\bigg(\frac{1}{n\ln n}\bigg) , \label{thm:rig:pn}\end{aligned}$$ which along with $P_n = \omega \big(n(\ln n)^3\big)$ is substituted into (\[thm:rig:ex\]) to derive $$\begin{aligned}
\mathbb{E}[X_n] \hspace{-1.1pt}& = \hspace{-1.1pt}\sqrt{\frac{\ln
n}{nP_n}} \cdot P_n \hspace{-1.1pt}=\hspace{-1.1pt} \sqrt{\frac{\ln
n}{n}} \cdot \omega \big(\sqrt{n(\ln n)^3}\hspace{1.5pt}\big)
\hspace{-1.1pt}= \hspace{-1.1pt}\omega \big((\ln n)^2\big),
\label{ex}\end{aligned}$$ inducing $\mathbb{E}[X_n] = \Omega (\sqrt{\ln n}\hspace{2pt})$, a condition of Theorem \[thm:grig\].
From (\[thm:rig:ex\]) (\[thm:rig:varx\]) (\[thm:rig:pn\]) and (\[ex\]), it follows that $$\begin{aligned}
\frac{\text{Var}[X_n]}{\{\mathbb{E}[X_n]\}^2} =
\frac{1-p_n}{\mathbb{E}[X_n]}
= o\bigg(\frac{1}{ (\ln
n)^2 }\bigg). \label{varx}\end{aligned}$$
In view of (\[expn\]) (\[ex\]) and (\[varx\]), we have proved that all conditions in Theorem \[thm:grig\] hold given conditions in Theorem \[thm:rig\] here, so we use Theorem \[thm:grig\] with graph $G(n,P_n,\mathcal {D}_n)$ setting as its instance $G_b(n,P_n,p_n)$. Then it is a simple matter to see that results of Theorem \[thm:rig\] follow.
The Proof of Theorem \[thm:grig\]
---------------------------------
Given Lemmas \[mono-gcp\] and \[lem:cp\] and the fact that both $k$-connectivity and the property of minimum node degree being at least $k$ are monotone increasing graph properties, we will show Theorem \[thm:grig\] once proving for any $\epsilon_n = o\left(\frac{1}{\ln n}\right)$ that $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \big[ G_u(n, P_n,
(1 \pm \epsilon_n)\mathbb{E}[X_n])
\text{
is $k$-connected}. \big] \nonumber \\
& =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$,} \\ e^{- \frac{e^{-\alpha ^*}}{(k-1)!}},
&\text{ if $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$,} \end{cases} \label{kconn}
\end{aligned}$$ and $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_u(n, P_n,
(1 \pm \epsilon_n)\mathbb{E}[X_n])\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \nonumber \\
& =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$,} \\ e^{- \frac{e^{-\alpha ^*}}{(k-1)!}},
&\text{ if $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$.} \end{cases} \label{kconn-mdn}
\end{aligned}$$
Under $\mathbb{E}[X_n] = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $\epsilon_n = o\left(\frac{1}{\ln n}\right)$, it follows that $(1 \pm \epsilon_n)\mathbb{E}[X_n]
= \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$. From Theorem \[thm:urig\], we will have (\[kconn\]) and (\[kconn-mdn\]) once we prove that sequences $\gamma_n^{+}$ and $\gamma_n^{-}$ defined through $$\begin{aligned}
\frac{\big\{(1\pm \epsilon_n)\mathbb{E}[X_n]\big\}^2}{P_n}
& = \frac{\ln n + {(k-1)}
\ln \ln n + {\gamma_n^{\pm}} }{n}\label{pe_epsilon4tac}
\end{aligned}$$ satisfy $$\begin{aligned}
\lim_{n \to \infty} \gamma_n^{\pm} & =
\begin{cases} -\infty, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ \infty, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$,} \\ \alpha ^* ,
&\text{ if $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$.} \end{cases} \label{pe_epsilon4tac2}
\end{aligned}$$ Now we establish (\[pe\_epsilon4tac2\]). From (\[thm:grig:pe\]) (\[pe\_epsilon4tac\]) and $\epsilon_n = o\left(\frac{1}{\ln n}\right)$, it follows that $$\begin{aligned}
\gamma_n^{\pm} & = n \cdot \frac{\big\{(1\pm \epsilon_n)\mathbb{E}[X_n]\big\}^2}{P_n} - [\ln n + {(k-1)}
\ln \ln n ] \nonumber \\
& = (1\pm \epsilon_n)^2 [\ln n + {(k-1)}
\ln \ln n +\alpha_n ] \nonumber \\
& \quad - [\ln n + {(k-1)}
\ln \ln n ] \nonumber \\
& = \alpha_n + \epsilon_n({\epsilon_n}\pm 2) [\ln n + {(k-1)}
\ln \ln n +\alpha_n ] \nonumber \\
& = \alpha_n \pm \bigg[o\bigg(\frac{\alpha_n}{\ln n}\bigg)+o(1)\bigg], \label{pe_epsilon4tac2tc}
\end{aligned}$$ where the last step uses $\epsilon_n = o\left(\frac{1}{\ln n}\right)$. Then (\[pe\_epsilon4tac2tc\]) clearly implies (\[pe\_epsilon4tac2\]). Therefore, as mentioned above, we establish (\[kconn\]) (\[kconn-mdn\]) and finally Theorem \[thm:grig\].
We demonstrate Theorem \[thm:grig\] with the help of Theorem \[thm:urig\], the proof of which is detailed in Section \[prf:thm:urig\].
For any $\epsilon_n = o\left(\frac{1}{\ln n}\right)$, it is clear that $$\begin{aligned}
(1\pm \epsilon_n)^2 & = 1\pm 2\epsilon_n + {\epsilon_n}^2
= 1\pm o\left(\frac{1}{\ln n}\right). \label{eps_2}
\end{aligned}$$ We recall conditions (\[thm:grig:pe\]) and $|\alpha_n| = o(\ln
n)$, which together with (\[eps\_2\]) and Fact \[fact\_ln\_n\_n\] yields $$\begin{aligned}
\frac{\big\{(1\pm \epsilon_n)\mathbb{E}[X_n]\big\}^2}{P_n}
& = \frac{\ln n + {(k-1)}
\ln \ln n + {\alpha_n} \pm o(1)}{n}.\label{pe_epsilon4}
\end{aligned}$$ With $\mathbb{E}[X] = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $\epsilon_n = o\left(\frac{1}{\ln n}\right)$, it follows that $(1 \pm \epsilon_n)\mathbb{E}[X]
= \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$, which along with (\[pe\_epsilon4\]) and $|\alpha_n| = o(\ln n)$ enables the use of Theorem \[thm:urig\] to derive
Since $k$-connectivity is a monotone increasing graph property [@zz], Theorem \[thm:grig\] is proved by (\[kconn\]) and Lemma \[lem:cp\].
The Proof of Theorem \[thm:urig\] {#prf:thm:urig}
=================================
As explained in Appendix \[seca:conf:unig\], we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in proving Theorem \[thm:urig\]. Then since a necessary condition for a graph to be $k$-connected is that the minimum degree is at least $k$, (\[uni-kon-0-mnd\]) implies (\[uni-kon-0\]), and we have $$\begin{aligned}
&\mathbb{P}
\big[\hspace{2pt}\text{Graph }G_u(n,P_n,K_n)\text{
is $k$-connected}.\hspace{2pt}\big] \nonumber
\\ & = \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_u(n,P_n,K_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \nonumber
\\ & \quad - \mathbb{P}\bigg[\hspace{-2pt}\begin{array}{l}G_u(n, P_n, K_n)\text{ has a minimum degree}\\\text{at least
}k,\text{ but is not $k$-connected}.\end{array}\hspace{-2pt}\bigg]. \label{kconvsmnd}
\end{aligned}$$ From (\[kconvsmnd\]), we know that (\[uni-kon-1\]) (resp., (\[uni-kon-e\])) will follow from the combination of Lemma \[lemma-1\] below and (\[uni-kon-1-mnd\]) (resp., (\[uni-kon-e-mnd\])), where Lemma \[lemma-1\] uses the extra condition $|\alpha_n| = o(\ln n)$ explained above. Also as mentioned before, (\[uni-kon-0-mnd\]) implies (\[uni-kon-0\]). Therefore, the proof of Theorem \[thm:urig\] will be completed once we demonstrate (\[uni-kon-0-mnd\]) (\[uni-kon-1-mnd\]) (\[uni-kon-e-mnd\]) and Lemma \[lemma-1\], where we also use the extra condition $|\alpha_n| = o(\ln n)$ in proving (\[uni-kon-0-mnd\]) (\[uni-kon-1-mnd\]) (\[uni-kon-e-mnd\]). We let $e^{-\infty} = 0$ and $e^{\infty} = \infty$, so $e^{- \frac{e^{-\lim_{n \to \infty}{\alpha_n}}}{(k-1)!}}$ equals $0$ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$, $1$ if $\lim_{n \to \infty}{\alpha_n}
=\infty$ and $e^{- \frac{e^{-\alpha ^*}}{(k-1)!}}$ if $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$. Then (\[uni-kon-0-mnd\]) (\[uni-kon-1-mnd\]) (\[uni-kon-e-mnd\]) under $|\alpha_n| = o(\ln n)$ can be compactly presented by Lemma \[lemma-2\] below. Hence, the proof of Theorem \[thm:urig\] finally reduces to proving Lemmas \[lemma-1\] and \[lemma-2\].
\[lemma-1\] For a uniform random intersection graph $G_u(n, P_n, K_n)$ under $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $ \frac{{K_n}^2}{P_n} = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n}$, where $\lim_{n \to \infty} \alpha_n$ exists and $|\alpha_n| = o(\ln n)$, it follows that $$\begin{aligned}
\lim_{n \to \infty} \hspace{-2pt} \mathbb{P}\bigg[\hspace{-4pt}\begin{array}{l}G_u(n, P_n, K_n)\text{ has a minimum degree}\\\text{at least
}k,\text{ but is not $k$-connected}.\end{array}\hspace{-4pt}\bigg] & \hspace{-1pt}=\hspace{-1pt} 0. \label{tbprof}
\end{aligned}$$
\[lemma-2\] For a uniform random intersection graph $G_u(n, P_n, K_n)$ under $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $ \frac{{K_n}^2}{P_n} = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n}$, where $\lim_{n \to \infty} \alpha_n$ exists and $|\alpha_n| = o(\ln n)$, it follows that $$\begin{aligned}
& \lim_{n \to \infty} \mathbb{P}
\left[\hspace{2pt}G_u(n, P_n, K_n)\text{ has a minimum degree at least
}k.\hspace{2pt}\right] \nonumber \\
& \quad = e^{- \frac{e^{-\lim_{n \to \infty}{\alpha_n}}}{(k-1)!}}. \nonumber \end{aligned}$$
To prove Lemma \[lemma-1\], we use the following Lemma \[lemma-1rp\] on $G_u(n, P_n, K_n) {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}G(n,\hat{p}_n)$, where $G(n,\hat{p}_n)$ is an Erdős–Rényi graph with $n$ nodes and edge probability $\hat{p}_n$, and the intersection of two graphs $G_A$ and $G_B$ defined on the same node set is constructed on the node set with the edge set being the intersection of the edge sets of $G_A$ and $G_B$.
\[\] \[lemma-1rp\] For a graph $G_u(n, P_n, K_n) {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}G(n,\hat{p}_n)$ under $P_n = \Omega (n)$, $\frac{K_n}{P_n} = o(1)$ and $ \frac{{K_n}^2}{P_n} \cdot \hat{p}_n = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n}$, where $\lim_{n \to \infty} \alpha_n$ exists and $|\alpha_n| = o(\ln n)$, it follows that $$\begin{aligned}
\lim_{n \to \infty} \hspace{-2pt} \mathbb{P} \bigg[\hspace{-4pt}\begin{array}{l}G_u(n, P_n, K_n){\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}G(n,\hat{p}_n)\text{ has a minimum}\\\text{degree at least }k,\text{ but is not $k$-connected}.\end{array}\hspace{-4pt}\bigg] & \hspace{-1pt}=\hspace{-1pt} 0.
\end{aligned}$$
Lemma \[lemma-1rp\] is from our work [@ZhaoYaganGligor Propositions 3 and 4]. Setting $\hat{p}_n = 1$, we have $G_u(n, P_n, K_n) {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}G(n,\hat{p}_n) = G_u(n, P_n, K_n)$ and obtain results on $G_u(n, P_n, K_n)$ from Lemma \[lemma-1rp\]:
- *For $G_u(n, P_n, K_n)$ under $P_n = \Omega (n)$, $\frac{K_n}{P_n} = o(1)$ and $ \frac{{K_n}^2}{P_n} = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n}$, where $\lim_{n \to \infty} \alpha_n$ exists and $|\alpha_n| = o(\ln n)$, result (\[tbprof\]) holds.*
Then clearly, Lemma \[lemma-1\] will be proved once we show conditions in Lemma \[lemma-1\] imply $P_n = \Omega (n)$ and $\frac{K_n}{P_n} = o(1)$. From conditions in Lemma \[lemma-1\], we have $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $ \frac{{K_n}^2}{P_n} = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n} \sim \frac{\ln n }{n} $ given $|\alpha_n| = o(\ln n)$. Then we further get $P_n = {K_n}^2 \big/\frac{{K_n}^2}{P_n} = \Omega \big(\ln n \big/\frac{\ln n }{n} \hspace{1pt}\big) = \Omega (n)$ and $\frac{K_n}{P_n} = \frac{{K_n}^2}{P_n} \big/ K_n = O\big(\frac{\ln n }{n}\big/ \sqrt{\ln n} \hspace{1pt}\big) = o(1)$. Hence, as mentioned above, Lemma \[lemma-1\] is established.
Now we prove Lemma \[lemma-2\]. We let $q_n$ be the edge probability in a uniform random intersection graph $G_u(n, P_n, K_n)$; i.e., two nodes in $G_u(n, P_n, K_n)$ have an edge in between with probability $q_n$. Under conditions of Lemma \[lemma-2\], given $|\alpha_n| = o(\ln n)$, we have $$\begin{aligned}
\frac{{K_n}^2}{P_n} = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_n}}{n} \sim \frac{\ln n }{n} . \label{PnKKst}
\end{aligned}$$ Hence, from [@ZhaoYaganGligor Lemma 8-Property (a)], it follows that $$\begin{aligned}
q_n = \frac{{K_n}^2}{P_n} \left[ 1 \pm O\bigg(\frac{{K_n}^2}{P_n}\bigg) \right] \sim \frac{\ln n }{n} . \label{eq_pe_lnnn}
\end{aligned}$$ Then, by [@mobihoca Section 3], Lemma \[lemma-2\] will follow once we show Lemma \[lemma-3\] below, where $\mathcal {V}_n = \{v_1, v_2,
\ldots, v_n \}$ is the set of nodes in graph $G_u(n, P_n, K_n)$.
\[lemma-3\] For a uniform random intersection graph $G_u(n, P_n, K_n)$ under $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $ q_n \sim \frac{\ln n }{n}$, it follows for integers $m\geq 1$ and $h \geq 0$ that $$\begin{aligned}
& \mathbb{P} [\text{Nodes }v_{1}, v_{2}, \ldots, v_{m}\text{ have
degree }h] \nonumber \\
& \quad \sim (h!)^{-m} (n q_n)^{hm} e^{-m n q_n}.
\label{eqn_node_v12n}\end{aligned}$$
The rest of this section is devoted to proving Lemma \[lemma-3\].
In a uniform random intersection graph $G_u(n, P_n, K_n)$, recalling that $\mathcal {V}_n = \{v_1, v_2,
\ldots, v_n \}$ is the set of nodes, we let $S_i$ be the set of $K_n$ distinct objects assigned to node $v_i \in \mathcal {V}_n$. We further define $\mathcal {V}_m$ as $ \{v_1, v_2, \ldots, v_m\}$ and $\overline{\mathcal {V}_m} $ as $ \mathcal {V}_n \setminus \mathcal {V}_m
$. Among nodes in $\overline{\mathcal {V}_m}$, we denote by $N_i$ the set of nodes neighboring to $v_i$ for $i=1,2,\ldots,m$. We denote $N_i {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}N_j$ by $N_{ij}$, and $S_i {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}S_j$ by $S_{ij}$.
We have the following two observations:
- If node $v_i$ has degree $h$, then $|N_{i}| \leq h$, where the equal sign holds if and only if $v_i$ is directly connected to none of nodes in $\mathcal{V}_m \setminus
\{v_i\}$; i.e., if and only if $\bigcap_{j \in \{1,2,\ldots,m\}\setminus\{i\}}
(S_{ij}=\emptyset)$ happens.
- If $|N_{i}| \leq h$ for any $i=1,2,\ldots,m$, then $$\begin{aligned}
& \bigg|\bigcup_{1\leq i \leq m} N_{i}\bigg| \leq \sum_{1\leq i \leq
m}N_{i} \leq hm , \label{Nileq}\end{aligned}$$ where the two equal signs in (\[Nileq\]) *both* hold if and only if $$\begin{aligned}
\bigg(\bigcap_{1\leq i <j \leq m} (N_{ij}=\emptyset)\bigg) {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\bigg(\bigcap_{1\leq i \leq m}(|N_{i}| = h)\bigg). \vspace{-1pt}\label{Nij}\end{aligned}$$
From i) and ii) above, if nodes $v_{1}, v_{2}, \ldots, v_{m}$ have degree $h$, we have either of the following two cases:
- Any two of $v_{1}, v_{2}, \ldots, v_{m}$ have no edge in between (namely, $\bigcap_{1\leq i <j \leq m}
(S_{ij}=\emptyset)$); and event (\[Nij\]) happens.
- $\big|\bigcup_{1\leq i \leq m} N_{i}\big|
\leq hm -1$.
In addition, if case (a) happens, then nodes $v_{1}, v_{2}, \ldots,
v_{m}$ have degree $h$. However, if case (b) occurs, there is no such conclusion. With $P_a$ (resp., $P_b$) denoting the probability of case (a) (resp., case (b)), we obtain $$\begin{aligned}
& P_a \leq \mathbb{P} [\text{Nodes }v_{1}, v_{2}, \ldots, v_{m}\text{ have
degree }h] \leq P_a + P_b, \vspace{-1pt}\nonumber\end{aligned}$$ where $$\begin{aligned}
P_a = \mathbb{P}\bigg[ \bigg(\bigcap_{1\leq i <j \leq m}
(S_{ij}=\emptyset)\bigg) & {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\bigg(\bigcap_{1\leq i <j \leq m}
(N_{ij}=\emptyset)\bigg) \nonumber \\
& {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\bigg(\bigcap_{1\leq i \leq
m}(|N_{i}| = h)\bigg)\bigg], \vspace{-1pt} \nonumber\end{aligned}$$ and $$\begin{aligned}
P_b & = \mathbb{P}\bigg[\bigg|\bigcup_{1\leq i \leq m} N_{i}\bigg|
\leq hm -1\bigg]. \vspace{-1pt} \nonumber\end{aligned}$$ Hence, (\[eqn\_node\_v12n\]) holds after we prove the following (\[prop2\]) and (\[prop1\]): $$\begin{aligned}
P_b & = o \left((nq_n)^{hm} e^{-m n q_n}\right). \label{prop2}\end{aligned}$$ and $$\begin{aligned}
P_a & \sim (h!)^{-m} (n q_n)^{hm} e^{-m n q_n} \cdot
[1+o(1)], \label{prop1}\end{aligned}$$
We will prove (\[prop2\]) and (\[prop1\]) below. We let $\mathbb{S}_m$ denote the tuple $(S_1,S_2,\ldots,S_m)$. The expression “${{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m = \mathbb{S}_m^*$” means “given $S_1=S_1^*,S_2=S_2^*,\ldots,S_m=S_m^*$”, where $\mathbb{S}_m^* =
(S_1^*,S_2^*,\ldots,S_m^*)$ with $S_1^*,S_2^*,\ldots,S_m^*$ being arbitrary $K_n$-size subsets of the object pool $\mathcal {P}_n$ (see Page in the graph definition for the meaning of $\mathcal {P}_n$). Note that $S_{ij}^{*} : = S_{i}^{*} \cap S_{j}^{*}$. For two different nodes $v$ and $w$ in $G_u(n, P_n, K_n)$, we use $v\leftrightarrow w$ to denote the event that there is an edge between $v$ and $w$; i.e., the symbol “$\leftrightarrow $” means “is directly connected with”.
The Proof of (\[prop2\])
------------------------
Let $w$ be an arbitrary node in $\overline{\mathcal{V}_m}$. The event $w \in \mathlarger{\cup}_{1\leq i \leq m} N_{i}$ means $w\leftrightarrow \text{ at least one of nodes in }\mathcal{V}_m$, which for different $w$ would be independent given $\mathbb{S}_m= \mathbb{S}_m^*$. Then we have $$\begin{aligned}
& \mathbb{P}\bigg[\bigg|
\bigcup_{1\leq i \leq m} N_{i}\bigg| = t {{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mid}}}}}}\mathbb{S}_m= \mathbb{S}_m^*\bigg] \label{t} \\
& = \frac{(n-m)!}{t!(n-m-t)!} \nonumber \\
& \times \big\{\mathbb{P}[w\leftrightarrow \text{ at least one of nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]\big\}^t \nonumber \\
& \times \big\{\mathbb{P}[w\leftrightarrow \text{none of
nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]\big\}^{n-m-t}.
\label{x}\end{aligned}$$
By the union bound, it holds that $$\begin{aligned}
& \mathbb{P}[w\leftrightarrow \text{at least one of
nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \leq \sum_{1\leq i \leq m}\mathbb{P}
[w\leftrightarrow v_i {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] = m q_n,\label{lll}\end{aligned}$$ which yields $$\begin{aligned}
& \mathbb{P}[w\leftrightarrow \text{none of nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] \geq 1 - m q_n. \label{sstar3}\end{aligned}$$ In addition, we find $$\begin{aligned}
& \mathbb{P}[w\leftrightarrow \text{none of nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\ & = \frac{\binom{P_n -
|\bigcup_{1\leq i \leq m} S_i^*|}{K_n}}{\binom{P_n}{K_n}} \nonumber \\
& \leq (1-q_n)^{{K_n}^{-1}{|\bigcup_{1\leq i \leq m} S_i^*|}} \quad
\text{(by \cite[Lemma 5.1]{yagan_onoff})} \nonumber \\
& \leq e^{-{K_n}^{-1}q_n{|\bigcup_{1\leq i \leq m} S_i^*|}} \quad
\text{(by $1+x \leq e^x$ for any real $x$)}.
\label{sstar4}\end{aligned}$$
We will prove $$\begin{aligned}
& \sum_{\mathbb{S}_m^*} \hspace{-2pt} \Bigg\{ \hspace{-1pt}\mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*] \hspace{-1pt} \times \hspace{-1pt} \bigg\{ \hspace{-1pt} \mathbb{P}\bigg[\begin{array}{l}w\leftrightarrow \text{none of}\\ \text{nodes in }\mathcal{V}_m\end{array} \hspace{-3pt} \bigg|\hspace{2pt} \mathbb{S}_m= \mathbb{S}_m^*\bigg]
\hspace{-1pt}\bigg\}^{n-m-hm}
\hspace{-1pt}\Bigg\} \label{sstar2} \\
& \quad \leq e^{-m
n q_n} \cdot
[1+o(1)] . \label{sstar}\end{aligned}$$ From (\[x\]) (\[lll\]) and (\[sstar\]), we derive $$\begin{aligned}
P_b & = \mathbb{P}\bigg[\bigg|\bigcup_{1\leq i \leq m} N_{i}\bigg|
\leq hm -1\bigg] \nonumber \\ &= \sum_{t=0}^{hm -1}
\sum_{\mathbb{S}_m^*} \Big\{ \mathbb{P}[\mathbb{S}_m = \mathbb{S}_m^*]
\cdot (\ref{t})
\Big\}
\nonumber \\
& \leq \sum_{t=0}^{hm -1} \Big[ n^t \cdot (m q_n)^t \cdot (\ref{sstar2})\Big]
\nonumber \\ & \leq (nq_n)^{hm} e^{-m n q_n} \cdot
[1+o(1)] \cdot m^{hm}
\sum_{t=0}^{hm -1} (mnq_n)^{t-hm}
. \label{bp}\end{aligned}$$ Applying (\[eq\_pe\_lnnn\]) to (\[bp\]), we obtain (\[prop2\]). Hence, we complete the proof of (\[prop2\]) once showing (\[sstar\]), whose proof is detailed below.
From (\[eq\_pe\_lnnn\]) (\[sstar3\]) and (\[sstar4\]), we have $$\begin{aligned}
\textrm{(\ref{sstar2})} \hspace{-1pt}& \hspace{-1pt}\leq\hspace{-1pt} (1 - m q_n)^{-m-hm} \nonumber \\
\hspace{-1pt}& \hspace{-1pt}\quad \times \sum_{\mathbb{S}_m^*}
\Big\{ \mathbb{P}[\mathbb{S}_m = \mathbb{S}_m^*] \cdot
e^{-{K_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m} S_i^*|}}
\Big\} \nonumber \\
\hspace{-1pt}& \hspace{-1pt}\leq\hspace{-1pt}[1+o(1)]\hspace{-1pt}\cdot\hspace{-1pt}\sum_{\mathbb{S}_m^*}
\Big\{ \mathbb{P}[\mathbb{S}_m \hspace{-1pt}=\hspace{-1pt} \mathbb{S}_m^*] \hspace{-1pt}\cdot\hspace{-1pt}
e^{-{K_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m} S_i^*|}} \Big\},\end{aligned}$$ so (\[sstar\]) holds once we demonstrate $$\begin{aligned}
& \sum_{\mathbb{S}_m^*} \Big\{ \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*] \cdot e^{-{K_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m}
S_i^*|}} \Big\} \nonumber \\
& \quad \leq e^{-m n q_n} \cdot
[1+o(1)] . \label{ms}\end{aligned}$$ We denote the left hand side of (\[ms\]) by $Z_{m,n}$. Dividing $\mathbb{S}_{m}^*$ into two parts $\mathbb{S}_{m-1}^*$ and $S_m^*$, we derive $$\begin{aligned}
Z_{m,n}
&= \sum_{\begin{subarray}{c}\mathbb{S}_{m-1}^*
\\S_m^*\end{subarray}} \Big\{ \mathbb{P}[(\mathbb{S}_{m-1} = \mathbb{S}_{m-1}^*)
{\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}(S_m = S_m^*)] \nonumber \\
& \quad\quad\quad\quad \times e^{-{K_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m}
S_i^*|}} \Big\}\nonumber \\ &= \sum_{\mathbb{S}_{m-1}^*}
\mathbb{P}[\mathbb{S}_{m-1} = \mathbb{S}_{m-1}^*] \bigg\{
e^{-{K_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\
& \quad\quad \times \sum_{S_m^* }
\mathbb{P}[ S_m = S_m^* ] e^{-{K_n}^{-1}nq_n{|S_m^* \setminus
\bigcup_{1\leq i \leq m-1} S_i^*|}}\bigg\} ,\label{HnmHnm1}\end{aligned}$$ where $$\begin{aligned}
& \sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{-{K_n}^{-1}nq_n{|S_m^*
\setminus \bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\ & \leq
e^{-n q_n}\sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{{K_n}^{-1}{ n
q_n}\big|S_m^* \cap
\big(\bigcup_{i =1}^{m-1} S_{i }^* \big) \big|} \nonumber \\ &
= e^{-n q_n} \sum_{r=0}^{K_n}
\mathbb{P}\bigg[\bigg|S_m{\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\bigg(\bigcup_{i =1}^{m-1}S_{i
}^*\bigg)\bigg| = r \bigg] e^{{K_n}^{-1}{n q_nr} } . \label{SS}\end{aligned}$$
Denoting $\big|\bigcup_{i=1}^{m-1}S_{i}^*\big|$ by $v$, then for $r$ satisfying the conditions $0 \leq r \leq |S_m^*| = K_n$ and $S_m^* {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}\big(\bigcup_{i=1}^{m-1}S_{i}^*\big) = K_n + v - r \leq P_n $ (i.e., for $r \in [\max\{0, K_n + v - P_n\} , K_n] $), we obtain $$\begin{aligned}
\hspace{-5pt} \mathbb{P}\bigg[\bigg|S_m \hspace{-1pt} {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\hspace{-1pt}
\bigg(\bigcup_{i=1}^{m-1}S_{i}^*\bigg)\bigg| \hspace{-1pt} = \hspace{-1pt} r \bigg] & =
\binom{v}{r} \binom{P_n - v}{K_n - r} \bigg/{\binom{P_n}{K_n}},
\label{probsm}\end{aligned}$$ which together with $ K_n \leq v \leq m K_n$ yields $$\begin{aligned}
& \text{L.H.S. of (\ref{probsm})} \nonumber \\ & \quad \leq \frac{(m K_n)^r}{r!} \cdot
\frac{(P_n - K_n)^{K_n - r}}{(K_n - r)!} \cdot \frac{K_n !}{(P_n - K_n)^{K_n}}\nonumber
\\& \quad \leq \frac{1}{r!} \bigg( \frac{m {K_n}^2}{P_n - K_n}\bigg)^r \text{ for $r \in [\max\{0, K_n + v - P_n\} , K_n] $}. \label{probsm2}\end{aligned}$$ Also, it is clear that $$\begin{aligned}
\text{L.H.S. of (\ref{probsm})} = 0 \text{ for $r \notin [\max\{0, K_n + v - P_n\} , K_n] $} .\label{probsm20clr}\end{aligned}$$
=====
For $r$ satisfying $$\begin{aligned}
0 &\leq r \leq |S_m|=K_n \nonumber\end{aligned}$$ and $$\begin{aligned}
r & = |S_m| + \bigg|\bigcup_{i =1}^{m-1}S_{i
}^*\bigg| - \bigg|S_m{\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}\bigg(\bigcup_{i =1}^{m-1}S_{i
}^*\bigg)\bigg| \nonumber \\
& \geq K_n + \bigg|\bigcup_{i =1}^{m-1}S_{i }^*\bigg|
- P_n , \nonumber\end{aligned}$$ as given in [@mobihocQ1 Eq. (36)], we have $$\begin{aligned}
\mathbb{P}\bigg[\bigg|S_m{\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\bigg(\bigcup_{i =1}^{m-1}S_{i
}^*\bigg)\bigg| = r \bigg] & \leq \frac{1}{r!} \bigg( \frac{m
{K_n}^2}{P_n - K_n}\bigg)^r. \label{probsm2}\end{aligned}$$
Applying (\[probsm2\]) and (\[probsm20clr\]) to (\[SS\]), we establish $$\begin{aligned}
& \sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{-{K_n}^{-1}nq_n{|S_m^*
\setminus \bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\ & \leq
e^{-n q_n} \sum_{r=0}^{K_n} \frac{1}{r!} \bigg( \frac{m {K_n}^2}{P_n - K_n}\bigg)^r
e^{{K_n}^{-1}{n q_nr} } \nonumber \\ &
\leq e^{-n q_n} \cdot e^{\frac{m {K_n}^2}{P_n - K_n}
\cdot e^{{K_n}^{-1}{n q_n}}} . \label{psnm}\end{aligned}$$ From $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and (\[PnKKst\]) (i.e., $\frac{{K_n}^2}{P_n}\sim \frac{\ln n}{n}$), we have $P_n = \omega(K_n)$ and further $$\begin{aligned}
\frac{m {K_n}^2}{P_n - K_n} & \sim \frac{m{K_n}^2}{P_n} \sim
\frac{m\ln n}{n}. \label{e1}\end{aligned}$$ For an arbitrary $\epsilon > 0$, from (\[eq\_pe\_lnnn\]), we obtain $q_n \leq (1+\epsilon)\frac{\ln n}{n}$ for all $n$ sufficiently large, which with $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big) \geq 2$ yields that for all $n$ sufficiently large, $$\begin{aligned}
e^{{K_n}^{-1}{n q_n}} & \leq e^{\frac{1}{2}(1+\epsilon)\ln n} =
n^{\frac{1}{2}(1+\epsilon)}. \label{e2}\end{aligned}$$ From (\[e1\]) and (\[e2\]), we get $$\begin{aligned}
\frac{m {K_n}^2}{P_n - K_n} \cdot e^{{K_n}^{-1}{n q_n}} & \leq m\ln n \cdot n^{\frac{1}{2}(\epsilon-1)} \cdot [1+o(1)] .
\label{e3}\end{aligned}$$ Since $\epsilon > 0$ is arbitrary, it follows from (\[e3\]) that for arbitrary $0<c<\frac{1}{2}$, then for all $n$ sufficiently large, it is clear that $$\begin{aligned}
\frac{m {K_n}^2}{P_n - K_n} \cdot e^{{K_n}^{-1}{n q_n}} & \leq
n^{-c}. \label{e4}\end{aligned}$$ Using (\[e4\]) in (\[psnm\]), for all $n$ sufficiently large, it follows that $$\begin{aligned}
\sum_{S_m^* } \mathbb{P}[ S_m = S_m^* ] e^{-{K_n}^{-1}nq_n{|S_m^*
\setminus \bigcup_{1\leq i \leq m-1} S_i^*|}} & \leq e^{-n q_n}
\cdot e^{n^{-c}} . \label{e5}\end{aligned}$$ Substituting (\[e5\]) into (\[HnmHnm1\]), for all $n$ sufficiently large, we obtain $$\begin{aligned}
&\hspace{-2pt} Z_{m,n} \nonumber \\
& \hspace{-3pt}\leq\hspace{-1.5pt} e^{-n q_n} \hspace{-1.5pt}\cdot\hspace{-1.5pt} e^{n^{-c}}
\hspace{-1.5pt}\cdot \hspace{-1.5pt}\sum_{\mathbb{S}_{m-1}^*} \hspace{-1pt}\mathbb{P}[\mathbb{S}_{m-1} \hspace{-1.5pt}=\hspace{-1.5pt}
\mathbb{S}_{m-1}^*]
e^{-{K_n}^{-1}nq_n{|\bigcup_{1\leq i \leq m-1} S_i^*|}} \nonumber \\
& \hspace{-3pt}\leq\hspace{-1pt} e^{-n q_n} \hspace{-1pt}\cdot \hspace{-1pt}e^{n^{-c}} \hspace{-1pt}\cdot\hspace{-1pt} Z_{m-1,n}.\end{aligned}$$ We then evaluate $Z_{2,n}$. By (\[ms\]), it holds that $$\begin{aligned}
& \hspace{-2pt}Z_{2,n} \nonumber \\
& \hspace{-3pt}=\hspace{-1.5pt}\sum_{S_1^*}\hspace{-1pt}\sum_{S_2^*} \hspace{-1pt}\Big\{ \hspace{-1pt}\mathbb{P}[(S_1 \hspace{-1pt}=\hspace{-1pt}
S_1^*)\hspace{-1pt}{\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\hspace{-1pt}(S_2 \hspace{-1pt}=\hspace{-1pt} S_2^*)] \hspace{-1.5pt}\cdot\hspace{-1.5pt}e^{-{K_n}^{-1}nq_n{|S_1^* {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}S_2^*|}} \Big\}\nonumber \\ & = \sum_{S_1^*} \mathbb{P}[ S_1 =
S_1^* ] \sum_{S_2^*} \mathbb{P}[ S_2 = S_2^* ]
e^{-{K_n}^{-1}nq_n{|S_1^* {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}S_2^*|}}. \label{mm1}\end{aligned}$$ Setting $m=2$ in (\[e5\]), for all $n$ sufficiently large, we derive $$\begin{aligned}
\sum_{S_2^*} \mathbb{P}[ S_2 = S_2^* ] e^{-{K_n}^{-1}nq_n{| S_2^*
\setminus S_1^*|}} & \leq e^{-n q_n} \cdot e^{n^{-c}} . \nonumber\end{aligned}$$ Then for all $n$ sufficiently large, it follows that $$\begin{aligned}
& \sum_{S_2^*} \mathbb{P}[ S_2 = S_2^* ] e^{-{K_n}^{-1}nq_n{|S_1^*
{\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}S_2^*|}} \nonumber \\
& = e^{-n q_n} \sum_{S_2^*} \mathbb{P}[ S_2 =
S_2^* ] e^{-{K_n}^{-1}nq_n{| S_2^* \setminus S_1^*|}} \nonumber \\
& \leq e^{-2n q_n} \cdot e^{n^{-c}} . \label{e6}\end{aligned}$$
From (\[mm1\]) and (\[e6\]), for all $n$ sufficiently large, we obtain $$\begin{aligned}
Z_{m,n} & \leq \big(e^{-n q_n} \cdot e^{n^{-c}}\big)^{m-2}
\cdot Z_{2,n} \nonumber \\
& \leq \big(e^{-n q_n} \cdot e^{n^{-c}}\big)^{m-2} \cdot e^{-2n
q_n} \cdot e^{n^{-c}}
\nonumber \\
& \leq e^{-mn q_n} \cdot e^{(m-1)n^{-c}} .\end{aligned}$$ Letting $n \to \infty$, we finally establish $$\begin{aligned}
Z_{m,n} & \leq e^{-m n q_n} \cdot
[1+o(1)] ; \nonumber\end{aligned}$$ i.e., (\[ms\]) is proved. As explained, (\[sstar\]) and then (\[prop2\]) follow.
The Proof of (\[prop1\])
------------------------
Again let $w$ be an arbitrary node in $\overline{\mathcal{V}_m}$. We have $$\begin{aligned}
& \mathbb{P}\bigg[\bigg(\bigcap_{1\leq i <j \leq m}
(N_{ij}=\emptyset)\bigg) {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}\bigg(\bigcap_{1\leq i \leq
m}(|N_{i}| = h)\bigg) {{\mathlarger{\mathlarger{\mathlarger{\mathlarger{\mid}}}}}}\mathbb{S}_m= \mathbb{S}_m^*\bigg]
\label{h} \\ & = \frac{(n-m)!}{(h!)^m(n-m-hm)!} \nonumber \\
& \quad \times \prod_{1\leq i \leq m}\left(\left\{\mathbb{P}
\left[\begin{array}{l}w\leftrightarrow v_i,\\\text{but }w\leftrightarrow\text{none of}\\\text{nodes
in }\mathcal{V}_m \setminus \{v_i\}\end{array}\Bigg|\hspace{3pt} \mathbb{S}_m= \mathbb{S}_m^*\right]\right\}^h\right) \nonumber \\
& \quad \times \big\{\mathbb{P}[w\leftrightarrow \text{none of
nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]\big\}^{n-m-hm}
\label{y}\end{aligned}$$ and $$\begin{aligned}
P_a & = \sum_{\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq
m} (S_{ij}^*=\emptyset)} \Big\{ \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*] \cdot (\ref{h})
\Big\} \label{pra} ,\end{aligned}$$ where $S_{ij}^{*} : = S_{i}^{*} \cap S_{j}^{*}$.
For $i=1,2,\ldots,m$, under $\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq m}
(S_{ij}^*=\emptyset)$, it follows that $$\begin{aligned}
& \mathbb{P}
[w\leftrightarrow v_i,\text{ but none of nodes
in }\mathcal{V}_m \setminus \{v_i\}{{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \geq \mathbb{P}
[w\leftrightarrow v_i {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]
\nonumber \\
& \quad - \sum_{\begin{subarray}{c} 1 \leq j \leq m \\ j\neq
i\end{subarray}} \mathbb{P}
[w\leftrightarrow \text{both }v_i\text{ and }v_j
{{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] , \label{ps2-nctacsdf}\end{aligned}$$ where we note $$\begin{aligned}
\mathbb{P} [w\leftrightarrow v_i {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] = q_n, \label{nsw0}\end{aligned}$$ and $$\begin{aligned}
& \mathbb{P} [w\leftrightarrow \text{both }v_i\text{ and }v_j
{{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]
\nonumber \\ & = \mathbb{P}
[w\leftrightarrow v_i {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]
+ \mathbb{P}
[w\leftrightarrow v_j {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]
\nonumber \\ & \quad - \mathbb{P}
[ (w\leftrightarrow v_i) {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}(w\leftrightarrow v_j) {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*]
\nonumber \\ & = q_n + q_n - \binom{P_n - 2K_n}{K_n} \bigg/ \binom{P_n}{K_n} \label{nsw1}\end{aligned}$$ given $\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq m}
(S_{ij}^*=\emptyset)$. From [@yagan_onoff Lemma 5.1], we get $\binom{P_n - 2K_n}{K_n} \big/ \binom{P_n}{K_n} \leq ( 1- {q_n })^2$, which with (\[nsw0\]) and (\[nsw1\]) are used in (\[ps2-nctacsdf\]) to derive $$\begin{aligned}
& \mathbb{P}
[w\leftrightarrow v_i,\text{ but none of nodes
in }\mathcal{V}_m \setminus \{v_i\}{{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \quad \geq q_n - (m-1) \cdot 2{q_n} ^2 . \label{ps2}\end{aligned}$$
Substituting (\[sstar3\]) and (\[ps2\]) to (\[y\]), and then from (\[pra\]), we obtain $$\begin{aligned}
P_a & \geq \frac{(n-m-hm)^{hm}}{(h!)^m} \cdot [q_n - 2(m-1) q_n
^2]^{hm} \nonumber \\
& \quad \times (1-mq_n)^{n-m-hm} \sum_{\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i
<j \leq m} (S_{ij}^*=\emptyset)} \mathbb{P}[\mathbb{S}_m =
\mathbb{S}_m^*]. \nonumber\end{aligned}$$ Then from (\[eq\_pe\_lnnn\]), it further hold that $$\begin{aligned}
P_a & \geq \frac{n^{hm}}{(h!)^m} \cdot (q_n)^{hm}
\cdot e^{- m n q_n} \nonumber \\
& \quad \times [1-o(1)] \cdot
\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg]
. \label{poa1}\end{aligned}$$
From (\[sstar4\]), under $\mathbb{S}_m^*:\hspace{2pt}\bigcap_{1\leq i <j \leq m}
(S_{ij}^*=\emptyset)$, it holds that $$\begin{aligned}
\mathbb{P}[w\leftrightarrow \text{none of nodes in }\mathcal{V}_m {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] & \leq e^{- m q_n} . \label{ps3a}\end{aligned}$$
For each $i=1,2,\ldots,m$, we have $$\begin{aligned}
& \mathbb{P}
[w\leftrightarrow v_i,\text{ but }w\leftrightarrow \text{none of nodes
in }\mathcal{V}_m \setminus \{v_i\}{{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] \nonumber \\
& \leq \mathbb{P}
[w\leftrightarrow v_i {{\mathlarger{\boldsymbol\mid}}\hspace{1pt}}\mathbb{S}_m= \mathbb{S}_m^*] =
q_n. \label{ps}\end{aligned}$$ Substituting (\[ps\]) and (\[ps3a\]) to (\[y\]), and then from (\[pra\]), we obtain $$\begin{aligned}
P_a & \hspace{-1.5pt}\leq\hspace{-1.5pt} \frac{n^{hm}}{(h!)^m} \hspace{-1.5pt}\cdot\hspace{-1.5pt} (q_n)^{hm}
\hspace{-1.5pt}\cdot\hspace{-1.5pt} e^{- m n q_n} \hspace{-1.5pt}\cdot\hspace{-1.5pt} \sum_{\mathbb{S}_m^*:\hspace{2pt}\hspace{-1.5pt}\bigcap_{1\leq i
<j \leq m} (S_{ij}^*=\emptyset)} \hspace{-1.5pt}\mathbb{P}[\mathbb{S}_m \hspace{-1.5pt}=\hspace{-1.5pt}
\mathbb{S}_m^*] \nonumber \\
& = \frac{n^{hm}}{(h!)^m} \cdot (q_n)^{hm}
\cdot e^{- m n q_n} \cdot
\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg]
. \label{poa2}\end{aligned}$$
From (\[poa1\]) and (\[poa2\]), we obtain $$\begin{aligned}
P_a & \sim \frac{n^{hm}}{(h!)^m} \cdot (q_n)^{hm}
\cdot e^{- m n q_n} \cdot
\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg].
\label{y2} \vspace{-1pt}\end{aligned}$$ By the union bound, it is clear that $$\begin{aligned}
& \mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m} (S_{ij}=\emptyset)\bigg]\vspace{-1pt}
\nonumber \\ & \geq 1 - \sum_{1\leq i <j \leq m}
\mathbb{P}[S_{ij}\neq \emptyset] = 1 - \binom{m}{2}q_n. \label{m2ps}\vspace{-1pt}\end{aligned}$$ From (\[eq\_pe\_lnnn\]) and (\[m2ps\]), since a probability is at most $1$, we get $$\begin{aligned}
\lim_{n \to \infty}\mathbb{P}\bigg[\bigcap_{1\leq i <j \leq m}
(S_{ij}=\emptyset)\bigg] & = 1 .\vspace{-1pt} \label{m2ps2}\end{aligned}$$ Using (\[m2ps2\]) in (\[y2\]), we establish (\[prop1\]).
================
We derive in [@mobihocQ1] the asymptotically exact probability and an asymptotic zero–one law for $k$-connectivity in graph $G(n,\tilde{p}_n){\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}G_u(n,P_n,K_n)$, which is the superposition of an Erdős–Rényi graph $G(n,\tilde{p}_n)$ on a uniform random intersection graph $G_u(n,P_n,K_n)$. Setting $\tilde{p}_n =
1$, graph $G(n,\tilde{p}_n){\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}G_u(n,P_n,K_n)$ becomes $ G_u(n,P_n,K_n)$. Then with $\tilde{p}_n = 1$, we obtain from [@mobihocQ1 Theorem 1] that if $P_n = \Omega (n)$ and $$\begin{aligned}
1- \binom{P_n- K_n}{K_n} \bigg/ \binom{P_n}{K_n} & = \frac{\ln n +
{(k-1)} \ln \ln n + {\beta_n}}{n}, \label{PnKnKn}\end{aligned}$$ then $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \left[\hspace{2pt}G_u(n,P_n,K_n)\text{
is $k$-connected}.\hspace{2pt}\right] \nonumber \\
& ~~ =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\beta_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\beta_n}
=\infty$,} \\ e^{- \frac{e^{-\beta ^*}}{(k-1)!}},
&\text{ if $\lim_{n \to \infty}{\beta_n}
=\beta ^* \in (-\infty, \infty)$.} \end{cases} \label{beta}
\end{aligned}$$ Note that if $\beta_n = \alpha_n \pm o(1)$, then (i) $\lim_{n \to
\infty}{\beta_n}$ exists if and only if $\lim_{n \to
\infty}{\alpha_n}$ exists; and (ii) when they both exist, $\lim_{n
\to \infty}{\beta_n} = \lim_{n \to \infty}{\alpha_n}$. Therefore, Theorem \[thm:urig\] is proved once we show $P_n =
\Omega (n)$ and (\[PnKnKn\]) with $\beta_n = \alpha_n \pm o(1)$ given conditions $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$, $|\alpha_n| = o(\ln n)$ and (\[thm:urig:pe\]).
From $|\alpha_n| = o(\ln n)$, (\[thm:urig:pe\]) and Fact \[fact1\], it holds that $$\begin{aligned}
\frac{{K_n}^2}{P_n} & \sim \frac{\ln n}{n} , \label{KKP}\end{aligned}$$ which along with $K_n = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ yields $$\begin{aligned}
P_n & \sim \frac{n{K_n}^2}{\ln n} = \Omega (n).\nonumber\end{aligned}$$
We derive in [@ZhaoYaganGligor Lemma 8] that $$\begin{aligned}
1\hspace{-.5pt}-\hspace{-.5pt} \binom{P_n- K_n}{K_n}
\hspace{-.5pt}\bigg/\hspace{-.5pt} \binom{P_n}{K_n} &
\hspace{-.25pt}=\hspace{-.25pt} \frac{{K_n}^2}{P_n}
\hspace{-.25pt}\cdot\hspace{-.25pt} \bigg[1 \pm O\bigg(
\frac{{K_n}^2}{P_n} \bigg) \bigg]. \label{PnKK}\end{aligned}$$ Applying (\[KKP\]) to (\[PnKK\]), $$\begin{aligned}
1\hspace{-.5pt}- \hspace{-.5pt}\binom{P_n- K_n}{K_n}
\hspace{-.5pt}\bigg/\hspace{-.5pt} \binom{P_n}{K_n} &\hspace{-.25pt}
= \hspace{-.25pt}\frac{{K_n}^2}{P_n}
\hspace{-.25pt}\cdot\hspace{-.25pt} \bigg[1 \pm o\left(\frac{1}{ \ln
n}\right)\bigg], \nonumber\end{aligned}$$ which together with (\[thm:urig:pe\]) and Fact \[fact\_ln\_n\_n\] leads to (\[PnKnKn\]) with condition $\beta_n = \alpha_n \pm
o(1)$. Since we have proved $P_n = \Omega (n)$ and (\[PnKnKn\]) with $\beta_n = \alpha_n \pm o(1)$, Theorem \[thm:urig\] follows from (\[beta\]).
Establishing Theorems \[thm:rig:rb\]–\[thm:grig:rb\] {#sec:thmprf:krb}
====================================================
Theorems \[thm:rig:rb\]–\[thm:grig:rb\] present results on $k$-robustness for binomial/uniform/general random intersection graphs. We prove Theorems \[thm:rig:rb\]–\[thm:grig:rb\] in this section and start with explaining the idea below. First, the zero-law of Theorem \[thm:rig:rb\] is established from the zero-law of Theorem \[thm:rig\] since $k$-robustness implies the property of minimum node degree being at least $k$ from Lemma \[lem-k-robu-mnd\] above, while the one-law of Theorem \[thm:rig:rb\] is proven from the coupling between binomial random intersection graphs and Erdős–Rényi graphs given by Lemma \[cp\_rig\_er\] of Section \[secfagcp\_rig\_er\]. Second, the zero-law of Theorem \[thm:urig:rb\] is demonstrated from the zero-law of Theorem \[thm:urig\] because $k$-robustness implies the property of minimum node degree being at least $k$ from Lemma \[lem-k-robu-mnd\] above, while the one-law of Theorem \[thm:urig:rb\] is established from the coupling between binomial random intersection graphs and uniform random intersection graphs given by Lemma \[rkgikg\] of Section \[secacp\_rkgikg\]. Finally, both the zero-law and one-law of Theorem \[thm:grig:rb\] are proved from the coupling between general random intersection graphs and uniform random intersection graphs given by Lemma \[lem:cp\] of Section \[secffalem:cp\].
Before proving Theorems \[thm:rig:rb\]–\[thm:grig:rb\], we first discuss the relationship between $k$-connectivity and $k$-robustness.
The Relationships of $k$-Robustness with $k$-Connectivity and Minimum Node Degree {#kconkrb}
---------------------------------------------------------------------------------
The references [@6481629; @7061412; @6425841] all claim that for any graph/network, $k$-robustness implies $k$-connectivity. However, we show that the above claim does not hold, using a simple example graph illustrated in Figure \[expgraph\]. Recalling the definition of $k$-robustness in Section \[sec:ConnectivityandRobustness\], it is straightforward to show that the graph in Figure \[expgraph\] is $2$-robust yet only $1$-connected (i.e., connected), not $2$-connected. Hence, $k$-robustness actually does not imply $k$-connectivity for general $k$. We also note that $1$-robustness (i.e., robustness) is clearly equivalent to $1$-connectivity (i.e., connectivity). Thus, with “$\Longleftrightarrow$” denoting equivalence and “$\centernot\Longrightarrow$” denoting the relation of not implying, we can write
- [*1-robustness $\Longleftrightarrow$ 1-connectivity, and*]{}
- [*$k$-robustness $\centernot\Longrightarrow$ $k$-connectivity for $k \geq 2$*]{}.
On the other hand, as shown in [@6425841 Figure 1], it holds that
- [*$k$-connectivity $\centernot\Longrightarrow$ $k$-robustness for $k \geq 2$*]{}.
In addition, $k$-robustness implies that the minimum node degree is at least $k$ as given by [@6425841 Lemma 1], which we put as Lemma \[lem-k-robu-mnd\] below for clarity. Note that $k$-connectivity also implies that the minimum node degree is at least $k$.
\[scale=1,auto=left\] (c) at (3,7) [$v_3$]{}; (a) at (1,8) [$v_1$]{}; (d) at (5,8) [$v_4$]{}; (b) at (1,6) [$v_2$]{}; (e) at (5,6) [$v_5$]{}; /in [a/b, a/c, b/c,c/d,c/e,d/e]{} () – ();
\[lem-k-robu-mnd\] For any graph/network, $k$-robustness implies that the minimum node degree is at least $k$.
In Section \[sec:thmprf:krb\], we provide a simple counterexample to disprove the above claim. The true relation between $k$-robustness and $k$-connectivity is that they both imply the property of minimum node degree being at least $k$.
As noted in Section \[sec:ConnectivityandRobustness\], several work [@6481629; @7061412; @6425841] all claim that for any network, $k$-robustness implies $k$-connectivity. Here we provide a simple counterexample to disprove the claim.
In addition,
The example graph is given in Figure \[expgraph\]. As we can see, the graph has a node set of $\mathcal{V} : = \{v_1,v_2,v_3,v_4,v_5\}$ and an edge set of $$\{v_1 \backsim v_2 ,v_1 \backsim v_3, v_2 \backsim v_3, ,v_3 \backsim v_4, v_3 \backsim v_5,v_4 \backsim v_5\},$$ where $\backsim$ means the adjacency between nodes.
We will explain that the graph in Figure \[expgraph\] is $2$-robust. To this end, recalling the definition of $k$-robustness in Section I, we need to show for non-empty subsets $S$ and $T$ of $\mathcal{V}$ satisfying $S {\hspace{2pt} \mathlarger{\cap}
\hspace{2pt}}T = \emptyset$ and $S {\hspace{2pt} \mathlarger{\cup}
\hspace{2pt}}T = \mathcal{V}$ ($S$ and $T$ constitute a partition of $\mathcal{V}$), at least one of the properties (a) and (b) below holds:
- there exists at least a node $v_a \in S$ such that $v_a$ has no less than $2$ neighbors inside $T$, and
- there exists at least a node $v_b \in T$ such that $v_b$ has no less than $2$ neighbors inside $S$.
We call $v_a$ in case (a) or $v_b$ in case (b) as a node having at least $2$ neighbors in the complement subset (the notion of the complement subset is relative to the subset that $v_a$ or $v_b$ belongs to); i.e., if this node belongs to $S$, it has at least $2$ neighbors in $T= \mathcal{V} \setminus S$, and if this node belongs to $T$, it has at least $2$ neighbors in $S= \mathcal{V} \setminus T$. As given in Table I, for each partition $(S,T)$ with non-empty sets $S$ and $T$ (a partition is unordered so $(S,T)$ and $(T,S)$ are the same), there always exists at least one node having no less than $2$ neighbors in the complement subset. In fact, from Table I, we have the following two cases (i) and (ii):
- If a partition has a set with a cardinality of $1$ (i.e., $|S|=1$ or $|T|=1$), then the node in the set with a cardinality of $1$ has at least $2$ neighbors in the complement subset. In other words, for a partition $\{v_i\}$ and $ \{v_1,v_2,v_3,v_4,v_5\} \setminus \{v_i\} $, where $i \in \{1,2,3,4,5\} $, node $v_i$ is a node having at least $2$ neighbors in the complement subset; namely, $\{v_i\}$ has at least $2$ neighbors in $ \{v_1,v_2,v_3,v_4,v_5\} \setminus \{v_i\} $. This clearly holds since each node in the graph in Figure \[expgraph\] has a degree at least $2$.
- For a partition that does not have a set with a cardinality of $1$, there might be multiple nodes having at least $2$ neighbors in the complement subset as we see from Table I. As explained below, node $v_3$ is always a node having at least $2$ neighbors in the complement subset. For a partition $(S,T)$ that does not have a set with a cardinality of $1$, since we also require $S$ and $T$ to be non-empty and the graph has $5$ nodes in total, then we have either $|S|=2, |T|=3$ or $|T|=2, |S|=3$. Since node $v_3$ is neighboring to any other node in the graph, then if node $v_3$ belongs to $S$, it has at least $|T|$ neighbors in the complement subset $T= \mathcal{V} \setminus S$, and if $v_3$ belongs to $T$, it has at least $|S|$ neighbors in the complement subset $S= \mathcal{V} \setminus T$. As we have either $|S|=2, |T|=3$ or $|T|=2, |S|=3$, then node $v_3$ always has at least $2$ neighbors in the complement subset.
From cases (i) and (ii) above, the graph in Figure \[expgraph\] is $2$-robust.
We now explain that the graph in Figure \[expgraph\] is only $1$-connected (i.e., connected), not $2$-connected. First, the graph is clearly $1$-connected since any node pair can find a path in between. Second, the graph is not $2$-connected since if node $v_3$ is removed, the remaining graph becomes disconnected.
To summarize, we have shown that the graph in Figure \[expgraph\] is $2$-robust, but only $1$-connected, not $2$-connected. Hence, the claim [@6425841; @6481629; @TCNS2015] that $k$-robustness implies $k$-connectivity for general $k$ is incorrect. Note that although $k$-robustness does not imply $k$-connectivity for general $k$, it should be emphasized that for $k=1$, $1$-robustness is equivalent to $1$-connectivity (i.e., connectivity). This can be seen straightforwardly from the definitions of $1$-robustness and $1$-connectivity. Therefore, with “$\Longleftrightarrow$” denoting equivalence and “$\centernot\Longrightarrow$” denoting the relation of no implying, we can write
- 1-robustness $\Longleftrightarrow$ 1-connectivity, and
- $k$-robustness $\centernot\Longrightarrow$ $k$-connectivity for $k \geq 2$.
On the other hand, it is simple to derive
- $k$-connectivity $\centernot\Longrightarrow$ $k$-robustness for $k \geq 2$.
To see this, we look at the graph in Figure 2. Clearly, the graph is $2$-connected since the removal of any single node does not disconnect the remaining graph. Yet, the graph is only $1$-robust, not $2$-robust. First, because 1-robustness is equivalent to 1-connectivity (i.e., being connected), the graph is clearly $1$-robust. Second, for a partition $\{v_1, v_2\}$ and $\{v_3, v_4\}$, there does not exist a node having at least $2$ neighbors in the complement subset, since $v_1$ has only one neighbor (i.e., $v_3$) in $\{v_3, v_4\}$, $v_2$ has only one neighbor (i.e., $v_4$) in $\{v_3, v_4\}$, $v_3$ has only one neighbor (i.e., $v_1$) in $\{v_1, v_2\}$, and $v_4$ has only one neighbor (i.e., $v_2$) in $\{v_1, v_2\}$. Therefore, the graph is not $2$-robust. To summarize, Figure 2 presents a graph that is $2$-connected but not $2$-robust.
Although for general $k$, $k$-robustness does not imply $k$-connectivity and $k$-connectivity does not imply $k$-robustness, $k$-robustness implies a property that is related to $k$-connectivity; specifically, $k$-robustness implies that the minimum node degree is at least $k$ as given in Lemma \[lem-k-robu-mnd\] below. Note that the property of the minimum node degree being at least $k$, and $k$-connectivity are related, since the former is a necessary condition of the latter [@ZhaoYaganGligor; @citeulike:505396].
Proof: Lemma \[lem-k-robu-mnd\] is also given as a result in [@6425841 Lemma 1]. Here for clarity, we detail the proof by discussing the case of $k=1$ and the case of $k\geq 2$, respectively.
First, for $k=1$, as explained above, $1$-robustness is equivalent to $1$-connectivity. Since $1$-connectivity implies the minimum node degree is at least $1$ (i.e., there is no isolated node), then $1$-robustness also implies the minimum node degree is at least $1$.
Second, we consider below the case of $k\geq 2$. Because the graph/network is $k$-robust with $k\geq 2$, clearly the graph/network has at least $3$ nodes. With $\mathcal{V}$ denoting the node set of the graph/network, it follows that $|\mathcal{V}| \geq 3$. Let $v$ be an arbitrary node in $\mathcal{V}$. We consider the partition of the node set $\mathcal{V}$ into the two subsets $\{v\}$ and $\mathcal{V}\setminus \{v\}$. Since the graph/network is $k$-robust, by definition, at least one of the properties (a) and (b) below holds:
- $v$ has at least $k$ neighbors in $\mathcal{V}\setminus \{v\}$, and
- there exists at least a node $v_b \in \mathcal{V}\setminus \{v\}$ such that $v_b$ has no less than $k$ neighbors inside $\{v\}$.
Due to $k\geq 2$, the property (b) above does not hold, so property (a) holds, which implies that node $v$ has at least $k$ neighbors in the graph/network. Since node $v$ is an arbitrary node in $\mathcal{V}$, then the minimum node degree of the graph/network is at least $k$.
To sum up, Lemma \[lem-k-robu-mnd\] is proved.
Similar to the process of showing Theorem \[thm:rig\] with the help of Theorem \[thm:urig\], we prove Theorem \[thm:rig:rb\] using Theorem \[thm:rig:rb\], the proof of which is given in Section \[prf:thm:rig:rb\].
As explained in Appendix \[seca:conf:bin\], we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in proving Theorem \[thm:rig:rb\]. Then from Lemma \[rkgikg\] and Theorem \[thm:urig:rb\], the proof of Theorem \[thm:rig:rb\] is completed once we show that with $K_{n,\pm}$ given by $$\begin{aligned}
K_{n,\pm} & = p_n P_n \pm \sqrt{3(p_n P_n + \ln n) \ln n} ,
\label{Kngeqsb:rb}\end{aligned}$$ under conditions of Theorem \[thm:rig:rb\] and $|\alpha_n| = o(\ln n)$, we have $K_{n,\pm} =
\Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and with $\alpha_{n,\pm}$ defined by $$\begin{aligned}
\frac{{K_{n,\pm}}^2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n +
{\alpha_{n,\pm}}}{n}, \label{thm:urig:peab:rb}
\end{aligned}$$ then $$\begin{aligned}
\alpha_{n,\pm} & = \alpha_{n} \pm O(1). \label{thm:urig:peaaph:rb}\end{aligned}$$ From conditions (\[thm:rig:pe:rb\]) and $|\alpha_n| = o(\ln n)$, and Fact \[fact1\], it is clear that $$\begin{aligned}
{p_n}^2 P_n & \sim
\frac{\ln n}{n} .\label{thm:rig:pe_sim:rb}\end{aligned}$$ Substituting (\[thm:rig:pe\_sim:rb\]) and condition $P_n = \Omega
\big(n(\ln n)^5\big)$ into (\[Kngeqsb:rb\]), we obtain $$\begin{gathered}
p_n P_n = \sqrt{{p_n}^2 P_n \cdot P_n} = \Omega
\bigg( \frac{\ln n}{n} \cdot n(\ln n)^5\bigg) = \Omega \big((\ln n)^3\big) , \nonumber
\\
K_{n,\pm} = \Omega \big((\ln n)^3\big) = \Omega \big(\sqrt{\ln
n}\hspace{2pt}\big) , \label{zh1:rb}\end{gathered}$$ and $$\begin{aligned}
\frac{{K_{n,\pm}}^2}{P_n} & = {p_n}^2 P_n \cdot \bigg[1 \pm \sqrt{3\bigg(1 + \frac{\ln n}{p_n P_n}\bigg) \frac{\ln n}{p_n P_n}}\bigg] \nonumber \\ & = {p_n}^2 P_n \cdot \bigg[1 \pm O\bigg(\frac{1}{\ln n}\bigg)\bigg]. \label{thm:urig:pea:rb}\end{aligned}$$ Then from (\[thm:rig:pe:rb\]) (\[thm:urig:peab:rb\]) (\[thm:urig:pea:rb\]) and Fact \[fact\_ln\_n\_n\], we obtain (\[thm:urig:peaaph:rb\]). As explained before, with (\[thm:urig:peab:rb\]) (\[thm:urig:peaaph:rb\]) and (\[zh1:rb\]), Theorem \[thm:rig:rb\] is proved from Lemma \[rkgikg\] and Theorem \[thm:urig:rb\].
The Proof of Theorem \[thm:rig:rb\] {#prf:thm:rig:rb}
-----------------------------------
Since $k$-robustness implies the property of minimum node degree being at least $k$ from Lemma \[lem-k-robu-mnd\], the zero-law of Theorem \[thm:rig:rb\] is clear from (\[uni-kon-0-mnd\]) of Theorem \[thm:rig\] in view that under conditions of Theorem \[thm:rig:rb\], if $\lim_{n \to \infty}{\alpha_n}
=-\infty$, $$\begin{aligned}
& \mathbb{P} \big[\hspace{2pt}G_b(n,P_n,p_n)\text{
is $k$-robust}.\hspace{2pt}\big] \nonumber \\
& \leq \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}G_b(n,P_n,p_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \to 0,\text{ as }n \to \infty.
\label{prf:thm:rig:rb1}
\end{aligned}$$ Note that Theorem \[thm:rig\] uses $P_n = \omega
\big(n(\ln n)^5\big)$ while Theorem \[thm:rig:rb\] uses $P_n = \Omega
\big(n(\ln n)^5\big)$. Above we can use (\[uni-kon-0-mnd\]) since (\[uni-kon-0-mnd\]) still holds under $P_n = \Omega
\big(n(\ln n)^5\big)$ as given in Remark \[rm\] after Theorem \[thm:rig\].
Below we prove the one-law of Theorem \[thm:rig:rb\]. As explained in Appendix \[seca:conf:bin\], we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in proving Theorem \[thm:rig:rb\]. Given (\[thm:rig:pe:rb\]) and $|\alpha_n| =
o(\ln n)$, we have $$\begin{aligned}
{p_n}^2 P_n & \sim
\frac{\ln n}{n} , \nonumber\end{aligned}$$ which together with condition $P_n = \Omega \big(n(\ln n)^5\big)$ leads to $$\begin{aligned}
p_n & \sim
\sqrt{\frac{\ln n}{nP_n}} = O\Bigg(\sqrt{\frac{\ln n}{n^2(\ln n)^5}}\hspace{2pt}\Bigg)
= O\bigg(\frac{1}{n(\ln n)^2}\bigg). \label{thm:rig:pnx}\end{aligned}$$ Noting that (\[thm:rig:pnx\]) implies condition $p_n = O\left( \frac{1}{n\ln n} \right)$ in Lemma \[cp\_rig\_er\], we apply Lemmas \[er\_robust\], \[mono-gcp\] and \[cp\_rig\_er\], and condition (\[thm:rig:pe:rb\]) to derive the following: there exists $\hat{p}_n = \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n} -
O(1)}{n}$ such that if $\lim_{n \to \infty}{\alpha_n} = \infty$, $$\begin{aligned}
& \mathbb{P}[\hspace{2pt}\text{Graph }G_b(n,P_n,p_n)
\text{ is $k$-robust.}
\hspace{2pt}] \nonumber
\\ & \geq \hspace{-1pt}
\mathbb{P}[\hspace{1.5pt}\text{Graph }G(n,\hat{p}_n)
\text{ is $k$-robust.}\hspace{1.5pt} ] \hspace{-1pt}- \hspace{-1pt} o(1)
\hspace{-1pt}\to \hspace{-1pt} 1,\text{ as }n \hspace{-1pt}\to \hspace{-1pt} \infty. \label{prf:thm:rig:rb2}
\end{aligned}$$ The proof of Theorem \[thm:rig:rb\] is completed via (\[prf:thm:rig:rb1\]) and (\[prf:thm:rig:rb2\]).
The Proof of Theorem \[thm:urig:rb\] {#prf:thm:urig:rb}
------------------------------------
As explained in Appendix \[seca:conf:unig\], we can introduce an extra condition $|\alpha_n| = o(\ln n)$ in proving Theorem \[thm:urig:rb\]. Since $k$-robustness implies that the minimum node degree is at least $k$ from Lemma \[lem-k-robu-mnd\], the zero-law of Theorem \[thm:urig:rb\] is clear from Lemma \[lemma-2\] in view that under conditions of Theorem \[thm:urig:rb\] with the extra condition $|\alpha_n| = o(\ln n)$, if $\lim_{n \to \infty}{\alpha_n}
=-\infty$, $$\begin{aligned}
& \mathbb{P} \big[\hspace{2pt}G_u(n,P_n,K_n)\text{
is $k$-robust}.\hspace{2pt}\big] \nonumber \\
& \leq \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}G_u(n,P_n,K_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg]\to 0,\text{ as }n \to \infty.
\label{prf:thm:urig:rb1}
\end{aligned}$$
Below we establish the one-law of Theorem \[thm:urig:rb\] with the help of Theorem \[thm:rig:rb\]. Given $K_n = \Omega \big((\ln
n)^3\big) = \omega\left( \ln n \right)$, we use Lemma \[cp\_urig\_rig\] to obtain that with $p_n$ set by $$\begin{aligned}
p_n & = \frac{K_n}{P_n}
\left(1 - \sqrt{\frac{3\ln
n}{K_n }}\hspace{2pt}\right), \label{pnexpr}
\end{aligned}$$ it holds that $$\begin{aligned}
& \mathbb{P}[\hspace{2pt}\text{Graph }G_u(n,P_n,K_n)
\text{ is $k$-robust.}
\hspace{2pt}] \nonumber
\\ & \quad \geq
\mathbb{P}[\hspace{2pt}\text{Graph }G_b(n,P_n,p_n)
\text{ is $k$-robust.}\hspace{2pt} ] - o(1). \label{robustcomp}
\end{aligned}$$
From (\[thm:urig:pe:rb\]) and $|\alpha_n| =
o(\ln n)$, we obtain $\frac{{K_n}^2}{P_n} \sim \frac{\ln n}{n}$, which together with $K_n = \Omega \big((\ln n)^3\big) $ results in $$\begin{aligned}
P_n & \sim \frac{n{K_n}^2}{\ln n} = \Omega \big(n(\ln n)^5\big),
\label{Pnlnn5}\end{aligned}$$ From $K_n = \Omega \big((\ln n)^3\big) $ and (\[pnexpr\]), it follows that $$\begin{aligned}
{p_n}^2 P_n & = \left[\frac{K_n}{P_n}
\left(1 - \sqrt{\frac{3\ln
n}{K_n }}\hspace{2pt}\right)\right]^2 \cdot P_n \nonumber \\ & =
\frac{{K_n}^2}{P_n}
\cdot \left[1 -
O\left(\frac{1}{ \ln n}\right)\right] .\label{pnPnlnn}
\end{aligned}$$ By (\[thm:urig:pe:rb\]) and (\[pnPnlnn\]), it is clear that $$\begin{aligned}
{p_n}^2 P_n & =
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n} - O(1)}{n}. \label{thm:rig:pe:rbcdx}\end{aligned}$$ Given (\[Pnlnn5\]) (\[thm:rig:pe:rbcdx\]) and $|\alpha_n| =
o(\ln n)$, we use Theorem \[thm:rig:rb\] and (\[robustcomp\]) to get that if $\lim_{n \to \infty}{\alpha_n} =\infty$, then $$\begin{aligned}
& \mathbb{P} \big[\hspace{2pt}G_u(n,P_n,K_n)\text{
is $k$-robust}.\hspace{2pt}\big] \to 1,\text{ as }n \to \infty.
\label{prf:thm:urig:rb2}
\end{aligned}$$ The proof of Theorem \[thm:urig:rb\] is completed via (\[prf:thm:urig:rb1\]) and (\[prf:thm:urig:rb2\]).
The Proof of Theorem \[thm:grig:rb\] {#prf:thm:grig:rb}
------------------------------------
Similar to the process of proving Theorem \[thm:grig\] with the help of Theorem \[thm:urig\], we demonstrate Theorem \[thm:grig:rb\] using Theorem \[thm:urig:rb\], which has been proved above.
Given Lemmas \[mono-gcp\] and \[lem:cp\] and the fact that $k$-robustness is a monotone increasing graph property, we will show Theorem \[thm:grig:rb\] once proving for any $\epsilon_n = o\left(\frac{1}{\ln n}\right)$ that $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \big[ G_u(n, P_n,
(1 \pm \epsilon_n)\mathbb{E}[X_n])
\text{
is $k$-robust}. \big] \nonumber \\
& =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$.} \end{cases} \label{kconn:rb}
\end{aligned}$$ Under $\mathbb{E}[X_n] = \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$ and $\epsilon_n = o\left(\frac{1}{\ln n}\right)$, it follows that $(1 \pm \epsilon_n)\mathbb{E}[X_n]
= \Omega \big(\sqrt{\ln n}\hspace{2pt}\big)$. From Theorem \[thm:urig:rb\], we will have (\[kconn\]) once we prove that sequences $\gamma_n^{+}$ and $\gamma_n^{-}$ defined through $$\begin{aligned}
\frac{\big\{(1\pm \epsilon_n)\mathbb{E}[X_n]\big\}^2}{P_n}
& = \frac{\ln n + {(k-1)}
\ln \ln n + {\gamma_n^{\pm}} }{n}\label{pe_epsilon4tac:rb}
\end{aligned}$$ satisfy $$\begin{aligned}
\lim_{n \to \infty} \gamma_n^{\pm} & =
\begin{cases} -\infty, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ \infty, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$.} \end{cases} \label{pe_epsilon4tac2:rb}
\end{aligned}$$ Note that (\[thm:grig:pe:rb\]) and (\[pe\_epsilon4tac:rb\]) are exactly the same as (\[thm:grig:pe\]) and (\[pe\_epsilon4tac\]), while (\[pe\_epsilon4tac2:rb\]) is a subset of (\[pe\_epsilon4tac2\]). Since (\[pe\_epsilon4tac2tc\]) follows from (\[thm:grig:pe\]) (\[pe\_epsilon4tac\]) and $\epsilon_n = o\left(\frac{1}{\ln n}\right)$, we use (\[thm:grig:pe:rb\]) (\[pe\_epsilon4tac:rb\]) and $\epsilon_n = o\left(\frac{1}{\ln n}\right)$ to obtain (\[pe\_epsilon4tac2tc\]), which further yields (\[pe\_epsilon4tac2:rb\]). Therefore, as mentioned above, we establish (\[kconn:rb\]) and finally Theorem \[thm:grig:rb\].
Note that condition (\[thm:grig:pe:rb\]) is the same as (\[thm:grig:pe\])
, and condition $|\alpha_n| = o(\ln n)$ holds. Then as shown in Theorem \[thm:grig\], for any $\epsilon_n =
o\left(\frac{1}{\ln n}\right)$, from (\[thm:grig:pe\]) (\[eps\_2\]), $|\alpha_n| = o(\ln n)$ and Fact \[fact\_ln\_n\_n\], we obtain (\[pe\_epsilon4\]) here. From $\mathbb{E}[X_n] = \Omega \big((\ln n)^3\big)$ and $\epsilon_n =
o\left(\frac{1}{\ln n}\right)$, it follows that $ (1 \pm \epsilon_n)\mathbb{E}[X_n] = \Omega \big((\ln n)^3\big) $, which along with (\[pe\_epsilon4\]) enables the use of Theorem \[thm:urig:rb\] to yield that for $\mathbb{E}[X_n] = \Omega
\big((\ln n)^3\big)$ and any $\epsilon_n = o\big(\frac{1}{\ln
n}\big)$, we have $$\begin{aligned}
& \lim_{n \to \infty}\mathbb{P} \big[ G_u(n, P_n,
(1 \pm \epsilon_n)\mathbb{E}[X_n])
\text{
is $k$-robust}. \big] \nonumber \\
& =
\begin{cases} 0, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=-\infty$}, \\ 1, &\text{ if $\lim_{n \to \infty}{\alpha_n}
=\infty$.} \end{cases} \label{kconn:rb}
\end{aligned}$$ Since $k$-robustness is a monotone increasing graph property according to [@6481629 Lemma 3], Theorem \[thm:grig:rb\] is proved by (\[kconn:rb\]) and Lemma \[lem:cp\].
Establishing Lemmas in Section \[sec:factlem\] {#sec:prf:fact:lem}
==============================================
Lemmas \[er\_robust\] and \[rkgikg\] are clear in Section \[sec:factlem\]. Below we prove Lemmas \[lem:cp\], \[cp\_rig\_er\] and \[cp\_urig\_rig\].
The Proof of Fact \[fact1\]
---------------------------
Since $k$ does not scale with $n$, for $|\alpha_n| = o(\ln n)$, it is easy to see $$\begin{aligned}
\lim_{n \to \infty} \frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{\ln n}
& = 1. \nonumber
\end{aligned}$$ Then it is trivial to derive (\[fact1eq\]) in Fact \[fact1\].
The Proof of Fact \[fact\_ln\_n\_n\]
------------------------------------
For $|\alpha_n| = o(\ln n)$, by Fact \[fact1\], it holds that $$\begin{aligned}
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}
& = \frac{\ln n}{n} \cdot [1 \pm o(1)],
\end{aligned}$$ yielding $$\begin{aligned}
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}
\cdot o\left(\frac{1}{ \ln n}\right)
&
= o\bigg(\frac{1}{n}\bigg),\nonumber
\end{aligned}$$ and $$\begin{aligned}
\frac{\ln n + {(k-1)} \ln \ln n + {\alpha_n}}{n}
\cdot O\left(\frac{1}{ \ln n}\right)
& = O\bigg(\frac{1}{n}\bigg).\nonumber
\end{aligned}$$ Then (\[fact\_ln\_n\_n\_eq1\]) and (\[fact\_ln\_n\_n\_eq2\]) clearly follow.
The Proof of Lemma \[lem:cp\] {#Couplinggeneraluniform}
-----------------------------
According to [@Rybarczyk Lemma 3], for any monotone increasing graph property $\mathcal {I}$ and any $|\epsilon_n|<1$, $$\begin{aligned}
& \mathbb{P} \big[ G(n,P_n,\mathcal {D}_n)\text{
has $\mathcal {I}$}. \big] - \mathbb{P} \hspace{-1pt} \big[
G_u(n,\hspace{-1.3pt}P_n,\hspace{-1.3pt}(1\hspace{-1.3pt}-\hspace{-1.3pt}
\epsilon_n)\mathbb{E}[X_n])\text{\hspace{-.2pt} has $\mathcal
{I}$}.\hspace{-.2pt} \big] \nonumber \\
& \geq \big\{ 1 - \mathbb{P}\hspace{-.2pt}[X_n \hspace{-1.3pt} <
\hspace{-1.3pt}
(1\hspace{-1.3pt}-\hspace{-1.3pt}\epsilon_n)\mathbb{E}[X_n]] \big\}^n
- 1, \label{coupling1}
\end{aligned}$$ and $$\begin{aligned}
& \mathbb{P} \big[ G(n,P_n,\mathcal {D}_n)\text{
has $\mathcal {I}$}. \big] - \mathbb{P} \hspace{-1pt} \big[
G_u(n,\hspace{-1.3pt}P_n,\hspace{-1.3pt}(1\hspace{-1.3pt}+\hspace{-1.3pt}
\epsilon_n)\mathbb{E}[X_n])\text{\hspace{-.2pt} has $\mathcal
{I}$}.\hspace{-.2pt} \big] \nonumber \\
& \leq 1 - \big\{1 - \mathbb{P}\hspace{-.2pt}[X_n \hspace{-1.3pt}
> \hspace{-1.3pt}
(1\hspace{-1.3pt}+\hspace{-1.3pt}\epsilon_n)\mathbb{E}[X_n]] \big\}^n.
\label{coupling2}
\end{aligned}$$
By (\[coupling1\]) (\[coupling2\]) and the fact that $\lim_{n \to
\infty}(1-m_n)^n = 1$ for $m_n = o\big(\frac{1}{n}\big)$ (this can be proved by a simple Taylor series expansion as in [@ZhaoYaganGligor Fact 2]), the proof of Lemma \[lem:cp\] is completed once we demonstrate that with $\text{Var}[X_n]
= o\Big(\frac{\{\mathbb{E}[X_n]\}^2}{ n(\ln n)^2 }\Big)$, there exists $\epsilon_n = o\big(\frac{1}{\ln n}\big)$ such that $$\begin{aligned}
{{\mathbb{P}}\left[{X_n <
(1 - \epsilon_n)\mathbb{E}[X_n]}\right]} & = o\bigg(\frac{1}{n}\bigg) , \label{leq1esp_lem}
\end{aligned}$$ and $$\begin{aligned}
{{\mathbb{P}}\left[{X_n >
(1 + \epsilon_n)\mathbb{E}[X_n]}\right]} & = o\bigg(\frac{1}{n}\bigg). \label{geq1esp_lem}
\end{aligned}$$ To prove (\[leq1esp\_lem\]) and (\[geq1esp\_lem\]), Chebyshev’s inequality yields $$\begin{aligned}
\mathbb{P} \big[\hspace{2pt} |X_n-\mathbb{E}[X_n]| >
\epsilon_n \mathbb{E}[X_n]\big] & \leq
\frac{\text{Var}[X_n]}{\big\{\epsilon_n \mathbb{E}[X_n]\big\}^2}.
\label{thm:grig:Xbound_lem}
\end{aligned}$$ We set $\epsilon_n$ by $\epsilon_n =
\sqrt[4]{\frac{n\text{Var}[X_n]}{\big\{\mathbb{E}[X_n]\big\}^2}} \cdot
\frac{1}{\sqrt{\ln n}} $. Then given condition $\text{Var}[X_n] = o\Big(\frac{\{\mathbb{E}[X_n]\}^2}{ n(\ln
n)^2 }\Big)$, we obtain $$\begin{aligned}
\epsilon_n & =
o\Bigg(
\sqrt[4]{\frac{1}{(\ln n)^2}} \hspace{2pt}\Bigg) \cdot \frac{1}{\sqrt{\ln n}}
= o\Big(\frac{1}{\ln
n}\Big), \label{eps_lem}\end{aligned}$$ and $$\begin{aligned}
\frac{\text{Var}[X_n]}{\big\{\epsilon_n \mathbb{E}[X_n]\big\}^2} & =
\sqrt{\frac{\text{Var}[X_n]}{n\big\{\mathbb{E}[X_n]\big\}^2}} \cdot
\ln n
= o\bigg(\frac{1}{n}\bigg).
\label{varX_lem}
\end{aligned}$$ By (\[thm:grig:Xbound\_lem\]) (\[eps\_lem\]) and (\[varX\_lem\]), it is straightforward to see that (\[leq1esp\_lem\]) and (\[geq1esp\_lem\]) hold with $\epsilon_n = o\big(\frac{1}{\ln
n}\big)$. Therefore, we have completed the proof of Lemma \[lem:cp\].
The Proof of Lemma \[cp\_rig\_er\] {#CouplingbinomialER}
----------------------------------
In view of [@zz Theorem 1], if ${p_n}^2 P_n < 1$ and $p_n=o\left(\frac{1}{ n}\right)$, with $\hat{p}_n: = {p_n}^2 P_n \cdot \left(1 - n{p_n} + 2 {p_n} -
\frac{{p_n}^2 P_n}{2} \right)$, then (\[cp\_res\_rig\_er\]) follows. Given conditions $p_n = O\left(
\frac{1}{n\ln n} \right)$ and ${p_n}^2 P_n =
O\left( \frac{1}{\ln n} \right)$ in Lemma \[cp\_rig\_er\], ${p_n}^2 P_n < 1$ and $p_n=o\left(\frac{1}{ n}\right)$ clearly hold. Then Lemma \[cp\_rig\_er\] is proved once we show $\hat{p}_n = {p_n}^2 P_n \cdot
\left[1-
O\left(\frac{1}{ \ln n}\right)\right]$, which is easy to see via $$\begin{aligned}
& - n{p_n} + 2 {p_n} - \frac{{p_n}^2 P_n}{2} \nonumber \\
& \hspace{1pt} = \hspace{-1pt} (-n\hspace{-1pt}+\hspace{-1pt}2)
\hspace{-1pt}\cdot\hspace{-1pt} O\left(
\frac{1}{n\ln n} \right) \hspace{-1pt}-\hspace{-1pt} \frac{1}{2}
\hspace{-1pt}\cdot\hspace{-1pt} O\left( \frac{1}{\ln n} \right)
\hspace{-1pt}=\hspace{-1pt} -
O\left(\frac{1}{ \ln n}\right) . \nonumber\end{aligned}$$ Hence, the proof of Lemma \[cp\_rig\_er\] is completed.
The Proof of Lemma \[cp\_urig\_rig\] {#Couplingcp_urig_rig}
------------------------------------
We use Lemma \[rkgikg\] to prove Lemma \[cp\_urig\_rig\]. From $K_n = \omega\left( \ln n \right)$ and $p_n =
\frac{K_n}{P_n}
\left(1 - \sqrt{\frac{3\ln
n}{K_n }}\hspace{2pt}\right)$, we first obtain $p_n P_n = \omega\left( \ln n \right)$ and then for all $n$ sufficiently large, $$\begin{aligned}
& K_n - \left[ p_n P_n + \sqrt{3(p_n P_n + \ln n) \ln n}
\hspace{1.5pt}\right] \nonumber \\ & = K_n \sqrt{\frac{3\ln n}{K_n
}} - \sqrt{3\left[ K_n \left(1 - \sqrt{\frac{3\ln n}{K_n
}}\hspace{2pt}\right) + \ln n\right] \ln n} \nonumber
\\ & = \sqrt{3K_n\ln n} -
\sqrt{3\left[K_n \hspace{-1pt}+ \hspace{-1pt} \sqrt{\ln n} \left(
\sqrt{\ln n} \hspace{-1pt}- \hspace{-1pt}
\sqrt{3K_n}\hspace{2pt}\right) \right ] \hspace{-1pt} \ln n}
\nonumber \\ & \geq 0. \nonumber\end{aligned}$$ Then by Lemma \[rkgikg\], Lemma \[cp\_urig\_rig\] is now established.
Conclusion and Future Work {#sec:Conclusion}
==========================
Under a general random intersection graph model, we derive sharp zero–one laws for $k$-connectivity and $k$-robustness, as well as the asymptotically exact probability of $k$-connectivity, where $k$ is an arbitrary positive integer. A future direction is to obtain the asymptotically exact probability of $k$-robustness for a precise characterization on the robustness strength.
[10]{} J. Zhao, O. Yağan, and V. Gligor, “On Connectivity and Robustness in Random Intersection Graphs,” 2016. Full version of this paper, available online at\
<https://sites.google.com/site/publicworkofzhao/TAC2016.pdf>
N. Abaid and M. Porfiri. Consensus over numerosity-constrained random networks. , 56(3):649–654, 2011.
S. Blackburn and S. Gerke. Connectivity of the uniform random intersection graph. , 309(16), 2009.
M. Bloznelis. Degree and clustering coefficient in sparse random intersection graphs. , 23(3):1254–1289, 2013.
M. Bloznelis, J. Jaworski, and K. Rybarczyk. Component evolution in a secure wireless sensor network. , 53:19–26, January 2009.
B. Bollob[á]{}s. . Cambridge Studies in Advanced Mathematics (No. 73). Cambridge University Press, 2001.
M. Bradonjić, A. Hagberg, N. Hengartner, and A. Percus, “Component evolution in general random intersection graphs,” in [*Workshop on Algorithms and Models for the Web Graph (WAW)*]{}, pp. 36–49, 2010.
H. Chan, V. Gligor, A. Perrig, and G. Muralidharan. On the distribution and revocation of cryptographic keys in sensor networks. , 2(3):233–247, 2005.
M. Deijfen and W. Kets. Random intersection graphs with tunable degree distribution and clustering. , 23:661–674, 2009.
D. Dolev. The byzantine generals strike again. , 3(1):14–30, 1982.
P. Erdős and A. Rényi. On random graphs, [I]{}. , 6:290–297, 1959.
P. Erdős and A. Rényi. On the strength of connectedness of random graphs. , pages 261–267, 1961.
L. Eschenauer and V. Gligor. A key-management scheme for distributed sensor networks. In [*ACM Conference on Computer and Communications Security (CCS)*]{}, 2002.
E. Godehardt and J. Jaworski. Two models of random intersection graphs for classification. , pages 67–81, 2003.
H. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram. Resilient asymptotic consensus in robust networks. , 31(4):766–781, April 2013.
H. LeBlanc, H. Zhang, S. Sundaram, and X. Koutsoukos. Resilient continuous-time consensus in fractional robust networks. In [*American Control Conference (ACC)*]{}, 2013.
K. Rybarczyk. Diameter, connectivity and phase transition of the uniform random intersection graph. , 311, 2011.
K. Rybarczyk. Sharp threshold functions for the random intersection graph via a coupling method. , 18:36–47, 2011.
K. [Rybarczyk]{}. . , 1301.0466 \[math.CO\], January 2013.
K. Singer-Cohen. . PhD thesis, Department of Mathematical Sciences, The Johns Hopkins University, 1995.
O. Yağan. . PhD thesis, Dept. of ECE, University of Maryland, June 2011.
O. Yağan. Performance of the [Eschenauer-Gligor]{} key distribution scheme under an on/off channel. , 58(6):3821–3835, June 2012.
O. Yağan. Zero-one laws for connectivity in inhomogeneous random key graphs. , August 2015. Available online at\
http://arxiv.org/pdf/1508.02407v1.pdf.
O. Yağan and A. M. Makowski. Zero-one laws for connectivity in random key graphs. , 58(5):2983–2999, May 2012.
H. Zhang, E. Fata, and S. Sundaram. A notion of robustness in complex networks. , 2015.
H. Zhang and S. Sundaram. Robustness of complex networks with implications for consensus and contagion. In [*IEEE Conference on Decision and Control (CDC)*]{}, pages 3426–3432, December 2012.
H. Zhang and S. Sundaram. Robustness of information diffusion algorithms to locally bounded adversaries. In [*IEEE American Control Conference (ACC)*]{}, pages 5855–5861, 2012.
H. Zhang and S. Sundaram. A simple median-based resilient consensus algorithm. In [*Annual Allerton Conference on Communication, Control, and Computing (Allerton)*]{}, pages 1734–1741, 2012.
J. Zhao, O. Yağan, and V. Gligor. On asymptotically exact probability of $k$-connectivity in random key graphs intersecting [Erdős–Rényi]{} graphs. 2014. Available online at http://arxiv.org/pdf/1409.6022v1.pdf
J. Zhao, O. Yağan, and V. Gligor. On the strengths of connectivity and robustness in general random intersection graphs. In [*IEEE Conference on Decision and Control (CDC)*]{}, pages 3661–3668, Dec 2014.
J. Zhao, O. Yağan, and V. Gligor. $k$-[C]{}onnectivity in random key graphs with unreliable links. , 61(7):3810–3836, July 2015.
J. Zhao, O. Yağan, and V. Gligor. On $k$-connectivity and minimum vertex degree in random $s$-intersection graphs. In [*ACM-SIAM Meeting on Analytic Algorithmics and Combinatorics (ANALCO)*]{}, pages 1–15, January 2015.
R. Di Pietro, L. Mancini, A. Mei, A. Panconesi, and J. Radhakrishnan. Sensor networks that are provably resilient. In [*International Conference on Security and Privacy for Emerging Areas in Communication Networks (Securecomm)*]{}, pages 1–10, Aug 2006.
R. Di Pietro, L. Mancini, A. Mei, A. Panconesi, and J. Radhakrishnan. Redoubtable sensor networks. , 11(3):13:1–13:22, 2008.
S. Blackburn, D. Stinson, and J. Upadhyay. On the complexity of the herding attack and some related attacks on hash functions. , 64(1-2):171–193, 2012.
[10]{}
K. Singer-Cohen, [*Random intersection graphs*]{}. PhD thesis, Department of Mathematical Sciences, The Johns Hopkins University, 1995.
M. Bloznelis, J. Jaworski, and K. Rybarczyk, “Component evolution in a secure wireless sensor network,” [*Networks*]{}, vol. 53, pp. 19–26, January 2009.
J. Zhao, O. Yağan, and V. Gligor, “Connectivity in secure wireless sensor networks under transmission constraints,” in [*Allerton Conference on Communication, Control, and Computing*]{}, 2014.
O. Yağan and A. M. Makowski, “Zero–one laws for connectivity in random key graphs,” [*IEEE Transactions on Information Theory*]{}, vol. 58, pp. 2983–2999, May 2012.
J. Zhao, O. Yağan, and V. Gligor, “Secure $k$-connectivity in wireless sensor networks under an on/off channel model,” in [*IEEE ISIT*]{}, pp. 2790–2794, 2013.
L. Eschenauer and V. Gligor, “A key-management scheme for distributed sensor networks,” in [*ACM CCS*]{}, 2002.
J. Zhao, O. Yağan, and V. Gligor, “$k$-[C]{}onnectivity in secure wireless sensor networks with physical link constraints — the on/off channel model,” [*arXiv*]{}, 1206.1531 \[cs.IT\], 2012.
M. Bradonjić, A. Hagberg, N. Hengartner, and A. Percus, “Component evolution in general random intersection graphs,” in [*Workshop on Algorithms and Models for the Web Graph (WAW)*]{}, pp. 36–49, 2010.
M. Bloznelis, “Degree and clustering coefficient in sparse random intersection graphs,” [*The Annals of Applied Probability*]{}, vol. 23, no. 3, pp. 1254–1289, 2013.
F. G. Ball, D. J. Sirl, and P. Trapman, “Epidemics on random intersection graphs,” [*The Annals of Applied Probability*]{}, vol. 24, pp. 1081–1128, June 2014.
M. [Bradonji[ć]{}]{}, A. [Hagberg]{}, N. [Hengartner]{}, N. [Lemons]{}, and A. [Percus]{}, “[The phase transition in inhomogeneous random intersection graphs]{},” [*arXiv*]{}, 1301.7320 \[cs.DM\], January 2013.
E. Godehardt and J. Jaworski, “Two models of random intersection graphs for classification,” [*Exploratory Data Analysis in Empirical Research*]{}, pp. 67–81, 2003.
K. [Rybarczyk]{}, “[The coupling method for inhomogeneous random intersection graphs]{},” [*arXiv*]{}, 1301.0466 \[math.CO\], January 2013.
K. Rybarczyk, “Sharp threshold functions for the random intersection graph via a coupling method,” [*The Electronic Journal of Combinatorics*]{}, vol. 18, pp. 36–47, 2011.
S. Blackburn and S. Gerke, “Connectivity of the uniform random intersection graph,” [*Discrete Mathematics*]{}, vol. 309, no. 16, 2009.
K. Rybarczyk, “Diameter, connectivity and phase transition of the uniform random intersection graph,” [*Discrete Mathematics*]{}, vol. 311, 2011.
V. Gligor, A. Perrig, and J. Zhao, “Brief encounters with a random key graph,” [*Lecture Notes in Computer Science*]{}, vol. 7028, 2013.
P. Erdős and A. Rényi, “On random graphs, [I]{},” [*Publicationes Mathematicae (Debrecen)*]{}, vol. 6, pp. 290–297, 1959.
K. Censor-Hillel, M. Ghaffari, G. Giakkoupis, B. Haeupler, and F. Kuhn, “ Tight bounds on vertex connectivity under vertex sampling,” in [*ACM-SIAM Symposium on Discrete Algorithms (SODA)*]{}, 2015.
J. Zhao, O. Yağan, and V. Gligor, “Sharp threshold functions for general random intersection graphs,” 2014. Available online at\
[[http://www.andrew.cmu.edu/user/junzhao/papers/CDCfull.pdf]{}]{}
D. Dolev, “The byzantine generals strike again,” 1981.
J. Zhao, O. Yağan, and V. Gligor, “On topological properties of wireless sensor networks under the $q$-composite key predistribution scheme with on/off channels,” in [*IEEE ISIT*]{}, 2014.
J. Zhao, O. Yağan, and V. Gligor, “Topological properties of wireless sensor networks under the $q$-composite key predistribution scheme with unreliable links,” Technical Report CMU-CyLab-14-002, Carnegie Mellon University, 2014.
J. Zhao, O. Yağan, and V. Gligor, “On $k$-connectivity and minimum vertex degree in random $s$-intersection graphs,” *arXiv e-prints*, 2014.
J. Zhao, “Minimum node degree and $k$-connectivity in wireless networks with unreliable links,” in [*IEEE ISIT*]{}, 2014.
F. Yavuz, J. Zhao, O. Yağan, and V. Gligor, “On secure and reliable communications in wireless sensor networks: [T]{}owards $k$-connectivity under a random pairwise key predistribution scheme,” in [*IEEE ISIT*]{}, 2014.
H. Zhang and S. Sundaram, “Robustness of complex networks with implications for consensus and contagion,” in [*IEEE CDC*]{}, pp. 3426–3432, December 2012.
N. Abaid, I. Igel, and M. Porfiri, “On the consensus protocol of conspecific agents,” [*Linear Algebra and its Applications*]{}, vol. 437, no. 1, pp. 221–235, 2012.
N. Abaid and M. Porfiri, “Consensus over numerosity-constrained random networks,” [*IEEE Transactions on Automatic Control*]{}, vol. 56, pp. 649–654, March 2011.
I. Matei, N. Martins, and J. Baras, “Almost sure convergence to consensus in markovian random graphs,” in [*IEEE CDC*]{}, 2008.
W. P. M. H. Heemels, A. Teel, N. van de Wouw, and D. Nesić, “Networked control systems with communication constraints: Tradeoffs between transmission intervals, delays and performance,” [*IEEE Transactions on Automatic Control*]{}, vol. 55, pp. 1781–1796, Aug 2010.
J. Qin, C. Yu, and S. Hirche, “Stationary consensus of asynchronous discrete-time second-order multi-agent systems under switching topology,” [*IEEE Transactions on Industrial Informatics*]{}, vol. 8, pp. 986–994, Nov 2012.
B. Touri and C. Langbort, “On indigenous random consensus and averaging dynamics,” in [*IEEE CDC*]{}, pp. 6208–6212, Dec 2013.
H. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram, “Resilient asymptotic consensus in robust networks,” [*IEEE Journal on Selected Areas in Communications (JSAC)*]{}, vol. 31, pp. 766–781, April 2013.
J. Zhao, O. Yağan, and V. Gligor, “On asymptotically exact probability of $k$-connectivity in random key graphs intersecting [Erdős–Rényi]{} graphs,” *arXiv e-prints*, 2014.
A lemma to confine $|\alpha_n|$ in Theorems \[thm:rig\] and \[thm:rig:rb\] as $o(\ln n)$ {#seca:conf:bin}
----------------------------------------------------------------------------------------
We present Lemma \[graph\_Hs\_cpln\] to confine $|\alpha_n|$ in Theorems \[thm:rig\] and \[thm:rig:rb\] as $o(\ln n)$; i.e., if Theorems \[thm:rig\] and \[thm:rig:rb\] hold under an extra condition $|\alpha_n| = o(\ln n)$, then they also hold regardless of this condition.
\[graph\_Hs\_cpln\]
**(a)** For graph $G_b(n,P_n,p_n)$ under $$\begin{aligned}
{p_n}^{2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n + {\beta_n}}{n} \label{al0-parta-Hs-od}\end{aligned}$$ with $\lim_{n \to \infty}\beta_n = -\infty$, there exists graph $G_b(n, P_n,\widetilde{p_n})$ under $$\begin{aligned}
{\widetilde{p_n}}^{2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n + {\widetilde{\beta_n}}}{n} \label{al0-parta-Hs}\end{aligned}$$ with $\lim_{n \to \infty}\widetilde{\beta_n} = -\infty$ and $\widetilde{\beta_n} = -o(\ln n)$ such that $G_b(n,P_n,p_n) \preceq G_b(n,P_n,\widetilde{p_n}) $.
**(b)** For graph $G_b(n,P_n,p_n)$ under $$\begin{aligned}
{p_n}^{2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n + {\beta_n}}{n} \label{al0-parta-Hs-pb-od}\end{aligned}$$ with $\lim_{n \to \infty}\beta_n = \infty$, there exists graph $G_b(n, P_n,\widehat{p_n})$ under $$\begin{aligned}
{\widehat{p_n}}^{2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n + {\widehat{\beta_n}}}{n} \label{al0-parta-Hs-pb}\end{aligned}$$ with $\lim_{n \to \infty}\widehat{\beta_n} = \infty$ and $\widehat{\beta_n} = o(\ln n)$ such that $G_b(n,P_n,\widehat{p_n}) \preceq G_b(n,P_n,p_n)$.
The proof of Lemma \[graph\_Hs\_cpln\] is given in Section \[sec:pro:graph\_Hs\_cpln\].
We now explain that given Lemma \[graph\_Hs\_cpln\], if Theorems \[thm:rig\] and \[thm:rig:rb\] hold under the extra condition $|\alpha_n| = o(\ln n)$, then they also hold regardless of the extra condition. Note that results (\[bin-kon-e\]) and (\[bin-kon-e-mnd\]) both have a condition $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$, which clearly implies $|\alpha_n| = o(\ln n)$. Hence, we only need to look at results (\[bin-kon-0\]) (\[bin-kon-1\]) (\[bin-kon-0-mnd\]) (\[bin-kon-1-mnd\]) (\[bin-krb-0\]) and (\[bin-krb-1\]). In particular, we will show that $$\begin{aligned}
\begin{array}{l}\text{if
(\ref{bin-kon-0}) (\ref{bin-kon-0-mnd}) and (\ref{bin-krb-0}) hold under condition
$\alpha_n = -o(\ln n)$,}\\\text{then they also hold regardless of the condition.}\end{array}\label{seecpl}\end{aligned}$$ and $$\begin{aligned}
\begin{array}{l}\text{if
(\ref{bin-kon-1}) (\ref{bin-kon-1-mnd}) and (\ref{bin-krb-1}) hold under condition
$\alpha_n = o(\ln n)$,}\\\text{then they also hold regardless of the condition.}\end{array}\label{seecplb}\end{aligned}$$ To see (\[seecpl\]), we use Lemma \[graph\_Hs\_cpln\]-Property (a) and Lemma \[mono-gcp\], and note that $k$-connectivity, the property of minimum node degree being at least $k$, and $k$-robustness are all monotone increasing graph properties. Then with $\mathcal{J}$ denoting any one of the above three properties, for graph $G_b(n,P_n,p_n)$ under (\[al0-parta-Hs-od\]) with $\lim_{n \to \infty}\beta_n = -\infty$, there exists graph $G_b(n, P_n,\widetilde{p_n})$ under (\[al0-parta-Hs\]) with $\lim_{n \to \infty}\widetilde{\beta_n} = -\infty$ and $\widetilde{\beta_n} = -o(\ln n)$ such that $$\begin{aligned}
& {{\mathbb{P}}\left[{G_b(n,P_n,p_n) \text{ has $\mathcal{J}$.}}\right]} \nonumber \\
& \leq {{\mathbb{P}}\left[{G_b(n, P_n,\widetilde{p_n})\text{ has $\mathcal{J}$.}}\right]}.
\label{seecpl1}
$$ If (\[bin-kon-0\]) and (\[bin-krb-0\]) hold under condition $\alpha_n=- o(\ln n)$, then we use them on graph $G_b(n, P_n,\widetilde{p_n})$ to obtain $$\begin{aligned}
&\lim_{n \to \infty} {{\mathbb{P}}\left[{G_b(n, P_n,\widetilde{p_n})\text{ has $\mathcal{J}$.}}\right]} = 0,
\label{seecpl3}
$$ Therefore, (\[seecpl1\]) and (\[seecpl3\]) yield $$\begin{aligned}
&\lim_{n \to \infty} {{\mathbb{P}}\left[{G_b(n, {P_n}, {p_n})\text{ has $\mathcal{J}$.}}\right]} = 0.
\nonumber\end{aligned}$$ In other words, for graph $G_b(n, {P_n}, {p_n})$, with $\alpha_n$ in Theorems \[thm:rig\] and \[thm:rig:rb\] replaced by $\beta_n$, (\[bin-kon-0\]) (\[bin-kon-0-mnd\]) and (\[bin-krb-0\]) hold. Note that we do not have any constraint on whether $\beta_n$ can be expressed as $ -o(\ln n)$. Hence, the arguments above establish (\[seecpl\]). The proof of (\[seecplb\]) using Lemma \[graph\_Hs\_cpln\]-Property (b) is similar to that of (\[seecpl\]) using Lemma \[graph\_Hs\_cpln\]-Property (a). We omit the details for simplicity.
To see (\[seecplb\]), we use Lemma \[graph\_Hs\_cpln\]-Property (b) and Lemma \[mono-gcp\], and note that $k$-connectivity, the property of minimum node degree being at least $k$, and $k$-robustness are all monotone increasing graph properties. Then for graph $G_b(n,P_n,p_n)$ under (\[al0-parta-Hs-pb-od\]) with $\lim_{n \to \infty}\beta_n = \infty$, there exists graph $G_b(n, P_n,\widehat{p_n})$ under (\[al0-parta-Hs-pb\]) with $\lim_{n \to \infty}\widehat{\beta_n} = \infty$ and $\widehat{\beta_n} = o(\ln n)$ such that $$\begin{aligned}
& {{\mathbb{P}}\left[{G_b(n,P_n,p_n) \text{ is $k$-connected.}}\right]} \nonumber \\
& \quad \geq {{\mathbb{P}}\left[{G_b(n, P_n,\widehat{p_n})\text{ is $k$-connected.}}\right]} ,
\label{seecpl1b} \\
& \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_b(n,P_n,p_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg] \nonumber \\
& \geq \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_b(n, P_n,\widetilde{p_n})\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg], \label{seecpl1b-mnd}\end{aligned}$$ and $$\begin{aligned}
& {{\mathbb{P}}\left[{G_b(n,P_n,p_n) \text{ is $k$-robust.}}\right]} \nonumber \\
& \quad \geq {{\mathbb{P}}\left[{G_b(n, P_n,\widehat{p_n})\text{ is $k$-robust.}}\right]}
\label{seecpl2b}\end{aligned}$$ If (\[bin-kon-1\]) (\[bin-kon-1-mnd\]) and (\[bin-krb-1\]) hold under condition $\alpha_n = o(\ln n)$, then we use them on graph $G_b(n, P_n,\widehat{p_n})$ to obtain $$\begin{aligned}
&\lim_{n \to \infty} {{\mathbb{P}}\left[{G_b(n, P_n,\widehat{p_n})\text{ is $k$-connected.}}\right]} = 1, \label{seecpl3b} \\
&\lim_{n \to \infty} \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_b(n, P_n,\widehat{p_n})\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg]= 1, \label{seecpl3b-mnd}\end{aligned}$$ and $$\begin{aligned}
& \lim_{n \to \infty}{{\mathbb{P}}\left[{G_b(n, P_n,\widehat{p_n})\text{ is $k$-robust.}}\right]} = 1.
\label{seecpl4b}\end{aligned}$$ Therefore, (\[seecpl1b\]) and (\[seecpl3b\]) give rise to $$\begin{aligned}
&\lim_{n \to \infty} {{\mathbb{P}}\left[{G_b(n, {P_n}, {p_n})\text{ is $k$-connected.}}\right]} = 1.
\nonumber\end{aligned}$$ (\[seecpl1b-mnd\]) and (\[seecpl3b-mnd\]) yield $$\begin{aligned}
&\lim_{n \to \infty} \mathbb{P} \bigg[\hspace{-2pt}\begin{array}{l}\text{Graph }G_b(n,P_n,p_n)\text{
has a}\\\text{minimum node degree at least $k$}.\end{array}\hspace{-2pt}\bigg]= 1.
\nonumber\end{aligned}$$ (\[seecpl2b\]) and (\[seecpl4b\]) induce $$\begin{aligned}
& \lim_{n \to \infty}{{\mathbb{P}}\left[{G_b(n, {P_n}, {p_n})\text{ is $k$-robust.}}\right]} = 1.
\nonumber\end{aligned}$$ In other words, for graph $G_b(n, {P_n}, {p_n})$, with $\alpha_n$ in Theorems \[thm:rig\] and \[thm:rig:rb\] replaced by $\beta_n$, (\[bin-kon-1\]) (\[bin-kon-1-mnd\]) and (\[bin-krb-1\]) hold. Note that we do not have any constraint on whether $\beta_n$ can be expressed as $ o(\ln n)$. Hence, the argument above proves (\[seecplb\]).
A lemma to confine $|\alpha_n|$ in Theorems \[thm:urig\] and \[thm:urig:rb\] as $o(\ln n)$ {#seca:conf:unig}
------------------------------------------------------------------------------------------
We present Lemma \[graph\_Gs\_cpl\] to confine $|\alpha_n|$ in Theorems \[thm:urig\] and \[thm:urig:rb\] as $o(\ln n)$; i.e., if Theorems \[thm:urig\] and \[thm:urig:rb\] hold under an extra condition $|\alpha_n| = o(\ln n)$, then they also hold regardless of this condition.
\[graph\_Gs\_cpl\]
**(a)** For graph $G_u(n, P_n, K_n)$ under $P_n = \Omega(n)$ and $$\begin{aligned}
\frac{{K_n}^{2}}{{P_n}} & = \frac{\ln n + {(k-1)} \ln \ln n + {\beta_n}}{n} \label{al1-parta}\end{aligned}$$ with $\lim_{n \to \infty}\beta_n = -\infty$, there exists graph $G_u(n,P_n, \widetilde{K_n})$ under $P_n = \Omega(n)$ and $$\begin{aligned}
\frac{{\widetilde{K_n}}^{2}}{{P_n}} & = \frac{\ln n + {(k-1)} \ln \ln n + {\widetilde{\beta_n}}}{n} \label{al0-parta}\end{aligned}$$ with $\lim_{n \to \infty}\widetilde{\beta_n} = -\infty$ and $\widetilde{\beta_n} = -o(\ln n)$, such that $G_u(n, P_n, K_n) \preceq G_u(n,P_n, \widetilde{K_n})$.
**(b)** For graph $G_u(n, P_n, K_n)$ under $P_n = \Omega(n)$ and $$\begin{aligned}
\frac{{K_n}^{2}}{{P_n}} & = \frac{\ln n + {(k-1)} \ln \ln n + {\beta_n}}{n} \label{al1}\end{aligned}$$ with $\lim_{n \to \infty}\beta_n = \infty$, there exists graph $G_u(n,P_n,\widehat{K_n})$ under $P_n = \Omega(n)$ and $$\begin{aligned}
\frac{{\widehat{K_n}}^{2}}{{P_n}} & = \frac{\ln n + {(k-1)} \ln \ln n + {\widehat{\beta_n}}}{n} \label{al0}\end{aligned}$$ with $\lim_{n \to \infty}\widehat{\beta_n} = \infty$ and $\widehat{\beta_n} = o(\ln n)$, such that $G_u(n,P_n,\widehat{K_n}) \preceq G_u(n, P_n, K_n)$.
The proof of Lemma \[graph\_Gs\_cpl\] is given in Section \[sec\_graph\_Gs\_cpl\].
We now explain that given Lemma \[graph\_Gs\_cpl\], if Theorems \[thm:urig\] and \[thm:urig:rb\] hold under the extra condition $|\alpha_n| = o(\ln n)$, then they also hold regardless of the extra condition. Note that results (\[uni-kon-e\]) and (\[uni-kon-e-mnd\]) both have a condition $\lim_{n \to \infty}{\alpha_n}
=\alpha ^* \in (-\infty, \infty)$, which clearly implies $|\alpha_n| = o(\ln n)$. Hence, we only need to look at results (\[uni-kon-0\]) (\[uni-kon-1\]) (\[uni-kon-0-mnd\]) (\[uni-kon-1-mnd\]) (\[uni-krb-0\]) and (\[uni-krb-1\]). In particular, we need to show that $$\begin{aligned}
\begin{array}{l}\text{if
(\ref{uni-kon-0}) (\ref{uni-kon-0-mnd}) and (\ref{uni-krb-0}) hold under condition
$\alpha_n = -o(\ln n)$,}\\\text{then they also hold regardless of the condition.}\end{array}\label{seecpluni}\end{aligned}$$ and $$\begin{aligned}
\begin{array}{l}\text{if
(\ref{uni-kon-1}) (\ref{uni-kon-1-mnd}) and (\ref{uni-krb-1}) hold under condition
$\alpha_n = o(\ln n)$,}\\\text{then they also hold regardless of the condition.}\end{array}\label{seecplbuni}\end{aligned}$$ The process of proving (\[seecpluni\]) and (\[seecplbuni\]) using Lemma \[graph\_Gs\_cpl\] is the same as the above process of proving (\[seecpl\]) and (\[seecplb\]) using Lemma \[graph\_Hs\_cpln\]. For brevity, we do not repeat the details here.
From Lemma \[lem-cpgraph-rigrig\], there exists a coupling under which $G_q(n, K_n, P_n)$ is an spanning supergraph of an Erdős–Rényi graph $G_{ER}(n,\frac{1}{q!} \cdot
\frac{{K_n}^{2q}}{{P_n} ^{q}} \cdot [1-o(1)] )$ with high probability, where an Erdős–Rényi graph [@citeulike:4012374] $G_{ER}(n, p_n)$ is defined on a set of $n$ nodes such that any two nodes establish an edge in between independently with probability $p_n$. Since connectivity is a monotone increasing graph property[^4], we obtain from Lemmas \[mono-gcp\] and \[lem-cpgraph-rigrig\] that $$\begin{aligned}
& \mathbb{P} \big[ G_q(n, K_n, P_n)\text{
is connected.}\big] \nonumber \\ & \quad \geq \mathbb{P} \big[
\textstyle{G_{ER}(n,\frac{1}{q!} \cdot \frac{{K_n}^{2q}}{{P_n} ^{q}}
\cdot [1-o(1)] )}\text{ is connected.}\big] - o(1). \label{bgqn}
\end{aligned}$$
Given condition $\frac{1}{q!}
\cdot \frac{{K_n}^{2q}}{{P_n} ^{q}} \geq \frac{c\ln n}{n}$ for all $n$ sufficiently large with constant $c>1$, we use Lemma \[erdsocon\] and have $$\begin{aligned}
& \mathbb{P} \big[
\textstyle{G_{ER}(n,\frac{1}{q!} \cdot \frac{{K_n}^{2q}}{{P_n} ^{q}}
\cdot [1-o(1)] )}\text{ is connected.}\big] \geq 1 - o(1),
\nonumber
\end{aligned}$$ which is substituted into (\[bgqn\]) to derive $$\begin{aligned}
& \mathbb{P} \big[ G_q(n, K_n, P_n)\text{
is connected.}\big] \geq 1 - o(1).\nonumber
\end{aligned}$$ Hence, $G_q(n, K_n, P_n)$ is connected with high probability; namely, $\textnormal{network}^{\textnormal{full visibility}}$ is connected with high probability.
Proof of Lemma \[graph\_Hs\_cpln\] {#sec:pro:graph_Hs_cpln}
----------------------------------
**Proving property (a):**
We set $$\begin{aligned}
\widetilde{\beta_n} = \max\{\beta_n, -\ln \ln n\}. \label{wPntdn2}\end{aligned}$$
Given (\[wPntdn2\]) and $\lim_{n \to \infty}\beta_n = -\infty$, we obtain $\lim_{n \to \infty}\widetilde{\beta_n} = -\infty$ and $\widetilde{\beta_n} = -o(\ln n)$. We use $\widetilde{\beta_n} = -o(\ln n)$ and (\[al0-parta-Hs\]) to have $ {\widetilde{p_n}}^{2}{P_n} \sim \frac{\ln n}{n}$, so it is clear for all $n$ sufficiently large that $\widetilde{p_n}$ is less than $1$ and can be used as a probability. Under $p_n \leq \widetilde{p_n}$, by [@zz Section 3], there exists a graph coupling under which $G_b(n,P_n,p_n)$ is a spanning subgraph of $G_b(n,P_n,\widetilde{p_n}) $; i.e., $G_b(n,P_n,p_n) \preceq G_b(n,P_n,\widetilde{p_n}) $.
**Proving property (b):**
We set $$\begin{aligned}
\widehat{\beta_n} = \min\{\beta_n, \ln \ln n\}. \label{wPntdn2-pb}\end{aligned}$$
Given (\[wPntdn2-pb\]) and $\lim_{n \to \infty}\beta_n = \infty$, we clearly obtain $\lim_{n \to \infty}\widehat{\beta_n} = \infty$ and $\widehat{\beta_n} = o(\ln n)$.
It holds from (\[wPntdn2-pb\]) that $\widehat{\beta_n} \leq \beta_n$, which along with (\[al0-parta-Hs-pb-od\]) and (\[al0-parta-Hs-pb\]) yields $p_n \geq \widehat{p_n}$. Under $p_n \geq \widehat{p_n}$, by [@zz Section 3], there exists a graph coupling under which $G_b(n,P_n,p_n)$ is a spanning supergraph of $G_b(n,P_n,\widehat{p_n}) $; i.e., $G_b(n,P_n,\widehat{p_n}) \preceq G_b(n,P_n,p_n)$.
The Proof of Lemma \[graph\_Gs\_cpl\] {#sec_graph_Gs_cpl}
-------------------------------------
**Proving property (a):**
We define $\widetilde{\beta_n}^*$ by $$\begin{aligned}
\widetilde{\beta_n}^* & = \max\{\beta_n, -\ln \ln n\}, \label{al2-parta}\end{aligned}$$ and define $\widetilde{K_n}^*$ such that $$\begin{aligned}
\frac{({\widetilde{K_n}^{*}})^{2}}{{{P_n}}} & = \frac{\ln n + {(k-1)} \ln \ln n + \widetilde{\beta_n}^*}{n}. \label{al3-parta}\end{aligned}$$ Note that $\widetilde{K_n}^*$ might or might not be an integer. We set $$\begin{aligned}
\widetilde{K_n} & : = \big\lfloor \widetilde{K_n}^* \big\rfloor, \label{al4-parta}\end{aligned}$$ where the floor function $\lfloor x \rfloor$ means the largest integer not greater than $x$.
From (\[al1\]) (\[al2-parta\]) and (\[al3-parta\]), it holds that $$\begin{aligned}
K_n \leq \widetilde{K_n}^*. \label{Kn1-parta}\end{aligned}$$ Then by (\[al4-parta\]) (\[Kn1-parta\]) and the fact that $K_n$ and $\widetilde{K_n}$ are both integers, it follows that $$\begin{aligned}
K_n \leq \widetilde{K_n}. \label{al6-parta}\end{aligned}$$ From (\[al6-parta\]), by [@Rybarczyk Lemma 3], there exists a graph coupling under which $G_u(n,P_n,K_n)$ is a spanning subgraph of $G_u(n,P_n,\widetilde{K_n})$; i.e., $G_u(n, P_n, K_n) \preceq G_u(n,P_n, \widetilde{K_n})$. Therefore, the proof of property (a) is completed once we show $\widetilde{\beta_n}$ defined in $(\ref{al0-parta})$ satisfies $$\begin{aligned}
\lim_{n \to \infty}\widetilde{\beta_n} & = - \infty, \label{al8-parta} \\
\text{and }\widetilde{\beta_n} & = - o(\ln n). \label{al7-parta}\end{aligned}$$
We first prove (\[al8-parta\]). From (\[al0-parta\]) (\[al3-parta\]) and (\[al4-parta\]), it holds that $$\begin{aligned}
\widetilde{\beta_n} \leq \widetilde{\beta_n}^*, \label{haa-parta}\end{aligned}$$ which together with (\[al2-parta\]) and $\lim_{n \to \infty}\beta_n = -\infty$ yields (\[al8-parta\]).
Now we establish (\[al7-parta\]). From (\[al4-parta\]), we have $\widetilde{K_n} > \widetilde{K_n}^* - 1$. Then from (\[al0-parta\]), it holds that $$\begin{aligned}
\widetilde{\beta_n} & = n \cdot \frac{{\widetilde{K_n}}^{2}}{{{P_n}}} - [\ln n + {(k-1)} \ln \ln n] \nonumber \\
& > n \cdot \frac{{(\widetilde{K_n}^*)^2 - 2\widetilde{K_n}^*}}{{{P_n}}} - [\ln n + {(k-1)} \ln \ln n]
. \label{aph1-parta}\end{aligned}$$ By $\lim_{n \to \infty}\beta_n =- \infty$, it holds that $\beta_n \leq 0$ for all $n$ sufficiently large. Then from (\[al2-parta\]), it follows that $$\begin{aligned}
\widetilde{\beta_n}^* = - O(\ln \ln n), \label{widetilde-al2-parta}\end{aligned}$$ which along with (\[al3-parta\]) yields $$\begin{aligned}
\frac{\widetilde{K_n}^*}{P_n} & \sim \sqrt{ \frac{ \ln n }{nP_n}} = O\bigg(\frac{\sqrt{\ln n}}{n}\bigg) .\label{aph5-parta}\end{aligned}$$ Applying (\[al3-parta\]) (\[aph5-parta\]) and $P_n = \Omega(n)$ to (\[aph1-parta\]), we obtain $$\begin{aligned}
\widetilde{\beta_n} & > \bigg\{n \cdot \frac{({\widetilde{K_n}^{*}})^{2}}{{{P_n}}} - [\ln n + {(k-1)} \ln \ln n] \bigg\} - 2n \cdot \frac{\widetilde{K_n}^*}{P_n} \nonumber \\
& = \widetilde{\beta_n}^* - O\big(\sqrt{\ln n}\hspace{2pt}\big). \label{widetilde-al-parta}\end{aligned}$$ Thus, from (\[haa-parta\]) (\[widetilde-al2-parta\]) and (\[widetilde-al-parta\]), clearly $ \widetilde{\beta_n} $ can be written as $- O\big(\sqrt{\ln n}\hspace{2pt}\big)$ and further $- o(\ln n)$; i.e., (\[al7-parta\]) is proved. Then as explained above, since we have shown (\[al8-parta\]) and (\[al7-parta\]), property (a) of Lemma \[graph\_Gs\_cpl\] is established.
**Proving property (b):**
We define $\widehat{\beta_n}^*$ by $$\begin{aligned}
\widehat{\beta_n}^* & = \min\{\beta_n, \ln \ln n\}, \label{al2}\end{aligned}$$ and define $\widehat{K_n}^*$ such that $$\begin{aligned}
\frac{({\widehat{K_n}^{*}})^{2}}{{{P_n}}} & = \frac{\ln n + {(k-1)} \ln \ln n + \widehat{\beta_n}^*}{n}. \label{al3}\end{aligned}$$ We set $$\begin{aligned}
\widehat{K_n} & : = \big\lceil \widehat{K_n}^* \big\rceil. \label{al4}\end{aligned}$$
From (\[al1\]) (\[al2\]) and (\[al3\]), it holds that $$\begin{aligned}
K_n \geq \widehat{K_n}^*. \label{Kn1}\end{aligned}$$ Then by (\[al4\]) (\[Kn1\]) and the fact that $K_n$ and $\widehat{K_n}$ are both integers, it follows that $$\begin{aligned}
K_n \geq \widehat{K_n}. \label{al6}\end{aligned}$$ From (\[al6\]), by [@Rybarczyk Lemma 3], there exists a graph coupling under which $G_u(n,P_n,K_n)$ is a spanning supergraph of $G_u(n,P_n,\widehat{K_n})$; i.e., $G_u(n,P_n,\widehat{K_n}) \preceq G_u(n, P_n, K_n)$. Therefore, the proof of property (b) is completed once we show $\widehat{\beta_n}$ defined in $(\ref{al0})$ satisfies $$\begin{aligned}
\lim_{n \to \infty}\widehat{\beta_n} & = \infty, \label{al8} \\
\text{and }\widehat{\beta_n} & = o(\ln n). \label{al7}\end{aligned}$$
We first prove (\[al8\]). From (\[al0\]) (\[al3\]) and (\[al4\]), it holds that $$\begin{aligned}
\widehat{\beta_n} \geq \widehat{\beta_n}^*, \label{haa}\end{aligned}$$ which together with (\[al2\]) and $\lim_{n \to \infty}\beta_n = \infty$ yields (\[al8\]).
Now we establish (\[al7\]). From (\[al4\]), we have $\widehat{K_n} < \widehat{K_n}^* + 1$. Then from (\[al0\]), it holds that $$\begin{aligned}
\widehat{\beta_n} & = n \cdot \frac{{\widehat{K_n}}^{2}}{{{P_n}}} - [\ln n + {(k-1)} \ln \ln n] \nonumber \\
& \leq n \cdot \frac{({\widehat{K_n}^{*}})^{2} + 3 {\widehat{K_n}^{*}} }{{{P_n}}} - [\ln n + {(k-1)} \ln \ln n] . \label{aph1}\end{aligned}$$ By $\lim_{n \to \infty}\beta_n = \infty$, it holds that $\beta_n \geq 0$ for all $n$ sufficiently large. Then from (\[al2\]), it follows that $$\begin{aligned}
\widehat{\beta_n}^* = O(\ln \ln n), \label{widehat-al2}\end{aligned}$$ which along with (\[al3\]) and condition $P_n = \Omega(n)$ induces $$\begin{aligned}
\frac{\widehat{K_n}^*}{P_n} & \sim \sqrt{ \frac{ \ln n }{nP_n}} = O\bigg(\frac{\sqrt{\ln n}}{n}\bigg) . \label{aph5}\end{aligned}$$ Hence, we have $\lim_{n \to \infty} \widehat{K_n}^* = \infty$ and it further holds for all $n$ sufficient large that $$\begin{aligned}
{(\widehat{K_n}^* + 1)}^{2}< ({\widehat{K_n}^{*}})^{2} + 3 {\widehat{K_n}^{*}} . \label{aph2}\end{aligned}$$ Applying (\[al3\]) (\[aph5\]) and $P_n = \Omega(n)$ to (\[aph1\]), we obtain $$\begin{aligned}
\widehat{\beta_n} & < \bigg\{n \cdot \frac{({\widehat{K_n}^{*}})^{2}}{{{P_n}}} - [\ln n + {(k-1)} \ln \ln n] \bigg\} + 3 n \cdot \frac{\widehat{K_n}^*}{P_n}
\nonumber \\
& = \widehat{\beta_n}^* + O\big(\sqrt{\ln n}\hspace{2pt}\big).\label{widehat-al}\end{aligned}$$ Thus, from (\[haa\]) (\[widehat-al2\]) and (\[widehat-al\]), clearly $ \widehat{\beta_n} $ can be written as $ O\big(\sqrt{\ln n}\hspace{2pt}\big)$ and further $ o(\ln n)$; i.e., (\[al7\]) is proved. Then as explained above, since we have shown (\[al8\]) and (\[al7\]), property (b) of Lemma \[graph\_Gs\_cpl\] is established.
[^1]: The material in this paper was presented in part at the 2014 IEEE Conference on Decision and Control (CDC), Los Angeles, CA, USA [@ZhaoCDC].
J. Zhao was, and O. Yağan and V. Gligor are with the Cybersecurity Lab (CyLab) and the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA. (Emails: [email protected], [email protected], [email protected]). J. Zhao is now with Arizona State University.
This research was supported in part by CyLab, and Department of Electrical & Computer Engineering at Carnegie Mellon University, and also by Grants CCF-0424422 and CNS-0831440 from the National Science Foundation to the Berkeley TRUST STC.
The research was supported in part by the U.S. Army Research Office, Durham, NC, USA, under Grant DAAD19-02-1-0389 and Grant W 911 NF 0710287 and in part by Carnegie Mellon University, Pittsburgh, PA, USA. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of any sponsoring institution, the U.S. government or any other entity.
J. Zhao was, and O. Yağan and V. Gligor are, with CyLab and the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA. J. Zhao is now with Arizona State University, Tempe, AZ 85287 (Emails: [email protected], [email protected], [email protected]).
[^2]: Under other parameter conditions, the conclusion may not hold as in the case of binomial random intersection graphs shown by Rybarczyk [@zz; @2013arXiv1301.0466R].
[^3]: Under other parameter conditions, the conclusion may not hold as in the case of binomial random intersection graphs shown by Rybarczyk [@zz; @2013arXiv1301.0466R].
[^4]: A graph property is called monotone increasing if it holds under the addition of edges in a graph.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Polaritons in metals, semimetals, semiconductors, and polar insulators, with their extreme confinement of electromagnetic energy, provide many promising opportunities for enhancing typically weak light-matter interactions such as multipolar radiation, multiphoton spontaneous emission, Raman scattering, and material nonlinearities. These highly confined polaritons are quasi-electrostatic in nature, with most of their energy residing in the electric field. As a result, these “electric” polaritons are far from optimized for enhancing emission of a magnetic nature, such as spin relaxation, which is typically many orders of magnitude slower than corresponding electric decays. Here, we propose using surface magnon polaritons in negative magnetic permeability materials such as MnF$_2$ and FeF$_2$ to strongly enhance spin-relaxation in nearby emitters in the THz spectral range. We find that these magnetic polaritons in 100 nm thin-films can be confined to lengths over 10,000 times smaller than the wavelength of a photon at the same frequency, allowing for a surprising twelve orders of magnitude enhancement in magnetic dipole transitions. This takes THz spin-flip transitions, which normally occur at timescales on the order of a year, and forces them to occur at sub-ms timescales. Our results suggest an interesting platform for polaritonics at THz frequencies, and more broadly, a new way to use polaritons to control light-matter interactions.'
author:
- 'Jamison Sloan$^{1*\dagger}$, Nicholas Rivera$^{1*}$, John D. Joannopoulos$^{1}$, Ido Kaminer$^{2}$, and Marin Soljačić$^{1}$'
bibliography:
- 'magnonbib.bib'
title: Extreme enhancement of spin relaxation mediated by surface magnon polaritons
---
Polaritons, collective excitations of light and matter, offer the ability to concentrate electromagnetic energy down to volumes far below that of a photon in free space [@atwater2007promise; @Basov:2016; @basov2017towards; @low2017polaritons; @iranzo2018probing; @ni2018fundamental], holding promise to achieve the long-standing goal of low-loss confinement of electromagnetic energy at the near-atomic scale. The most famous examples are surface plasmon polaritons on conductors, which arise from the coherent sloshing of surface charges accompanied by an evanescent electromagnetic field. These collective excitations are so widespread in optics that their manipulation is referred to as “plasmonics.” Plasmons enjoy a myriad of applications, particularly in spectroscopy due to their enhanced interactions with matter. This enhancement applies to spontaneous emission, Raman scattering, optical nonlinearities, and even dipole-“forbidden” transitions in emitters [@moskovits1985surface; @albrecht1977anomalously; @jeanmaire1977surface; @nie1997probing; @kauranen2012nonlinear; @andersen2011strongly; @takase2013selection; @rivera2016shrinking]. Beyond plasmons in metals, polaritons in polar dielectrics, such as phonon polaritons [@caldwell2013low; @caldwell2014sub; @dai2014tunable; @caldwell2015low] are now being exploited for similar applications due to their ability to concentrate electromagnetic energy on the nanoscale in the mid-IR/THz spectral range.
{width="\linewidth"}
The ability of nano-confined polaritons to strongly enhance electromagnetic interactions with matter can ultimately be understood in terms of electromagnetic energy density. An electromagnetic quantum of energy $\hbar\omega$, confined to a volume $V$, leads to a characteristic root-mean-square electric field of order $\sqrt{\frac{\hbar\omega}{\epsilon_0 V}}$. In the case of field interaction with an electron in an emitter, this characteristic field drives spontaneous emission, and thus concentration of energy to smaller volumes leads to enhanced emission. This well-studied phenomenon is best known as the Purcell effect [@purcell1946purcell]. Interestingly, if one looks at the electromagnetic energy distribution of a highly confined plasmon- or phonon- polariton, one finds that an overwhelming majority of this energy resides in the electric field. For a polariton with a wavelength 100 times smaller than that of a photon at the same frequency, the energy residing in the magnetic field is of the order of a mere $0.01\%$ of the total energy $\hbar\omega$. This largely suggests that such excitations are relatively inefficient for enhancing spontaneous emission processes which couple to the magnetic field, such as spin-flip transitions or magnetic multipole decays. Nevertheless, enabling magnetic decays at very fast rates represents a rewarding challenge, as increasing rates of spontaneous emission can provide new opportunities for detectors, devices, and sources of light. The Purcell enhancement of magnetic dipole transitions has been approached by a few basic means: the use of highly confined resonances at optical frequencies [@rolly2012promoting; @hussain2015enhancing], metamaterials [@slobozhanyuk2014magnetic; @mahmoud2014wave] and for microwave frequencies, materials with simultaneously very high quality factor and highly confined fields. These advances are reviewed in Ref. [@baranov2017modifying]. Many of these methods have the benefit of compatibility with well known materials and use at optical frequencies, but the Purcell enhancements in these cases are typically very far from maximal Purcell enhancements that can be achieved with “electric” polaritons at similar frequencies [@koppens2011graphene; @kumar2015tunable; @rivera2016shrinking; @miller2017limits; @rivera2017making; @kurman2018control]. This prompts the question: what kind of electromagnetic response allows one to achieve a similar degree of very strong enhancement for magnetic decays?
The duality between electric and magnetic phenomena, combined with ideas from plasmonics and nano-optics, suggests a new pathway for achieving strong magnetic transition enhancement: highly confined magnetic modes in materials with negative magnetic *permeability*. In particular, plasmon- and phonon-polaritons are associated with a negative dielectric permittivity $\epsilon(\omega)$. By electromagnetic duality, if one replaces $\epsilon(\omega)$ with the magnetic permeability $\mu(\omega)$, then the electric field $\mathbf{E}$ in the dielectric structure becomes the magnetic field $\mathbf{H}$ in the dual magnetic structure. Thus, to very efficiently enhance magnetic decays, one desires a material with negative $\mu(\omega)$ which supports modes dual to “electric” surface polaritons. While likely not the only example, antiferromagnetic resonance is a well-studied example of a phenomenon which can provide precisely this permeability, and the corresponding modes are surface magnon polaritons.
Here, we propose enhancement of spin relaxation in emitters using highly confined magnon polaritons. We show that the interaction of a magnetic dipole with modes that are primarily magnetic in nature can level the playing field between electric and magnetic processes. Specifically, we find that these systems can shrink the wavelength of light by factors over 10,000, and predict speedups of magnetic dipole spontaneous emission processes on the order of $10^{12}$. Such enhancements could enable extremely slow magnetic decays with radiative lifetimes on the order of a year to occur at sub-millisecond timescales.
The organization of this manuscript is as follows: in section I, we review the electrodynamics of surface magnon polaritons, and derive the dispersion relation and mode profile of magnon polariton modes for the example of an antiferromagnet. In section II, we develop the theory of spin relaxation of emitters into these modes, and in section III, we provide quantitative results for the spontaneous emission by spin systems near existing magnon-polaritonic materials, such as MnF$_2$ and FeF$_2$.
Surface magnon polariton modes
==============================
We begin by reviewing the confined modes which exist on thin films of materials with negative permeability. The modes we describe are surface magnon polaritons [@mills1974polaritons; @almeida1988dynamical; @camley1982surface; @shu1982surface] with $\text{Re}\,\mu(\omega) \leq 0$. For the specific case of an antiferromagnetic material near resonance, the frequency-dependent permeability takes the form of a Lorentz oscillator which depends on the microscopic magnetic properties of the antiferromagnetic crystal. Studies of the crystal structures of important antiferromagnetic materials can be found in [@stout1954crystal]. The magnetic permeability function for antiferromagnetic resonance (AFMR) derived in [@kittel1951theory] is $$\mu_{xx} = \mu_{yy} = 1 + \frac{2\gamma^2 H_A H_M}{\omega_0^2 - (\omega + i\Gamma)^2},
\label{eq:permeability_model}$$ with coordinates shown in Figure \[fig:schematic\]. In Equation \[eq:permeability\_model\], $\omega_0$ is the resonance frequency, $H_A$ is the anisotropy field, $H_M$ is the magnetization field, $\gamma$ is the gyromagnetic ratio, and $\Gamma = 1/\tau$ is a phenomenological damping parameter inversely proportional to the loss relaxation time $\tau$. Furthermore, in the approximation of low damping, the resonant frequency is given as $\omega_0 = \gamma\sqrt{2H_A(H_A + H_E)}$, where $H_E$ is the exchange field which is representative of the magnetic field required to invert neighboring spin pairs. For antiferromagnetic materials such as MnF$_2$ and FeF$_2$, the resonance frequencies $\omega$ takes values 1.69$\times 10^{12}$ and 9.89$\times 10^{12}$ rad/s respectively, and have negative permeability over a relatively narrow bandwidth on the scale of a few GHz. Most importantly for our purposes, $\text{Re}\mu(\omega) < 0$ for $\omega < \omega_0 < \omega_{\text{max}}$, which will permit surface-confined modes. Table \[tab:parameter\_table\] shows values of material parameters for a variety of antiferromagnetic materials.
Material $H_A (T)$ $H_E$ (T) $M_S$ (T) $\omega_0$ ($\times 10^{12}$ rad/s) $\tau$ (s)
----------- ----------- ----------- ----------- ------------------------------------- -----------------------
MnF$_2$ 0.787 53.0 0.06 1.69 $7.58\times 10^{-9}$
FeF$_2$ 19.745 53.3 0.056 9.89 $1.06\times 10^{-10}$
GdAlO$_3$ 0.365 1.88 0.062 0.23 –
: Anisotropy fields, exchange fields, sublattice magnetization, resonance frequencies, and damping constants (where known) for antiferromagnetic materials that can support SMPs. Parameters are taken from Refs. [@dumelow1997continuum; @luthi1983surface]. []{data-label="tab:parameter_table"}
{width="\linewidth"}
Antiferromagnetic fluorides exhibit a uniaxial permeability structure with two orthogonal components of the permeability tensor given by $\mu(\omega)$ above, and the other orthogonal component as unity. We start by focusing on crystal orientations in which $\mu = (\mu(\omega), \mu(\omega), 1)$. It is also worthwhile to note that experiments, specifically on nonreciprocal optical phenomena [@remer1986nonreciprocal], have been performed on these materials in a less conventional geometry where $\mu = (\mu(\omega), 1, \mu(\omega))$. The in-plane anisotropy of this configuration substantially complicates the dispersion relation and propagation structure of the modes. As such, we focus primarily on the former case, but present results for the latter at the end of the text.
For concreteness, we focus on MnF$_2$, a material which has been studied in depth both in theory and experiment [@greene1965observation; @stamps1991nonreciprocal], and also exhibits a relatively low propagation loss. We note that FeF$_2$ is also a promising candidate with higher resonance frequency, but also higher loss [@hutchings1970spin; @brown1994nonreciprocal]. We solve for surface magnon polaritons supported by optically very thin (here, sub-micron) MnF$_2$ films surrounded by air. For the confined modes we consider, the effect of retardation is negligible, and thus we can find the magnon modes using a quasi-magnetostatic treatment [@camley1980long]. In the absence of retardation, the electric field is negligible, and the magnetic field, since there are no free currents, satisfies $\nabla\times\mathbf{H} = 0$. Thus the magnetic field can be written as the gradient of a scalar potential $\mathbf{H} = \nabla\psi_H$. This scalar potential then satisfies a scalar Laplace equation $$ \partial_i \mu_{ij}(\omega) \partial_j \psi_H = 0,$$ where we have used repeated indices to denote summation. Applying boundary conditions for continuity of the magnetic potential at the two interfaces of a film of thickness $d$ gives the dispersion relation $$q_n = \frac{1}{d\sqrt{-\mu(\omega)}}\left[\tan^{-1}\left(\frac{1}{\sqrt{-\mu(\omega)}}\right) + \frac{n\pi}{2}\right],$$ where $n$ is an integer, $q_n$ is the in-plane wavevector of mode $n$, and $\mu(\omega)$ is the permeability given in Equation \[eq:permeability\_model\]. We see that $q_n$ is inversely proportional to the thickness of the slab $d$. Identically to confined modes on thin films of plasmonic materials (silver and gold for instance), a thinner film results in a smaller wavelength. Figure \[fig:dispersion\]c shows plots of the scalar potential $\psi_H$ associated with SMP modes on MnF$_2$, which is proportional to the magnetic field in direction of propagation. The scalar potential solutions to the Laplace equation take the form $$\psi_H^{n}(\mathbf{r},\omega) = \begin{cases}
e^{i\mathbf{q}_n\cdot\rm{\rho}}e^{-q_n|z|} & |z| > d/2 \\
\left(\frac{e^{-q_nd}}{f(q_nd)}\right) e^{i\mathbf{q}_n\cdot\rm{\rho}}f(q_nz) & |z| < d/2
\end{cases},$$ where $\rho = (x,y)$ is the in-plane position, $f(x) = \cos(x)$ for even modes, and $f(x) = \sin(x)$ for odd modes. Taking the gradient of the scalar potential gives the fully vectorial magnetic field, which reveals that the surface magnon polariton mode propagates in the in-plane direction $\hat{q}$ with circular polarization $\hat{\varepsilon}_{\mathbf{q}} = (\hat{q} + i\hat{z})/\sqrt{2}$. This polarization is well known to be typical of quasistatic surface polariton modes, whether they are the transverse magnetic modes associated with quasielectrostatic excitations or transverse electric modes associated with quasimagnetostatic excitations.
{width="\linewidth"}
In Figure \[fig:dispersion\]b, we plot the material-thickness-invariant dispersion relation $\omega(qd)$. The dimensionless wavevector $qd$ indicates how the size of the in-plane wavevector compares to the thickness of the film. The dispersion plot shows the first four bands – the fundamental mode as well as three higher harmonics. Due to the the reflection symmetry of the geometry in the $z$-direction, two of these modes are even parity, and two are odd parity. We can interpret the mode index as the number of half oscillations which the magnetic field makes in the $z$-direction of the film. Higher order modes will have larger wavevectors. Once again, we can further understand the dispersion relation of these modes through analogy to existing polaritonic systems. Specifically, MnF$_2$ is a hyperbolic material since $\mu_\perp > 0$ while $\mu_\parallel < 0$ (where the directions $\perp$ and $\parallel$ are taken with respect to the $z$ axis). This is much like the naturally occurring hyperbolic material hexagonal boron nitride, which has one component of its permittivity negative, while another component is positive. As a result of this, these systems have a multiply-branched dispersion, and the electromagnetic fields are guided inside the crystal. The most impressive figure of merit of these modes is the size of their wavelength in comparison to the free space wavelength at a given frequency, also known as a confinement factor or effective index of the mode. Figure \[fig:properties\]b highlights this, showing the confinement factor $\eta = qc/\omega = \lambda_0/\lambda_{\text{SMP}}$ for the first four modes ($n=0,1,2,3$) on $d=200$ nm MnF$_2$ as a function of frequency. We see that the fundamental mode reaches a peak confinement of $\eta = 2\times 10^4$, while the first harmonic is confined to twice that with $\eta = 4\times 10^4$.
These values exceed by nearly two orders of magnitude the maximum confinement values that have been observed in common plasmonic media such as thin films of silver, gold, or titanium nitride, or doped graphene. Furthermore, since the confinement scales linearly with $q\sim 1/d$, decreasing the material thickness increases the achievable range of confinement factors. As a simple example of this, consider that a material thickness of $d=50$ nm would correspond to a wavevector 4 times larger than for $d=200$ nm, in other words a maximum fundamental mode confinement of $8\times 10^4$, and a confinement above $10^4$ for much of the surface magnon band.
An explanation for this high confinement in terms of most basic principles is that the frequencies at which surface magnon polaritons exist (GHz-THz) are orders of magnitude lower than for plasmons which typically exist in IR to optical regimes. Simultaneously, the scale of the wavevector $q$ in both plasmonic and magnonic media is set by the film thickness $d$ (such that plasmons and magnons will have similar wavevectors). In other words, at a fixed material thickness, lower frequency surface magnons have substantially higher potential for geometrical squeezing than surface plasmons. We note that this is not of purely formal interest, as when considering the enhancement of spontaneous emission, one finds that the enhancement is proportional to a power of precisely this confinement factor.
In addition to understanding the confinement of magnon polaritons, it is also important to understand their propagation characteristics, such as propagation quality factor, and group velocity. Figure \[fig:properties\]a,c shows the quality factor $Q = \text{Re}(q)/\text{Im}(q)$, as well as the normalized group velocity $v_g/c$ as a function of frequency for the first four modes. We see that propagation losses are lowest toward the middle of the allowed frequency band, showing quality factors greater than 20 for $n=0$. Additionally, we see that the group velocity $v_g$ reaches its maximum near the lower portion of the allowed frequency range, and goes toward zero at the other end.
Theory of spin relaxation into magnon polaritons
================================================
We now discuss the mechanisms that can allow an emitter to couple to highly confined SMPs, and then calculate rates of emitter relaxation associated with SMP emission. A magnetic field can couple to both the electron spin angular momentum and orbital angular momentum, as both angular momenta contribute to the electron’s magnetic moment. We describe this interaction quantum mechanically with a interaction Hamiltonian $H_{\text{int}}$ between an emitter and a magnetic field $$H_{\text{int}} = -\bm{\mu}\cdot\mathbf{B} = -\frac{\mu_B(\mathbf{L} + g\mathbf{S})}{\hbar}\cdot\mathbf{B},
\label{eq:hamiltonian}$$ where $\bm{\mu}$ is the total magnetic moment of the atom, $\mathbf{S} = \hbar\bm{\sigma}$ is the spin angular momentum operator, $\mathbf{L}$ is the orbital angular momentum operator, $g \approx 2.002$ is the Landé g-factor. In this Hamiltonian, we note that $\mathbf{B}$ is the quantized magnetic field operator associated with SMP modes.
In order to provide a fully quantum mechanical description of the interactions, we use the formalism of macroscopic QED to express the magnetic field operator as a mode expansion over SMP modes. This approach is similar to that in [@glauber1991quantum], which was applied to quantize electromagnetic fields in dielectric structures. We consider a geometry of a negative $\mu$ material which is translation invariant (i.e., a slab geometry). In this case, the modes are labeled by an in-plane wavevector $\mathbf{q}$. We find then that the magnetic field Schrodinger operator at time $t=0$ takes the form: $$\mathbf{B}(\mathbf{r}) = \sum_{\mathbf{q}}\sqrt{\frac{\mu_0\hbar\omega}{2 A C_q}}\left(\hat{\varepsilon}_{\mathbf{q}}e^{i \mathbf{q}\cdot\mathbf{\rho}}e^{-qz}a_{\mathbf{q}} + \hat{\varepsilon}_{\mathbf{q}}^*e^{-i \mathbf{q}\cdot\mathbf{\rho}}e^{-qz}a^\dagger_{\mathbf{q}}\right).
\label{eq:bfield_operator}$$ where $a_{\mathbf{q}}^\dagger$ and $a_{\mathbf{q}}$ are creation and annihilation operators for the SMP modes satisfying the canonical commutation relation $[a_{\mathbf{q}},a_{\mathbf{q}'}^\dagger] = \delta_{\mathbf{qq}'}$, $\hat{\varepsilon}_{\mathbf{q}}$ is the mode polarization, $A$ is the area normalization factor, and $C_q = \int dz\,\mathbf{H}^*(z)\cdot \frac{d(\mu\omega)}{d\omega}\cdot\mathbf{H}(z)$ is a normalization factor ensuring that the mode $\mathbf{H} = \nabla\psi_H$ has an energy of $\hbar\omega_{\mathbf{q}}$. The energy has been calculated according to the Brillouin formula for the electromagnetic field energy in a dispersive medium in a transparency window [@landau2013electrodynamics; @archambault2010quantum]. In this expression for the energy, we have also used the fact that the modes are magnetostatic in nature, and that the electric energy associated with them is negligible. To understand the strength of the coupling between an emitter’s spin and SMPs, we calculate spontaneous emission of a spin into a thin negative $\mu$ material such as an antiferromagnet, using Fermi’s golden rule. The rate of spin relaxation by emission of a magnon of wavevector $\mathbf{q}$ is given as $$\Gamma_{\mathbf{q}}^{(eg)} = \frac{2\pi}{\hbar^2}|\braket{g, \mathbf{q}|H_{\text{int}}|e,0}|^2\delta(\omega_{\mathbf{q}} - {\omega_{eg}})
\label{eq:fermi_golden_rule}$$
We specify the initial and final states of the system as $\ket{e,0}$ and $\ket{g,\mathbf{q}}$ respectively, where $e$ and $g$ index the excited and ground states of the emitter, $\mathbf{q}$ is the wavevector of the magnon resulting from spontaneous emission, $\omega_{\mathbf{q}}$ is its corresponding frequency, and ${\omega_{eg}}$ is the frequency of the spin transition.
Substituting Equation \[eq:bfield\_operator\] into the Hamiltonian of Equation \[eq:hamiltonian\], and then applying Fermi’s golden rule as written in Equation \[eq:fermi\_golden\_rule\], we find that the spontaneous emission rate $\Gamma^{(eg)}$ per unit magnon in-plane propagation angle $\theta$ is given by: $$\frac{d\Gamma^{(eg)}_{\text{dipole}}}{d\theta} = \frac{\mu_B^2\mu_0{\omega_{eg}}}{2\pi\hbar}\frac{q^3({\omega_{eg}})}{C_q({\omega_{eg}})|v_g({\omega_{eg}})|}e^{-2q({\omega_{eg}})z_0}|M_{eg}|^2
\label{eq:dipole_transition_rate},$$ where $|v_g| = |\nabla_{\mathbf{q}}\omega|$ is the magnitude of the SMP group velocity and $M_{eg} = \braket{g|\hat{\epsilon}_{\mathbf{q}}\cdot(\mathbf{L} + g\mathbf{S})|e}$ is the matrix element which describes the transition. In cases where the transition corresponds only to a change of spin of the electron, this matrix element is simply proportional to $\bm{\sigma}_{eg} = \braket{\downarrow|\bm{\sigma}\cdot\hat{\varepsilon}_{\mathbf{q}}|\uparrow}$. Here, the angular dependence can come solely from the magnon polarization. For a spin transition oriented along the $z$ (ie. out-of-plane) axis, the transition strength into modes at different $\theta$ will be the same, and thus the distribution of emitted magnons isotropic. Spin transitions along a different axis will break this symmetry, resulting in angle dependent emission. In any case, the total rate of emission is obtained by integrating over all angles as $\Gamma^{(eg)}_{\text{dipole}} = \int_0^{2\pi} \left(\frac{d\Gamma^{(eg)}}{d\theta}\right)\,d\theta$.
This formalism can be extended to include losses using the methodology established in [@scheel2008macroscopic]. It was found explicitly in [@rivera2016shrinking] that in general the presence of losses does not drastically change the total decay rate of the emitter, unless the emitter is at distances from the material much smaller than the inverse wavevector of the modes that are emitted. In the particular case of SMPs of MnF$_2$, the modes have quality factors of $Q \sim 20-30$, and the distances chosen are fairly large, so neglecting material losses is justified. Having presented the general framework for analyzing SMP emission, we now present specific results for SMP emission into a thin film of MnF$_2$.
Transition Rate Results
=======================
Magnetic Dipole Transition Rates
--------------------------------
We first discuss the transition rates and associated Purcell factors of magnetic dipole emitters. For a $z$-oriented spin flip of frequency ${\omega_{eg}}$ placed a distance $z_0$ from the surface of a negative $\mu$ film, the spontaneous emission rate is given as $$\Gamma_{\text{dipole}}^{(eg)} = \frac{\mu_B^2 \mu_0 {\omega_{eg}}}{\hbar} \frac{q^2({\omega_{eg}})}{C'({\omega_{eg}})|v_g({\omega_{eg}})|}e^{-2q({\omega_{eg}})z_0},
\label{eq:z_dipole_transition_rate}$$ where $C'(\omega) = C(\omega)/q(\omega)$ is a quantity is introduced to remove the wavevector dependence from the normalization. We also note that the group velocity $|v_g(\omega)| \propto 1/q(\omega)$, and thus the whole expression carries a wavevector dependence of $\Gamma_{\text{dipole}}^{(eg)} \propto q^3({\omega_{eg}})$.
![**Dipole transition rate enhancement by SMPs.** (a) Dipole transition rate for a $z$-oriented spin flip as a function of normalized frequency and distance $z_0$ from the emitter to the surface of a $d=200$ nm MnF$_2$ film. The transition rates decay exponentially with increasing distance from the surface. (b) Line cuts of the information shown in (a) for different fixed distances $z_0$. The axis on the left shows the total transition rate, while the axis on the right shows the Purcell factor, in other words, the transition rate normalized by the free space transition rate.[]{data-label="fig:dipole_rates"}](dipole_transition_rates_paper.png){width="\linewidth"}
Figure \[fig:dipole\_rates\] shows the emission rate as a function of frequency $\omega$ and emitter distance $z_0$ for a $d=200$ nm MnF$_2$ film. Panel (b) shows line cuts of the dipole transition rate at various emitter distances $z_0$. In this geometry we find that for the highest supported magnon frequencies, the total rate of emission may exceed $10^5$ s$^{-1}$, which corresponds to a decay time of 10 $\mu$s. This is eleven orders of magnitude of improvement over the free space decay lifetime of more than a week. We see that for sufficiently close distances $z_0$, the decay rate increases with $\omega$, spanning many orders of magnitude over a small frequency bandwidth. Furthermore, we see that with increasing distance $z_0$, the total decay rate is suppressed exponentially by the evanescent tail of the surface magnon. More specifically, we see in the exponential dependence $e^{-2q({\omega_{eg}})z_0}$ that in order for rate enhancement to be effective, $z_0$ should be comparable to or ideally smaller than $1/q \sim d$. For a 200 nm film, enhancement begins to saturate for $z_0 < 20$ nm. In terms of a potential experiment, these are promising parameters which could result in a total transition rate of $10^4$ s$^{-1}$. Finally, we note that at distances $z_0$ extremely near to the surface, effects such as material losses or nonlocality may cause the behavior of the transition rate to deviate slightly from the predicted behavior.
Thinner films offer even more drastic capabilities for enhancement. The dipole transition rate and Purcell factor scale as $\eta^3$, which means that shrinking the film thickness $d$ even by conservative factors can result in a rapid increase in the maximum transition rate achievable. This $\eta^3$ scaling is exactly the same scaling found for Purcell factors of electric dipole transition enhancement in the vicinity of highly confined electrostatic modes such as SPPs [@rivera2016shrinking; @caldwell2013low; @rivera2017making].
It is also worthwhile to consider not only the total transition rates, but also the Purcell factors. The right side axis of Figure \[fig:dipole\_rates\](b) shows the Purcell factor for spin relaxation into SMPs, computed as the ratio between the enhanced transition rate and the free space transition rate, and denoted as $F_p(\omega) = \Gamma_{\text{dipole}}/\Gamma_0$. We note that while the transition rate in the magnonic environment is technically the sum of the SMP emission rate and the radiative rate, in our systems the radiative rate is so small that it need not be considered.
Having established the duality between electric and magnetic surface polaritonics in the context of Purcell enhancement, other important conclusions about the scope and utility of SMPs follow. Most notably, Purcell factors for higher order magnetic processes should scale with mode confinement identically to those for the corresponding electric processes. Given an emitter-material system that can support such processes, it should be possible to compute transition rates of higher order processes such as magnetic quadrupole transitions and multi-magnon emission processes. Electromagnetic duality implies that a magnetic quadrupole transition Purcell factor, for instance, should scale as $\propto \eta^5$. For emission into modes confined to factors of 1000 or more, this enhancement factor could easily exceed $10^{15}$, eluding to the possibility of making highly forbidden magnetic quadrupole processes observable.
Emission with in-plane anisotropy
---------------------------------
Thus far, we have considered geometries of MnF$_2$ in which the anisotropy axis of the crystal is out of the plane of a thin film (in the $z$ direction). Past work has brought both theoretical interest as well as experimental studies on antiferromagnetic surface interfaces in which the magnetic permeability anisotropy axis lies in-plane. In other words, the material has negative permeability in the out-of-plane direction as well as one in-plane direction, while having a permeability of 1 in the other in-plane direction. This geometry gives rise to an rich anisotropic dispersion relation of SMP modes, which in turn result in a nontrivial angular dependence for processes of spontaneous emission. We summarize those findings here.
![**Dispersion for anisotropic modes.** Isofrequency contours for MnF$_2$ of thickness $d=200$ nm. The frequency labels are given as $\omega/\omega_0$, where $\omega_0$ is the resonance frequency of the material. The first type I modes are shown in red, while the type II modes with $n=1$ are shown in blue.[]{data-label="fig:anisotropic_isofreq"}](isofreq.png){width="40.00000%"}
{width="\textwidth"}
For the in-plane anisotropic geometry with $\mu = (\mu(\omega), 1, \mu(\omega))$, the dispersion (obtained again by solving Maxwell ’s equations for a quasimagnetostatic scalar potential) is given by solutions to: $$e^{qd\sqrt{\beta(\theta,\omega)}} = \frac{1 - \mu(\omega)\sqrt{\beta(\theta,\omega)}}{1 + \mu(\omega)\sqrt{\beta(\theta,\omega)}},$$ where $\beta(\theta,\omega) = \cos^2\theta+\sin^2\theta/\mu(\omega)$ and $\theta$ is the in-plane propagation angle measured with respect to the $x$-axis. When $\beta > 0$, the mode function has a $z$-dependence of $\cosh(qz)$ or $\sinh(qz)$, dependent on the parity of the solution. When $\beta < 0$, the modes have a $\cos(qz)$ or $\sin(qz)$ dependence. We note that the $\beta < 0$ solutions have a multiply branched structure which correspond to higher harmonic modes, just as with the in-plane isotropic case discussed throughout the text. Furthermore, recalling that $\mu < 0$ and examining $\beta(\theta,\omega)$, we see that for angles of propagation near 0, $\beta$ will be positive, while for angles of propagation near $\pi/2$, $\beta$ is negative. Based on the sign of $\beta$, we can classify the modes into two distinct types. We refer to $\beta > 0$ modes as type I modes, and $\beta < 0$ modes as type II modes. The fundamental type I modes propagate in the range $\theta \in (0,\theta_x)$, where $\theta_x = \tan^{-1}(\sqrt{-\mu(\omega)})$, while the type II modes with $n=1$ propagate in the range $\theta \in (\theta_y, \pi/2)$, with $\theta_y = \cos^{-1}(1/\sqrt{-\mu(\omega)})$. The angular propagation ranges for the type I modes and the lowest order type II mode are non-overlapping, and the gap between $\theta_x$ and $\theta_y$ increases with $\omega$.
The dispersion for even type I and type II modes are respectively given as: $$\begin{aligned}
q_{\text{I}} &= -\frac{1}{d\sqrt{\beta(\theta,\omega)}}\tanh^{-1}\left(\frac{1}{\mu(\omega)\sqrt{\beta(\theta,\omega)}}\right), \\
q_{\text{II}}^n &= \frac{1}{d\sqrt{-\beta(\theta,\omega)}}\tan^{-1}\left(\frac{1}{\mu(\omega)\sqrt{-\beta(\theta,\omega)}} + n\frac{\pi}{2}\right),
\end{aligned}$$ where $n$ is an integer. We see that for even type I modes, only a single band of surface polariton modes exists, while for type II modes, a richer structure with harmonics exists due to the multivalued nature of the arctangent, just as in the in-plane isotropic case. In Figure \[fig:anisotropic\_isofreq\], we see the isofrequency contours for the dispersion in the case of in-plane anisotropy. We clearly observe that the mode structure is anisotropic, in that type I modes behave differently than type II modes. We comment briefly on the polarization of the modes. The in-slab $\mathbf{H}$-field polarization of the type I and II modes are respectively given as $$\hat{\varepsilon}_q = \begin{cases}
\displaystyle\frac{\hat{q}\cosh(qz) + i\sinh(qz)\hat{z}}{\sqrt{2}}, & \text{type I} \\
\displaystyle\frac{\hat{q}\cos(qz) + i\sin(qz)\hat{z}}{\sqrt{2}}, & \text{type II}
\end{cases}.$$
Applying the same formalism as before, the rate of emission into SMPs per unit angle by a $z$-oriented spin flip of strength $\mu_B$ is given by
$$\frac{d\Gamma^{(eg)}}{d\theta} = \frac{\mu_B^2 \mu_0 {\omega_{eg}}}{2\pi\hbar}\frac{q^3(\theta,{\omega_{eg}})|\bm{\sigma}_{eg}\cdot\hat{\epsilon}_{\mathbf{q}}|^2}{C_q(\theta,{\omega_{eg}})|v_g(\theta,{\omega_{eg}})|}e^{-2q(\theta,{\omega_{eg}})z_0}.$$
The total rate is as per usual obtained by integrating over all angles: $$\Gamma^{(eg)} = \frac{\mu_B^2 \mu_0 {\omega_{eg}}}{2\pi\hbar} \int_0^{2\pi} d\theta \frac{q^3(\theta,{\omega_{eg}})|\bm{\sigma}_{eg}\cdot\hat{\epsilon}_{\mathbf{q}}|^2}{C_q(\theta,{\omega_{eg}})|v_g(\theta,{\omega_{eg}})|}e^{-2q(\theta.{\omega_{eg}})z_0}.$$
In Figure \[fig:polar\_dGammadtheta\] we see the lossless differential decay rate $d\Gamma^{(eg)}/d\theta$ plotted as a function of polar angle $\theta$ for a $z$-oriented spin flip transition at different emitter frequencies $\omega$. We see that with increasing frequency, the angular spread of type I modes narrows, while the angular spread of type II modes increases. We can understand this behavior in terms of the availability and confinement of modes for different propagation angles $\theta$. The most highly confined modes are the type I modes near the angular cutoff. As $\omega$ increases the confinement of type I modes at low angles increases, while the confinement of type II modes decreases. This system exhibits the interesting property that tuning the frequency of the emitter over a narrow bandwidth dramatically shapes the angular spectrum of polariton emission. An interesting consequence is that inhomogeneous broadening of the emitter could play a strong role in determining the observed angular spectrum of magnons emitted.
![**Magnetic dipole transition rate for in-plane anisotropic MnF$_2$.** Magnetic dipole transition rate for a $z$-oriented dipole transition a distance $z_0=5$ nm from the surface into two different SMP modes in a $d=200$ nm thick anisotropic slab of MnF$_2$. The type I mode emits most strongly but over a narrower range of frequencies. The cutoff frequency is the frequency at which the first type I mode no longer satisfies the boundary conditions. The first order type II mode is emitted more weakly but is supported over the entire range of frequencies for which $\mu(\omega) < 0$.[]{data-label="fig:anisotropic_rates"}](anisotropic_rates.png){width="50.00000%"}
In Figure \[fig:anisotropic\_rates\], we see the total transition rate $\Gamma^{(eg)}$ for a dipole emitter above MnF$_2$ oriented with the anisotropy axis in the $y$ direction. While the transition rates of both modes are greatly enhanced compared to the free space transition rate of order $10^{-6}$ $s^{-1}$, the type I mode benefits approximately two orders of magnitude more than the first type II mode. The Purcell factors for the type I mode in particular ranges from $10^{10}$ to $10^{12}$, and is thus quite comparable to Purcell factors obtained for the in-plane isotropic discussed previously. In this sense, we see that extreme enhancement of MD transition rates is achievable in both crystal orientations. The dispersion relation, however, is notably different in these cases. Further alterations to the dispersion in either geometry can be made using an external applied field, resulting in nonreciprocal propagation of modes. The net result is a highly flexible platform for ultrafast interaction between magnetic transitions and matter.
Conclusions and Outlook
=======================
We have shown that exceptionally confined surface magnon polaritons, such as those on antiferromagnetic materials, could speed up magnetic transitions by more than 10 orders of magnitude, bridging the inherent gap in decay rates which typically separates electric and magnetic processes. We predict that these confined magnetic surface modes in systems with realizable parameters may exhibit confinement factors in excess of $10^4$. We developed the theory of magnon polaritons and their interactions with emitters in a way that unifies this set of materials with other more well-known polaritonic materials, casting light on opportunities to use these materials to gain unprecedented control over spins in emitters. To push the field of magnon polaritonics at THz frequencies forward, it will be necessary to identify an ideal experimental platform for manipulating these modes and interfacing them with matter. As these modes exist at very low frequencies, experiments probing emitter interactions will need to take place under cryogenic conditions. Another question is what class of emitters may be well-suited to interact with these polaritonic modes. This is made challenging by the very narrow polaritonic bands of antiferromagnetic materials, as well as the few existing materials. This latter problem of course can be solved, as there are many more antiferromagnetic materials, which may support a negative permeability. Another interesting direction is the consideration of 2D antiferromagnetic materials. In terms of existing materials, a potential emitter system which can interact with these magnons is ErFeO$_3$, which has several electric and magnetic dipole transitions in the range between 0.25 and 1.5 THz [@mikhaylovskiy2017selective]. It could also prove interesting to consider GHz-THz orbital angular momentum transitions between high energy levels in Rydberg atoms. Processes involving the emission of multiple surface magnons, or mixed processes with the emission of a magnon polariton in addition to one or more excitations of another nearby material, could also be considered. In any case, surface magnon polaritons provide an interesting new degree of control over magnetic degrees of freedom in matter as well as a means to consider magnetic analogs at THz frequencies of many famous effects in plasmonics and polaritonics.
The authors thank Charles Roques-Carmes and Nicolas Romeo for help reviewing the manuscript. Research supported as part of the Army Research Office through the Institute for Soldier Nanotechnologies under contract no. W911NF-18-2-0048 (photon management for developing nuclear-TPV and fuel-TPV mm-scale-systems). Also supported as part of the S3TEC, an Energy Frontier Research Center funded by the US Department of Energy under grant no. DE-SC0001299 (for fundamental photon transport related to solar TPVs and solar-TEs). I.K. is an Azrieli Fellow, supported by the Azrieli Foundation, and was partially supported by the Seventh Framework Programme of the European Research Council (FP7-Marie Curie IOF) under grant no. 328853-MC-BSiCS. N.R. recognizes the support of the DOE Computational Science Graduate Fellowship (CSGF) Number DE-FG02-97ER25308.
|
{
"pile_set_name": "ArXiv"
}
|
\#1\#2[ .5ex -.1em/-.15em .25ex]{}
Abstract {#abstract .unnumbered}
========
This master thesis presents a new type of Positron Emission TOF Apparatus using Liquid xenOn (PETALO). The detector is based in the Liquid Xenon Scintillating Cell (LXSC). The cell is a box filled with liquid xenon (LXe) whose transverse dimensions are chosen to optimize packing and with a thickness optimized to contain a large fraction of the incoming photons. The entry and exit faces of the box (relative to the incoming gammas direction) are instrumented with large silicon photomultipliers (SiPMs), coated with a wavelength shifter, tetraphenyl butadiene (TPB). The non-instrumented faces are covered by reflecting Teflon coated with TPB. In this thesis we show that the LXSC can display an energy resolution of 5% FWHM, much better than that of conventional solid scintillators such as LSO/LYSO. The LXSC can measure the interaction point of the incoming photon with a resolution in the three coordinates of 1 mm. The very fast scintillation time of LXe (2 ns) and the availability of suitable sensors and electronics permits a coincidence resolution time (CRT) in the range of 100-200 ps, again much better than any current PET-TOF system. The LXSC constitutes the core of a high-sensitivity, nuclear magnetic resonance compatible, PET device, with enhanced Time Of Flight (TOF) sensitivity.
Introduction {#sec.intro}
============
Positron Emission Tomography (PET) is a non invasive imaging technique that produces a three-dimensional image of functional processes in the body. The system detects pairs of gamma rays emitted indirectly by a positron-emitting radionuclide (tracer), which is introduced into the body on a biologically active molecule.
PET technology has evolved rapidly during the last decade, thanks to the introduction of high-yield, high-resolution, and fast-response solid scintillator detectors, such as LSO/LYSO. At the same time, a new type of sensors, the so-called silicon-photomultipliers or SiPMs, are quickly replacing conventional PMTs as readout devices. SiPMs have large gains, comparable to that of PMTs, excellent photon detection efficiency (PDE), close to 50% around 420 nm can be fabricated in a variety of small size dices, allowing the construction of very modular pixelated systems. Furthermore, the use of SiPMs makes PET compatible with technologies that require very intense magnetic fields, such as Magnetic Resonance Imaging (MRI).
However, a limitation of this excellent technology is the high cost of the scanner. The driving factor is the high cost of crystals such as LSO or LYSO, but the large number of channels needed for good energy and spatial resolution are also relevant.
This master thesis presents a new type of Positron Emission TOF Apparatus using Liquid xenOn (PETALO). The use of liquid xenon (LXe) improves the energy resolution (ER) and the coincidence resolution time (CRT) which can be achieved by conventional PET systems based in LSO/LYSO. At the same time, LXe is much cheaper than LSO/LYSO and the high yield and homogeneity of the liquid allows a sparser instrumentation. As a consequence, PETALO may show better performance at lower cost than conventional PET scanners, thus offering a potential break-through of the technology.
PETALO is based in the Liquid Xenon Scintillating Cell (LXSC). The cell is a box filled with liquid xenon (LXe) whose transverse dimensions are chosen to optimize packing and with a thickness optimized to contain a large fraction of the incoming photons. The entry and exit faces of the box (relative to the incoming gammas direction) are instrumented with large silicon photomultipliers (SiPMs), coated with a wavelength shifter, tetraphenyl butadiene (TPB). The non-instrumented faces are covered by reflecting Teflon coated with TPB.
In this thesis we show that the LXSC can display an energy resolution better than 5% FWHM, much better than that of conventional solid scintillators such as LSO/LYSO. The LXSC can measure the interaction point of the incoming photon with a resolution in the three coordinates of 1–2 mm. The very fast scintillation time of LXe and the availability of suitable sensors and electronics permits a CRT in the range of 100-200 ps, again much better than any current PET-TOF system.
This document is organized as follows. In section \[sec.pet\] we introduce the basic ideas of Positron Emission Tomography. In section \[sec.ppet\], we discuss the performance characteristics of PET scanners. Section \[sec.ssd\] compares the conventional solid scintillating detectors used in conventional PET scanners with LXe. The PETALO concept is introduced in section \[sec.petalo\]. Section \[sec.sipm\] describes the concept and main performance parameters of silicon photomultipliers. The LXSC is discussed in section \[sec.lxsc\] and the study of the LXSC properties is discussed in section \[sec.mc\]. We sketch possible PETALO scanners in section \[sec.pets\]. Finally in section \[sec.conclu\] we present our conclusions.
Conclusions {#sec.conclu}
===========
PETALO is a new technology for TOF-PET systems based on the LXSC, a fully hermetic, homogenous box filled with liquid xenon and equipped with SiPM coated with TPB. PETALO offers the following advantages:
1. Light yield higher than any conventional SSD.
2. Excellent intrinsic energy resolution (3.5 – 4.5 % FWHM depending on configuration of LXSC and of light yield).
3. Excellent spatial resolution (1-2 mm in the three coordinates).
4. Potentially capable of detecting multi-site Compton events. Thus suitable as Compton telescope.
5. Very fast time response, resulting in enhanced sensitivity (reduced number of random coincidences) and making it possible breakthrough TOF application.
6. Fully MRI compatible.
7. Competitive cost. The sparse version of the LXSC2 would cost today roughly 2-3 times less than the cost of the equivalent LSO unit. With the cost of SiPMs falling, the cost of PET scanners will soon be fully dominated by the material of choice. Xenon is much cheaper than LSO and thus a large-scale apparatus (full body PET) is conceivable.
With respect to the pioneer work of the Waseda group, PETALO introduces the concept of the fully reflective, hermetic and homogenous LXSC, capable to detect the VUV light emitted by xenon with high efficiency and very small geometrical effects, thanks to the use of SiPMs.
[99]{}
D. L. Snyder, L.J. Thomas, M.M. Ter-Pogossian, [*A Matheematical Model for Positron-Emission Tomography Systems Having Time-of-Flight Measurements*]{}, IEEE Trans. NS-28, 3575 (1981). [doi:10.1109/TNS.1981.4332168](http://doi.org/10.1109/TNS.1981.4332168).
T.F. Budinger, [*Time-of-flight positron emission tomography: status relative to conventional PET*]{}, J. Nucl. Med. 24, 73 (1983).
E. Aprile and T. Doke, [*Liquid Xenon Detectors for Particle Physics and Astrophysics*]{}, Rev. Mod. Phys. 82 2053–2097 (2010). [doi:10.1103/RevModPhys.82.2053](http://doi.org/10.1103/RevModPhys.82.2053)
T. Doke, A. Hitachi, J. Kikuchi, K. Masuda, H. Okada and E. Shibamura. [*Absolute Scintillation Yields in Liquid Argon and Xenon for Various Particles*]{} Jpn. J. Appl. Phys. 41 1538 (2002). [doi:10.1143/JJAP.41.1538](http://dx.doi.org/10.1143/JJAP.41.1538)
V. Chepel, M.I. Lopes, V. Solovov, R. Ferreira Marques and A.J.P.L. Policarpo, [*Development of liquid xenon detectors for medical imaging*]{}, in Proceedings on The International Workshop on Technique and Application of Xenon Detectors (Xenon 01), University of Tokyo, December 2001. [doi:10.1142/9789812705075\_0003](http://doi.org/10.1142/9789812705075_0003).
V.Y. Chepel, M.I. Lopes, R. Ferreira Marques and A.J.P.L. Policarpo, [*Purification of liquid xenon and impurity monitoring for a PET detector*]{}, Nucl. Instrum. Methods A 349, 500 (1994) [doi:10.1016/0168-9002(94)91217-3](http://doi.org/10.1016/0168-9002(94)91217-3).
V.Y. Chepel, , M.I. Lopes, H.M. Araújo, M.A. Alves, R. Ferreira Marques and A.J.P.L Policarpo, [*Liquid xenon multiwire chamber for positron tomography*]{}, Nucl. Instrum. Methods A 367, 58 (1995). [doi:10.1016/0168-9002(95)00528-5](http://doi.org/10.1016/0168-9002(95)00528-5).
M.I. Lopes, V.Y. Chepel, J.C. Carvalho, R. Ferreira Marques and A.J.P.L. Policarpo, [*Performance analysis based on a Monte Carlo simulation of a liquid xenon PET detector*]{}, IEEE Trans. NS-42, 2298 (1995). [doi:10.1109/23.489431](http://dx.doi.org/10.1109/23.489431)
V.Y. Chepel, M.I Lopes, A Kuchenkov, R Ferreira Marques and A.J.P.L Policarpo, [*Performance study of liquid xenon detector for PET*]{}, Nucl. Instrum. Methods A 392, 427 (1997). [doi:10.1016/S0168-9002(97)00196-4](http://doi.org/10.1016/S0168-9002(97)00196-4).
P. Crespo, V.Y. Chepel, M.I. Lopes, L. Janeiro, R.F. Marques and A.J.P.L. Policarpo, [*Pulse shape analysis in the liquid xenon multiwire ionisation chamber for PET*]{}, IEEE Trans. NS-45, 561 (1998). [doi:10.1109/23.682448](http://doi.org/10.1109/23.682448 ).
V.Y. Chepel, V. Solovov, J. Van Der Marel, M.I. Lopes, P. Crespo, L. Janeiro; D. Santos, R.F. Marques and A.J.P.L. Policarpo, [*The liquid xenon detector for PET: recent results*]{}, IEEE Trans NS-46, 1038 (1999). [doi:10.1109/23.790822](http://dx.doi.org/10.1109/23.790822).
P. Crespo, J. van der Marel, V.Y. Chepel, M.I. Lopes, D. Santos, L. Janeiro, V. Solovov, R.F. Marques and A.J.P.L. Policarpo, [*Pulse processing for the PET liquid xenon multiwire ionisation chamber*]{}, IEEE Trans, NS-47, 2119 (2000). [doi:10.1109/23.903858](http://doi.org/10.1109/23.903858).
Lavoie L., [*Liquid xenon scintillators for imaging of positron emitters*]{}, Medical Physics 3, No. 5, 283 (1976).
Doke T., J. Kikuchi, and F. Nishikido, [*Time-of-flight positron emission tomography using liquid xenon scintillation*]{}, Nucl. Instrum. Methods A 569, 863 (2006). [doi:10.1016/j.nima.2006.07.067](http://doi.org/10.1016/j.nima.2006.07.067)
F. Nishikido, T. Doke, J. Kikuchi, T. Mori, H. Murayama, T. Ooshita and H. Takahashi, [*Performance of Prototype Liquid Xenon Scintillation Detector System for Time-of-Flight Type Positron Emission Tomography with Improved Photomultipliers*]{}, Jpn J. Appl. Phys. 44, 5193 (2005). [doi:10.1143/JJAP.44.5193](http://dx.doi.org/10.1143/JJAP.44.5193)
F. Nishikido, T. Doke, J. Kikuchi, T. Mori, K. Takizawa and M. Yamamoto, [*Performance of a Prototype of Liquid Xenon Scintillation Detector System for Positron Emission Tomography*]{}, Jpn J. Appl. Phys. 43, 779 (2004). [doi:10.1143/JJAP.43.779](http://dx.doi.org/10.1143/JJAP.43.779)
<http://next.ific.uv.es/next/> (Retrieved on 4 September 2015).
NEXT Collaboration, [*Operation and first results of the NEXT-DEMO prototype using a silicon photomultiplier tracking array*]{}, JINST 8 P09011 (2013). [doi:10.1088/1748-0221/8/09/P09011](http://dx.doi.org/10.1088/1748-0221/8/09/P09011).
NEXT-100 Collaboration, [*Design and characterization of the SiPM tracking system of the NEXT-100 demonstrator*]{}, e-Print: [arXiv:1206.6199](http://arxiv.org/abs/1206.6199) (2012).
J.J. Gomez-Cadenas, [*The NEXT experiment*]{}, e-Print: [arXiv:1411.2433](http://arxiv.org/abs/1411.2433) (2014).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'I will present an overview of identified particle spectra at high $p_T$ ($p_T{}^{>}_{\sim}$ 5 GeV/$c$) in both p+p collisions and AA collisions at RHIC. In p+p collisions, summary of particle ratios of K, $\eta$, $\omega$, $\rho$, $\phi$, $p$, $\bar{p}$, $\Lambda$ and heavy-flavor (open charm, $J/\Psi$) to $\pi$ at high-pt will be compiled and compared to the ratios of integrated yields. The spectra are used in $x_t$ scaling study and compared to pQCD calculations. These will help us establish particle composition in jets and the quark and gluon contributions to hadron production at high $p_T$. Similar jet chemistry has been extracted in Au+Au data in search for a quantitative measure of color charge dependence of jet energy loss.'
address: 'Physics Department, Brookhaven National Laboratory, Upton, NY 11973, USA'
author:
- Zhangbu Xu
title: 'High $p_T$ Identified Particle Spectra\'
---
Introduction
============
Fundamentals in QCD
-------------------
There are two fundamental questions in QCD that have motivated the relativistic heavy-ion collisions: quark confinement and symmetry breaking [@lee95]. There is asymptotic freedom, where the coupling ($\alpha_s$) of the strong interaction becomes weaker at shorter distance and higher energy, while the required energy to pull color objects apart grows with distance(quark confinement). The fact that gluons carry color charge also has profound consequences: gluons can interact strongly among themselves and with quarks, which generates $>98\%$ of the masses of hadrons(symmetry breaking). SU(3) is the right group for QCD. This provides an accurate account of color-charge factor (“interaction strength”) of the following three Leading Order processes: quark emitting a gluon ($\alpha_sC_F$), gluon splitting into two gluons ($\alpha_sC_A$), and gluon splitting into quark-anti-quark pair ($\alpha_sT_F$). The SU(3) predicts $C_A/C_F=9/4$ while experiments with $e^+e^-$ collisions at 91 GeV at LEP obtain a value of $2.29\pm0.06(stat.)\pm0.14(syst.)$ [@lepcolor] from multiple jet analyses. This provides a stringent constraint that SU(3) is the right group theory of QCD. In a four-jet event in $e^+e^-$ analyzed for the observation of color-charge factor, each jet has an average energy of 23 GeV, not very different from jet produced by p+p and A+A collisions at RHIC [@starppjet].
Monifests of QCD properties on high-$p_T$ hadron spectra
--------------------------------------------------------
How would the color-charge factor manifest in heavy-ion collisions? In jet quenching scenario, an energetic quark (gluon) emits a gluon due to interaction with QGP [@jetquenching]. To leading order, the diagram is the same as quark (gluon) emits a gluon in $e^+e^-$ collisions. Therefore, the energy loss is proportional to $\alpha_sC_{[A,F]}<\hat{q}>L^2$, where $C_A(C_F)$ is for quark (gluon) jet. It is obvious that if pQCD is applicable to the energy loss, the different energy loss between quark and gluon has to be $C=C_A/C_F=9/4$. Heavy quarks are a different beast (a nice feature though) since the gluon radiation is suppressed at small angle because the emission rate is inversely proportional to $((M/E)^2+\theta^2)$ (dead cone effect) [@deadcone]. In $5^{<}_{\sim}p_T{}^{<}_{\sim}20$ GeV/c, the charm quarks show similar behavior as light quarks ($u,d,s$) while bottom quarks are much less suppressed due to dead cone effect. This provides excellent observables to test one of the basic ingredient of QCD: whether SU(3) in QCD is still the most relevant effective group in strongly interacting Quark-Gluon Plasma. We emphasize the [**“effectiveness”**]{} since the author doesn’t believe we are testing the correctness of the SU(3) for QCD. The anolog is that there are many effective theories for QED in the condense matter even though the QED is the correct theory for electromagnetic interaction. In reality, it is not as simple as this, there are geometry and pathlength fluctuation in jet quenching. The WHDG model has taken this into account, and shows that jet quenching at parton level still proportional to this color-charge factor [@WHDG; @xnwang; @renk]. In addition, an energetic quark (gluon) can have Compton-like scattering with the partons in QGP, providing a flavor-changed ($q\rightarrow q$ or $g\rightarrow q$) leading parton [@liuconversion] (details in discussion section \[discussion\]). However, due to confinement, we are not able to directly observe quark or gluon in an experiment. Energetic partons fragment into cluster of hadrons (jet). Jet reconstruction therefore provides the closest observable to single energetic partons. Indeed, this is the basic tool for the measurements of color-charge factor and for many observations and discoveries related to QCD or beyond QCD [@lepcolor; @CDF; @starppjet] in high-energy $e^+e^-$ and hadron collisions.
The production of hadrons from jet can be separated into three distinct terms in a naive picture: parton distribution function(PDF), parton interaction cross-section at $2\rightarrow2$($gg\rightarrow gg$,$qq\rightarrow qq$,$qg\rightarrow qg$, $gg\rightarrow q\bar{q}$, $q\bar{q}\rightarrow gg$, and $q\bar{q}\rightarrow q\bar{q}$), and parton fragmentation function [@jetquenching]. In principle, PDF is provided by DIS e+p collisions, $2\rightarrow2$ amplitude is provided by pQCD theory, and FF is from $e^+e^-$. If we assume: those three terms can be factorized, PDF and FF are measured with sufficient accuracy, and FF has universality, we can predict what hadron spectra at high $p_T$ should be in p+p collisions at RHIC. Are these ingredients sufficient for p+p collisions? how will measurements in p+p collisions provide additional information for our understanding of QCD and for model development? What will be modified in A+A collisions: PDF, FF, $\alpha_s$ and the effective color-charge factor?
In heavy-ion collisions at RHIC, the energy is sufficient to produce a well-defined jet at initial stage as in p+p collisions [@starppjet]. However, the soft processes, which are necessary for QGP creation, produce overwhelming background and prevent a meaningful full-jet reconstruction to date. Instead, we rely on leading hadrons to identify and study the effect of jet quenching. This inevitably requires a detailed understanding of the fragmentation function (FF) of the partons to hadrons. We know from detailed measurements in $e^+e^-$ at LEP and SLAC that gluon and quark jets have distinct features when fragmenting. In general, gluons produce more soft particles and more leading baryons than quarks do. Flavor separated fragmentation functions from quark and gluon into identified pions have been provided by $e^+e^-$ data, and have been tested extensively at hadron colliders [@CDF; @phenixpi0; @starPIDpapers].
![Identified charged hadron spectra ($\pi^{\pm},p,\bar{p}$) in p+p collisions at $\sqrt{s}=200$ GeV. The curves are pQCD calculations with different fragmentation functions for pion and proton.[]{data-label="ppspectra"}](pionpbarpSpectra_v4.eps)
In this talk, I have presented the recent development in constraints of fragmentation functions by measurements of identified hadrons in p+p collisions at RHIC, and new theoretical development of fragmentation functions with global fit to data provided by both $e^+e^-$ and hadron collisions [@AKK; @KKP; @DSS; @AKK2]. The single inclusive hadron spectra are limited to be [**$p_T{}^{>}_{\sim}5$**]{} GeV/c to avoid complication of quark coalescence effect in A+A collisions and to provide a more reliable test bench for pQCD. I would then use this information to guide us in interpreting nuclear modification functions $R_{AA}$ of leading identified hadrons in central Au+Au collisions at RHIC.
Fragmentation Functions and Jet Chemistry in p+p collisions at RHIC
===================================================================
The most uncertain part of the three terms relevant to RHIC physics is fragmentation function. In the $5<p_T<20$ GeV/c range at RHIC energy, most of the pions come from $qg\rightarrow qg$ and $qq\rightarrow qq$ processes (PYTHIA). Different pion fragmentation functions provided by DSS, AKK and KKP (abbreviations of the authors’s last names) parameterizations can satisfactorily describe $\pi^{0}$ and $\pi^{\pm}$ [@AKK; @AKK2; @KKP; @DSS]. Although baryon production may be more difficult to interpret/implement in QCD models (e.g. popcorn mechanism in Lund Model), it is no different from mesons in terms of parameterization of fragmentation functions provided enough data points with good accuracy in the relevant region. However, fragmentation functions from quark and gluon to leading baryons (proton, $\Lambda$, etc.) are poorly constrained. This is especially true at high-z (fraction of leading hadron to the jet energy), where it is most relevant in jet quenching at RHIC. In Ref. [@DSS], the authors performed a global fit of proton and anti-proton fragmentation functions with data from SLAC, LEP and STAR data [@starPIDpapers]. They concluded that “at the presently accessible range of transverse momenta ($p_T<7$ GeV/c) and at mid-rapidities the production of single-inclusive hadrons is mainly driven by gluon-induced processes and fragmentation, turning these data into the best constraint on the gluon fragmentation function $D^{p}_{g}$ at large value of $z$ currently available”. The authors stated that, with extended measurements in $p_T$ to where the quark fragmentation becomes significant, these data will allow to separate quark-to-proton and anti-quark-to-proton fragmentation functions in the global fit.
PHENIX is able to reconstruct in the EMC mesons which decay to final-state photons and in the future with TOF and Aerogel Cherenkov Detector [@phenixPIDQM06] for charged hadron identification. STAR/RHIC extends the particle identification of charged hadrons in Time Projection Chamber (TPC) from $p_T{}^{<}_{\sim}7$ GeV/c to $p_T{}^{<}_{\sim}15$ GeV/c. There are several improvements over the years in STAR, which make this extension possible:
- Momentum and distortion calibration of TPC[@GeneTPC]
- Ionalization energy loss (dE/dx) calibration of TPC [@PIDtech]
- Jet-triggered data by EMC to increase the statistics at high $p_T$ [@starppjet]
![$\bar{p}/p$ ratio as function of $p_T$ in p+p collisions at $\sqrt{s}=200$ GeV. In comparison is the ratio from pQCD+DSS and PYTHIA. Noted that other fragmentation models (KKP, AKK, Kretzer) did not distinguish particle and anti-particle and therefore the ratio is unity.[]{data-label="pbarpratio"}](pbarpRatioJp2_v7.eps)
Fig. \[ppspectra\] shows the $\pi^{\pm}$, p and $\bar{p}$ spectra from minbias and jet-triggered p+p collisions. The results from jet-triggered event sample have been corrected for trigger bias [@QMproceedings], and are consistent with results from minbias data where two overlap. Most of the models (DSS, KKP and AKK) can reproduce the charged pion spectra quite well. We see slightly decrease of $\pi^{-}/\pi^{+}$ from unity at $p_T{}^{<}_{\sim}5$ GeV/c to 0.8 at $p_T\simeq15$ GeV/c. The difference is due to the contribution of valence-quark fragmentation. In the current KKP, AKK and Kretzler models, there is no distinction between $u\rightarrow\pi^{+}$ and $u\rightarrow\pi^{-}$. Neither is there difference between $u,d\rightarrow p$ and $u,d\rightarrow\bar{p}$. This means that $\pi^{-}/\pi^{+}=\bar{p}/p=1$. The DSS model includes this difference, and is found to describe the $\pi^{-}/\pi^{+}$ ratio as a function of $p_T$ as shown in Fig. \[pbarpratio\]. AKK and DSS can describe the proton yields reasonably well (within 20%). However, DSS model tends to overpredict the ratio of $\bar{p}/p$ at high $p_T$ while AKK doesn’t distinguish proton and anti-proton. The overprediction of $\bar{p}$ yields at high $p_T$ from DSS means that the fragmentation of quark-to-antiproton was over-estimated. The new AKK [@AKK2] fragmentation functions implement the flavor dependence of parton fragmentation and predict a 10–20% percent difference of $\pi^{-}$ to $\pi^{+}$ at $p_T=10$ GeV/c due to valence quark contribution. These ratios should provide a stringent constraint on quark and gluon contributions to identified hadrons at this $p_T$ range. Fig. \[jetchemistry\] shows the measured particle ratio in p+p collisions by STAR and PHENIX collaborations [@QMproceedings; @phenixeta; @starcharm; @starstrange; @phenixpi0].
![The jet chemistry of Particle ratios at high $p_T$ in p+p collisions at $\sqrt{s}=200$ GeV. Particle ratios are obtained from particle yields with $p_T>5$ GeV/c except $\Lambda(\bar{\Lambda})/\pi$.[]{data-label="jetchemistry"}](ratiosChempp.eps)
Effects of Jet Quenching on Jet Chemistry
=========================================
The light mesons are: $\pi^{\pm,0}$, $\eta$, $K^{0}_{S}$, $\rho^{0}$ and $\omega$; baryons are: $\Lambda(\bar{\Lambda})$, $p(\bar{p})$; in addition, heavy-flavor measurements are: $(Q\rightarrow e)$ and $J/\Psi$. The jet chemistry from particle ratios is very different from particle ratios from the total integrated yields at low $p_T$ [@starwhitepaper]. For example, pions are the dominant source of soft hadrons (bulk) while $\rho$ yields at high $p_T$ are high than pion yields [@starrho]. The change of chemistry in the bulk provides a chemometer for assessing the chemical potential, temperature and strangeness equilibrium. Jet quenching changes the parton composition and fragmentation, resulting in possible change of hadron chemistry in out-going jets. Nuclear modification factor $R_{AA}$ is used to quantify the suppression of hadron yields in Au+Au collisions relative to $N_{bin}$-scaled yields in p+p collisions. We observed that light mesons ($\pi,\eta,\rho,K^{0}_{S}$) at high $p_T$ and electrons from heavy-flavor semileptonic decay have similar $R_{AA}$ while $p(\bar{p})$ $R_{AA}$ are systematically above the pion $R_{AA}$.
In AKK model and PYTHIA simulation as well, 60% of $\pi^{\pm}$ are from quark fragmentation and 40% from gluon fragmentation at $p_T=10$ GeV/c while that partition for ($p+\bar{p}$) is 10% and 90%, and that for $K^{\pm},K^{0}_{S}$ is 20% and 80%. This provides a tool for studying quark and gluon color-charge factor in jet quenching. In WHDG model [@WHDG], the $R_{AA}$ charm quark resulting from the radiative energy loss is very similar to that of the light quarks in the $p_T$ range accessible to us. If we separate the $R_{AA}$ of hadrons at parton sources, $$R_{AA}^{\pi}=0.6R_{AA}^{q}+0.4R_{AA}^{g},
R_{AA}^{p}=0.1R_{AA}^{q}+0.9R_{AA}^{g},
R_{AA}^{K}=0.2R_{AA}^{q}+0.8R_{AA}^{g},
R_{AA}^{c\rightarrow e}=R_{AA}^{Q}.$$ Since $R_{AA}^{q}\simeq C\times R_{AA}^{g}$ [@WHDG] where $C$ is the effective color-charge factor, which is 9/4 in pQCD model [@xnwang]. If double ratio of hadron $R_{AA}$ is taken, this results in: $$R_{AA}(p/\pi)={{0.1C+0.9}\over{0.6C+0.4}},
R_{AA}(K/\pi)={{0.2C+0.8}\over{0.6C+0.4}},
R_{AA}(Q\rightarrow e/\pi)={{C}\over{0.6C+0.4}}.$$ In the analog to the measurement of $C_A/C_F$ in $e^{+}e^{-}$ collisions, we plot in Fig. \[colorvalue\] the effective color-charge factor ($C$) extracted from this very simple approach. One can see that the value is systematically lower than 9/4. When the WHDG model [@WHDG] was used to fit the $R_{AA}$ of $\pi^{0}$ in $5<p_T<20$ GeV/c [@phenixqhat], it is evident that there is a stronger increase of $R_{AA}$ as function of $p_T$ in the model than that exists in the data. The $R_{AA}^{\pi^{0}}$ data points essentially show no $p_T$ dependence. At parton level, the $R_{AA}$ of light quarks and gluons in WHDG model show very little $p_T$ dependence. Since $R_{AA}^{q}>R_{AA}^{g}$ and quark contribution to final-state pion increases with $p_T$, the resulting $R_{AA}^{\pi}$ increases with $p_T$. On the other hand, if the energy loss were not sensitive to the different color-charge factor of quark and gluon ($R_{AA}^{q}=R_{AA}^{g}$) in the model, the resulting $R_{AA}^{\pi}$ would have shown little $p_T$ dependence. This seems to be consistent with the jet chemistry analysis of $R_{AA}$ among the identified hadrons, where the effective color-charge factor $C\simeq1$.
![Effective color-charge factor from $R_{AA}$ due to energy loss between gluon and quark energetic parton jets. The algebra is detailed in the text. This shows that the difference of energy loss between different partons is small, in contrast to the expectation of $C_A/C_F=9/4$ and dead cone effect. The band in the figure is the measurement of $C_A/C_F$ from $e^+e^-$ at LEP.[]{data-label="colorvalue"}](RHIC_C2F.eps)
Discussions {#discussion}
===========
What does it mean when $C\neq9/4$? As mentioned earlier, there are several effects which can destroy the proportionality of energy loss and color-charge factor. We attempt to select only the particles with $p_T>5$ GeV/c to avoid the potential contributions from quark coalescence at hadronization to the jet chemistry. It is not clear whether that is a sufficient high $p_T$. The color-charge factor was measured and compared to leading order (LO) and NLO at small $\alpha_{S}$ where pQCD is applicable. In a strongly interaction QGP, although the out-going jet energy is comparable to jets produced in 4-jet events at LEP energy in $e^+e^-$ collisions, the radiation or collision happens in a strongly interacting medium and with much smaller $q^{2}$.
Ref. [@liuconversion; @liuFries] show that higher order contribution can change the in-coming leading quark into an out-going leading gluon. This effectively decreases the value of $C$. With a larger interaction cross-section than expected from pQCD, the conversion can describe the $\bar{p}/\pi^{-}$ ratio from decrease from p+p to Au+Au, to a reverse trend. The authors [@liuFries] further suggested a test using $K/\pi$ ratio. The current $K^{0}_{S}/\pi$ measurement (p+p data from PHENIX and Au+Au from STAR) is not yet able to distinguish the models.
Recently, there was a proposal [@brodskypaper] of “color transparency and direct hadron production” to explain the baryon/meson production at intermediate $p_T$. I thought that this is another fancy term for quark coalescence. On the other hand, there may be a significant difference between this mechanism and quark coalescence at higher $p_T(>5$ GeV/c) and with vector mesons ($\phi,J/\Psi$) as presented by [@QMproceedings] where the effect of quark coalescence is small but the direct hadron production should be significant from this prediction. The proposed mechanism seems to be consistent with the measurements as noted in the Ref. [@brodskypaper]. Another possibility is that the ridge, which carries the chemistry of the bulk with high baryon/meson ratio [@QMproceedings], extends to high $p_T$ and contaminates the jet spectra. Extending the study of ridge chemistry at lower $p_T$ ($<5$ GeV/c) to higher $p_T$ will be crucial to see how much the ridge contributes to the baryon yields at high $p_T$. We noted that there are many examples of effective field theory in QED condensed matter physics when a strong field/interaction exist in the medium. For example, the perfect 2D electron liquid in Fractional Quantum Hall Effect in QED has Chern-Simon theory. We hope that a better measurement of effective color-charge factor may provide us with important information about the effective interactions in QGP.
Summary
=======
In summary, RHIC has provided the first measurements of jet chemistry in both p+p collisions and Au+Au collisions at high $p_T$. Comparison of $R_{AA}$ among different particles shows that light mesons have similar $R_{AA}$ as expected from pQCD and quark/gluon fragmentation while baryons have similar (slightly larger but within systematical errors) $R_{AA}$ as light mesons unexpected from jet quenching and quark/gluon fragmentation. The gluon jets are not more suppressed than light-quark jets or heavy-quark jets. Therefore, the effective color-charge factor ($C_A/C_F$) is consistent with unity.
Acknowledgement
===============
The author would like to thank the organizer for invitation, and thank Adam Kocoloski, Yichun Xu, Zebo Tang, Lijuan Ruan, Patricia Fachini, Anne Sickles, Bedanga Mohanty, Simon Albino, Xin-nian Wang, Simon Wicks, Thorsten Renk and Werner Vogelsang for valuable discussions and providing the data points and pQCD curves. This work is supported in part by the 2004 Presidential Early Career Award (PECASE) and DOE.
References {#references .unnumbered}
==========
[10]{} T.D. Lee, The proceedings of Quark Matter 1995. ALEPH Collaboration: R. Barate [*et al.*]{}, ; OPAL Collaboration: G. Abbiendi [*et al.*]{}, ; DELPHI Collaboration: P. Abreu [*et al.*]{}, Zeit. Phys. C 59, 357 (1993);. B.I. Abelev [*et al.*]{}, . R. Baier [*et al.*]{}, Ann. Rev. Nucl. Part. Sci. 50, 37 (2000); M. Gyulassy [*et al.*]{}, nucl-th/0302077 Yu. L. Dokshitzer and D.E. Kharzeev, . S. Wicks [*et al.*]{}, arXiv: nucl-th/0512076. X.N.Wang [*et al.*]{}, ; ; X.N. Wang [*et al.*]{}, ; . T. Renk and K. Eskola, W. Liu, C.M. Ko and B.W. Zhang, nucl-th/0607047 A. Abulencia [*et al.*]{}(CDF Collaboration), . S.S. Adler [*et al.*]{}(PHENIX Collaboration), . B. Mohanty [*et al.*]{}, nucl-ex/0705.9053; L Ruan [*et al*]{}, nucl-ex/0701070; J. Adams [*et al.*]{}, ; J. Adams [*et al.*]{}, ; B.I. Abelev [*et al.*]{},; J. Adams [*et al.*]{}, . S. Albino [*et al.*]{}, . B. A. Kniehl, G. Kramer and B. Potter, . Daniel de Florian, Rodolfo Sassot and Marco Stratmann arXiv: 0707.1506; ; calculations provided by W. Vogelsang. S. Albino, B.A. Kniehl and G. Kramer, e-Print: arXiv:0803.2768 \[hep-ph\]. M. Kunno [*et al.*]{}, Quark Matter 2006. G. Van Buren [*et al.*]{}, . M. Shao [*et al.*]{}, ; H. Bichsel, ; Yichun Xu [*et al.*]{}, DNP 2007, paper in preparation. P. Fachini, these proceedings; B. Mohanty, these proceedings, arXiv:0804.4760 \[nucl-ex\]; C. Suarez, these proceedings. S.S. Adler [*et al.*]{}, ; S.S. Adler [*et al.*]{}, ; ; S.S. Adler [*et al.*]{}, ; B.I. Abelev [*et al.*]{},; J. Adams [*et al.*]{}, ; B.I. Abelev [*et al.*]{}, arXiv:0805.0364. J. Adams [*et al.*]{}, nucl-ex/0601042; STAR Preliminary QM06. J. Adams [*et al.*]{} (STAR Collaboration), . J. Adams [*et al.*]{} (STAR Collaboration), . A. Adare, [*et al.*]{}(PHENIX Collaboration), arXiv:0801.4020 W. Liu and R.J. Fries, arXiv:0801.0453 \[nucl-th\] S. J. Brodsky, RHIC Online news, January 15, 2008; S. J. Brodsky and A. Sickles, arXiv:0804.4608.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Jonathan Baxter[^1]\
Department of Mathematics and Statistics\
The Flinders University of South Australia
title: Learning Internal Representations
---
[^1]: This work was supported by a Shell Australia Postgraduate Fellowship and an Australian Postgraduate Research Award.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we focus on finding all the factorials expressible as a product of a fixed number of $2k$-nacci numbers with $k \geq 2$. We derive the 2-adic valuation of the $2k$-nacci sequence and use it to establish bounds on the solutions of the initial equation. In addition, we specify a more general family of sequences, for which we can perform a similar procedure. We also investigate a possible connection of these results with $p$-regular sequences.'
address: 'Jagiellonian University, Faculty of Mathematics and Computer Science, Institute of Mathematics, [Ł]{}ojasiewicza 6, 30 - 348 Kraków, Poland'
author:
- Bartosz Sobolewski
title: 'The 2-adic valuation of generalized Fibonacci sequences with an application to certain Diophantine equations.'
---
Introduction {#sec:introduction}
============
For a fixed integer $r \geq 2$ define the generalized Fibonacci ($r$-nacci) sequence $\{t_n\}_{n\geq 0}$ as follows: $$\begin{aligned}
\label{def:2k-nacci}
t_n =
\begin{cases}
0 & \text{ for } n = 0,\\
1 & \text{ for } 1 \leq n \leq r-1, \\
\sum\limits_{i=1}^{r} t_{n-i} & \text{ for } n \geq r-1.
\end{cases}\end{aligned}$$ Notice, that for $r=2$ we obtain the usual Fibonacci sequence, which has already been studied extensively. In this paper we will mostly restrict ourselves to the case of even $r \geq 4$ and write $r=2k$ for some $k \geq 2$. The main motivation for our considerations is to completely solve the equation $$\begin{aligned}
\label{eq:product_of_terms}
m! = \prod_{i=1}^d t_{n_i}\end{aligned}$$ in positive integers $m, n_1, ..., n_d$.
For $p$ prime define the $p$-adic valuation of a non-zero integer $s$ as $\nu_p(s) = \max \{l \geq 0 : p^l | s \}$ and $\nu_p(0) = \infty$. Equation (\[eq:product\_of\_terms\]) for the case of $r=3$ and $d=1$ was solved by Lengyel and Marques in [@Lengyel_Marques] by means of computing $\nu_2(t_n)$ and then applying this result to obtain an effective upper bound on $m$ and $n_1$. In this paper we will follow a similar procedure for $2k$-nacci sequences with $k \geq 2$.
To begin with, in Theorem \[thm:equation\_solution\] we specify a more general family of integer sequences $\{s_n\}_{n\geq 0}$ for which we are able to solve equation (\[eq:product\_of\_terms\]) and show a general procedure to achieve this goal. Informally speaking, we need the term $s_n$ to grow at least exponentially and $\nu_p(s_n)$ – at most polynomially with an exponent less than 1, for some $p$ prime.
Theorem \[thm:2-adic\_order\] provides a simple expression for $\nu_2(t_n)$ when $r=2k \geq 4$ and the subsequent corollary shows that the sequence $\{t_n\}_{n\geq 0}$ satisfies the conditions of Theorem \[thm:equation\_solution\]. We then find all the solutions of equation (\[eq:product\_of\_terms\]) for small values of $k$ and $d$.
We also briefly discuss how our results are related to $p$-regular sequences. Recall that a sequence $\{s_n\}_{n\geq 0}$ with rational values is $p$-regular iff its $p$-kernel $$\mathcal{N}_p(a) = \left\{ \{s_{p^l n + j} \}_{n\geq 0} : \; l\geq 0, \; 0 \leq j < p^l\right\}$$ is contained in a finitely generated $\mathbb{Z}$-module. More details on regular sequences can be found in [@Allouche_Shallit] and [@Allouche_Shallit_2]. As we note later, the formula given in Theorem \[thm:2-adic\_order\] implies $2$-regularity of $\{\nu_2(t_n)\}_{n \geq 0}$. However, it turns out that exponential growth of a sequence $\{s_n\}_{n \geq 0}$ and $p$-regularity of $\{\nu_p(s_n)\}_{n \geq 0}$ still do not guarantee that the assumptions of Theorem \[thm:equation\_solution\] are met. In this case we cannot determine, using the shown method, whether the equation (\[eq:product\_of\_terms\]) has only a finite number of solutions.
Main results {#sec:main}
============
As we mentioned before, we start with describing a general situation in which equation (\[eq:product\_of\_terms\]) can be completely solved. For two sequences $\{a_n\}_{n \geq 0}$ and $\{b_n\}_{n \geq 0}$ we denote $a_n = O(b_n)$ if there exists a positive constant $K$ such that $|a_n| \leq K |b_n|$ for sufficiently large $n$. Similarily, we write $a_n = \Omega(b_n)$ if there exists a positive constant $K$ such that $|a_n| \geq K |b_n|$ for sufficiently large $n$. First, we give an auxiliary lemma, also used in [@Lengyel_Marques], which is an easy corollary from Legendre’s formula for $\nu_p(m!)$.
\[lem:legendre\_ineq\] For any integer $m \geq 1$ and prime $p$, we have $$\frac{m}{p-1} - \left \lfloor \frac{\log m}{\log p} \right \rfloor -1 \leq \nu_p(m!) \leq \frac{m-1}{p-1}.$$
\[thm:equation\_solution\] Let $\{s_n\}_{n\geq 0}$ be a sequence of positive integers such that $$\begin{aligned}
\label{eq:from_below}
\log s_n = \Omega(n).\end{aligned}$$ Let $p$ be a prime. Assume that $$\begin{aligned}
\label{eq:from_above}
\nu_p(s_n) = O\left(n^C\right)\end{aligned}$$ for some constant $C < 1$. Then for each fixed positive integer $d$ the equation $$\begin{aligned}
\label{eq:general_product}
m! = \prod_{i=1}^d s_{n_i}\end{aligned}$$ has only a finite number of solutions in $m, n_1, n_2, \ldots, n_d$.
We adjust the method used in [@Lengyel_Marques] to a more general setting. Roughly speaking, we will show that if (\[eq:general\_product\]) is satisfied and we let both sides grow, then the $p$-adic valuation of the right hand side increases slower than $\nu_p(m!)$. For each value of $p$ we proceed in the same way, so for simplicity assume that $p=2$. By our assumptions, there exist some positive constants $K_1, K_2$ and an integer $n_0 \geq 0$ such that $\nu_2(s_n) \leq K_1 n^C$ and $\log_2 s_n \geq K_2 n$ for $n \geq n_0$.
Suppose that $n_i \geq n_0$ for $i=1,2,\ldots,d$. There is only a finite number of solutions with $m < 6$ because $s_n$ grows at least exponentially. By Lemma \[lem:legendre\_ineq\], for $p = 2$ we get $$\begin{aligned}
\label{eq:m_lower_bound}
\frac{1}{2}m \leq m - \lfloor \log_2 m \rfloor - 1 \leq \nu_2(m!), \end{aligned}$$ where the leftmost inequality is true for $m \geq 6$. On the other hand, $$\begin{aligned}
\nu_2 \left( \prod_{i=1}^d s_{n_i} \right ) = \sum_{i=1}^{d} \nu_2(s_{n_i}) \leq K_1 \sum_{i=1}^{d} n_i^C
\leq d K_1 \left(\max_{1 \leq i \leq d} n_i \right)^C.\end{aligned}$$ Hence, for $m \geq 6$ $$\begin{aligned}
\label{eq:m_below_n}
m \leq 2dK_1 \left(\max_{1 \leq i \leq d} n_i \right)^C.\end{aligned}$$ We need another inequality with $n_i$ bounded from above by $m$. By the assumption (\[eq:from\_below\]) $$\begin{aligned}
\log_2 \left(\prod_{i=1}^d s_{n_i} \right) = \sum_{i=1}^{d} \log_2 s_{n_i} \geq K_2 \sum_{i=1}^{d} n_i \geq K_2 \max_{1 \leq i \leq d} n_i.\end{aligned}$$ Furthermore, for $m \geq 5$ the following inequality holds: $$\begin{aligned}
\label{eq:m_fact_from_above}
m! < \left ( \frac{m}{2} \right)^m,\end{aligned}$$ and hence $$\begin{aligned}
\label{eq:n_below_m}
m \log_2 \frac{m}{2} > K_2 \max_{1 \leq i \leq d} n_i.\end{aligned}$$ Combining inequalities (\[eq:m\_below\_n\]) and (\[eq:n\_below\_m\]) yields $$\begin{aligned}
\label{eq:m_below_m}
\log_2 \frac{m}{2} > K m^{\frac{1}{C} -1},\end{aligned}$$ where $K = K_2/ (2dK_1)^{1/C}$. But $C < 1$ implies that the exponent in (\[eq:m\_below\_m\]) is strictly greater than 0, so (\[eq:m\_below\_m\]) holds only for finitely many $m$. Thus, by inequality (\[eq:n\_below\_m\]) we obtain an upper bound for all $n_i$.
Now assume without loss of generality that $n_i < n_0$ for $i=j+1,j+2,\ldots,d$. Observe that for each fixed $n_{j+1},n_{j+2},\ldots,n_d$ the problem is equivalent to solving the equation (\[eq:general\_product\]) with $d=j$ and the left hand side divided by a positive integer constant $M$. Then in (\[eq:m\_lower\_bound\]) and (\[eq:m\_fact\_from\_above\]) we need to replace $m!$ with $m! /M$ which only changes the set of $m$ for which both of those inequalities hold. But this leads to the same conclusion as before.
Observe that the method of Theorem \[thm:equation\_solution\] does not work if we let $d$ be unbounded. Indeed, the constant $K$ in (\[eq:m\_below\_m\]) becomes arbitrarily small as $d$ increases, so we cannot use the subsequent argument. Informally speaking, the $p$-adic valuation of the expression $\prod_{i=1}^d s_{n_i}$ might grow too fast for the method to work.
The condition (\[eq:from\_below\]) is satisfied for sequences expressible in Binet form. Hence, for linear recurrence sequences, we usually need to check only the condition (\[eq:from\_above\]) for some $p$. One might also ask whether we can replace it with some other assumption. Shu and Yao proved in [@Shu_Yao] a condition on a binary recurrence sequence, which guarantees $p$-regularity of $\{\nu_p(s_n)\}_{n \geq 0}$, and mentioned a possible generalization to recurrences of higher order. It is known that $p$-regular sequences grow at most polynomially, which is a result by Allouche and Shallit [@Allouche_Shallit]. Unfortunately, this does not give a bound on $C$ in (\[eq:from\_above\]) and the proof of Theorem \[thm:equation\_solution\] fails if $C \geq 1$. Therefore, some additional information besides regularity needs to be known about $\{\nu_p(s_n)\}_{n \geq 0}$ in order to put the theorem to use.
The reasoning in Theorem \[thm:equation\_solution\] provides an upper bound on the solutions of the equation (\[eq:general\_product\]) if we are able to find the values of $C,K_1, K_2$ and $n_0$. We will show that it is indeed the case for our sequence $\{t_n\}_{n \geq 0}$. We start by determining the $2$-adic valuation of each $t_n$. A similar characterization of $\{\nu_2(t_n)\}_{n\geq 0}$ for $r=2$ is given by Lengyel in [@Lengyel_order] and [@Lengyel_multi] (by a different method).
\[thm:2-adic\_order\] If $k \geq 2$, then the sequence $\{\nu_2(t_n)\}_{n\geq 0}$ satisfies the following conditions: $$\nu_2(t_n) =
\begin{cases}
0 & \text{ for } n \equiv 1,2,...,2k \pmod{2k+1}, \\
1 & \text{ for } n \equiv 2k+1 \pmod{2(2k+1)}, \\
\nu_2(n) + \nu_2(k-1) + 2 & \text{ for } n \equiv 0 \pmod{2(2k+1)}.
\end{cases}$$
The proof of the theorem is presented in Section \[sec:proof\].
Theorem \[thm:2-adic\_order\] implies that $\{\nu_2(t_n)\}_{n\geq 0}$ is a $2$-regular sequence. Note that this conclusion does not follow directly from the results of [@Shu_Yao], as we consider recurrence of any even order.
Now we proceed to show that Theorem \[thm:equation\_solution\] can be applied to our sequence $\{t_n\}_{n \geq 0}$. The following lemma establishes a lower bound on $t_n$.
\[lem:term\_size\] For all $n \geq 1$ we have $$\begin{aligned}
\label{eq:exp_growth}
t_n \geq \phi^{n-r-1},\end{aligned}$$ where $\phi$ is the unique real root of the equation $x^{r} = x^{r-1} + ... + x + 1$ lying in the interval $(1,2)$.
By lemma 3.6 in [@Wolfram] there is indeed exactly one such $\phi$, which in addition lies in the interval $(2(1- 2^{-r}), 2)$. For $n = 1, 2, ..., r$ the inequality (\[eq:exp\_growth\]) follows from starting conditions for $t_n$ and the fact that $t_r = 2(r-1) \geq 2 > \phi$. Then we proceed easily by induction.
\[cor: finite\_solutions\] If $r=2k \geq 4$ then the equation (\[eq:product\_of\_terms\]) has only a finite number of solutions in $m, n_1, n_2, \ldots, n_d$ and this number can be effectively bounded from above.
By theorem \[thm:2-adic\_order\] we have $$\nu_2(t_n) \leq \nu_2(n) + \nu_2(k-1) + 2 \leq \log_2(n) + \nu_2(k-1) + 2 \leq \sqrt{n},$$ where the last inequality is true for example for $n \geq 2^{2 \max \{2, \nu_2(k-1)\}}$. By lemma \[lem:term\_size\] $$\log t_n \geq (n-2k-1) \log \phi \geq \frac{1}{2} n \log \phi,$$ for $n \geq 2(2k+1)$, because $\phi > 1$. Hence, we can take $n_0 = \max \{2(2k+1), 2^{2 \max \{2, \nu_2(k-1)\}}\}$ and apply the method used in Theorem \[thm:equation\_solution\] to find an upper bound on $m, n_1, n_2, \ldots, n_d$.
We could apply Corollary \[cor: finite\_solutions\] to find all non-trivial solutions (with $t_{n_i} > 1$ for each $i$) of the equation (\[eq:product\_of\_terms\]) for given $k$ and $d$. However, using the explicit form of $\nu_2(t_n)$, one can make the bounds much more precise. We will once again follow the approach shown in [@Lengyel_Marques]. The computations are quite similar to those in Theorem \[thm:equation\_solution\], so we omit the details.
By Lemma \[lem:legendre\_ineq\] and Theorem \[thm:2-adic\_order\], we obtain $$\begin{aligned}
\label{eq:m_below_n_2}
m - \lfloor \log_2 m \rfloor - 1 &\leq \nu_2(m!) = \sum_{i=1}^{d} \nu_2(t_{n_i}) \leq \sum_{i=1}^{d} [\nu_2(n_i) + \nu_2(k-1) + 2 ] \nonumber \\ &= d[\nu_2(k-1) + 2] + \nu_2 \left( \prod_{i=1}^d n_i \right ) \nonumber \\
&\leq d[\nu_2(k-1) + 2] + \log_2 \left( \prod_{i=1}^d n_i \right ) .\end{aligned}$$ On the other hand, from Lemma \[lem:term\_size\], and inequality (\[eq:m\_fact\_from\_above\]), we get for $m \geq 5$ $$\begin{aligned}
\label{eq:n_below_m_2}
\left [\sum_{i=1}^{d} n_i - d(2k-1) \right ] {\log_2 \phi_k} \leq \sum_{i=1}^{d} \log_2(t_{n_i}) = \log_2(m!) < m (\log_2 m -1),\end{aligned}$$ where $\phi_k$ is the value of $\phi$ in Lemma \[lem:term\_size\] corresponding to $r=2k$. The AM–GM inequality applied to all $n_i$, together with (\[eq:m\_below\_n\_2\]) and (\[eq:n\_below\_m\_2\]), yields $$\begin{aligned}
\label{eq:only_m}
m - \lfloor \log_2 m \rfloor - 1 - d[\nu_2(k-1) + 2] < d \log_2
\left [ \frac{m}{d\log_2 \phi_k} (\log_2 m -1) + (2k-1) \right ].\end{aligned}$$ This gives an upper bound on $m$ and, consequently, on each $n_i$. As an example, in the table below we give the upper bound on $m$ obtained for $2k$-nacci sequences with $2 \leq k \leq 5$ and $1 \leq d \leq 10$.
\[my-label\]
1 2 3 4 5 6 7 8 9 10
--- ---- ---- ---- ---- ---- ---- ---- ---- ---- -----
2 11 19 27 35 43 51 59 67 75 84
3 13 22 31 40 50 59 68 77 87 96
4 11 19 28 36 44 52 60 68 76 84
5 14 25 35 46 56 67 77 88 98 109
Using this result we find that the only non–trivial solution of the equation (\[eq:product\_of\_terms\]) with $2 \leq k \leq 5$ and $1 \leq d \leq 10$ appears in the $4$-nacci sequence and is the single term $t_5 = 3!$.
Proof of Theorem \[thm:2-adic\_order\] {#sec:proof}
======================================
In order to study the 2-adic valuation of $t_n$ it is enough to focus on $n$ divisible by $2k+1$ as all the other terms are odd. In this case we can write $n = s 2^l (2k+1)$ for $s$ odd and $l \geq 0$. We will divide our proof into two main parts.
First, we will show by induction on $l$ and $s$ that $2k$ consecutive terms of the sequence $\{t_n\}_{n \geq 0}$, starting with $t_{s 2^l (2k+1)}$, satisfy a particular system of congruences, given in Lemma \[lem:later\_terms\]. However, as it will turn out, this argument works only for $l \geq l_0$, where $l_0$ depends on $\nu_2(k-1)$. Moreover, the initial system of congruences for $l=l_0$ involves some constants, which need to be computed.
Therefore, we will have to employ another method for $l \leq l_0$. We will show how to obtain the values of $t_{n + 2k+1}, \ldots, t_{n+ 4k}$ in terms of $t_{n}, \ldots, t_{n+ 2k-1}$ for any $n$ and then proceed by induction. As a result, we will be able to express $t_n$ in quite concrete form, given in Lemma \[lem:easy\_expression\], involving binomial coefficients weighted by powers of 2.
For simplicity, we introduce the following matrix notation: $$\begin{aligned}
T_n = \begin{bmatrix}
t_n \\
t_{n+1} \\
\vdots \\
t_{n+2k-1}
\end{bmatrix},
\qquad
B_n =
\begin{bmatrix}
t_n & t_{n+1} & \hdots & t_{n+ 2k-1} \\
t_{n+1} & t_{n+2} & \hdots & t_{n+2k}\\
\vdots& \vdots & &\vdots\\
t_{n+ 2k-1} &t_{n+2k} & \hdots & t_{n+4k-2}
\end{bmatrix},\end{aligned}$$ where $n \geq 0$. By $C$ we will denote the companion matrix of $t_n$ which has the form $$\begin{aligned}
\label{eq:matrix_C}
C=
\begin{bmatrix}
0 & 1 & 0 &\hdots & 0 & 0 \\
\vdots & 0 & 1 & \ddots & & 0\\
& \vdots & \ddots &\ddots & \ddots & \vdots \\
& & & \ddots & 1 & 0 \\
0 & 0 & 0 & \cdots & 0 & 1 \\
1 & 1 & 1 &\hdots & 1 & 1
\end{bmatrix},\end{aligned}$$ where the entries above the diagonal and in the bottom row are equal to 1 and all other entries are zero. It is easy to check that $CT_n = T_{n+1}$ and $CB_n = B_{n+1}$, so for any positive integers $n$ and $w$ we have $$\begin{aligned}
C^n T_w &= T_{n+w}, \label{eq:pushing_the_vector} \\
C^n B_w &= B_{n+w}. \label{eq:pushing_the_matrix}\end{aligned}$$
First, we state an identity involving the terms of the sequence $\{t_n\}_{n \geq 0}$.
\[lem:reduction\_formula\] The matrix $B_0$ is invertible and its determinant is odd. Moreover, for all positive integers $n, w$ we have the formula $$\begin{aligned}
\label{eq:reduction_formula}
t_{n+w} = T_n^T B_0^{-1} T_w.\end{aligned}$$
It is easy to see that $t_n$ is even iff $n$ is divisible by $2k+1$. Therefore, $$\det B_0 \equiv \det
\begin{bmatrix}
0 & 1 &1 & \hdots & & 1 \\
1 & 1 & \hdots &&& 1 \\
1 & \vdots& &&\udots & 0 \\
\vdots&& & \udots & \udots & 1 \\
& &\udots & \udots & \udots & \vdots \\
1 & 1 & 0 & 1 & \cdots & 1
\end{bmatrix} \pmod{2},$$ where zeros in the latter matrix appear only at positions corresponding to $t_0$ and $t_{2k+1}$ in $B_0$, that is, at $(1,1)$ and $(i,j)$ such that $i+j = 2k + 2$. By subtracting the first row from all the others, we easily obtain $\det B_0 \equiv 1 \pmod{2}$, which proves the first part of the statement.
From (\[eq:pushing\_the\_vector\]) and (\[eq:pushing\_the\_matrix\]), we get $$T_{n+w} = C^n T_{w} = B_n B_0^{-1} T_w.$$ The first coordinate gives us the formula for $t_{n+w}$.
The identity (\[eq:reduction\_formula\]) might seem difficult to apply without an explicit expression for $B_0^{-1}$. However, it plays a major role in deriving the congruence relations in the following lemma.
\[lem:later\_terms\] For any $l \geq 0$ the following congruence relation holds: $$\begin{aligned}
\label{eq:easy_divis}
T_{2^l (2k+1)} \equiv T_0 \pmod{2^{l+1}}.\end{aligned}$$ Moreover, if a column vector $A \in \mathbb{Z}^{2k}$ satisfies $$\begin{aligned}
T_{2^{l_0} (2k+1)} \equiv 2^{l_0+1} A + T_0 \pmod{2^{l_0+\nu_2(k-1)+3}},\end{aligned}$$ where $l_0 = \nu_2(k-1)+2$, then for any $l \geq l_0$ and $s \geq 1$ we also have $$\begin{aligned}
\label{eq:the_second_form_2}
T_{s2^l (2k+1)} \equiv s2^{l+1} A + T_0 \pmod{2^{l+\nu_2(k-1)+3}}.\end{aligned}$$
Obviously (\[eq:easy\_divis\]) is true for $l=0$. Now assume that (\[eq:easy\_divis\]) holds for some $l \geq 0$. We can write $$\begin{aligned}
\label{eq:the_second_form}
t_{2^l (2k+1)+ j} = 2^{l+1} a_{l,j} + t_j,\end{aligned}$$ where $a_{l,j}$ are some positive integers for $j=0,1, \ldots, 2k-1$. Define also $a_{l,2k}, a_{l,2k+1}, \ldots,a_{l,4k-2}$ by the same recurrence as $\{t_n\}_{n \geq 0}$. Then (\[eq:the\_second\_form\]) is satisfied for $j=0,1, \ldots, 4k-2$. For convenience denote by $e_j \in \mathbb{Z}^{2k}$ the vector with 1 on the $j$-th position (counting from 0) and 0 on the other positions, and additionally define $$\begin{aligned}
A_{l,j} =
\begin{bmatrix}
a_{l,j} &
a_{l,j+1} &
\cdots &
a_{l,j+2k-1}
\end{bmatrix}^T\end{aligned}$$ for $j=0,1, \ldots, 2k-1$. It follows from the definition of $B_0$ that $B_0^{-1} T_j = e_j$.
Fix any $0 \leq i \leq 2k-1$. The formula (\[eq:reduction\_formula\]) yields $$\begin{aligned}
t_{2^{l+1}(2k+1) + i} = T_{2^{l}(2k+1)}^T B_0^{-1} T_{2^{l}(2k+1) + i}\end{aligned}$$ Therefore, using (\[eq:the\_second\_form\]) we get $$\begin{aligned}
\label{eq:reduction_6}
t_{2^{l+1}(2k+1) + i} &= 2^{2l+2} A_{l,0}^T B_0^{-1} A_{l,i} + 2^{l+1} \left( A_{l,0}^T e_i +e_0^T A_{l,i}
\right ) + T_0^T e_i \nonumber \\
&= 2^{2l+2}c_{l,i} + 2^{l+2} a_{l,i} + t_i\end{aligned}$$ for some rational $c_{l,i}$ such that $\det(B_0) c_{l,i}$ is an integer. In Lemma \[lem:reduction\_formula\], however, we showed that $\det B_0$ is odd which means that $c_{l,i}$ must be an integer. Thus, $$\begin{aligned}
2^{l+2} |t_{2^{l+1}(2k+1) + i} - t_i,\end{aligned}$$ from which (\[eq:easy\_divis\]) follows. If we choose $l \geq l_0 = \nu_2(k-1)+2$ then the term $2^{2l+2} c_{l,i}$ in (\[eq:reduction\_6\]) is reduced modulo $2^{l+\nu_2(k-1)+4}$. We can take $A = A_{l_0,0}$ to complete the proof of (\[eq:the\_second\_form\_2\]) for $s=1$.
To proceed by induction on $s$ notice that the index $s 2^l (2k+1) + i$ can be expressed as a sum of indices in the following way: $$s 2^l (2k+1) + i = (2^l (2k+1) + i) + (s-1) 2^l (2k+1).$$ We can then perform a similar computation as in (\[eq:reduction\_6\]) to get the desired result.
Our specific choice of the divisor in (\[eq:the\_second\_form\_2\]) equal to $2^{l+\nu_2(k-1)+3}$ is based on the observation of $t_n$ and is indeed effective in proving the formula for $\nu_2(t_n)$. The numbers $a_{l,j}$ in (\[eq:the\_second\_form\]) are determined uniquely modulo $2^{\nu_2(k-1)+2}$. We are particularly interested in finding the value of $a_{l_0,0}$ which will directly give us the 2-adic valuation of $t_{2^l(2k+1)}$ for $l \geq l_0 = \nu_2(k-1)+2$, provided that $a_{l_0,0} \leq \nu_2(k-1)+1$. However, as mentioned before, we need to develop another method to analyze the case when $l \leq l_0$.
We start with deriving a formula for expressing $t_{n + 2k+1}, \ldots, t_{n+ 4k}$ in terms of $t_{n}, \ldots, t_{n+ 2k-1}$.
\[lem:pushing\_the\_sequence\] Define $C$ as in (\[eq:matrix\_C\]). Then $$C^{2k+1} = 2
\begin{bmatrix}
1 & 1 & \hdots & 1 \\
2 & 2 & \hdots & 2 \\
\vdots & \vdots & & \vdots \\
2^{2k-1} & 2^{2k-1} & \hdots & 2^{2k-1}
\end{bmatrix}
-
\begin{bmatrix}
1 & 0 & \hdots & 0 & 0 \\
2 & 1 & \ddots & & 0 \\
\vdots & \vdots & \ddots & \ddots & \vdots \\
2^{2k-2} & 2^{2k-3}& \hdots & 1& 0 \\
2^{2k-1} & 2^{2k-2} & \hdots & 2 & 1
\end{bmatrix}.$$
Using the identity $t_{n+2k+1} = 2t_{n+2k} - t_{n}$, one can show by induction that for any $0 \leq i \leq 2k-1$ $$\begin{aligned}
\label{eq:matrix_rows}
t_{n +2k+1+i} = 2^{i+1} t_{n+ 2k} - \sum_{j=0}^{i} 2^{i-j} t_{n+j} = 2 \cdot 2^i \sum_{j=0}^{2k-1} t_{n+j}- \sum_{j=0}^{i} 2^{i-j} t_{n+j}.\end{aligned}$$ We know that $D = C^{2k+1}$ is the only matrix satisfying $T_{n+2k+1} = DT_n$ for all $n \geq 0$. Thus, for each $0 \leq i \leq 2k-1$ the coefficients at $t_{n+j}$ in (\[eq:matrix\_rows\]) correspond to the $i$-th row of $C^{2k+1}$ (counting from 0).
We are also going to need two standard identities involving binomial coefficients. For the convenience of the reader we include the proof.
\[binomial\_formulas\] For all positive integers $m, w$ we have
[(a)]{} $\sum\limits_{i=0}^{w} \binom{m+i}{m} = \binom{m+w+1}{m+1}$,
[(b)]{} $\sum\limits_{i=0}^{w} \binom{m+i}{m} 2^i = (-1)^{m+1} + 2^{w+1} \sum\limits_{j=0}^{m} \binom{m+w+1}{m-j} (-2)^j.$
For any fixed $m \geq 0$ the formula (a) follows easily from induction on $w$.
To prove (b) take any $w,m \geq 0$ and consider the function $$f(x) = \frac{1}{m!}\sum\limits_{i=0}^{w} x^{i+m} = \frac{1}{m!} \; \frac{x^{m+w+1}-x^m}{x-1}$$ for $x \neq 1$. It is easy to see that the left side of (b) is equal to $f^{(m)}(x)$ evaluated at $x=2$. Applying the Leibniz formula we get $$\begin{aligned}
f^{(m)}(x) &= \frac{1}{m!} \sum_{j=0}^{m} \binom{m}{j} \left(\frac{1}{x-1}\right)^{(j)} \left(x^{m+w+1}-x^m \right)^{(m-j)} \\
&=\sum_{j=0}^{m} \frac{(-x)^{j}}{(x-1)^{j+1}} \left [ \binom{m+w+1}{m-j} x^{w+1} - a\binom{m}{j} \right ] \\
&= x^{w+1} \sum_{j=0}^{m} \binom{m+w+1}{n-j} \frac{(-x)^{j}}{(x-1)^{j+1}} +\left( \frac{-1}{x-1} \right)^{m+1}.\end{aligned}$$ Substituting $x=2$ we get the desired result.
The following lemma gives us an easily computable expression for $2k$ subsequent terms $t_n$.
\[lem:easy\_expression\] Define the following column vectors in $\mathbb{N}^{2k}$: $$\begin{aligned}
w = \begin{bmatrix}
1 & 1 & \hdots & 1
\end{bmatrix}^T,
\quad
v_m = \begin{bmatrix}
\binom{m}{m} & \binom{m+1}{m} \cdot 2^1 & \binom{m+2}{m} \cdot 2^2 &
\hdots & \binom{m+2k-1}{m} 2^{2k-1}
\end{bmatrix}^T\end{aligned}$$ for $m \geq 0$. Then for any $m \geq 1$ we have $$\begin{aligned}
\label{eq:whole_push}
T_{m(2k+1)} \equiv w + (-1)^{m+1} \cdot 4(k-1) \sum_{i=0}^{m-1}v_i + (-1)^{m+1} v_{m-1} \pmod{2^{2k + 1}}.\end{aligned}$$
Using the form of $C^{2k+1}$ given in Lemma \[lem:pushing\_the\_sequence\], it is easy to see that $$\begin{aligned}
\label{eq:first_push}
C^{2k+1} w = 2 \cdot 2k v_0 - (2v_0 - w) = w + 2(2k-1)v_0.\end{aligned}$$ Now fix $m \geq 0$. Applying Lemma \[binomial\_formulas\] to each coordinate gives us
$$\begin{aligned}
\label{eq:later_push}
C^{2k+1} v_m \equiv 2\cdot(-1)^{m+1} v_0 - v_{m+1} \pmod{2^{2k+1}}.\end{aligned}$$
One can check that $$C^{2k+1} T_0 = w + 2(2k-1)v_0 - v_0 = w + 4(k-1)v_0 + v_0,$$ so (\[eq:whole\_push\]) is true for $m=1$. Now assume that (\[eq:whole\_push\]) holds for some $m \geq 1$. Using (\[eq:first\_push\]) and (\[eq:later\_push\]) we get $$\begin{aligned}
T_{(m+1)(2k+1)} &= C^{2k+1} T_{m(2k+1)} \\
&\equiv C^{2k+1} \left[w + (-1)^{m+1} \cdot 4(k-1) \sum_{i=0}^{m-1}v_i + (-1)^{m+1} v_{m-1}\right] \\
&\equiv w + 2(2k-1)v_0 + (-1)^{m+1} \cdot 4(k-1) \sum_{i=0}^{m-1} \left[2 \cdot (-1)^{i+1} v_0 - v_{i+1}\right] \\
&\qquad +(-1)^{m+1} \left[ 2 \cdot (-1)^{m} v_0 - v_{m}\right] \\
&\equiv w + \left[4(k-1) - 8(k-1) \epsilon_m\right]v_0 + (-1)^{m} \cdot 4(k-1) \sum_{i=1}^{m} v_{i} \\
&\qquad + (-1)^{m} v_m \pmod{2^{2k+1}},\end{aligned}$$ where $\epsilon_m$ is equal to $m$ modulo 2. Thus, the coefficient at $v_0$ is equal to $(-1)^m \cdot 4(k-1)$, so we can incorporate it into the sum. Finally, we obtain $$T_{(m+1)(2k+1)}
\equiv
w + (-1)^{m} \cdot 4(k-1) \sum_{i=0}^{m} v_{i} + (-1)^{m} v_m \pmod{2^{2k+1}}.$$
We are now ready to prove Theorem \[thm:2-adic\_order\].
The term $t_n$ is even iff $n$ is divisible by $2k+1$, which proves that $\nu_2(t_n) = 0$ for $n \equiv 1,2,...,2k \pmod{2k+1}$. Observe that if $k \geq 2$, then $2 \nu_2(k-1) + 5 \leq 2k+1$, so by Lemma \[lem:pushing\_the\_sequence\] for $m \geq 1$ we have $$T_{m(2k+1)} \equiv
w + (-1)^{m+1} \cdot 4(k-1) \sum_{i=0}^{m-1}v_i + (-1)^{m+1} v_{m-1} \pmod{2^{2 \nu_2(k-1) + 5}}.$$ Looking at the first entry of this vector, we obtain $$\begin{aligned}
\label{eq:super_formula}
t_{m(2k+1)} \equiv 1 + (-1)^{m+1} \cdot 4m(k-1) + (-1)^{m+1} \pmod{2^{2 \nu_2(k-1) + 5}}.\end{aligned}$$ Thus, for odd $m$ we get $t_{m(2k+1)} \equiv 2 \pmod{4}$, hence $\nu_2(t_n) = 1$ for $n \equiv 2k+1 \pmod{2(2k+1)}$.
Now let $n = s(2k+1) 2^l$ for odd $s$ and $l \geq 1$, so that $n \equiv 0 \pmod{2(2k+1)}$. We will further split the third case into two subcases, depending whether $l \leq l_0 = \nu_2(k-1) +2$ or $l > l_0$. If $l \leq l_0$ then from (\[eq:super\_formula\]) we obtain $$\begin{aligned}
\label{eq:subcase_1}
t_n \equiv s2^{l+2} (k-1) \pmod{2^{2 \nu_2(k-1) + 5}},\end{aligned}$$ so $\nu_2(t_n) = l+ 2 + \nu_2(k-1) = \nu_2(n) + \nu_2(k-1) + 2$.
We cannot extend the same argument to $l > l_0$ because we only know the congruence modulo $2^{2 \nu_2(k-1) + 5}$. However, substituting $s=1$ and $l = l_0$ in (\[eq:subcase\_1\]) gives us a possible value $a_{l_0,0} = 2(k-1)$, as defined in Lemma \[lem:later\_terms\]. Using Lemma \[lem:later\_terms\] for any $l > l_0$, we get in the first coordinate $$t_n \equiv 2^{l+1} a_{l_0,0} \equiv 2^{l+2} (k-1) \pmod{2^{l + \nu_2(k-1) + 3}},$$ which again yields $\nu_2(t_n) = \nu_2(n) + \nu_2(k-1) + 2$.
[1]{} J.-P. Allouche and J. Shallit, The ring of $k$-regular sequences, *Theor. Comput. Sci.* [**98**]{}: 163–197 (1992). J.-P. Allouche and J. Shallit, The ring of $k$-regular sequences II, *Theor. Comput. Sci.* [**307**]{}: 3–29 (2003). T. Lengyel, The order of Fibonacci and Lucas numbers, *Fibonacci Quart.* [**33**]{}: 234–239 (1995). T. Lengyel, Divisibility properties by multisection, *Fibonacci Quart.* [**41**]{}: 72–79 (2003). T. Lengyel and D. Marques, The 2-adic order of the Tribonacci Numbers and the equation $T_n = m!$, *J. Integer Seq.* [**17**]{}: Article 14.10.1 (2014). D. Marques, The order of appearance of product of consecutive Fibonacci numbers, *Fibonacci Quart.* [**50**]{}: 132–139 (2012). Z. Shu and J.-Y. Yao, Analytic functions over $\mathbb{Z}_p$ and $p$-regular sequences, *C. R. Math.* [**349**]{}: 947–952 (2011). D. A. Wolfram, Solving generalized Fibonacci recurrences, *Fibonacci Quart.* [**36**]{}: 129–145 (1998)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A discussion is presented on the crossover of the susceptibility from mean-field to Ising critical behavior upon approaching the critical point from below and from above $T_c$, both for $^3$He and Xe. Fits of the experimental susceptibility data are made to curves from Monte Carlo simulations, and the corresponding Ginzburg numbers $G_i$ for each measured property are deduced. Also the first correction amplitudes for the confluent singularities are obtained from the fit of the data. The respective ratios of these numbers and those obtained for the coexistence curves for $^3$He and Xe, presented elsewhere, are discussed in terms of predictions.'
address: 'Department of Physics, Duke University, Durham, NC 27708-0305\'
author:
- Horst Meyer
title: 'Susceptibility Crossover Behavior in $^3$He and Xe Near Their Liquid-Vapor Critical Point - A Progress Report[^1]'
---
Introduction
============
The interest in crossover phenomena from the asymptotic to mean-field critical behavior in fluids has been described in detail in a recent review article by Anisimov and Sengers [@Anisimov:S:00], which lists many references and where different theoretical approaches and also a comparison with some experiments are presented. The subject of a recent paper[@Luijten:M:00] was a comparison between predictions for the crossover from Monte Carlo calculations[@Luijten:B:98] and experimental data of simple fluids. The susceptibility $\chi^+$ above $T_c$ (or compressibility) and the liquid and vapor densities along the coexistence curve (CXC) for Xe and $^3$He were studied. From a fit of the data to the predicted curves, the corresponding Ginzburg numbers G could be estimated. For the CXC, the exit of the fluid from the critical regime into a background behavior could be clearly seen by a systematic departure from the predicted curve, well before the regime of mean-field critical behavior could be reached. For $\chi^+$, the behavior of Xe, a “classical fluid", agreed well with predictions, but there were systematic differences for $^3$He, and a qualitative discussion was made in terms of the interplay between quantum and critical fluctuations for this fluid.
The purpose of this progress report is to extend the same analysis to the susceptibility $\chi^-$ data below $T_c$ for Xe and $^3$He, and also to give a status report on this program. After a background review, the susceptibility data of several experimental groups, namely along the critical isochore ($T>T_c$) and along the liquid and the vapor side of the CXC ($T<T_c$), are discussed. Comparison is made with curves from Monte Carlo calculations [@Luijten:B:98], and the corresponding Ginzburg numbers G($\chi^-$) are estimated. The internal consistency for the Ginzburg numbers so obtained is checked by determining $G$ from the fit of data to a 2-term series expansion representing part of the curve obtained from the MC calculations. From the collection of Ginzburg numbers obtained so far \[$G(\chi^+)$,$G(\chi^-)$ and $G(CXC)$\], their ratios are discussed in the light of predictions. In spite of the uncertainties in the G($\chi^-$) below $T_c$ due sparsity of data and experimental scatter, and also in the G($\chi^+$) for $^3$He, some preliminary conclusions can be made. This progress report is to draw attention to the interest of such results, to their present incomplete understanding and to the great need of better data.
A short review
==============
Properties considered
---------------------
We now list the properties discussed in this paper, and introduce the definitions of reduced temperature and density, $t \equiv (T-T_c)/T_c$ and $\Delta \rho \equiv (\rho - \rho_c)/\rho_c$. The coexistence curve is expressed by $$\label{coex}%-----------------------
\Delta \rho_{LV} =(\rho_{\rm liq}- \rho_{\rm vap})/\rho_c = B_0
(-t)^{\beta}[1 + B_1(-t)^{\Delta_1} +B_2(-t)^{\Delta_2} ...]$$where $\rho_{\rm liq} , \rho_{\rm vap}$ and $\rho_c $ are the densities in the coexisting liquid and vapor phases, and at the critical point. Furthermore $\beta$ = 0.326 is the critical exponent and the bracket includes the correction-to-scaling confluent singularity terms. Here the $B_i$’s are amplitudes characteristic of the fluid, and $\Delta_1$ = 0.52, $\Delta_2$ = 1.04 are the exponents obtained by Wegner [@Wegner:72] and by Newman and Riedel[@Newman:R:84].
The susceptibility $\chi$ of the fluid, namely the analog of the susceptibility of a magnet, is given by $\chi \equiv (\partial
\rho/\partial
\mu)_T =
\rho^2\beta_T$, where $\beta_T$ is the isothermal compressibility and $\mu$ is the chemical potential. As discussed by Sengers and Levelt Sengers[@Sengers:LS:78], the 3-D lattice-gas model (which corresponds to the 3-D Ising model in magnets) has properties that adequately describe real fluids. One particular aspect is that of symmetry in the $\mu-\Delta
\rho$ plane. (In this respect, $^3$He is the fluid that best conforms to this model. See Appendix) As a consequence, the derivative $\chi$ is a symmetric function of $\Delta \rho$ along an isotherm. Hence below $T_c$, one obtains $\chi_{Liq} = \chi_{Vap}$, where the susceptibilities are measured on both sides of the coexistence curve, and therefore we expect consistency between the data on both liquid and the vapor sides. Here we introduce the reduced quantity $\chi^* \equiv
\chi(P_c/\rho_c^2)$, where the critical parameters have been listed in ref.[@Luijten:M:00]. Above $T_c$ and along the critical isochore, $\chi^* = \beta_T P_c$.
Similarly to Eq.1, the expansion for the susceptibility $\chi^{*(+,-)}$ from the asymptotic critical regime is given by $$\label{Wegner}%-----------------------
\chi^{*(+,-)} = \Gamma^{(+,-)}_0|t|^{-\gamma}[1 + \Gamma^{(+,-)}_1
|t|^{\Delta_1} + \Gamma^{(+,-)}_2 |t|^{\Delta_2} +.......]$$where the indices + and - indicate the region $t>0$ along the critical isochore and $t<0$ along the coexistence curve, respectively. Here again the $\Gamma_i$’s are amplitudes characteristic of the fluid and $\gamma$ = 1.24 is the critical exponent. The ratio $\Gamma_0^+/\Gamma_0^-$ for a given fluid has been calculated from series expansion by Liu and Fisher to be[@Liu:F:89] $$\label{liu/Fisher}%-----------------------
\Gamma^+_0 / \Gamma^-_0 = 4.95 \pm 0.15$$This compares with the value of 4.82 obtained from the ratio $(\gamma/\beta)[(1-2\beta)\gamma/2\beta(\gamma-1)]^{\gamma-1}$ predicted from the parametric representation of the equation of state[@Schofield:L:H:69], where $\gamma$ = 1.24 and $\beta$ = 0.327 were used. The most recent values of this ratio (see [@Bervillier]) are very close to 4.77. Predictions for the ratio of the amplitudes $B_1$, $\Gamma^+_1$ and $\Gamma^-_1$ will be presented below.
Ginzburg numbers and amplitude ratios
-------------------------------------
The Ginzburg criterion and Ginzburg number have been discussed in detail in the article by Anisimov, Kiselev, Sengers and Tang[@Anisimov:K:S:T:92] on the crossover approach to global critical phenomena in fluids. The Ginzburg number $G$ is seen as a dimensionless temperature, obtained from the criterion $t \gg G$ which gives an estimate for the range of $t$, where the classical critical theory is valid, this is where the fluctuation contribution is small. For fluids, an order of magnitude estimate[@Anisimov:K:S:T:92] of $G$ leads to $\approx 10^{-2}$, and furthermore for a 3-D fluid, one finds $G \propto R^{-6}$, where $R$ is the normalized molecular interaction range [@Anisimov:K:S:T:92]. Hence the asymptotic critical behavior takes place for $t \ll G$ while the classical critical behavior is expected for $1 \gg t
\gg G$. However as pointed out in ref.[@Anisimov:S:00], in ordinary fluids the crossover is never completed in the critical domain ($t
\ll 1$) since $R$ is of the same order as the distance between molecules. The Monte Carlo algorithm developed by Luijten and Bloete[@Luijten:B:95] allows the full crossover region in 3-D Ising models to be covered. The calculation then gives a curve of a given singular property $f_i$ = $f_i( |t|/G_i)$ covering $\approx$ 8 or more decades in $|t|/G_i$, and is clearly more complete than the expansion series expressed in correction-to scaling terms of series. A simple check for the internal consistency in determining G can be made by the expected expansion in terms of the corrections-to-scaling confluent singularities as $$\label{Wegner}%-----------------------
f_i = A_{0,i}|t|^{-\lambda_i}[1 + A_{1,i}(|t|/G_i)^{\Delta_1} +.......]$$where the amplitude $A_{0,i}$ is non-universal but where the numerical coefficients $A_{1,i}$, and the exponent $\lambda_i$ are universal for all fluids and characteristic of the property (susceptibility, CXC etc...). A fit of the curves $f_i$ = $f_i(
|t|/G_i)$, calculated by the Monte Carlo approach[@Luijten:B:98], to Eq. 4 (restricted to the region of $|t|/G_i$ where higher terms are negligible) gives $A_1(\chi^+)$ = 0.10, $A_1(\chi^-)$ = 0.65, and $A_1(CXC)$ = 0.23.
Of particular interest here is the calculation of the susceptibility $\chi$, both below and above $T_c$, represented in a very sensitive way by the plot [@Luijten:B:98] of the effective exponent $\gamma_{eff}$, which is given by the derivative $$\label{differential}%-----------------------
\gamma_{\rm eff} \equiv -{{d\ln \chi_T^* }
\over {d\ln |t|}}.$$In Fig.1, the plots of $\gamma^{(+,-)}_{eff}$ versus $ |t|/G^{(+,-)}(\chi)$ for $t>0$ and $t<0$, as obtained from Monte Carlo calculations, are presented side-by-side for comparison. The various symbols denote the successively larger values of the interaction range $R$ that were used to generate the master curve as the distance from $T_c$ is increased. The lines labeled “BK, BB and SF" for $t>0$ and “App" for $t<0$ are theoretical curves described in ref.[@Luijten:B:98]. The slope of the exponent,$-\partial\gamma_{\rm
eff}/\partial\ln|t|$, gives information on the crossover width between the asymptotic Ising value of $\gamma = 1.24$ and the mean-field one $\gamma = 1$. This width is shown to be much narrower for the region $t<0$ than for $t>0$, as was pointed out in ref.[@Luijten:B:98]. We shall see that this difference in crossover width is reflected in the data for $^3$He.
\[fig:fig1\]
The relations between the G’s and the first Wegner terms in the correction to scaling follow from Eqs, 1, 2 and 4. For instance in the case of the susceptibility above $T_c$, one has $$\label{ginzburgWegner}%-----------------------
\Gamma_1^+ = A_1 (\chi^+)[G(\chi^+)]^{-\Delta_1}$$with $\Delta_1$ = 0.5, which will be used in the discussion of the data analysis.
\[fig:fig2\]
Aharony and Ahlers[@Aharoni:A:80] have discussed the ratios of the amplitudes of the “correction-to-scaling confluent singularity" terms in expressions such as Eqs. 1 and 2 for different properties, and in particular for the order parameter (here CXC) and the susceptibility above $T_c$ versus $|t|$. They expressed thermodynamic quantities with singularities at the critical point as $$\label{Aharony}%-----------------------
f_i = A_{0,i}|t|^{-\lambda_i}[1
+ a_i |t|^{\Delta_1} + O|t|^{2\Delta_1}]$$ where $\lambda_i$ is the (asymptotic) critical exponent of the property $i$, such as susceptibility, specific heat, order parameter etc.. and the $a_i$’s are the amplitudes of the first correction-to-scaling term of the confluent singularity (already introduced here as $\Gamma_1$ and $B_1$). Among the relations they derived, one which considers the ratio of the correction term amplitudes for two properties $i$ and $j$ of a fluid is of particular interest to us, namely $$\label{effectiveexponent}%-----------------------
(\lambda_{i, eff} - \lambda_i)/(\lambda_{j, eff} - \lambda_j) = a_i/a_j$$Here $\lambda_{i,{\rm eff}}$ is the effective exponent and $\lambda_{i\, {\rm or}\, j}$ is the asymptotic exponent with $\lambda_i = \gamma = 1.24$ and $-\lambda_j = \beta = 0.326$. The numerical values for $\lambda_{i,{\rm eff}}$ are obtained by fitting experimental data to a simple power law over the same range of $|t|$ where the fit to Eq. 7 has been made. A prediction of the ratio of Ginzburg numbers via Eqs.4 and 7 can therefore be made and compared with that from experiments. One has $$\label{Ginzburgratios}%-----------------------
( G_j/G_i) = [a_iA_1(j)/a_jA_1(i)]^{1/\Delta_1}.$$Here $A_1(j)/A_1(i)$ is the ratio of the numerical coefficients in Eq.4 for the properties $j$ and $i$, listed after Eq.4.
Bagnuls, Bervillier, Meiron and Nickel[@Bagnuls:B:M:N:87] have calculated the ratios $a_i/a_j$ using “massive field theory" for the $\Phi^4$ model in 3-D for the n=1 class. These ratios are universal and are found to be $a(\chi^+)/a(\chi^-)$ = $\Gamma_1^+/\Gamma^-_1$ = 0.315 $\pm$ 0.013 and $a(\chi^+)/a(CXC)$ = $\Gamma^+/B_1$ = 0.9$\pm$ 0.2. (See Tables VIII and IX of ref[@Bagnuls:B:M:N:87]). This implies that the ratio of the Ginzburg numbers is universal too. From Eq.9 one then obtains G($\chi^+$)/G($\chi^-$) = 0.23$\pm$0.01 and G($\chi^+$)/G(CXC) = 0.15 $\pm$0.07.
Experimental data and determination of $T_c$
============================================
In the experiments, $\chi$ has been determined either from the intensity of light scattering (Xe)[@Guettinger:C:81; @Smith:G:B:71] or from the measurements of the density versus pressure along isotherms in Xe[@Michels:W:L:54], and in $^3$He[@Wallace:M:70; @Chase:Z:76; @Agosta:W:C:M:87; @Hahn:B:Z:00] and from the vertical density gradient in the gravity field for $^3$He[@Pittman:D:M:79] above $T_c$. Here we briefly comment on the various measurements and their respective scatter.
As will be seen below, the $\chi$ data for $t>0$ have an appreciably higher accuracy than those below $T_c$. In the light scattering measurements, this might possibly be due to the added difficulty of sending the laser beam alternatively into the superposed liquid and vapor phases of a cell with a small height, and where the meniscus will become concave as $T_c$ is approached, because of the decreasing surface tension. This contrasts with measurements above $T_c$ where the beam is positioned near mid-height of the cell, and where the maximum light scattering intensity is observed by slowly scanning the vertical position. In the present analysis, the data by Smith et al[@Smith:G:B:71] above $T_c$ have not been used, because their scatter is larger than that of the more recent data by Güttinger and Cannell[@Guettinger:C:81]. However we note that in ref[@Smith:G:B:71] the amplitude ratio is $\Gamma^+_0/\Gamma^-_0$ = 4.1, which is not far from the predictions. If the data of ref[@Guettinger:C:81] above $T_c$ are combined with those below $T_c$[@Smith:G:B:71], a ratio of 5.8$\pm$ 0.4 is obtained by fitting both sets of data to Eq. 2. The amplitudes are listed in Table I.
The maximum value of $\chi^+$ at $\rho_c$ from isotherm data above $T_c$ is obtained with a higher precision than is the extrapolation of $\chi^-$ to the coexistence curve below $T_c$. As mentioned above, it is expected from the Ising model that $\chi_{vap} = \chi_{liq}$. Yet there can be appreciable scatter in both determinations which reflects the uncertainties both in the differentiations of the $\rho(P)$ data sets and also in the factor $\rho^2_{Vap}$, or $\rho^2_{Liq}$. The results are presented as reduced quantities $\chi*
\equiv \chi P_c/\rho_c^2$, where the critical parameters have been listed in ref.[@Luijten:M:00]
In the experiments with Xe where optical methods were used, the determination of the critical temperature has been achieved for the $\chi$ measurements by observing the disappearance of the meniscus, and by a fit to Eq.2 [@Guettinger:C:81]. The coexistence curve data were fitted to Eq.1 [@Naerger:B:90]. On the average these determinations were made with an uncertainty of $\delta t\approx \pm 5
\times 10^{-6}$.
By contrast in the experiments with $^3$He without optical access, the uncertainty in $T_c$ is more important. It is probably smallest in the experiments by Pittman [@Pittman:D:M:79] where $T_c$ was determined principally from measurements below $T_c$ (coexistence curve), and where $T_c$ was obtained from a fit of Eq.1, as described in that paper. Here the claimed uncertainty is $\delta T_c/T_c \approx \pm 9\times 10^{-6}$. In the older measurements of $\chi$ from isotherms by Wallace and Meyer[@Wallace:M:70], the choice of $T_c$ was obtained by extrapolation of both coexistence curve data and compressibility data, and based on a simple power law with effective exponents. The uncertainty was claimed to be $\delta T_c/T_c \approx
\pm 6 \times 10^{-5}$. The use of this power law led to a systematic error in $T_c$ which was evidenced by deviations from the compressibility data of ref.[@Pittman:D:M:79], as shown in Fig.1 of [@Pittman:D:M:79]. In the $\chi$ measurements by Chase and Zimmerman[@Chase:Z:76], also from isotherms, the determination of $T_c$ was done in a similar way as in ref[@Wallace:M:70]. In the most recent measurements of $\chi$ from isotherms by the MISTE team at JPL[@Hahn:B:Z:00], the value of $T_c$ was determined from a fit of the $\chi$ data above $T_c$ to Eq. 6 with a systematic uncertainty of $\delta
T_c/T_c
\approx
\pm 1.5 \times10^{-5}$ in $T_c$ (F. Zhong, private communication).
Data presentation and Monte Carlo calculations
==============================================
In Fig.2 the susceptibility data for $^3$He from refs [@Wallace:M:70; @Pittman:D:M:79; @Hahn:B:Z:00; @Agosta:W:C:M:87] are presented, both above and below $T_c$, scaled by the leading singularity $|t|^{-\gamma}$. The data of ref[@Chase:Z:76] lie systematically up to 20% below the other data sets and have not been included in the plot to avoid overcrowding the figure. The data of ref[@Pittman:D:M:79] can be made to agree well with those of ref.[@Hahn:B:Z:00], if their respective values of $T_c$ are slightly shifted well within the stated uncertainty mentioned above. The shifts are as follows: $\delta T_c/T_c = + 6 \times 10^{-6}$ for data of ref.[@Hahn:B:Z:00] and $\delta T_c/T_c = - 6 \times
10^{-6}$ for data of ref.[@Pittman:D:M:79]. By combining the two sets of data with the mutually shifted $T_c$, a fit to Eq. 2 gives $\Gamma^+_0$ = 0.145. For a given set of experiments,where data above and below $T_c$ were obtained, the same choice of $T_c$ was implemented. The error in $T_c$ in the experiments of ref[@Wallace:M:70] was corrected by an appropriate shift $\delta T_c$ withing the stated uncertainty, which resulted in the data above $T_c$ to lie uniformly $\approx$ 5% below those of refs.[@Pittman:D:M:79] and [@Hahn:B:Z:00].
Table I lists the amplitudes of the leading terms $\Gamma^+_0$, $\Gamma^-_0$ and $B_0$, and of the first correction terms $\Gamma^+_1$, $\Gamma^-_1$ and $B_1$ obtained by a fit of the data to Eq.1 resp. Eq. 2. The errors listed are all systematic, not statistical. The corresponding sources of data are listed in the last column. The data fits for $\Gamma^+_1$ and $\Gamma^-_1$ in $^3$He and for $\Gamma^-_1$ in Xe (with very scant data and appreciable scatter close to $T_c$) were made by setting the higher terms in Eq. 1 to zero. To this purpose, the fitting was restricted to the range $|t| < 2\times 10^{-2}$, where presumably the higher terms in Eqs. 1, 2 (and therefore also in Eq. 4) can be neglected. Because of the strong correlation between the amplitudes in the data fitting procedure to Eqs 1 and 2, an uncertainty of only say O($\pm $5%) in $\Gamma_0$ can produce a much larger one of O($\pm$50%) in $\Gamma_1$. For the CXC of $^3$He and Xe, and for $\Gamma^+_1$ of Xe, the amplitudes listed in refs[@Guettinger:C:81; @Pittman:D:M:79; @Naerger:B:90] were used.
The $^3$He data by Pittman , which result from measuring the density difference between two superposed sensors, show a smaller scatter close to $T_c$ than those ref.[@Hahn:B:Z:00], but they are restricted to the range $t> 5\times 10^{-4}$, below which the vertical density profile becomes sharply non-linear as stratification from gravity increases. Above $t> 5\times 10^{-2}$, where $\chi$ has become small, this method is no longer sensitive, as shown by the rapidly increasing scatter. One notes that for the $^3$He $\chi^-$ data, the leading amplitude $\Gamma^-_0$ is consistent well within the large scatter with the expected $\Gamma^-_0 = \Gamma^+_0$/4.95 = 0.029 where the factor 4.95 was given in Eq. 3. Here we have taken $\Gamma^+_0$ = 0.145, the value from the combined set of data from refs. [@Pittman:D:M:79] and [@Hahn:B:Z:00]. In spite of the data scatter below $T_c$, the difference in the change of $\chi\,|t|^{1.24}$ with $|t|$ in the regime above and below $T_c$ is quite striking : $\mathcal{R}$(t=0.1)/$\mathcal{R}$(t$\rightarrow 0)$ = 1.25 compared to $\mathcal{R}$(-t=0.1)/$\mathcal{R}$(t$\rightarrow 0)$ = 1.93 , where $\mathcal{R}$(t) $\equiv
\chi|t|^{1.24}$. This is consistent with the finding that $\gamma_{eff}(t<0)$ is smaller than $\gamma_{eff}(t>0)$ over the common experimental range $10^{-3}<|t|< 10^{-1}$. As was mentioned before, MC calculations predict that the crossover width is narrower for $t<0$ than for $t>0$.
In Figs. 3 and 4, plots of the scaled reduced and normalized susceptibility $ \bar \chi\,
|t|^{1.24}$ versus $|t|/G(\chi^{(+,-)})$ are shown. Here $\bar \chi = \chi^* / \Gamma^{(+,-)}_0$, as in ref.[@Luijten:M:00]. In each figure the subscripts indicate the phase (vapor or liquid) along the coexistence curve. In the limit $|t| \rightarrow 0$ the ratio $\bar\chi\, |t|^{\gamma}$ becomes unity. However the MC calculations for $t<0$ have more scatter than for $t>0$ as seen from Fig.1. In spite of this, their trajectory for $[(-t/G] < 1\times 10^{-2}$ can be estimated quite well, since with decreasing $|t|$ the curves will follow Eq.4, and tend to unity. The values of the $\Gamma$’s and $G(\chi^+)$ obtained via Eq. 6 and similarly of $G(\chi^-)$ and $G(CXC)$ for both fluids are shown in Table 1.
In Fig.3 both the susceptibilities[@Guettinger:C:81; @Michels:W:L:54; @Smith:G:B:71] above and below $T_c$ for Xe are presented, - the first one already shown in ref[@Luijten:M:00]. Below $T_c$, there are few data points, and the scatter prevents a precise determination of $\Gamma^-_0$ and therefore the resulting value of G($\chi^-)$ is much more uncertain than that of G$(\chi^+)$. It should be mentioned, as was done in ref.[@Luijten:M:00], that the fit for $t>0$ was made taking $\Gamma_0$ = 0.0594 instead of 0.0577 obtained in ref.[@Guettinger:C:81]. This choice of $\Gamma_0$, determined by the adopted value of $\gamma$ = 1.240, is no doubt responsible for the different values of G($\chi^+$) obtained from the data fit to Eq.2 and to the MC curve, respectively 0.006 amd 0.018, as reported in ref[@Luijten:M:00].
Fig.4 shows the plots for $^3$He of refs [@Pittman:D:M:79; @Hahn:B:Z:00; @Wallace:M:70; @Agosta:W:C:M:87] of Fig.2, with $\Gamma^+_0$ = 0.145. For $\chi^-$, the value $\Gamma^-_0$ = 0.029 mentioned above was used, which “anchors" the data presentation $\bar \chi\,t^{1.24}$ in the asymptotic regime. The fit according to Eq.2 is restricted to $|t|< 3 \times 10^{-2}$. The top plot is different from that of ref[@Luijten:M:00] as it combines the data of refs[@Hahn:B:Z:00] and [@Pittman:D:M:79] as has been described above. The new value of $\Gamma^+_0$ = 0.145 (instead of 0.139) then leads to a larger value of G($\chi^+$), listed in Table 2. \[By accident, in the bottom plot for $t<0$, the symbols WM(vapor) and JPL (liquid) on one hand, and WM(liquid) and JPL(vapor) on the other, are undistinguishable\]. It is clear that the restricted data range in $|t|/$G$(\chi^-)$ for $^3$He below $T_c$ does not enable confirming the variation of the quantum fluctuations in the crossover region, proposed for $t>0$[@Luijten:M:00], where the data extend beyond $t=10^{-1}$.
\[fig:fig3\]
\[fig:fig4\]
Discussion
==========
In Table 1, the values of the first correction term amplitudes, obtained from a fit of the experimental data to Eqs.1 and 2 are obtained and their ratios are compared with predictions [@Bagnuls:B:M:N:87]. The large (systematic) error bars reflect the data scatter and fit quality. A fair consistency within the large uncertainties is obtained, indicating that the result from the data analysis appears consistent with the universality prediction based on the $\Phi^4$ model.
In Tables 1 and 2, the Ginzburg numbers and relevant ratios are listed, and the results are now briefly discussed. Here again the error bars are guesses based on the fitting uncertainties, since a satisfactory error calculation could not be done. In Table 2, the G’s are those obtained by a fit of the data to the curve obtained from Monte Carlo calculations. Obviously one of the great merits of MC calculations is to give a much wider range of $|t|/G$ where the data can be fit to predictions than can a 2-term expansion such as Eqs 1 and 2. At the same time, it is instructive to compare the resulting G’s obtained from both methods.
\[fig:tab2\]
\[fig:tab1\]
This is done by comparing the numbers in the last column of Table 1 with those on the first three columns in Table 2. On the whole, there is acceptable consistency between the determination from both methods, with the exception for Xe. This might be caused by the different choices of $\Gamma_0$ for the the $\chi^+$ data, as mentioned before.
We note that the Ginzburg numbers have uncertainties that reflect the degree of difficulty in fitting the data to the curve obtained from Monte Carlo calculations. Yet, even with this caveat, certain tentative conclusions can be reached. First, the order of magnitude of the $G_i$ is as expected[@Anisimov:S:00], namely $O(10^{-2})$. Second, the ratios $G(\chi^+)/G(\chi^-)$ are roughly the same for Xe and $^3$He, within the stated uncertainty. From [@Aharoni:A:80] it is not clear whether there should be universality for this ratio or for $G(\chi^+)/G(CXC)$, where the experimental values for both fluids are different. But as noted before, universality for the ratios of the first correction term amplitudes is predicted from the $\Phi^4$ model[@Bagnuls:B:M:N:87], and therefore the ratios of the corresponding Ginzburg numbers, obtained via Eq.9, are universal too.
We now compare the ratios with those expected, based on ref[@Aharoni:A:80]. For $^3$He, the measured effective exponents for $\chi$ were $\gamma_{eff}^+$ = 1.19 [@Wallace:M:70; @Chase:Z:76; @Pittman:D:M:79] and $\gamma_{eff}^-$ = 1.08 [@Wallace:M:70; @Chase:Z:76] over the range $5\times10^{-4}<|t|< 5\times 10^{-2}$ . When these $\chi$ data were published, this result was very surprising, because it was expected that the exponents should be the same both above and below $T_c$. However in the light of the Monte Carlo calculations that show the crossover to be quite different on both sides of $T_c$ (See Fig.1), this discrepancy in the values of $\gamma_{eff}$ can be understood. Interestingly the susceptibility data for Xe[@Smith:G:B:71] do not show this difference, and both effective exponents are listed[@Smith:G:B:71] as $\gamma_{eff}^{(+,-)}$ = 1.21 over the range $2 \times 10^{-4} <|t|< 8\times 10^{-3}$. For the coexistence curve, the effective exponent has been reported to be $\beta_{eff}$ = 0.360 for $^3$He [@Wallace:M:70; @Chase:Z:76; @Pittman:D:M:79] and 0.355 for Xe[@Cornfeld:C:72].
The predicted ratios $G(\chi^+)/ G(CXC)$ and $G(\chi^+)/G(\chi^-)$ from Eqs.8 and 9, and from Bagnuls [@Bagnuls:B:M:N:87] via Eq.9 are listed in Table II. Starting with the results from ref.[@Aharoni:A:80], $G(\chi^+)/ G(CXC)
\approx$ 0.10 is determined from the effective exponents and lies in between the values listed for Xe and $^3$He. This prediction, which is consistent with the value obtained by Bagnuls , is then good to within say $\pm$ 20%. There is agreement within the combined uncertainties for $^3$He , but not so for Xe. The prediction of the ratio $G(\chi^+)/G(\chi^-)$ from ref.[@Aharoni:A:80] with Eq.8, is uncertain : if $\gamma^+_{eff} = \gamma^-_{eff}$ is taken, as appears to be the case for the Xe data, the ratio is 0.2. However when the values for $\gamma^+_{eff}$ and $\gamma^-_{eff}$ for $^3$He are used, as listed above, the ratio becomes 0.02 ! The first value is consistent with the predictions by Bagnuls . Overall the experimental data analysis in terms of the correction term amplitudes $a_i$ and the $G_i$’s is still in a preliminary state and further progress is needed.
Conclusion
==========
A status report has been presented of the program describing the crossover from asymptotic to mean-field behavior in different properties for two simple fluids. So far, our understanding is incomplete, since the accuracy of several sets of experimental data needs substantial improvement. By contrast, MC calculations[@Luijten:B:98] are making precise predictions of the crossover for the susceptibility $\chi^+$ and $\chi^-$ as well as the coexistence curve in terms of Ginzburg numbers. Also there are quantitative predictions of the ratios of the correction term amplitudes[@Bagnuls:B:M:N:87].
In spite of the uncertainty in experimental data, some conclusions can be reached. The Ginzburg numbers for $\chi$ and CXC in $^3$He and Xe and their ratios have been obtained from a data analysis. The latter were compared with predictions and discussed. Also from the ratios of the first correction term amplitudes $B_1$, $\Gamma_1(\chi^+)$ and $\Gamma_1(\chi^-$), there appears confirming evidence of their predicted universality within the large uncertainties. Further progress can be expected when better experimental data of the susceptibility of Xe and $^3$He below $T_c$, and over a larger temperature range have been obtained.
Acknowledgments
===============
This research was supported by NASA grant NAG 3-1838. The greatest debt of gratitude goes to E. Luijten for the very stimulating and informative interactions with him and for his generous effort in preparing several plots. Furthermore he made a detailed criticism of this paper. I am also indebted to A. B. Kogan for his help with the plots in Fig.2, with data fitting and for his technical help with the formatting, to G.O Zimmerman for supplying an original figure of the Chase and Zimmerman $\chi$ data, to M. Barmatz and F. Zhong for permission to use the unpublished $\chi$ data (labeled Zhong and JPL in the figures) obtained in the JPL MISTE project, and to M. Giglio for supplying tabulations of $\chi^-$ in Xe. I am very grateful to J.M.H. Levelt Sengers for correspondence on the V.d.W. model and to F. Zhong for comments on this report and for very useful suggestions. Finally I am indebted to C. Bervillier for correcting some references, and to him and to E. Vicari and A. Pelissetto for bringing those in [@Bervillier] to my attention.
Appendix
========
Extension of the rectilinear diameter above $T_c$ ?
---------------------------------------------------
In the course of the data analysis for obtaining the $\chi^+$ of Xe from ref.[@Michels:W:L:54], the location of the maximum for $\chi^+$ with respect to the critical isochore was determined. This line of points might be thought to extend the trajectory of the rectilinear diameter as T increases and passes the critical point. Over the whole range of the data ($t<$ 0.47) it was found to have a slope $\Delta \rho/t$ = - 0.049, to be compared with the slope of - 0.725 for the rectilinear diameter[@Cornfeld:C:72]. In the Ising model, the slope is zero for both lines. Similarly the Van der Waals model predicts the rectilinear diameter slope as -2/5 [@Sengers:LS:78], and from an expansion above $T_c$ the slope for the maximum of $\chi^+$ to be zero [@Levelt:70]. Hence there is a slope discontinuity in the same direction as observed in Xe. For $^3$He, experiments give a rectilinear diameter slope of +0.022 [@Pittman:D:M:79]. Above $T_c$ from an inspection of the the maximum location in $\chi^+$ in the data analysis of various experiments [@Meyer:99], the slope is found to be zero within experimental error over the range $t< 2\times 10^{-1}$. Beyond this range, the $\chi^+$ versus $\rho$ curve along an isotherm is no longer symmetric with respect to $\rho_c$ and as $t$ increases, the maximum of $\chi^+$ shifts to larger densities.
M.A. Anisimov and J.V. Sengers “ Critical and Crossover Phenomena in Fluids and Fluid Mixtures” To appear in “Supercritical Fluids-Fundamentals and Applications” E. Kiran, P.G. Debenedetti and C.J. Peters, Eds, (Kluwer, Dordrecht).
E. Luijten and H. Meyer, Phys. Rev. E [**62**]{} 3257 (2000).
E. Luijten and K. Binder, Phys. Rev. E [**58**]{} 4060(R) (1998).[**59**]{} 7254(E) (1999).
F. Wegner, Phys. Rev. B [**5**]{} 4529 (1972).
K.E. Newman and E.K. Riedel, Phys. Rev. B [**30**]{} 6615 (1984).
J.V. Sengers and J.M.H. Levelt Sengers, “Critical Phenomena in Classical Fluids”, in Progress in Liquid Physics, C.A. Croxton, ed. (Wiley, Chichester, UK, 1978) p. 103.
A. Liu and M.E. Fisher, Physica [**A 156**]{} (35) 1989.
P. Schofield, J.D. Litster and J.T. Ho, Phys.Rev. Lett. [**23**]{}, 1098 (1969).
Further values for the ratio $\Gamma^+_0/\Gamma^-_0$ are 4.77$\pm$0.3 by C. Bagnuls et al. (ref.13); 4.75$\pm$0.03 by M. Caselle and M. Hasenbusch, J. Phys. [**A 30**]{}, 4963 (1997); 4.73$\pm$0.16 and 4.79$\pm$0.1 by R. Guida and J. Zinn-Justin, J. Phys. [**A31**]{}, 8103 (1998), 4.77$\pm$0.02 by M. Campostrini, A. Pelissetto, P. Rossi and E. Vicari, Phys. Rev. [**E 60**]{} 3526 (1999) and 4.762$\pm$0.008 by P. Butera and M. Comi, Phys. Rev. [**B 62**]{} 14837 (2000).
M.A. Anisimov, S.B. Kiselev, J.V. Sengers and S. Tang, Physica A, 188, 487 (1992).
E. Luijten and H.W.J. Bloete, Int. J. Mod. Phys.C[**6**]{} 359 (1995).
A. Aharony and G. Ahlers, Phys. Rev. Lett. [**44**]{}, 782 (1980).
C. Bagnuls, C. Bervillier, D.I. Meiron and B.G. Nickel, Phys. Rev. B [**35**]{} 3585 (1987).
H. Güttinger and D.S. Cannell, Phys. Rev. B [**24**]{} 3188 (1981).
I.W. Smith, M. Giglio and G.B. Benedek, Phys. Rev.Lett. [**27**]{}, 1556 (1971).
A. Michels, T. Wassenaar and P. Louwerse, Physics [**20**]{} 99 (1954).
B.A. Wallace and H. Meyer, Phys. Rev. A [**2** ]{} 1536 (1970). A tabulation of the susceptibility data is given in a technical report, Duke University (1972), unpublished.
C.C. Agosta, S. Wang, L.H. Cohen and H. Meyer, J. Low Temp. Phys. [**67**]{}, 237 (1987).
M. Barmatz, I. Hahn and F. Zhong, Proceedings of the 2000 NASA/JPL Investigators Workshop on Fundamental Physics in Microgravity, edited by D. Strayer, Solvang,CA, June 2000.
C. Pittman, T. Doiron and H. Meyer, Phys. Rev. B [**20**]{} 3678 (1979) The fit of the $\chi^+$ data was done to a series with two correction terms in Eq.2, namely leading to $\Gamma^+_0$ = 0.139 , $\Gamma^+_1$ = 3.2$\pm$1 and $\Gamma^+_2 = - 12\pm$ 14 (Their Table III). However in this fit the second term is very important, and through the strong correlation in the least squares fit, this term impacts on the value of $\Gamma^+_0$. Above the fit range of $t$ = 2$\times 10^{-2}$, the agreement with the data deteriorates very sharply. A fit with only the first term, more reasonable, gives $\Gamma^+_0$ = 0.141 and $\Gamma^+_1$ = 1.5$\pm$0.2 over the same range of $t$.
C.E.Chase and G.O. Zimmerman, J. Low Temp. Phys.,315 (1976)
U. Naerger and D. Balzarini Phys. Rev. B [**42**]{} 6651 (1990)
A.B. Cornfeld and H.Y Carr, Phys. Rev. Lett. [**29**]{}, 28 (1972).
J.M.H. Levelt Sengers, Industrial and Engineering Fundamentals [**9**]{}, 470 (1970)
From an unpublished analysis of $\rho$ versus $P$ data along isotherms from Agosta et al.(ref. 16)
[^1]: To appear in: Proceedings 2000 NASA/JPL Investigators Workshop on Fundamental Physics in Microgravity. D. Strayer, Editor
|
{
"pile_set_name": "ArXiv"
}
|
1.5em
\[t\][c]{} .5em **Mikhail A. Chmykhov and Nikolai A. Kudryashov**
1.5em
**APPROXIMATE SOLUTIONS OF NONLINEAR HEAT EQUATION FOR GIVEN FLOW[^1]**
1.5em [Department of Applied Mathematics,Moscow Engineering and Physics Institute (State university), 31 Kashirskoe Shosse, 115409 Moscow, Russian Federation]{} [[*E-mail:*]{} [email protected], [email protected] ]{}
1.5em
> The one-dimensional problem of the nonlinear heat equation is considered. We assume that the heat flow in the origin of coordinates is the power function of time and the initial temperature is zero. Approximate solutions of the problem are given. Convergence of approximate solutions is discussed.
The problem statement.
======================
The nonlinear heat equation has the form [@r1]
$$\label{Eq_main}
\frac{\partial u}{\partial t} = {\varkappa} \frac{\partial}{\partial
r} \left( u^n \frac{\partial u}{\partial r} \right), \qquad r>0,
\qquad t>0$$
where $ u(r,t) $ is temperature in the point $ r $; $ t $ is time; $ \varkappa $ is coefficient of the heat conductivity.
Suppose the boundary condition has the form
$$\label{Gr_u}
\left. u^n \frac{\partial u}{\partial r} \right|_{\theta = 0} = - q_0 t^k, \qquad t>0$$
Where $ q_0 $ is constant. Equation (\[Gr\_u\]) corresponds to the energy flow in the origin of coordinates.
We also take the initial condition in the form
$$\label{U_0}
u(r,t=0) = 0, \qquad r>0$$
The boundary value problem (\[Eq\_main\])—(\[U\_0\]) was considered [@r1; @r2; @r3; @r4]. When the temperature specified on the boundary the problem for equation (\[Eq\_main\]) were considered in [@r3; @r4; @r5; @r6; @r7]. Numerical solution of the problem (\[Eq\_main\])—(\[U\_0\]) can be found using difference method [@7].
In this work we are going to look for approximate solutions of the problem (\[Eq\_main\])—(\[U\_0\]).
Using
$$\label{Zamena}
u = v^{\frac{1}{n}}$$
we get
$$\label{Eq_new}
v_t = \varkappa v v_{r r} + \frac{\varkappa}{n} v_r^2, \qquad r>0,
\qquad t>0$$
Taking the problem (\[Eq\_main\])—(\[U\_0\]) into consideration we have the boundary and the initial conditions for equation (\[Eq\_new\]) in the form
$$\label{Gr_Usl_new}
\left. v^{\frac{1}{n}} \frac{\partial v}{\partial r} \right|_{\theta
= 0} = -q_0 n t^k, \qquad t>0$$
$$\label{N_Usl_new}
v(r,t=0) = 0, \qquad r>0$$
To solve the problem (\[Eq\_new\]) — (\[N\_Usl\_new\]) we can use the self-similar variables [@r1; @r2; @r3; @r4; @r5; @r6; @r7; @8]
$$\label{S-S_var}
v(r,t)=A t^m f(\theta), \qquad \theta = \frac{B r}{t^p}$$
where $f(\theta )$ is a function of $ \theta $; $ A $, $ B $, $m$ and $p$ are constants.
Assuming
$$\label{Param}
p = \frac{m+1}{2}, \qquad \varkappa A B^2 = 1$$
and substituting (\[S-S\_var\]) into equation (\[Eq\_new\]) we obtain
$$\label{ObshEq}
f f_{ \theta \theta } + {\frac {1}{n} f_{\theta}^2} +{\frac {m+1}{2}
\theta f_{\theta}} - mf =0$$
Substituting (\[S-S\_var\]) into the boundary condition (\[Gr\_Usl\_new\]) and taking into account
$$\label{Param11}
\frac{m(n+1)}{n} - \frac{(m+1)}{2} = k$$
$$\label{Param12}
q_0 B^{-1} n = A^{\frac{n+1}{n}}$$
we have the boundary condition for equation (\[ObshEq\]) in the form
$$\left. f^{\frac{1}{n}} \frac{d f}{d \theta} \right|_{\theta = 0} =
-1$$
From equation (\[N\_Usl\_new\]) we find the second boundary condition
$$f(\theta \to \infty) = 0$$
Using conditions (\[Param\]), (\[Param11\]) and (\[Param12\]) we get constants $ A, \ B, \ m $ in form
$$A = \varkappa^{\frac{n}{n+2}} q_0^{\frac{2 n}{n+2}} n^{\frac{2
n}{n+2}}, \qquad B = \varkappa^{- \frac{n+1}{n+2}} q_0^{-
\frac{n}{n+2}} n^{- \frac{n}{n+2}}$$
$$\label{m0k}
m = \frac{n (2k+1)}{n+2}$$
Taking expressions (\[Param\]), (\[Param11\]) and (\[Param12\]) into account we have the boundary value problem (\[Eq\_main\])—(\[U\_0\]) in the form
$$\label{ObshEq_nu}
f f_{ \theta \theta } + {\frac {1}{n} f_{\theta}^2} +{\frac {m+1}{2}
\theta f_{\theta}} - m f =0$$
$$\label{Gr_lim}
\left. f^{\frac{1}{n}} \frac{d f}{d \theta} \right|_{\theta =0} = -1$$
$$\label{Gr_inf}
f(\theta \to \infty) = 0$$
To find the solution of the problem (\[Eq\_main\]) — (\[U\_0\]) we have to solve the problem (\[ObshEq\_nu\]) — (\[Gr\_inf\]). This is the aim of this work.
Method applied.
===============
It is known that the velocity of the boundary for the nonlinear heat conductivity is finite. Let us assume that $ \theta = \alpha $ is the boundary. Therefore, at the point $ \alpha $ temperature is equal to zero $( \ f(\alpha) = 0 \ )$, but derivative is non-zero $\left( \
\dfrac{{df}} {{d\theta}} \ne 0 \ \right) $.
We look for approximate solution of the problem (\[ObshEq\_nu\]) — (\[Gr\_inf\]) in the form
$$\label{Ser_01}
f(\theta ) = \sum\limits_{i = 0}^N {\beta _i \left( {\alpha - \theta
} \right)^i }$$
From equation (\[Ser\_01\]) we get
$$\label{Proizvodnii}
\left. {\frac{{df}} {{d\theta }}} \right|_{\theta = \alpha } = -
\beta _1 , \quad \left. {\frac{{d^2 f}} {{d\theta ^2 }}}
\right|_{\theta = \alpha } = 2\beta _2 , \quad \ldots \quad, \left.
{\frac{{d^i f}} {{d\theta ^i }}} \right|_{\theta = \alpha } =
i!\left( { - 1} \right)^i \beta _i .$$
Taking $ f(\alpha ) = 0 $ into account we get $ \beta_{{0}}=0 $
Substituting (\[Proizvodnii\]) into equation (\[ObshEq\]) we have the coefficients $
\beta_i $.
$$\begin{gathered}\label{BetaO_p1}
\beta_1 = \frac{1}{2} \alpha n \left( m + 1 \right) \\
\beta_2 = \frac{1}{4} {\frac { \left( m - 1 \right) n}{n+1}} \\
\beta_3 = - \frac{1}{12} {\frac {n (m-1) \left( nm+n+2m \right)}{
\left( n+1 \right) ^{2}\alpha\, \left( 1+2\,n
\right) \left( m+1 \right) }} \\
\beta_4 = \frac{1}{48} {\frac {n (m-1) \left( nm+n+2m \right)
P_4^{(m,n)}}{{\alpha}^{2} \left( n+1 \right) ^{3} \left( 1+2\,n
\right) \left( m+1 \right) ^{2 } \left( 3\,n+1 \right) }} \\
\beta_5 = - \frac{1}{240} {\frac {n (m-1) \left( nm+n+2m \right)
P_5^{(m,n)}}{{\alpha}^{3} \left( 3\,n+1 \right) \left( m+1 \right)
^{3} \left( 1+2\,n \right) ^{2} \left( n+1 \right) ^{4} \left(
1+4\,n \right) }}
\end{gathered}$$
where $ P_4^{(m,n)} $ and $ P_5^{(m,n)} $ are polynomials
$$\begin{gathered}
P_4^{(m,n)} = 5\,nm-n+7\,m-3 \\
\end{gathered}$$
$$\begin{gathered}
P_5^{(m,n)} =
303\,n{m}^{2}+82\,{m}^{2}+102\,{n}^{3}{m}^{2}+317\,{n}^{2}{m}^{2}-204
\,{n}^{2}m \\
-238\,nm -70\,m-48\,{n}^{3}m+12+7\,{n}^{2}-6\,{n}^{3}+31\,n
\end{gathered}$$
Using expression (\[BetaO\_p1\]) we can obtain some exact solutions of the problem (\[ObshEq\_nu\]) — (\[Gr\_inf\]).
If $ n + 2 m + n m = 0 $ then $ \beta_i = 0 \quad (i \geq 3 $). Taking $ m = - \frac{n}{n+2} $ into consideration we have exact solution
$$\label{Resh_0}
f(\theta) = \frac{1}{2} \alpha n \left( m + 1 \right) \left( {\alpha
- \theta } \right) + \frac{1}{4} {\frac { \left( m - 1 \right)
n}{n+1}} \left( {\alpha - \theta } \right)^2$$
However this exact solution does not satisfy the boundary condition (\[Gr\_lim\]). This solution can be used to solve the Cauchy problem for equation (\[Eq\_main\]).
If $ m = 1 $ then $ \beta_i = 0 \quad (i \geq 2 $) and we have another exact solution for equation (\[ObshEq\_nu\]).
$$\label{Eq_2_12}
f(\theta) = \alpha n \left( {\alpha - \theta } \right)$$
Exact solution (\[Eq\_2\_12\]) satisfies the boundary condition (\[Gr\_lim\]).
Solutions of the problem (\[ObshEq\_nu\]) — (\[Gr\_inf\]).
==========================================================
Consider the case $ n > 0 $ and $ k = {1}/{n} $, ($ m = 1 $).
$$\label{Solv_pl_1n}
f(\theta) = \alpha n \left( \alpha - \theta \right)$$
Substituting (\[Solv\_pl\_1n\]) into the boundary condition (\[Gr\_lim\]) we obtain
$$\label{Find_a}
\frac{{\left( n+1 \right)} {\left( \alpha^2 n
\right)}^{\frac{n+1}{n}}}{n \alpha}=1$$
From condition (\[Find\_a\]) we get the parameter $ \alpha $
$$\alpha = \left( n^{\frac{1}{n}}(n+1) \right)^{\frac{n}{n+2}}$$
The values of $ \alpha $ and $ n $ at $ k={1}/{n} $ are given in table \[t:pl\].
n 1 $ {4}/{3} $ $ 2 $ $ {5}/{2} $ 3 4
------------ ------------- ------------- -------------- ------------- -------------- --------
$ \alpha $ 1.2599 1.5299 2.0598 2.4586 2.8619 3.6840
n $ {9}/{2} $ $ 5 $ $ {11}/{2} $ 6 $ {13}/{2} $ $ 7 $
$ \alpha $ 4.1025 4.5256 4.9528 5.3838 5.8184 6.2561
: []{data-label="t:pl"}
Exact solutions of the boundary value problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]) at $ k = {1}/{n} $ ($ n > 0 $) are described by formula (\[Solv\_pl\_1n\]).
Consider the case $ {n = 1} $ and $ k > 0 $. Approximate solution of the boundary value problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]) can be written in the form
$$\label{f3_10}
\begin{gathered}
f(\theta) = \frac{1}{2} \alpha \left( m + 1 \right) \left(
\alpha-\theta \right) + \frac{1}{8} {\left( m - 1 \right) } \left(
\alpha-\theta \right) ^{2 } \\
- \frac{1}{144} {\frac { \left( 3\,m+1 \right) \left( m-1 \right)
}{ \left( m+1 \right) \alpha}} {\left( \alpha-\theta \right) ^{3}}
+\frac{1}{1152} {\frac { \left( 3\,m+1 \right) \left( m-1 \right)
\left( 3\,m-1 \right) }{ \left( m+1 \right)
^{2}{\alpha}^{2}}} { \left( \alpha-\theta \right) ^{4}}\\
- \frac{1}{172800} {\frac { \left( 3\,m+1 \right) \left( m-1
\right) \left( 201\,{m}^{2 }-140\,m+11 \right) }{ \left( m+1
\right) ^{3}{\alpha}^{3}}} { \left( \alpha-\theta
\right) ^{5}} + \ldots
\end{gathered}$$
From the boundary condition (\[Gr\_lim\]) we obtain
$$\label{01}
\begin{gathered}
{\frac {{\alpha}^{3} }{2985984000}} \, \frac {\left(
105147\,{m}^{4}+ 384822\,{m}^{3}+519188\,{m}^{2}+307082\,m+66161
\right)} { \left( m+1 \right) ^{6}}\\
{ \left( 9491+
24237\,{m}^{4}+84342\,{m}^{3}+54602\,m+103808\,{m}^{2} \right) } = 1
\end{gathered}$$
From equation (\[01\]) we find
$$\begin{gathered}
\alpha = 1440\, { \left( m+1 \right)^{2}} / \left(
2548447839\,{m}^{8}+18195239088\,{m}^{7} \right. \\
+55955316456\,{m}^{6} +96920939400\,{m}^{5}+103409323126\,{m}^{4}\\
\left. { +69458768096\,{m}^{3} +28562945760\,{m}^{2} +6527038184\,m
+627934051} \right) ^{\frac{1}{3}}
\end{gathered}$$
k 0 $ 1 $ $ 2 $ $ 3 $ 4 5
------------ -------- -------- -------- -------- -------- --------
$ m $ 1/3 1 5/3 7/3 3 11/3
$ \alpha $ 1.1762 0.7937 0.6222 0.5211 0.4532 0.4039
: []{data-label="t:pl_k_n1"}
The values of $ \alpha $, $ m $ and $ k $ at $ n=1 $ are given in table \[t:pl\_k\_n1\]. Approximate solutions of the problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]) at $ n=1 $ are expressed by formula (\[f3\_10\]).
For the case $ k=0 $ approximate solution of the boundary value problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]) takes the form
$$\label{A-sol_p}
f(\theta) = \left\{
\begin{gathered}
\frac{2 \alpha(\alpha- \theta)}{3} -\frac{(\alpha-\theta)^2}{12} +\frac{(\alpha-\theta)^3}{144\alpha}
-\frac{(\alpha-\theta)^5}{23040\alpha^3}, \qquad \\
\hfill 0 < \theta < \alpha; \\
0, \quad \hfill \alpha < \theta; \\
\end{gathered}
\right.$$
where $ \alpha $ is $ \alpha = 1.1762 $.
To check approximate solution (\[A-sol\_p\]) we have compared it with numerical solution of the boundary value problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]). The comparison of approximate solution (\[A-sol\_p\]) and numerical one at $ k=0 $ and $ n=1 $ is given on Fig.\[fig:Zsol\_p\]. Solid line is approximate solution and circles correspond to numerical solution of the problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]). From Fig.\[fig:Zsol\_p\] we can see that these solutions are similar.
Consider the case $ {n = 4/3} $ and $ k>0 $. Approximate solution of the boundary value problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]) is given by formula
$$\label{f33_10}
\begin{gathered}
f(\theta) = \frac{2}{3} \alpha \left(m+1 \right) \left(
\alpha-\theta \right) + \frac{1}{7} \left(m-1\right) \left(
\alpha-\theta \right) ^{2 } \\
-{\frac {2}{539}}\,{\frac { \left( 5\,{m}^{2}-2-3\,m \right) \left(
\alpha-\theta \right) ^{3}}{\alpha\, \left( m+1 \right) }} \\
+{\frac {1}{ 37730}}\,{\frac { \left(
205\,{m}^{3}-188\,{m}^{2}-43\,m+26 \right) \left( \alpha-\theta
\right) ^{4}}{ \left( m+1 \right) ^{2}{\alpha}^{
2}}} \\
-{\frac { 9 \left( 29055\,{m}^{4}+7078\,m+
1199\,{m}^{2}-464-36868\,{m}^{3} \right) \left( \alpha-\theta
\right) ^{5}}{137997475 {\alpha}^{3} \left( m+1 \right) ^{3}}} +
\ldots
\end{gathered}$$
Some values of $ \alpha $, $ m $ and $ k $ at $ n= 4/3 $ are given in table \[t:pl\_k\_n43\].
k 0 $ 1 $ $ 2 $ $ 3 $ 4 5
------------ -------- -------- -------- -------- -------- --------
$ m $ 2/5 6/5 2 14/5 18/5 22/5
$ \alpha $ 0.9256 0.6000 0.4608 0.3807 0.3278 0.2897
: []{data-label="t:pl_k_n43"}
Consider the case $ {k = 0} $ and $ n > 0 $. From the expression (\[m0k\]) we have
$$m=\frac{n}{n+2}$$
Approximate solutions of the boundary value problem (\[ObshEq\_nu\]) – (\[Gr\_inf\]) at $ {k = 0} $, $ n > 0 $ take the form
$$\label{Resh_l1}
\begin{gathered}
f(\theta)=\frac{1}{2}\left( {\frac
{\alpha\,{n}^{2}}{n+2}}+\,\alpha\,n \right) \left( \alpha-\theta
\right) -\frac{1}{2}\,{\frac {n \left( \alpha-\theta
\right) ^{2}}{ \left( n+1 \right) \left( n+2 \right) }}+ \\
+\frac{1}{6}\,{\frac {{n}^{2} \left( \alpha-\theta \right)
^{3}}{\alpha\, \left( n+1 \right) ^{3} \left( 1+2\,n \right) }}
-\frac{1}{24}\,{\frac { \left( 2\,{n}^{2 }+n-3 \right) {n}^{2}
\left( \alpha-\theta \right) ^{4}}{ \left( n+1
\right) ^{5} \left( 1+2\,n \right) {\alpha}^{2} \left( 3\,n+1
\right) }}+ \\
+{\frac {1}{120}}\,{\frac {{n}^{2} \left( 12+8\,n-75\,{n}^{
2}+12\,{n}^{5}-77\,{n}^{3} \right) \left( \alpha-\theta \right)
^{5}} {{\alpha}^{3} \left( n+1 \right) ^{7} \left( 1+2\,n \right)
^{2} \left( 3\,n+1 \right) \left( 1+4\,n \right) }} + \ldots
\end{gathered}$$
Using boundary condition (\[Gr\_lim\]) we obtain the value of $ \alpha $. Some values of $ \alpha $ and $ n $ at $ k=0 $ are given in table \[t:pl\_k0\]
n 1 $ {4}/{3} $ $ 2 $ $ {5}/{2} $ 3 4
------------ ------------- ------------- -------------- ------------- -------------- --------
$ \alpha $ 1.4819 1.1578 0.7889 0.6283 0.5178 0.3775
n $ {9}/{2} $ $ 5 $ $ {11}/{2} $ 6 $ {13}/{2} $ $ 7 $
$ \alpha $ 0.3307 0.2934 0.2632 0.2382 0.2172 0.1994
: []{data-label="t:pl_k0"}
Conclusion.
===========
The boundary value problem of the nonlinear heat equation for the given flow was considered. This problem was solved using the both numerical and analytical approaches. Some exact solutions were found. Approximate solutions of the boundary value problem were obtained. Comparison of the numerical and the approximate solutions was given.
This work was supported by the International Science and Technology Center (project B1213).
[99]{}
[^1]: in Proceedings of the XXXIII Summer School Advanced Problems in Mechanics APM’2005, St.Petersburg, 2005
|
{
"pile_set_name": "ArXiv"
}
|
[Entropic Trapping of Additive Particles at Polymer Surfaces and Interfaces]{}
[ Galen T. Pickett ]{}
\#1
[*Department of Physics and Astronomy, California State University Long Beach, 1250 Bellflower Blvd., Long Beach, CA 90840\
*]{}
[**Abstract**]{}
I consider the possibility that Gaussian random walk statistics are sufficient to trap nanoscopic additives at either a polymer interface or surface. When an additive particle goes to the free surface, two portions of the polymer surface energy behave quite differently. The purely enthalpic contribution increases the overall free energy when the additive protrudes above the level of the polymer matrix. The entropic part of the surface energy arising from constraints that segments near a surface can’t cross it, is partly relaxed when the additive moves to the free surface. These two portions of the polymer surface energy determine the equilibrium wetting angle formed between the additive and the polymer matrix, the measurement of which in an experiment would allow an independent determination of each piece of the polymer surface energy.
Introduction
============
The addition of small, nanoscopic additives to polymer matrices is quite important technologically and scientifically.[@bigfirst] Even when the particles have no special properties other than hard exclusions of polymer chains, they are useful in improving the toughness of plastic components by preventing large-scale disruptions of the matrix, for instance near a crack tip [particle\_tough]{}. They can also pin local interfaces in a phase segregating polymer blend, providing a mechanism for halting the growth of domains and making blends behave more compatibly [anna\_blend\_part]{}.
Using polymer patterns as templates for inorganic and metallic depositions is another example involving polymers and hard additive particles. Here, microsegregation or other patterns in a polymer film can be selectively decorated by the evaporation of metal [pattern\_deposition]{}. I am particularly interested in these deposition processes with high surface-energy metal particles that reside at the free surface of a polymer thin film [hmj,shull\_prl,find1,find2]{}. Usually, patterned films cooled below their glass transition temperatures are used for such purposes, with the goal of keeping the polymer pattern static and allowing the additive particles to decorate the top of an impenetrable polymer layer [glassy\_decorate]{}. The vicinity of the free surface of such patterned films can often have a deep supression of $T_g$, giving the metallic elements a chance to interact with a rubbery or molten polymer layer [bimetallic]{}. However, even materials that would be expected to be wet with polymers in bulk situations at temperatures well above the bulk value of $T_g$ seem upon deposition to remain pinned at the free polymer surface rather than to penetrate to the interior of the films [find1,pinned\_free]{}. When a further layer of polymer is floated upon these particles, these trapped polymers are free to diffuse along the film normal. When the floated layer is a different homopolymer, the particles can remain trapped at the interface [shull\_trap\_blend]{}. When the particles wet neither of the polymer involved, they act as a compatibilizer, replacing polymer-polymer contacts with polymer-particle contacts [anna\_particle\_compatibilize]{}.
To understand the fundamental physical effects driving these particles toward free surfaces, the basic contributions to the bare polymer surface energy will be crucial to understand. These contributions have typically two parts, giving partial surface energies usually the same order of magnitude [silberberg]{}. Firstly, there is the enthalpic contribution to the polymer surface energy, arising from short-range dispersion effects, contributing $\gamma_p^o \equiv w_p /\sigma^2$ to the overall polymer surface energy. Roughly speaking, every segment at the interface costs an interaction energy $w_p$ compared to the reference of pure melt polymer. Here, $\sigma$ is a typical polymer segment length scale. Another interaction is $\gamma_{ps}^o \equiv w_{ps}/\sigma^2$ as well as $\gamma_s^0 \equiv w_s/\sigma^2$, giving the enthalpic contribution to the surface energy between polymer and the sphere, and the sphere and vacuum respectively. Usually, $w_p \approx w_{ps}$, with both much smaller than the energy scale for sphere-vacume contacts, $w_s$. There is a strong thermodynamic driving force keeping the additive particle surrounded with at least a thin skin of polymer segments [@find1; @find2].
The other main contribution to the polymer surface energy is entirely entropic in nature, and arises from the constraint placed on a random walk polymer configuration in the vicinity of an impenetrable interface [silberberg]{}. Given a polymer segment near an interface, the adjoining polymer segments must avoid crossing the forbidden surface. This constraint can be thought of as an external bias applied to the chain random walks, with the bias energy approximately equal to $k_B T$ per constrained surface monomer. Here, $k_B$ is Boltzmann’s constant, and $T$ is the temperature. Silberberg determined the correct weighting from an ingenious symmetry argument, but the overall effect is entirely entropic, and contributes to the surface energy of $\gamma_e \equiv k_B T/\sigma^2$, The total polymer surface energy, $\gamma_p$, is then given by $\gamma_p = \gamma_p^o + \gamma_e$. Typically, $\gamma_p^o \approx \gamma_e$, so that entropic and enthalpic effects are equally important in determining $\gamma_p$. It may be possible to separate these two effects, using one to drive the creation of new surface area, while using the other to stabilize the particle position. Analyzing how these contributions affect the deposition or evaporation of metallic particles onto a polymer melt is the main goal of this paper.
As such, the scenario of evaporating metallic particles onto a polymer film is far from a new one, and has been studied in mean field models [wetting\_noolandi\_ref\_32]{} and interaction-site monte-carlo simulations [interaction\_site]{}. Melt-particle interactions have been studied in very detailed bead-spring simulations [bead-spring-nist]{}. While the constraints placed on polymer conformations are dealt with explicitly in an early mean field theory [wetting\_noolandi\_ref\_32]{}, that theory is most applicable to broad interfaces which dilutes the importance of the segment-swapping contribution to the surface energy, and therefore underestimates the entropic stabilizing effect. Interestingly, this theory does predict that fully wet particles should be stabilized near a polymer free surface at a range of approximately the particle radius. Studies aimed at studying the formation of nanoclusters of metallic particles after being embedded in the upper reaches of a polymer film also show a marked tendency for the growing metallic particles to cluster at the free surface as they grow, but it should be noted that this particular simulation does not treat the polymer conformational entropy well, and therefore a layer of entropically active wetting segments is absent in the equilibrated system. The very through simulations in [bead-spring-nist]{} investigate the interactions of moderately long polymer bead-spring chains with hard additive icosehedral particles. They find that polymer chains in contact with the the surfaces of these particles are indeed aligned along the particle surfaces, indicating that Silberberg’s notion of the entropic contributions to the polymer surface energy are operating.
In a much simpler theory focusing on the Silberberg idea, I consider impenetrable additive particles wetted with polymer segments.[find1,find2]{} The question is [*do such particles have an enhanced surface activity*]{}? The quantitative prediction is that generally wetting particles are surface active, and to a degree that will allow the measurement of the relative size of the entropic and enthalpic contributions to the system free energy.
![ The surface of a typical polymer melt. Chains with at least one monomer at the surface are shown as bold lines. Each monomer within the shaded region at the free surface represeents a free energy cost of $k_B T$ to maintain the polymer conformations so that they do not cross the free surface. The bold line represents enthalpic, and the shaded region represents entropic, polymer-vacuum contacts. A) bare melt, B) particle in the bulk, and C) particle surface stabilized by entropic contacts. []{data-label="fig.1"}](figure1.eps){width="5in"}
I will consider two simple, though related cases. The first is a single spherical wetting particle in the vicinity of the free surface of a polymer melt, as in fig. \[fig.1\] part C. The second case has a wetting, though partially selective particle near the interface between incompatible polymer melts. Even when the particles are preferentially wet with a single one of the polymers involved, there are indeed situations in which entropic gains keep the particles trapped at the interface. In the special case in which the spherical particles are deposited upon a polymer melt, and then covered with an amorphous layer of the same polymer, the entropic pinning of the single particles is released, driving the particles away from the original interface, as is observed experimentally [shull\_prl]{}. It should be noted that the argument here is entirely equilibrium in character, and does not depend on the presence of any other additive particles. Thus, the effect described here does not depend on bridging interactions between adjacent particles [tom\_bridge]{}.
After describing the effect for a single surface, I will treat the case of a polymer-polymer interface. After a short discussion, I will draw my conclusions.
Free Surface
============
As in fig. \[fig.1\] part A a polymer melt surface is characterized by a strong constraint on the conformations of all the polymer segments laying within a polymer segment size, $\sigma$, of the interface. Roughly speaking, for each such segment at the interface, the random walk configuration of its chain must be biased so that both the preceding and the subsequent monomers on the chain are prevented from crossing the interface. In fig. \[fig.1\] part A the darkly drawn polymers are those which have at least one monomer segment at the surface. All of the monomer segments within the shaded region represent monomers located at the surface, and each of these incurs an entropic cost of $k_BT$ in order to bias the random walk of its parent chain to respect the interface. Each of the segments at the free surface also are in contact with vacuum on one side, and an semi-infinite sea of monomer on the other. If the total surface area of the free surface of the melt is $L^2$, then the free energy associated with maintaining this interface is: $$F_{surf} = \frac{L^2}{\pi \sigma^2} k_B T \left(1 + \frac{w_p}{k_B T}\right)
\equiv L^2 \gamma,$$ where $w_p$ is the energy to have a single monomer segment in contact with vacuum, and $\gamma$ is the surface energy of the polymer. Clearly $$\gamma = \gamma_{entropic} + \gamma_{enthalpic},$$ with $$\gamma_{entropic} \equiv \gamma_e =\frac{k_B T}{\pi \sigma^2} \mbox{ and }
\gamma_{enthalpic} \equiv \gamma_p^o =\frac{w_p}{\pi \sigma^2}.$$ The free energy involving the presence of the additive particle, a sphere of radius $R$ is: $$F_{sphere} = \frac{4 \pi R^2}{\pi \sigma^2} k_B T \left( 1 +
\frac{w_{ps}}{k_B T}\right).
\label{five}$$ Here, $w_{ps}$ is the energy cost per unit monomer to make a polymer-sphere contact. In principle, there is another energy in the problem, $w_s$, the energy needed to make a vacuum-particle contact for each polymer-segment sized patch of surface. We assume, as is generally the case when the additive particles are metallic [@find2], that $$w_p \approx w_{ps} \ll w_s$$ and the additive particle are always wet with polymer segments.
Polymer segments near the sphere also endure entropic constraints. Up to terms of order $\sigma/R$ this entropic cost is exactly the same for the polymer free surface, and represents the term in $F_{sphere}$ that scales strictly as $k_B T$, as in eq. \[five\]. Altogether, the total polymer $+$ sphere free energy is: $$F_{in} = F_{surf} + F_{sphere} = L^2 (\gamma_{e} +
\gamma_{p}^o ) + 4 \pi R^2 ( \gamma_{e} + \gamma_{ps}^o),$$ when the sphere is free to explore the interior of the matrix..
Now, consider a situation as in fig. \[fig.1\] part C, where the sphere protrudes some distance above the level of the polymer matrix. For definiteness, let us suppose that the sphere protrudes a distance $0< h < 2 R$ above the polymer matrix, so that the portion of the sphere protruding covers an azimuthal angle $\theta$ as in fig. \[fig.2\].
![ (A) Definitions of several quantities in the text are shown here schematically. This height of protrusion, $h$, the particle radius $R$ and the wetting angle $\theta$ are shown. (B) Young’s law as applied at the polymer-metal contact. []{data-label="fig.2"}](figure2.eps){width="5in"}
When $\theta =0$, the particle has just been brought to contact with the free surface, and $h = R (1- \cos \theta)=0$. When $\theta = \pi$, $h=2R$, and the particle has been expelled from the matrix. The particle occupies a circular patch at the original interface with a radius equal to $R \sin \theta$, so the contribution from the surface of the unperturbed polymer matrix is: $$F_{surf} = (L^2 - \pi R^2 \sin^2 \theta) (\gamma_e + \gamma_p^o),$$ or, using elementary trigonometry, $$F_{surf}(h) = (L^2 - \pi 2 h R + \pi h^2) (\gamma_e + \gamma_p^o).$$ This represents the entropic and enthalpic contributions from the flat polymer-vacuum interface.
The contribution to the system free energy depending on the presence of the sphere is also easy to estimate. Using the fact that the sphere is still completely wet with polymer segments, and that each of these segments still represents constraints on the chain conformations of $k_B T$ per monomer, the entropic contribution to the sphere interaction is still [*exactly the same*]{} as if it were completely immersed in the polymer melt. The polymer-sphere contacts are exactly the same as well, but there are now quite a few more polymer-vacuum contacts to account for. In all: $$F_{sphere}(h) = 4 \pi R^2 (\gamma_e + \gamma_{ps}^o) + 2 \pi R h \gamma_p^o$$ Thus, the total free energy is $$F_{out}(h) = F_{surf}(h) + F_{sphere}(h),$$ the free energy associated with having the particles a distance $h$ “out” of the polymer melt.
It only remains to determine the equilibrium value for $h$, which can be found by minimizing $\Delta F(h) = F_{in} - F_{out}(h)$: $$\Delta F(h) = (\gamma_e + \gamma_p^o) \pi (h^2 - 2 h R) +
2 \pi R h \gamma_p^o.$$ This relation can be essentially read off from the schematic in fig. \[fig.1\] part C. The dark line at the surface of the polymer melt marks polymer segment-vacuum contacts, and the shaded region represents the volume in which entropic biases are required to keep the chains from crossing a surface. There is less shaded area when the particle protrudes above the melt, but with a cost in more polymer-vacuum contacts.
Minimizing $\Delta F$ with respect to $h$ yields the equilibrium value for $h$: $$h_{eq} = R \frac{ \gamma_e}{ \gamma_e + \gamma_p^o}.$$ The wetting angle that the particle makes with the polymer matrix is simply $\theta$, with an equilibrium value of: $$\cos \theta_{eq}= \frac{\gamma_p^o}{\gamma_e + \gamma_p^o},
\label{youngs}$$ specifying the usual equilibrium contact angle. Thus, the polymer coated particle acts just as Young’s Law says it should, as applied to the effective three-component interface consisting of the polymer melt (with surface tension $\gamma_1=\gamma_p^o+\gamma_e$), the portion of the particle immersed in the melt with an effective surface tension $\gamma_2=\gamma_{ps}^o + \gamma_e$ and the portion of the particle protruding from the surface of the melt with a surface tension of $\gamma_3=\gamma_p^o +\gamma_e + \gamma_{ps}$. According to Young’s law, the wetting angle $\theta$ obeys: $$\cos \theta = \frac{ \gamma_2 - \gamma_3}{\gamma_1},$$ agreeing with the result in eq. \[youngs\].
Thus, measurements of $h$ and $R$ and $\gamma$ disentangle the two contributions to the [*bare polymer surface energy*]{}, and allows us to resolve its two parts. In terms of the measurable quantities, we have $$\begin{aligned}
\gamma_p^o & = & \gamma \frac{h}{R} \mbox{ (enthalpic), and}\\
\gamma_e & = & \gamma \left(1- \frac{h}{R}\right) \mbox{ (entropic).}\label{observable}\end{aligned}$$ Indeed, in the system of [find1]{}, the protrusion of silver nanoparticles indicates that the $\gamma_e/\gamma_p^o$ is on the order of 2. From the form of eq. \[observable\], a number of interesting predictions are possible. First, the observable range of $h$ is $0 < h < R$, so that the additive particle can either truly wet with the polymer, and have virtually no surface activity ($h=0, \theta =0$) up to $h=R, \theta = \pi/2$. The most deweting configuration possible here is when $\gamma_p =0$ so that there is no enthalpic penalty for increasing $h$. In this limit, we see that the additive particles act effectively as [*neutrally wetting*]{} particles.
For nonzero $h$, the amount of free energy pinning the particles to the free surface can be estimated: $$F_{pin} = \Delta F (h_{eq}) = - \pi R^2 \gamma
\left|\frac{\gamma_e}{\gamma}\right|^2 = - \pi R^2 \gamma (1-\mu)^2.$$ Clearly, when the enthalpic contribution to $\gamma$ dominates and $\gamma_e \ll \gamma$, (so that $\mu \approx 1$) the energy pinning the spheres to the free surface can become quite small, so that particle can readily leave the surface and explore the interior of the melt. If the melt is a film with thickness $H$, the translational entropy gained by the sphere when it leaves the surface and enters the films is $$F_{trans} \approx \ln \frac{H}{R},$$ so that the fraction of bound spheres vs. interior spheres can be calculated for any given $H$ through: $$\begin{aligned}
N_{surf} & = & N_{tot} \frac{\lambda^2}{1 + \lambda^2} \\
N_{free} & \sim & N_{tot} \frac{1}{1+\lambda^2},\end{aligned}$$ where $$\lambda = \frac{H}{R} \exp - \frac{F_{pin}}{k_BT},$$ and $N_{tot}$ is the total number of deposited particles in the system. In the discussion, below, I describe the basic sizes of these effects. It should be kept in mind, however, that these particle are driven toward the interface by an entirely entropic effect, which would be removed if the the vacuum above the sample were replaced with an identical layer of molten polymer. Also, it should be kept in mind that I am interested here in the ultimate equilibrium situation, and this argument completely ignores [*how long*]{} this equilibrium might take to be effected. Long-lived metastable states can change the experimentally observable phenomena (for example, the situation in which polymer segments have not yet had enough time to envelop the additive particle which is initially placed on the surface). Also, this is a single-particle argument. Interactions between the additives can lead to further surface-active effects that have been discussed elsewhere [shull]{}.
Interface
=========
Now, I shall turn to the complimentary situation in which the additive particle is placed at the interface between incompatible polymer melts, $A$ and $B$. The surface energy between these two polymers is $\gamma_{AB}$. The bare particle surface energy is not relevant in this case, but we now have two interactions between polymers and the particle: $\gamma_{As}$ is the surface energy between polymer $A$ and the particle, and likewise for $B$. This energy has entropic and enthalipic parts as well: $$\gamma_{As} = \frac{k_B T}{\sigma^2} + \frac{w_{As}}{\sigma^2}
\equiv \gamma_e + \gamma_{As}^o$$ Similar quantities can be defined for $B$ polymer, and for the $AB$ interfacial tension.
There are two physically distinct situations to consider, the case in which the particle wets completely with, say $B$ type monomers, and protrudes slightly into the upper $A$ region, and the more usual case in which the particle is partially wet with both $A$ and $B$ monomers. To determine which is the case requires a simple comparison of free energies.
When the particle partially wets with $A$ and $B$ monomers, I again define $h$ to be the distance the particle protrudes into the $A$ region. The free energy of the partially wetting parrticle is: $$\begin{aligned}
F_{partial}(h) & = & (\gamma_{AB}^o + \gamma_e) \pi (h^2 - h R) + \nonumber \\
& &4 \pi R^2 \gamma_e + 2 \pi h R \gamma_{As}^o + \nonumber \\
& & (4 \pi R^2 - 2 \pi h R ) \gamma_{Bs}^o.\end{aligned}$$ These terms correspond to the contribution from the $AB$ interface occupied by the particle, the entropic surface energy of the particle, the $A$-particle and $B$-particle surface contacts. The equilibrium position of the particle, $h$, is found from minimizing $F_{partial}(h)$: $$h_{eq} = R \frac{ \gamma_{AB}^o - \gamma_A^o + \gamma_B^o}{\gamma_{AB}^o + \gamma_e} =
R \frac{\gamma_{AB} - \gamma_{As} + \gamma_{Bs}}{\gamma_{AB}},$$ again a manifestation of Young’s Law. As usual, we can define a wetting parameter, $$\epsilon = \frac{\gamma_A - \gamma_B}{\gamma_{AB}}$$ so that $h_{eq} = (1- \epsilon) R$. The free energy for this partially wetting case is thus: $$F_{partial} = -\pi R^2 \gamma_{AB}(1-\epsilon)^2 + 4 \pi R^2 \gamma_{Bs}.$$ Now, assume that the particles are wet with $B$-type fluid. The free energy to maintain the particle protruding a distance $h$ into the upper $A$ fluid is: $$F_{B-wet}(h) = \gamma_{AB} \pi (h^2 - R h) + 4 \pi R^2 \gamma_{Bs} +
2 \pi h R \gamma_{As}^o.$$ This represents exactly the same physical situation as is present in Section 2, with the $A$ polymer playing the role of the vacuum. Immediately, then, we have: $$h_{eq} = R \frac{ \gamma_e}{\gamma_{AB}^o + \gamma_e}.$$ The equilibrium free energy in this case is therefore: $$F_{B-wet} = - \pi R^2 \gamma_{AB}(1-\mu)^2 + 4 \pi R^2 \gamma_{BS}$$ where the appropriate wetting parameter here is $$\mu = \frac{\gamma_{AB}^o}{\gamma_{AB}^o + \gamma_e}.$$ Thus, there is a transition between the wetting and non-wetting situations when $$(\epsilon-1)^2 = (\mu-1)^2,
\label{condition}$$ the only relevant root of which is $\mu=\epsilon$. That is, as long as $0<\epsilon<\mu$, the particle will be partially wet, and $\mu< \epsilon<1$ will cause the particle to wet with $B$ monomers, although it is still confined to the $AB$ interface by the entropic effect. Clearly, when $\epsilon=0$, the equilibrium situation is that of neutrally wetting particles, half covered with $A$ monomers, half covered with $B$ monomers, exactly straddling the $AB$ interface. As $\epsilon$ is increased, the particle moves toward the $B$ domain with $h < R$, until a transition to the particle wet with $B$ monomers is achieved, when the requirement of eq. \[condition\] is met.
![ Given the entropic wetting parameter $\mu \equiv \cos \theta$ for the fully-$B$ wet sphere, if the partial wetting parameter $\epsilon$ satisfies $|\epsilon| < |\mu|$, then the particle partially wets. Otherwise, and any further increase in $|\epsilon|$ is irrelevant for determining the energy binding the particle to the interface. []{data-label="fig.3"}](figure3_new.eps){width="5in"}
As in the case of particles residing at a free surface, an estimate can be made of the equilibrium density of particles at the interface vs. the equilibrium density of particles dispersed into the $B$ domain. When $\epsilon<0$, the situation is exactly the same, except the transition is toward the particle wetting with $A$ monomers, with $\gamma_{BS}
\rightarrow \gamma_{AS}$ in the reference energy. The situation is shown schematically in fig. \[fig.3\].
Thus, the sequence of events is as follows. When high surface energy particles are evaporated or deposited on the surface of a polymer melt, they will first decorate themselves with a mass of restricted monomer, creating a skin of $B$ monomer, say. If the monomer-particle interaction is not strong (so that the entropic energy dominates the equilibrium) the particles will be essentially trapped at the surface, even upon annealing. If more $B$ monomer is spun onto the system in a thick blanket, or if a thin film is floated onto the existing $B$ surface, the entropic stabilization of the particles toward the interface will disappear, and the particles will start to engage in restricted diffusion normal to the original interface (at possibly a very small rate, as the particles in equilibrium resemble a star polymer, with many protruding loops and long arms of bound polymer). If, on the other hand, a different type of monomer is deposited on the surface, the particles may or may not still be trapped at the surface. Very incompatible polymers ($\gamma_{AB}$ is very large) that neutrally wet the particles ($\gamma_{As} \approx \gamma_{Bs}$) will be entropically trapped at the interface. The particles act as compatibilizers. This state of affairs can continue up to and even past the point where the particle wets with $B$ polymer.
Discussion
==========
The scale of the energy trapping a typical particle at the free surface could easily lead to practical trapping of even wetted particles. The basic scale of the effect is $R^2 \gamma_{AB}$. For a $R=10 nm$ gold sphere at a free polystyrene surface, this energy scale is $ R^2 \gamma_{PS} \approx (10 nm)^2 k_B T /(1 nm)^2$, or approximately $100 k_B T$. Therefore, a particle is stabilized by a factor of $100 k_B T (1-\epsilon)^2$. For a typical flexible polymer, $\gamma_e \approx \gamma_p$, so that $\epsilon \approx 1/2$. Such a particle is therefore trapped by an energy of approximately $25 k_B T$ per particle, and they protrude approximately $h=1/2 R= 5nm$ above the polystyrene surface. Under normal circumstances, it would be virtually impossible to observe such a particle leave the $A$ surface. The particles are thus, for all intents, irreversibly attached to the free surface, although their lateral motion is unrestricted. As mentioned above, coating the sphere-polymer system with another polystyrene layer will result in the particle being freed from the surface, and engaging in diffusive motion along the film normal.
However, if a different polymer is spread on the surface of the metal polymer system, we have to compare the wetting parameter, $\epsilon$ for the $A-B$ particle system to the entropic wetting parameter $\mu$ for the $AB$ interface. Ideal conditions for unpinning of the spheres from the $AB$ interface would require $\epsilon \approx \mu \approx 1$, and we can expect such conditions to occur regularly for very selective interactions that are dominated by $w_{AS}$ and $w_{BS}$. Note that the critical element here has nothing to do with the incompatibility of the $A$ and $B$ fluids per-se, but is sensitive to the interactions of the polymer with the embedded sphere.
The entropic effect here stabilizing the particles at the interface can be thought of an as extreme form of the depletion attraction of colloidal systems [depletion]{}. In that case, a solid sphere excludes the colloids from a spherical shell of size $r$ around the sphere of radius $R$, where the colloid radius is $r$. When two such spheres approach, the overlap of their exclusion shells results in [*fewer constraints*]{} that need to be maintained in the system. In the current context, the colloidal particles are represented by the polymer segments at the interfaces. The particles sitting at the interface remove a circular patch of constraints that would otherwise have to be enforced, at the expense of making energetically unfavorable contacts. Thus, we can expect that spheres that have managed to escape the surface and explore the interior of the film will be attracted by an effective potential potential with range $\approx \sigma$ which counts up effectively the number of redundant constraints when two spheres are in contact. The size of the attraction is therefore $$F_{attract} \approx \sqrt{ \frac{R}{\sigma}} k_B T.$$ Thus, for the $10nm$ gold spheres considered above, the size of the attraction is approximately $k_B T$, resulting in moderately sized clusters of particles, or indeed a much weaker adsorption of these particles at a solid-polymer interface.
The basic assumption here is that the polymer segments have sufficient time to envelop a particle deposited atop a polymer film [@find1; @find2]. Even when the film itself is glassy, the segregation of free ends in the melt [kumar\_surface]{} and the extra free volume expected at a free surface can result in a significant reduction in both the glass transition temperature and the effects of polymer chain topology [glass\_film]{}.
Conclusion
==========
I have demonstrated an entropic attraction resulting from constrained chain trajectories in the presence of a hard surface which can result in the trapping of polymer particles at both a free surface and at a polymer-polymer interface. The size of the effect can easily be several tens of $k_B T$ per particle for nanoscopic particles, and can even result in a few $k_B T$ attractive short-ranged contact interaction between particles wholly within the polymer matrix.
[20]{} Anna C. Balazs Todd Emrick, Thomas P. Russell, [*Science*]{} [**314**]{}, 1107-1110 (2006).
G.-M. Kim and D.-H. Lee, [*J. Appl. Polym. Sci.*]{}[**82**]{}, 88, (2001).
D. Suppa, O. Kuksenok, A. C. Balazs, and J. M. Yeomans, [*MRS Symp. Proceedings*]{}, [**710**]{}, 61, (2002).
T. L. Morkved, P. Wiltzius, H. M. Jaeger, D. G. Grier and T. A. Witten [*Appl. Phys. Lett.*]{}, [**64**]{}, 422 (1994).
WA Lopes, HM Jaeger, [*Nature*]{}, [**414**]{}(6865), 735 (2001).
D. H. Cole, K. R. Shull, L. E. Rehn and P. Baldo [*Phys. Rev. Letts*]{}, [**78**]{}, 5006, (1997).
Ranjan D. Deshmukh and Russell J. Composto,[*Chem. Mater.*]{}, [**19**]{}(4), 745 (2007).
Ranjan D. Deshmukh and Russell J. Composto,[*Langmuir*]{}, [**23**]{}(26), 13169 (2007).
V. M. Rudoy, O. V. Dement’eva, I. V. Yaminskii, V. M. Sukhov, M. E. Kartseva, and V. A. Ogarev, [*Colloid Journal*]{}, [**64**]{}, 746, (2002).
S. Deki, K. Akamatsu, Y. Hatakenaka, M. Mizuhata, and A. Kajinami [*Nanostructured Mater.*]{}, [ **11**]{}, 59, (1999).
V. Zaporojtchenko, T. Strunskus, and K. Behnke [*J. Adhes. Sci. Technol.*]{}, [ **14**]{}, 250 (2000).
M. S. Kunz, K. R. Shull, and A. Kellock, [*J. Coll. Int. Sci*]{}, [**156**]{}, 240 (1993).
J.-Y. Lee, R. B. Thompson, D. Jasnow, A. C. Balazs, [*Phys. Rev. Letts*]{}, [**89**]{}, 155503 (2002).
A. Silberberg, [*J. Colloid Interface Sci.*]{}, [**90**]{}, 86 (1982).
G. J. Kovacs and P. S. Vincent [*J. Colloid Interface Sci.*]{}, [**90**]{}, 467, (1982).
B. D. Silverman, [ *Macromolecules*]{}, [**24**]{}, 2467, (1991).
F. W. Starr, T. B. Schroder, and S. C. Glotzer, [*Macromolecules*]{}, [**35**]{}, 4481, (2002).
D. H. Cole, K. R. Shull, L. Rehn, P. Baldo, [*Macromolecules*]{}, [**32**]{}, 771 (1999).
S. T. Milner and T. A. Witten, [*Macromolecules*]{}, [**25**]{}, 5495 (1992).
S. Ramakrishan, M. Fuchs, K. S. Schweizer and C. F. Zukoski, [*Langmuir*]{}, [**18**]{}, 1082 (2002).
S. K. Kumar, M. Vacatello, and D. Y. Yoon, [*Macromolecules*]{}, [**23**]{}, 2189 (1990).
J. Q. Pham and P. F. Green [*J. Chem. Phys.*]{}, [**116**]{}, 5801 (2002).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the spectral and timing properties of the accreting millisecond X-ray pulsar observed by , and during the X-ray outburst occurred in April 2018. The spectral properties of the source are consistent with a hard state dominated at high energies by a non-thermal power-law component with a cut-off at $\sim70$ keV. No evidence of iron emission lines or reflection humps has been found. From the coherent timing analysis of the pulse profiles, we derived an updated set of orbital ephemerides. Combining the parameters measured from the three outbursts shown by the source in the last $\sim11$ years, we investigated the secular evolution of the spin frequency and the orbital period. We estimated a neutron magnetic field of $3.1\times 10^{8}\,\,\, \textrm{G}<B_{PC}<4.5\times 10^{8}\,\,\, \textrm{G}$ and measured an orbital period derivative of $-4.1\times 10^{-12}$ s s$^{-1}$ $<\dot{P}_{orb}<7.1\times 10^{-12}$ s s$^{-1}$. We also studied the energy dependence of the pulse profile by characterising the behaviour of the pulse fractional amplitude in the energy range 0.3–80 keV. These results are compared with those obtained from the previous outbursts of and other previously known accreting millisecond X-ray pulsars.'
author:
- |
A. Sanna$^{1}$[^1], F. Pintore$^{2}$, A. Riggio$^{1}$, S. M. Mazzola$^{3}$, E. Bozzo$^{4}$, T. Di Salvo$^{3}$ C. Ferrigno$^{4}$, A. F. Gambino$^{3}$, A. Papitto$^{5}$, R. Iaria$^{3}$, L. Burderi$^{1}$\
$^{1}$Dipartimento di Fisica, Università degli Studi di Cagliari, SP Monserrato-Sestu km 0.7, 09042 Monserrato, Italy\
$^{2}$INAF-Istituto di Astrofisica Spaziale e Fisica Cosmica - Milano, via E. Bassini 15, I-20133 Milano, Italy\
$^{3}$Università degli Studi di Palermo, Dipartimento di Fisica e Chimica, via Archirafi 36, 90123 Palermo, Italy\
$^{4}$ISDC, Department of Astronomy, University of Geneva, Chemin d’Écogia 16, CH-1290 Versoix, Switzerland\
$^{5}$INAF-Osservatorio Astronomico di Roma, Via di Frascati 33, I-00078, Monte Porzio Catone (Roma), Italy
bibliography:
- 'biblio.bib'
date: 'Accepted 2018 August 18. Received 2018 August 18; in original form 2018 June 17'
title: ': spectral and timing properties of its 2018 outburst'
---
\[firstpage\]
Keywords: X-rays: binaries; stars:neutron; accretion, accretion disc,
Introduction
============
is a low-mass X-ray binary discovered on 2007 June 7 during an X-ray outburst observed by the . Follow-up observations carried out with the and the [*Rossi X-ray Timing Explorer*]{} () provided the localisation of the source with an arcsec accuracy and led to the discovery of pulsations at a frequency of $\sim182$ Hz, classifying the source as an accreting millisecond X-ray pulsar [AMXP, see e.g. @Patruno12b for a review], in a 54.7 minutes orbit [@Krimm07]. A second outburst was recorded in July 2009 and the result of the observational campaign carried out with and was reported in @Patruno10b [hereafter P10]. In both occasions, the source displayed a spectral energy distribution compatible with the so-called “island/extreme island state” of an atoll source [see e.g. @Hasinger1989aa and reference therein] and reasonably well described by a model comprising a power-law with a photon index of $\Gamma$=1.8-2.0 with no high-energy cut-off and a black-body component with a temperature of $kT$=0.4-0.7 keV [@Linares08]. Based on the upper limits derived on the spin-down torque, the neutron star magnetic field was constrained in a range compatible with values expected for an AMXP and observed from other sources of this class [0.4$\times$10$^8$ G$<$B$<$9$\times$10$^8$ G; @Patruno10a]. The source was discovered to undergo a new outburst by INTEGRAL on 2018 April 1 [@mereminskiy18]. The event was confirmed by , and follow-up observations provided the detection of pulsations at the known spin period of the source and a preliminary description of its broad-band X-ray spectrum [@krimm18; @bult18a; @bult18b; @cha18; @mazzola18; @kuiper18; @Bult2018c].
In this work, we carried out spectral and coherent timing analysis of the 2018 outburst of , using , , and observations of the source. We updated the source ephemerides and investigated the orbital period evolution over a baseline of almost 11 years by combining the current results with those reported from previous outbursts. We also discuss the broad-band spectral properties of .
Observations and data reduction
===============================
XMM-Newton {#sec:XMM}
----------
observed on 2018 April 8 (Obs.ID. 0830190401) for a total exposure time of $\sim$ 66 ks. During the observation, the EPIC-pn (hereafter PN) camera was operated in [timing]{} mode and [burst]{} mode for $\sim$ 49 ks and $\sim$ 10 ks, respectively. The RGS instrument observed in spectroscopy mode during the entire observation, while the EPIC-MOS1 and EPIC-MOS2 were operated in [full frame]{} and [timing]{} mode, respectively. To perform spectral and timing analysis of the source we focused on the PN and MOS2 data (the limited statistics and time resolution of the MOS1 data did not provide a significant improvement in any of the results presented here and in the following sections). These were processed using the Science Analysis Software (SAS) v. 16.0.0 with the up-to-date calibration files and the RDPHA calibrations [see e.g. @Pintore15a]. We filtered events within the energy range 0.3-10.0 keV, retaining single and double pixel events only (<span style="font-variant:small-caps;">pattern$\leq$4</span>). We extracted the source events for the PN and MOS2 using RAWX=\[29:45\] and RAWX=\[285:325\], respectively. We filtered background events for the PN selecting RAWX=\[3:5\] and we checked that the selected background was not contaminated by the emission from the source. For the MOS2, we extracted the background using an empty circular region of radius 150” from the MOS1 dataset. The mean PN and MOS2 observed count rates during the observation were $\sim22$ cts/s and $\sim4.5$ cts/s, characterised by a slow decreasing trend. The background mean count rate in the PN selected RAWX range is of the order of $\sim0.5$ cts/s (0.3-10.0 keV). Thermonuclear (Type-I) X-ray burst episodes [see e.g. @Strohmayer2010aa for a review] were not detected in the EPIC data. We extracted RGS data with standard procedures. We checked that the RGS1 and RGS2 spectra were consistent and then we merged them with the task [rgscombine]{}.
Fig. \[fig:lc\] shows the monitoring light curve of the 2018 outburst of as seen by (black points) and obtained from the on-line data products tool [@Evans2009a]. The green star represents the beginning of the observation taken few days after the outburst peak. To perform the timing analysis, we reported the PN photon arrival times to the Solar System barycentre by using the <span style="font-variant:small-caps;">barycen</span> tool (DE-405 solar system ephemeris). We applied the best available X-ray position of the source (reported in Tab. \[tab:solution\]) estimated performing astrometric analysis to the available observation of the source [@Evans2009a]. The new source coordinates are compatible, to within the associated uncertainties, with the position reported by @Krimm07.
![ light curve (black points) of the 2018 outburst of the accreting millisecond X-ray pulsar . Data are shown from MJD = 58210 (2018-04-02). Upper limits on the count rate are shown with empty triangles. The green star, red squares, blue diamonds, and purple square represent the starting times of the , , and observations, respectively.[]{data-label="fig:lc"}](lc){width="48.00000%"}
observed twice during its 2018 outburst. The first observation (Obs.ID. 90402313002) started at 08:31 `UT` on 2018 April 8 for an elapsed time of $\sim 85$ks, resulting in a total effective exposure time $\sim43$ks. The second observation (Obs.ID. 90402313004) started at 02:56 `UT` on 2018 April 14 for an elapsed time of $\sim 125$ks, corresponding to a total effective exposure time of $\sim68$ks. The epochs at which observed are shown as red squares in Fig. \[fig:lc\]. We screened and filtered the events with the data analysis software (<span style="font-variant:small-caps;">nustardas</span>) version 1.5.1. We extracted the source events from the FPMA and FPMB focal planes within a circular region of radius 90$''$ centered on the source position. A similarly extended region shifted to a position not contaminated by the source emission was used for the extraction of the background events. For each of the two observations, we obtained the background-subtracted light curves. These are characterised by an average count rate per FPM of $\sim10$ and $\sim 0.001$ counts/s, respectively. During the second observation the source was not significantly detected, and we thus discard these data for the remaining analysis. We corrected the photon arrival times for the motion of the Earth-spacecraft system with respect to the Solar System barycentre with the [barycorr]{} tools (using DE-405 solar system ephemeris), in analogy to what was done for the data.
observed seven times during its 2018 outburst (see Tab. \[tab:obs\] for more details). We extracted events across the 0.2-12 keV band applying standard screening criteria using the HEASOFT version 6.24 and NICERDAS version 4.0. Observations 105023105/6/7 showed the presence of high-energy background features. To further proceed with the timing analysis we excluded (when available) data 50 s before the raise and 100 s after the decay of the flares. We then barycentered the NICER photon arrival times with the [barycorr]{} tool using DE-405 Solar system ephemeris and adopting the source coordinates reported in Tab. \[tab:solution\].
------------ -------------- ------------ ----------
Instrument Obs.ID. Date Exp. (s)
(revolution)
-PN 0830190401 2018-04-08 49072
90402313002 2018-04-08 43457
90402313004 2018-04-14 65763
(1939) 2018-04-07 85000
1050230101 2018-04-03 6716
1050230102 2018-04-04 6424
1050230103 2018-04-07 2201
1050230104 2018-04-08 9490
1050230105 2018-04-09 3861
1050230106 2018-04-10 6141
1050230107 2018-04-11 4470
------------ -------------- ------------ ----------
: Log of the observations of used to perform the spectral and timing analysis.[]{data-label="tab:obs"}
INTEGRAL {#sec:integral}
--------
was observed with [@wink] from 2018 April 7 at 18:58 to 2018 April 8 at 19:56 (UTC), during the satellite revolution 1939. We analysed all data by using version 10.2 of the Off-line Scientific Analysis software (OSA) distributed by the ISDC [@courvoisier03]. The observations are divided into science windows (SCWs), i.e. pointings with typical durations of $\sim$2-3 ks. We analysed a total of 25 SCWs in which the source was located to within an off-axis angle of 3.5 deg from the center of the JEM-X [@lund03] field of view (FoV) and within an off-axis angle of 12 deg from the center of the IBIS/ISGRI [@ubertini03; @lebrun03] FoV. These choices allowed us to minimise the instruments calibration uncertainties[^2].
We extracted first the IBIS/ISGRI and JEM-X mosaics. was detected in the IBIS/ISGRI 20-40 keV and 40-80 keV mosaics at a significance of $20\sigma$ and $13\sigma$, respectively. The corresponding fluxes estimated from the mosaics were 15.3$\pm$0.8 mCrab (roughly 1.2$\times$10$^{-10}$ erg cm$^{-2}$s$^{-1}$) and 9.5$\pm$0.8 mCrab (roughly 7$\times$10$^{-11}$ erg cm$^{-2}$s$^{-1}$). The source was relatively faint for JEM-X and detected at $11\sigma$ in the 3-10 keV mosaic obtained by combining all JEM-X data. The correspondingly estimated flux was 26$\pm$3 mCrab (roughly 4.0$\times$10$^{-10}$ erg cm$^{-2}$s$^{-1}$). We extracted the JEM-X light curves of the source with a bin time of 2 s to search for type-I X-ray bursts, but no significant detection was found.
Data analysis and results
=========================
Spectral analysis
-----------------
We performed a broad-band spectral analysis combining all the available data. In particular, we selected the 2.0-10 keV, 1.2–2.0 keV, 3–70 keV, 30–90 keV and 4–40 keV for PN/MOS2, RGS, and /ISGRI, and JEMX, respectively.
We first fitted these spectra simultaneously adopting a simple [tbabs\*(cutoffpl)]{} model, with the addition of a multiplicative constant to take into account differences in the inter-calibrations of the instruments and the non-simultaneity between the datasets. The fit with this model did not provide an acceptable result ($\chi^2/dof=2903.49/1349$), showing a marked discrepancy between the spectral slopes of the PN and data. This is a well know issue and was already reported in the past [see e.g. @Sanna2017b]. We thus allowed the photon indexes of the PN and spectra to vary independently in the fit. Although the fit was statistically improved ($\chi^2/dof=1872.42/1348$), some residuals were still present and visible especially at the lower energies. We added a soft component (a multicolour black-body disc, [diskbb]{} in [xspec]{}; @Mitsuda84) to the spectral model, which provided an additional significant improvement to the fit ($\chi^2/dof=1648.36/1346$). Assuming a distance of 8.5 kpc (based on the proximity toward the direction of the Galactic center) and an inclination angle of $\leq60\deg$ inferred taking into account the lack of dips and eclipses in the X-ray light-curve [see e.g. @Frank02], we estimated an implausible inner disc radius of $\leq1.2$ km. We thus replaced the [diskbb]{} component with a single-temperature [bbodyrad]{}. The quality of the fit did not change significantly and we measured a black-body temperature of $0.85\pm0.03$ keV. The corresponding emitting radius was estimated at $1.8\pm0.2$ km, compatible with the size of an hot spot on the NS surface. The broad-band spectrum of is shown in Figure \[spec11\], together with the best fit model and the residuals from the fit. All parameters of the best fit model are listed in Tab. \[tab:spectrum\]. We note that no emission lines were detected in the spectra, at odds with the findings reported from the analysis of X-ray data collected during the previous outbursts from the source (P10). The 3$\sigma$ upper limit that we obtained on the equivalent width of an iron emission feature centred at 6.5 keV and characterised by a width of 0.3 keV (see P10) for the 2018 outburst of the source is 5 eV.
{height="12cm"}
. \[spec11\]
Model Parameter
----------- ------------------- ---------------------------
[tbabs]{} nH ($10^{22}$) $8.14^{+0.14}_{-0.15}$
kT (keV) $0.85^{+0.03}_{-0.04}$
Norm $4.5^{+0.8}_{-0.7}$
$\Gamma$(pn) $1.44^{+0.04}_{-0.04}$
$\Gamma$ (NuSTAR) $1.65^{+0.03}_{-0.03}$
$\Gamma$ (MOS2) $1.54^{+0.04}_{-0.04}$
E$_{cut}$ (keV) $75^{+13}_{-10}$
Norm $0.042^{+0.003}_{-0.003}$
$\chi^2/dof$ 1646/1345
: Spectral parameters obtained from the best fit to the data of the 2018 outburst with the model described in the text ([tbabs$\times$(bbodyrad+cutoffpl)]{}).[]{data-label="tab:spectrum"}
Timing analysis {#sec:ta2018}
---------------
To investigate the timing properties of we corrected the delays of the PN photon time of arrivals caused by the X-ray pulsar orbital motion under the hypothesis of a circular orbit. As a starting point, we considered the orbital period ($P_{\text{orb}}=3282.32(3)$ s and the projected semi-major axis (a$\sin(i)$/c=0.00598(2) lt/s) corresponding to the ephemerides obtained from the 2009 outburst of the source (see Tab. 2 in P10). To investigate possible shifts on the time of passage from the ascending node ($T_{\text{NOD}}$), we extrapolated the closest value to the PN observation starting from the value reported in P10 and assuming a constant orbital period. We then explored a grid of parameters spaced by 1 s within a range of few kilo-seconds around the expected value. We searched for pulsations by exploiting the epoch-folding technique on the entire observations using 16 phase bins, starting with $\nu_0$ = 182.065803 Hz and exploring around $\nu_0$ with steps of $10^{-7}$ Hz, for a total of 10001 steps. The pulse profile with the largest signal-to-noise ratio was found at $\nu=182.065803(1)$ Hz and $T_{\text{NOD}}=58216.18423(1)$ MJD.
Starting from the latter orbital solution, we corrected the photon arrival times in the PN and observations and we created pulse profiles by epoch-folding 500 s-long data segments using 8 phase bins. As a starting point, we used the mean spin frequency $\nu=182.065803(1)$ Hz obtained from the preliminary analysis of the PN data. Close to the tail of the outburst, we increased the length of the data segments in order to obtain statistically significant pulse profiles. Each pulse profile was modelled with a sinusoid from which we measured the amplitude and the fractional part of the phase residual. We retained only profiles with ratio between the sinusoidal amplitude and the corresponding $1\sigma$ uncertainty equal or grater than 3. The addition of a second harmonic did not improve the fit to the pulse profiles, being statistically significant in less than 20% of the intervals.
To improve the source ephemeris, we carried out a coherent timing analysis on the combined PN and data by fitting the time evolution of the pulse phase delays with the model: $$\begin{aligned}
\label{eq:ph}
\label{eq:ph_fit}
\Delta \phi(t)=\phi_0+\Delta \nu\,(t-T_0)+\frac{1}{2}\dot{\nu}\,(t-T_0)^2+R_{orb}(t),\end{aligned}$$ where $\phi_0$ represents a constant phase, $\Delta \nu$ is a correction factor on the frequency used to epoch-fold the data, $\dot{\nu}$ represents the spin frequency derivative determined with respect to the reference epoch ($T_0$), and $R_{orb}(t)$ is the residual orbital modulation caused by discrepancies between the *real* set of orbital parameters and those used to correct the photon time of arrivals [see e.g., @Deeter81].
For each new set of orbital parameters obtained from this analysis, we applied the corrections to the photon arrival times and created new pulse phase delays that we modelled with Eq. \[eq:ph\]. We iteratively repeated this process until no significant improvements were found for any of the model parameters. We reported the best-fit parameters in the left column of Tab. \[tab:solution\], while in Fig. \[fig:phase\_fit\] we show the pulse phase delays for the PN and with the best-fitting models. We note that the aforementioned timing solution is compatible within the errors with that reported by @Bult2018c from the analysis of the dataset only.
![*Top panel:* Evolution of the pulse phase delays obtained by epoch-folding 500 s-long intervals of PN and data (shown in blue and green, respectively). Data are shown from MJD $\simeq$ 58211.6 (2018-04-03 14:24:00.0 UTC). The red dotted line represents the best-fit model described in the text, while the light-blue shaded area delimited by the black dotted lines represents the 95% confident region. *Bottom panel:* Residuals in ms with respect to the best-fitting model for the pulse phase delays.[]{data-label="fig:phase_fit"}](phase_xmm_nicer_2){width="48.00000%"}
Using the updated set of ephemerides reported in Tab. \[tab:solution\], we corrected the times of the events and we epoch-folded 800 s-long intervals. We modelled the pulse profiles with a sinusoidal model and we investigated the evolution of the pulse phase delays using Eq. \[eq:ph\]. The best-fit parameters, compatible within the uncertainties with those obtained from the phase-connected timing analysis of the PN and observations, are shown in the right column of Tab. \[tab:solution\].
Parameters PN-
---------------------------------------------------- ------------------------- ----------------------------
R.A. (J2000)
DEC (J2000)
Orbital period $P_{orb}$ (s) 3282.40(4) 3282.4(6)
Projected semi-major axis a sin*i/c* (lt-ms) 5.96(2) 5.98(5)
Ascending node passage $T_{\text{NOD}}$ (MJD) 58216.18433(10) 58216.1841(2)
Eccentricity (e) < 2$\times 10^{-2}$ < 5$\times 10^{-2}$
Spin frequency $\nu_0$ (Hz) 182.06580377(11) 182.065803(1)
Spin frequency 1st derivative $\dot{\nu}_0$ (Hz/s) $<|1.4|\times 10^{-12}$ $-4.3(2.1)\times 10^{-11}$
$\chi^2$/d.o.f. [131.2/126]{} 109.1(65)
![ pulse profiles obtained by epoch-folding the (top panel), (medium panel), and (bottom panel) data. The updated set of ephemerides reported in Tab. \[tab:solution\] have been used together with a sampling of 32 bins. The best-fitting model, obtained by combining two sinusoids with harmonically related periods, is reported in red. Green dot-dashed and blue dot-dot-dashed lines represent the fundamental and the second harmonic pulse profile components, respectively. We show in all cases two pulse cycles for clarity.[]{data-label="fig:pulse_prof"}](pulse_profile){width="48.00000%"}
In Fig.\[fig:pulse\_prof\] we report the best pulse profiles obtained by epoch-folding the PN (top panel), (medium panel), and (bottom panel) data after correcting for the best-fitting parameters reported in Tab. \[tab:solution\]. The average pulse profile differs significantly from a sinusoidal function. It is well described by using a combination of two sinusoids shifted in phase. The () fundamental and second harmonic have background-corrected fractional amplitudes of $\sim$5.6% (4%) and $\sim$3.4% (1.6%), respectively. For the average profile we obtain fractional amplitudes of $\sim$4.7% and $\sim$3.1% for the fundamental and second harmonic (not corrected for the background), respectively.
We also studied the energy dependence of the pulse profile by slicing the PN energy range (0.3–10 keV) in 20 intervals, and the energy range (1.6–80 keV) in 10 intervals. Energy bins have been selected in order to contain the same number of events. For each energy interval, we epoch-folded the events at the spin frequency values reported in Tab. \[tab:solution\] and we approximated the background-subtracted pulse profiles with a model consisting of two sinusoidal components (fundamental and second harmonic) for which we determined the fractional amplitudes. In Fig. \[fig:amp\_vs\_energy\], we show the pulse profile energy dependence of the PN (blue) and (green) fractional amplitude for the fundamental (filled points) and second harmonic (filled squares) components. The PN fundamental component shows an increase from $\sim4\%$ at around 1 keV up to $\sim7\%$ at 6 keV, followed by a plateau around $\sim6\%$ above 10 keV. The second harmonic shows a decreasing trend (almost anti-correlated with the fundamental component) from $\sim4.5\%$ at 1 keV down to $\sim2\%$ at 10 keV. The fundamental component shows an increasing trend between $\sim5\%$ at around 2 keV and $\sim7\%$ up to 15 keV and then it stabilises up to 80 keV. Similarly to the PN data, the second harmonic decreases from $\sim5\%$ at 2 keV down to $\sim2\%$ at 10 keV where it starts increasing up to $\sim4\%$ at 80 keV. We note, however, that the statistics of the data is far from optimal and that future observations with an improved statistic combined with a finer sampling of the high energy region are needed to better investigate both the fundamental and second harmonic pulse fraction trend in this region.
![Energy dependence of the pulse profile fractional amplitude for the fundamental (dots) and second harmonic (squares) components used to model the profiles obtained from the PN (blue) and the (green) datasets.[]{data-label="fig:amp_vs_energy"}](energy_profile){width="48.00000%"}
Discussion
==========
We presented the spectral and timing properties of the AMXP pulsar obtained combining observations collected by , , and during its 2018 outburst.
Spectral properties
-------------------
The spectral results indicate that the source is highly absorbed ($N_{\rm H}$$\sim8.1\times10^{22}$ cm$^{-2}$), in line with previous findings [@Krimm07] and the data taken at the beginning of the outburst [@mereminskiy18]. The $N_{\rm H}$ value we measured is slightly higher than that reported from the data [@bult18a]. Based on the comparison with other known AMXPs, we consider unlikely such large variability in the absorption column density and assume that the most reliable measure is provided by our spectral analysis (which includes data collected by the RGS instrument on-board ).
The photon index of the power-law has changed from $2.04\pm0.03$ during the first days of the outburst (as measured from the data), to $\sim1.5$ during the observation. The detection of a cut-off at $\sim70$ keV strongly indicates that the source was in a hard state, as usually observed for AMXPs in outburst [e.g. @Patruno12b; @Burderi13 for a review]. Note that such a cut-off was not reported in the previous outburst of , when the source spectrum displayed an hard tail extending up to 100 keV [@Linares08].
At odds with previous outbursts (see P10), no significant iron lines were observed in the and data collected during the event in 2018. We note, however, that the poor energy resolution of the data from the previous outbursts did not allow P10 to reliably constrain the line energy, the emissivity index, the inner disc radius, as well as the inclination of the system and the properties of the Compton reflection hump. This makes any comparison with the 2018 outburst particularly challenging. Assuming a line energy of 6.5 keV and a width of 0.3 keV (extrapolated from the spectral residuals reported in Fig. 5 of P10), we estimated an upper limit on the equivalent width of any iron line not detected during the 2018 outburst of the order of 5 eV. We note that no evidence of iron emission lines or reflection humps has been reported also in the cases of the AMXPs IGR J16597$-$3704 [@Sanna2018a], IGR J17379$-$3747 [@Sanna2018b], XTE J1807$-$294 [@Falanga05a], and XTE J1751-305 [@Miller03]. Finally, we measured an absorbed 0.3–70 keV flux of $(2.88\pm 0.01)\times 10^{-9}$ erg cm$^{-2}$ s$^{-1}$ (compatible with the flux values measured few days after the peak of 2007 and 2009 outbursts) and a luminosity of $2.5\times10^{36}$ erg s$^{-1}$, assuming a distance of 8.5 kpc (i.e., about 2$\%$ of the Eddington limit).
Pulse profile and energy dependence
-----------------------------------
We investigated the properties of the pulse profile of as a function of energy by analysing the observations collected with the PN (0.3-10 keV) and (3-80 keV). Since the pulse profile is well described by a combination of two sinusoids (see Fig. \[fig:pulse\_prof\]), we independently studied the fractional amplitude of the fundamental and second harmonic components.
As reported in the top panel of Fig. \[fig:amp\_vs\_energy\], the pulse fractional amplitude estimated from the fundamental component shows a clear increasing trend with energy, varying from 4% to 7% in the energy range 1-6 keV, followed by a slight decrease between 6 and 8 keV that at higher energies levels to $\sim 6$%. This trend is very similar to that reported by P10 for the 2009 outburst, although we notice that the high energy behaviour of the fundamental component inferred from the data suggests a monotonic increase while the observation from the 2018 outburst clearly shows a constant tendency above 10 keV. Similar energy dependence of the fractional amplitude has been already reported for other AMXPs such as Aql X$-$1 [@Casella08], [@Patruno09a; @Sanna2016a], IGR J00291+5934 [with a bit more complex energy dependence in the range 3-10 keV @Falanga05b; @Sanna2017b] and XTE J1807$-$294 [@Kirsch04]. No consensus has been reached in terms of the process responsible for the hard spectrum of the pulsation detected in these sources. However, mechanisms such as a strong Comptonisation of the beamed radiation seems to well describe the properties of a few sources [@Falanga07b]. Nonetheless, it is worth noting that a completely opposite energy dependence of the pulse profile have been observed in other AMXPs such as XTE J1751$-$305 [@Falanga07b], IGR J17511$-$3057 [@Papitto10; @Falanga11; @Riggio11a] and [@Cui98b; @Falanga07b; @Sanna:2017ab].
Finally, the fractional amplitude of the second harmonic (Fig. \[fig:amp\_vs\_energy\] bottom panel), shows a clear decreasing trend from $\sim$4% at 1 keV to $\sim$2% at 10 keV, after which it starts increasing and reaches the values 4% in the highest energy bin. We notice that the corresponding trend reported by P10 for the 2009 outburst shows a slightly weaker fractional amplitude below 10 keV.
Secular spin evolution {#sec:spin}
----------------------
has been observed in outburst three times since its discovery [see @Krimm07; @Patruno10b]. To investigate the secular evolution of the spin frequency we considered the estimates from the 2007 and the 2009 [reported by @Patruno10b see Tab 2 and 3], that we combined with the 2018 spin frequency value reported in Tab. \[tab:solution\].
![[*Top panel:*]{} Differential corrections to the time of passage at the ascending node for the three outbursts of . Data are shown from MJD = 54000 (2006-09-22). The delays are calculated with respect to the first outburst of the source (see Sec. \[sec:orb\_ev\] for more details). Black dots represent the delays from the first two outbursts reported by P10, while the green star is the value obtained combining the PN and observations. The dot-dot-dashed line represents the quadratic model that better describes the data. [*Bottom panel:*]{} Secular evolution of the spin frequency of within the $\sim$11 yr baseline. Frequencies are rescaled with respect to the value $\nu_0 = 182.0658039$ Hz). Black points represent the frequency measurements of the previous outbursts estimated by P10, while the green star represents the spin values obtained from the combined timing analysis of the PN and observations. The dashed line represents the best-fitting linear model, corresponding to a spin down derivative $\dot{\nu}_{SD}=-4.8(6)\times 10^{-16}$ Hz/s, where uncertainties are reported at 1$\sigma$ confidence level.[]{data-label="fig:evo"}](secular_evolution){width="48.00000%"}
We modelled the three spin frequency values with a linear function (see bottom panel in Fig. \[fig:evo\]), obtaining a best-fit with $\chi^2=0.03$ with 1 d.o.f. and a spin frequency derivative of $\dot{\nu}_{sd}=-4.8(6)\times 10^{-16}$ Hz/s, consistent with the upper limit reported by P10. Combining the rotational-energy loss rate to the rotating magnetic dipole emission, we can derive the magnetic field strength at the NS polar caps. Assuming a rotating dipole in presence of matter, the NS magnetic dipole moment can be approximated as $$\mu \simeq 1.05\times10^{26}\left(\frac{1}{1+\sin^2{\alpha}}\right)^{-1/2} I_{45}^{1/2}\nu_{182}^{-3/2}\dot{\nu}_{-16}^{1/2}\,\,\,\, \text{G cm}^3,
\label{eq:mag}$$ where $\alpha$ is the angle between the rotation and magnetic axes [see e.g. @Spitkovsky2006a for more details], $I_{45}$ is the NS moment of inertia in units of $10^{45}$ g cm$^2$, $\nu_{182}$ is the NS spin frequency rescaled for , $\dot{\nu}_{-16}$ is the spin-down frequency derivative in units of $10^{-16}$ Hz/s. Adopting our estimates of the spin frequency and its secular spin-down derivative, and assuming the extreme values $\alpha=0$ deg and $\alpha=90$ deg we can limit the NS magnetic moment to be $2.3\times10^{26}\,\,\, \textrm{G\,cm$^3$}<\mu < 3.3\times10^{26}\,\,\, \textrm{G\,cm$^3$}$. Defining the magnetic field strength at the magnetic caps as $B_{PC}= 2 \mu/R_{NS}^{3}$, and considering a NS radius of $R_{NS}=1.14\times10^{6}$ cm [corresponding to the FPS equation of state for a 1.4 M$_\odot{}$ NS, see e.g., @Friedman1981a; @Pandharipande1989a], we obtain $3.1\times 10^{8}\,\,\, \textrm{G}<B_{PC}<4.5\times 10^{8}\,\,\, \textrm{G}$, consistent with the value reported by @Mukherjee2015 and similar to what has been derived for other AMXPs. It is worth noting that the estimate presented here is likely a lower limit on the magnetic field strength. As a matter of fact, even though no significant spin-up derivative has been reported in the observed outbursts [see e.g. @Krimm07; @Patruno10b], matter has been transferred and accreted on the NS surface likely accelerating the compact object.
Orbital period evolution {#sec:orb_ev}
------------------------
To investigate the secular evolution of the orbital period we used the epoch of passage from the ascending node ($T_{\text{NOD}}$) measured in each of the three outbursts observed from the source, and the corresponding number of elapsed orbital cycles (N) determined with respect to a certain reference time at a specific orbital period. Under the assumption of constant orbital period, the predicted passages from the ascending node $T_{\text{NOD}_{{\text PRE}}}(N)=T_{\text{NOD}_{{\text 0}}}+N\, P_{orb}$ can be determined and compared with the measured ones to calculate the corresponding differential corrections [see e.g. @Papitto05; @diSalvo08; @Hartman08; @Burderi09; @Burderi2010a; @Iaria2015a; @Iaria2014a; @Iaria2018a; @Sanna2016a]. In order to be able to perform a coherent (orbital) timing, we need to verify that we can unambiguously determine the number of elapsed orbital cycles for each $T_{\text{NOD}}$. The condition is thus the following: $$\left(\sigma^2_{T_{\text{NOD}}}+\sigma^2_{P_{orb}} N^2_{\text MAX}+\frac{1}{4}P^2_{orb} \dot{P}^2_{orb} N^4_{\text MAX}\right)^{1/2} < \frac{P_{orb}}{2},
\label{eq:orb_cycles}$$ where $\sigma_{T_{\text{NOD}}}$ and $\sigma_{P_{orb}}$ are the uncertainties on the time of passage from the ascending node and the orbital period used as a reference for the timing solution, respectively. $\dot{P}_{orb}$ is the secular orbital derivative and $N_{\text MAX}$ is the integer number of orbital cycles elapsed by the source during the time interval covered by the three outbursts observed. Specifically, during the period 2007-2018, elapsed $N_{\text MAX}=[(T_{\text{NOD}_{2018}}-T_{\text{NOD}_{2007}})/P_{orb}]\sim 104000$ orbital cycles.
Even considering the most accurate orbital period reported in Tab. \[tab:par\_fit\_orb\], it is clear that despite the possible effects of an orbital period derivative, the condition reported in Eq.\[eq:orb\_cycles\] cannot be satisfied. It is thus not possible to unambiguously associate a number $N$ to all the $T_{\text{NOD}}$ values within the baseline 2007-2018. Instead, we can tentatively phase connect the first two outbursts separated in time by only 2.1 years, corresponding to $N_{\text MAX}\sim 2\times 10^4$. To test Eq.\[eq:orb\_cycles\], we took the 2009 orbital period as a reference and we considered its uncertainty at the 95% confidence level (0.06 s). Given the lack of knowledge on the orbital period derivative, we assumed as a fiducial value the average obtained combining the estimates from the only two AMXPs for which this quantity has been inferred: $\dot{P}_{orb}=3.6(4)\times 10^{-12}$ s/s for [see e.g. @diSalvo08; @Patruno2016a; @Sanna:2017ab] and $\dot{P}_{orb}=1.1(3)\times 10^{-10}$ s/s for [@Sanna2016a], that corresponds to $\dot{P}_{orb}\sim 6\times 10^{-11}$ s/s. Substituting the values into Eq.\[eq:orb\_cycles\], we find the uncertainty on the time of passage from the ascending node to be of the order of $0.35P_{orb}$, which satisfies the possibility to apply coherent orbital timing on the two outbursts. Assuming $\dot{P}_{orb}\sim 6\times 10^{-11}$ s/s, we obtain an improved orbital period $P_{orb,2.1yr}=3282.3503(12)$ s.
Substituting the more accurate estimate of the orbital period into Eq.\[eq:orb\_cycles\] and using the same prescription for $\dot{P}_{orb}$, we obtain that the propagated uncertainty on $T_{\text{NOD}}$ for the 2007-2018 baseline is below the $0.5P_{orb}$ threshold and we are then allowed to coherently phase connect the orbital parameters among the three outbursts. As we can unambiguously associate the number of elapsed orbital cycles to each $T_{\text{NOD}}$, we calculate the correction on the NS passage from the ascending node $\Delta T_{\text{NOD}}$, with respect to the beginning of the 2007 outburst. For each outburst we determine the quantity $T_{\text{NOD}_{{\text obs}}}-T_{\text{NOD}_{{\text PRE}}}$ estimated with respect to $P_{orb,2.1yr}=3282.3503(12)$ s, and we plot it as a function of corresponding elapsed cycles (top panel in Fig. \[fig:evo\]). Using the quadratic function: $$\begin{aligned}
\label{eq:fit_tstar}
\Delta T_{\text{NOD}} = \delta T_{\text{NOD},2007} + N\, \delta P_{orb,2.1yr}+0.5\,N^2\, \dot{P}_{orb}P_{orb,2.1yr},\end{aligned}$$ we determine the unique set of parameters that approximate the $T_{\text{NOD}}$ values shown in Fig. \[fig:evo\], where $\delta T_{\text{NOD} 2007}$ represents the correction to the adopted time of passage from the ascending node, $\delta P_{orb,2.1yr}$ is the correction to the orbital period, and $\dot{P}_{orb}$ is the orbital-period derivative. In the bottom part of Tab. \[tab:par\_fit\_orb\], we report the combined orbital solution of the source and the corresponding uncertainties reported at the $1\sigma$ confidence level.
---------- ------------------ -------------- -------------------------
Outburst $T_{\text{NOD}}$ $P_{orb}$ $\dot{P}_{orb}$
(MJD) (s) ($10^{-12}$ s s$^{-1}$)
2007 54265.28087(10) 3282.41(15) -
2009 55026.03431(3) 3282.32(3) -
2018 58216.18433(10) 3282.40(4) -
Combined 54265.28087(10) 3282.3519(5) $1.5 \pm 2.8$
---------- ------------------ -------------- -------------------------
: [*Top:*]{} best-fitting orbital parameters derived for each individual outburst. The values of the first two outbursts was obtained from P10. [*Bottom:*]{} Best-fitting orbital parameters derived combining the source outbursts observed between 2007 and 2018 (see text for more details). Uncertainties are reported at $1\sigma$ confidence level.[]{data-label="tab:par_fit_orb"}
The uncertainty on the orbital period derivative is such that we cannot determine whether the orbit is secularly expanding or shrinking. However, the longer baseline with respect to P10 allow us to improve by few orders of magnitude the constraint on the strength of the orbital derivative. Already at this stage, we can exclude an orbital evolution similar to that of the AMXP for which an extremely fast expansion has been reported [@Sanna2016a]. On the other hand, considering the 95% confidence level interval $-4.1\times 10^{-12}$ s/s $<\dot{P}_{orb}<7.1\times 10^{-12}$ s/s [see also @Bult2018c], we note that the secular evolution of is still compatible with the fast expansion reported for [see e.g. @diSalvo08; @Patruno2016a; @Sanna:2017ab] as well as with very slow evolution suggested for [see e.g. @Patruno2016b; @Sanna2017b]. Future outbursts will allow us to further constrain the orbital period derivative and the secular evolution of the system.
Conclusions
===========
We reported on the spectral and timing properties of the 2018 outburst of the AMXP observed with , , and . From the phase-connected timing analysis of the and observations, we obtained an updated set of the source ephemerides, compatible within the errors with those obtained from the dataset. Owing to the multiple observations performed during the source outburst, we obtained, for the first time since the decommission of , a reliable constraint on the spin frequency derivative (|$\dot{\nu}|<1.4\times 10^{-12}$ Hz/s) of an AMXP during the accretion state. Combing the timing properties from the previous two outbursts, we estimated a secular spin-down frequency derivative $\dot{\nu}_{sd}=4.8(6)\times 10^{-16}$ Hz/s, compatible with a magnetic field (at the polar caps) of $3.1\times 10^{8}\,\,\, \textrm{G}<B_{PC}<4.5\times 10^{8}\,\,\, \textrm{G}$. Furthermore, we obtained a secular orbital period derivative in the range $-4.1\times 10^{-12}$ s/s $<\dot{P}_{orb}<7.1\times 10^{-12}$ s/s (95% confidence level), suggesting that more outbursts are required to further constrain the orbital evolution of the system. We also investigated the pulsation spectral energy distribution of in the energy range 0.3–10 keV and 3–80 keV, using the and datasets, respectively. The pulse fractional amplitude trend shown by the fundamental and second harmonic components present similarities with those reported for other AMXPs likely suggesting a Comptonisation origin.
Finally, we found that the broad-band (3–90 keV) energy spectrum of observed during its 2018 outburst is well described by an absorbed cut-off power law plus a soft thermal component. A photon index of $\sim$1.5 combined with a cut-off at $\sim$ 70 keV strongly suggest that the source was observed in a hard state. Contrary to previous outbursts, we detected no significant reflection features, with a constraining upper limit on the iron line equivalent width ($\sim$ 5 eV).
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank N. Schartel for providing us with the possibility to perform the ToO observation in the Director Discretionary Time, and the team for the technical support. We also use Director Discretionary Time on , for which we thank Fiona Harrison for approving and the team for the technical support. We acknowledge financial contribution from the agreements ASI-INAF I/037/12/0 and ASI-INAF 2017-14-H.O. This work was partially supported by the Regione Autonoma della Sardegna through POR-FSE Sardegna 2007-2013, L.R. 7/2007, Progetti di Ricerca di Base e Orientata, Project N. CRP-60529. AP acknowledges funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement 660657-TMSP-H2020-MSCA-IF-2014, as well as the International Space Science Institute (ISSIBern) which funded and hosted the international team “The disk magnetosphere interaction around transitional millisecond pulsar”.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: http://www.isdc.unige.ch/integral/analysis
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Optimization techniques are of great importance to effectively and efficiently train a deep neural network (DNN). It has been shown that using the first and second order statistics (e.g., mean and variance) to perform Z-score standardization on network activations or weight vectors, such as batch normalization (BN) and weight standardization (WS), can improve the training performance. Different from these existing methods that mostly operate on activations or weights, we present a new optimization technique, namely gradient centralization (GC), which operates directly on gradients by centralizing the gradient vectors to have zero mean. GC can be viewed as a projected gradient descent method with a constrained loss function. We show that GC can regularize both the weight space and output feature space so that it can boost the generalization performance of DNNs. Moreover, GC improves the Lipschitzness of the loss function and its gradient so that the training process becomes more efficient and stable. GC is very simple to implement and can be easily embedded into existing gradient based DNN optimizers with only one line of code. It can also be directly used to fine-tune the pre-trained DNNs. Our experiments on various applications, including general image classification, fine-grained image classification, detection and segmentation, demonstrate that GC can consistently improve the performance of DNN learning. The code of GC can be found at https://github.com/Yonghongwei/Gradient-Centralization.'
author:
- Hongwei Yong
- Jianqiang Huang
- Xiansheng Hua
- Lei Zhang
title: 'Gradient Centralization: A New Optimization Technique for Deep Neural Networks'
---
Introduction
============
The broad success of deep learning largely owes to the recent advances on large-scale datasets [@russakovsky2015imagenet], powerful computing resources (e.g., GPUs and TPUs), sophisticated network architectures [@he2016deep; @huang2017densely] and optimization algorithms [@bottou1991stochastic; @kingma2014adam]. Among these factors, the efficient optimization techniques, such as stochastic gradient descent (SGD) with momentum [@qian1999momentum], Adagrad [@duchi2011adaptive] and Adam [@kingma2014adam], make it possible to train very deep neural networks (DNNs) with a large-scale dataset, and consequently deliver more powerful and robust DNN models in practice. The generalization performance of the trained DNN models as well as the efficiency of training process depend essentially on the employed optimization techniques.
There are two major goals for a good DNN optimizer: accelerating the training process and improving the model generalization capability. The first goal aims to spend less time and cost to reach a good local minima, while the second goal aims to ensure that the learned DNN model can make accurate predictions on test data. A variety of optimization algorithms [@qian1999momentum; @duchi2011adaptive; @kingma2014adam; @duchi2011adaptive; @kingma2014adam] have been proposed to achieve these goals. SGD [@bottou1991stochastic; @bottou2010large] and its extension SGD with momentum (SGDM) [@qian1999momentum] are among the most commonly used ones. They update the parameters along the opposite direction of their gradients in one training step. Most of the current DNN optimization methods are based on SGD and improve SGD to better overcome the gradient vanishing or explosion problems. A few successful techniques have been proposed, such as weight initialization strategies [@glorot2010understanding; @he2015delving], efficient active functions (e.g., ReLU [@nair2010rectified]), gradient clipping [@pascanu2012understanding; @pascanu2013difficulty], adaptive learning rate optimization algorithms [@duchi2011adaptive; @kingma2014adam], and so on.
In addition to the above techniques, the sample/feature statistics such as mean and variance can also be used to normalize the network activations or weights to make the training process more stable. The representative methods operating on activations include batch normalization (BN) [@ioffe2015batch], instance normalization (IN) [@ulyanov2016instance; @huang2017arbitrary], layer normalization (LN) [@lei2016layer] and group normalization (GN) [@wu2018group]. Among them, BN is the most widely used optimization technique which normalizes the features along the sample dimension in a mini-batch for training. BN smooths the optimization landscape [@Santurkar2018How] and it can speed up the training process and boost model generalization performance when a proper batch size is used [@Zhang2016Deep; @he2016deep]. However, BN works not very well when the training batch size is small, which limits its applications to memory consuming tasks, such as object detection [@he2017mask; @ren2015faster], video classification [@karpathy2014large; @abu2016youtube], etc.
Another line of statistics based methods operate on weights. The representative ones include weight normalization (WN) [@salimans2016weight; @huang2017centered] and weight standardization (WS) [@qiao2019weight]. These methods re-parameterize weights to restrict weight vectors during training. For example, WN decouples the length of weight vectors from their direction to accelerate the training of DNNs. WS uses the weight vectors’ mean and variance to standardize them to have zero mean and unit variance. Similar to BN, WS can also smooth the loss landscape and speed up training. Nevertheless, such methods operating on weight vectors cannot directly adopt the pre-trained models (e.g., on ImageNet) because their weights may not meet the condition of zero mean and unit variance.
Different from the above techniques which operate on activations or weight vectors, we propose a very simple yet effective DNN optimization technique, namely gradient centralization (GC), which operates on the gradients of weight vectors. As illustrated in Fig. \[F:gradient\](a), GC simply centralizes the gradient vectors to have zero mean. It can be easily embedded into the current gradient based optimization algorithms (e.g., SGDM [@qian1999momentum], Adam [@kingma2014adam]) using only one line of code. Though simple, GC demonstrates various desired properties, such as accelerating the training process, improving the generalization performance, and the compatibility for fine-tuning pre-trained models. The main contributions of this paper are highlighted as follows:
- We propose a new general network optimization technique, namely gradient centralization (GC), which can not only smooth and accelerate the training process of DNN but also improve the model generalization performance.
- We analyze the theoretical properties of GC, and show that GC constrains the loss function by introducing a new constraint on weight vector, which regularizes both the weight space and output feature space so that it can boost model generalization performance. Besides, the constrained loss function has better Lipschitzness than the original one, which makes the training process more stable and efficient.
Finally, we perform comprehensive experiments on various applications, including general image classification, fine-grained image classification, object detection and instance segmentation. The results demonstrate that GC can consistently improve the performance of learned DNN models in different applications. It is a simple, general and effective network optimization method.
Related Work
============
In order to accelerate the training and boost the generalization performance of DNNs, a variety of optimization techniques [@ioffe2015batch; @wu2018group; @salimans2016weight; @qiao2019weight; @qian1999momentum; @pascanu2012understanding] have been proposed to operate on activation, weight and gradient. In this section we briefly review the related work from these three aspects.
**Activation:** The activation normalization layer has become a common setting in DNN, such as batch normalization (BN) [@ioffe2015batch] and group normalization (GN) [@wu2018group]. BN was originally introduced to solve the internal covariate shift by normalizing the activations along the sample dimension. It allows higher learning rates [@Bjorck2018Understanding], accelerates the training speed and improves the generalization accuracy [@luo2018towards; @Santurkar2018How]. However, BN does not perform well when the training batch size is small, and GN is proposed to address this problem by normalizing the activations or feature maps in a divided group for each input sample. In addition, layer normalization (LN) [@lei2016layer] and instance normalization (IN) [@ulyanov2016instance; @huang2017arbitrary] have been proposed for RNN and style transfer learning, respectively.
**Weight:** Weight normalization (WN) [@salimans2016weight] re-parameterizes the weight vectors and decouples the length of a weight vector from its direction. It speeds up the convergence of SGDM algorithm to a certain degree. Weight standardization (WS) [@qiao2019weight] adopts the Z-score standardization to re-parameterize the weight vectors. Like BN, WS can also smooth the loss landscape and improve training speed. Besides, binarized DNN [@rastegari2016xnor; @courbariaux2016binarized; @courbariaux2015binaryconnect] quantifies the weight into binary values, which can improve the generalization capability for certain DNNs. However, a shortcoming of those methods operating on weights is that they cannot be directly used to fine-tune pre-trained models since the pre-trained weight may not meet their constraints. As a consequence, we have to design specific pre-training methods for them in order to fine-tune the model.
**Gradient:** A commonly used operation on gradient is to compute the momentum of gradient [@qian1999momentum]. By using the momentum of gradient, SGDM accelerates SGD in the relevant direction and dampens oscillations. Besides, $L_2$ regularization based weight decay, which introduces $L_2$ regularization into the gradient of weight, has long been a standard trick to improve the generalization performance of DNNs [@krogh1992simple; @zhang2018three]. To make DNN training more stable and avoid gradient explosion, gradient clipping [@pascanu2012understanding; @pascanu2013difficulty; @abadi2016deep; @kim2016accurate] has been proposed to train a very deep DNNs. In addition, the projected gradient methods [@gupta2018cnn; @larsson2017projected] and Riemannian approach [@cho2017riemannian; @vorontsov2017orthogonality] project the gradient on a subspace or a Riemannian manifold to regularize the learning of weights.
Gradient Centralization
=======================
Motivation
----------
BN [@ioffe2015batch] is a powerful DNN optimization technique, which uses the first and second order statistics to perform Z-score standardization on activations. It has been shown in [@Santurkar2018How] that BN reduces the Lipschitz constant of loss function and makes the gradients more Lipschitz smooth so that the optimization landscape becomes smoother. WS [@qiao2019weight] can also reduce the Lipschitzness of loss function and smooth the optimization landscape through Z-score standardization on weight vectors. BN and WS operate on activations and weight vectors, respectively, and they implicitly constrict the gradient of weights, which improves the Lipschitz property of loss for optimization.
Apart from operating on activation and weight, can we directly operate on gradient to make the training process more effective and stable? One intuitive idea is that we use Z-score standardization to normalize gradient, like what has been done by BN and WS on activation and weight. Unfortunately, we found that normalizing gradient cannot improve the stability of training. Instead, we propose to compute the mean of gradient vectors and centralize the gradients to have zero mean. As we will see in the following development, the so called gradient centralization (GC) method can have good Lipschitz property, smooth the DNN training and improve the model generalization performance.
Notations
----------
We define some basic notations. For fully connected layers (FC layers), the weight matrix is denoted as $\mathbf{W}_{fc} \in \mathbb{R}^{C_{in}\times C_{out}}$, and for convolutional layers (Conv layers) the weight tensor is denoted as $\mathbf{W}_{conv} \in \mathbb{R}^{C_{in}\times C_{out}\times(k_1k_2)}$, where $C_{in}$ is the number of input channels, $C_{out}$ is the number of output channels, and $k_1$, $k_2$ are the kernel size of convolution layers. For the convenience of expression, we unfold the weight tensor of Conv layer into a matrix/tensor and use a unified notation $\mathbf{W} \in \mathbb{R}^{M\times N}$ for weight matrix in FC layer ($\mathbf{W}\in \mathbb{R}^{C_{in}\times C_{out}}$) and Conv layers ($\mathbf{W}\in \mathbb{R}^{(C_{in}k_1k_2)\times C_{out}}$). Denote by $\mathbf{w}_i\in\mathbb{R}^{M}$ ($i=1,2,...,N$) the $i$-th column vector of weight matrix $\mathbf{W}$ and $\mathcal{L}$ the objective function. $\nabla_{\mathbf{W}}\mathcal{L}$ and $\nabla_{\mathbf{w}_i}\mathcal{L}$ denote the gradient of $\mathcal{L}$ w.r.t. the weight matrix $\mathbf{W}$ and weight vector $\mathbf{w}_i$, respectively. The size of gradient matrix $\nabla_{\mathbf{W}}\mathcal{L}$ is the same as weight matrix $\mathbf{W}$. Let $\mathbf{X}$ be the input activations for this layer and $\mathbf{W}^T\mathbf{X}$ be its output activations. $\mathbf{e}=\frac{1}{\sqrt{M}}\mathbf{1}$ denotes an $M$ dimensional unit vector and $\mathbf{I}\in\mathbb{R}^{M\times M}$ denotes an identity matrix.
Formulation of GC
-----------------
For a FC layer or a Conv layer, suppose that we have obtained the gradient through backward propagation, then for a weight vector $\mathbf{w}_{i}$ whose gradient is $\nabla_{\mathbf{w}_{i}}\mathcal{L}$ ($i=1,2,...,N$), the GC operator, denoted by $\Phi_{GC}$, is defined as follows: $$\begin{aligned}
\Phi_{GC}(\nabla_{\mathbf{w}_{i}}\mathcal{L})=\nabla_{\mathbf{w}_{i}}\mathcal{L}-\mu_{\nabla_{\mathbf{w}_{i}}\mathcal{L}}
\end{aligned}\label{GZ_vector}
\vspace{-2mm}$$ where $\mu_{\nabla_{\mathbf{w}_{i}}\mathcal{L}}=\frac{1}{M}\sum_{j=1}^M \nabla_{\mathbf{W}_{i,j}}\mathcal{L}$. The formulation of GC is very simple. As shown in Fig. \[F:gradient\](b), we only need to compute the mean of the column vectors of the weight matrix, and then remove the mean from each column vector. We can also have a matrix formulation of Eq. (\[GZ\_vector\]): $$\begin{aligned}
\Phi_{GC}(\nabla_{\mathbf{W}}\mathcal{L})=\mathbf{P}\nabla_{\mathbf{W}}\mathcal{L}, \ \ \ \ \ \mathbf{P}=\mathbf{I}-\mathbf{e}\mathbf{e}^{T}
\end{aligned}\label{GZ_matrix}
\vspace{-1mm}$$
The physical meaning of $\mathbf{P}$ will be explained later in Section \[Imp\_Gen\]. In practical implementation, we can directly remove the mean value from each weight vector to accomplish the GC operation. The computation is very simple and efficient.
Embedding of GC to SGDM/Adam
----------------------------
GC can be easily embedded into the current DNN optimization algorithms such as SGDM [@qian1999momentum; @bottou2010large] and Adam [@kingma2014adam]. After obtaining the centralized gradient $\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})$, we can directly use it to update the weight matrix. Algorithm \[alg1\] and Algorithm \[alg2\] show how to embed GC into the two most popular optimization algorithms, SGDM and Adam, respectively. Moreover, if we want to use weight decay, we can set $\mathbf{\widehat{g}}^t=\mathbf{P}(\mathbf{g}^t+\lambda\mathbf{w})$, where $\lambda$ is the weight decay factor. It only needs to add one line of code into most existing DNN optimization algorithms to execute GC with negligible additional computational cost. For example, it costs only $0.6$ sec extra training time in one epoch on CIFAR100 with ResNet50 model in our experiments (71 sec for one epoch).
[2]{}
Weight vector $\mathbf{w}^0$, step size $\alpha$, momentum factor $\beta$, $\mathbf{m}^0$
$\mathbf{g}^t=\nabla_{\mathbf{w}^t}\mathcal{L}$ $\mathbf{\widehat{g}}^t=\Phi_{GC}(\mathbf{g}^t)$ $\mathbf{m}^t=\beta\mathbf{m}^{t-1}+(1-\beta)\mathbf{\widehat{g}}^t$ $\mathbf{w}^{t+1}=\mathbf{w}^t-\alpha\mathbf{m}_t$
\[alg1\]
[2]{}
Weight vector $\mathbf{w}^0$, step size $\alpha$, $\beta_1$, $\beta_2$, $\epsilon$, $\mathbf{m}^0$,$\mathbf{v}^0$
$\mathbf{g}^t=\nabla_{\mathbf{w}^t}\mathcal{L}$ $\mathbf{\widehat{g}}^t=\Phi_{GC}(\mathbf{g}^t)$ $\mathbf{m}^t=\beta_1\mathbf{m}^{t-1}+(1-\beta_1)\mathbf{\widehat{g}}^t$ $\mathbf{v}^t=\beta_2\mathbf{v}^{t-1}+(1-\beta_2)\mathbf{\widehat{g}}^t\odot\mathbf{\widehat{g}}^t$ $\mathbf{\widehat{m}}^t=\mathbf{m}^t/(1-(\beta_1)^t)$ $\mathbf{\widehat{v}}^t=\mathbf{v}^t/(1-(\beta_2)^t)$ $\mathbf{w}^{t+1}=\mathbf{w}^t-\alpha\frac{\mathbf{\widehat{m}}^t}{\sqrt{\mathbf{\widehat{v}}^t}+\epsilon}$
\[alg2\]
Properties of GC
================
As we will see in the section of experimental result, GC can accelerate the training process and improve the generalization performance of DNNs. In this section, we perform theoretical analysis to explain why GC works.
Improving Generalization Performance {#Imp_Gen}
------------------------------------
One important advantage of GC is that it can improve the generalization performance of DNNs. We explain this advantage from two aspects: weight space regularization and output feature space regularization.
**Weight space regularization:** Let’s first explain the physical meaning of $\mathbf{P}$ in Eq.(\[GZ\_matrix\]). Actually, it is easy to prove that: $$\begin{aligned}
\mathbf{P}^2=\mathbf{P}=\mathbf{P}^T, \ \ \ \ \ \ \mathbf{e}^T\mathbf{P}\nabla_{\mathbf{W}}\mathcal{L}=0.
\end{aligned}
\vspace{-2mm}$$ The above equations show that $\mathbf{P}$ is the projection matrix for the hyperplane with normal vector $\mathbf{e}$ in weight space, and $\mathbf{P}\nabla_{\mathbf{W}}\mathcal{L}$ is the projected gradient.
![The geometrical interpretation of GC. The gradient is projected on a hyperplane $\mathbf{e}^T(\mathbf{w}-\mathbf{w}^t)=0$, where the projected gradient is used to update the weight.[]{data-label="F:Project"}](projectedgrad.png){width="65.00000%"}
The property of projected gradient has been investigated in some previous works [@gupta2018cnn; @larsson2017projected; @cho2017riemannian; @vorontsov2017orthogonality], which indicate that projecting the gradient of weight will constrict the weight space in a hyperplane or a Riemannian manifold. Similarly, the role of GC can also be viewed from the perspective of projected gradient descant. We give a geometric illustration of SGD with GC in Fig. \[F:Project\]. As shown in Fig. \[F:Project\], in the $t$-th step of SGD with GC, the gradient is first projected on the hyperplane determined by $\mathbf{e}^T(\mathbf{w}-\mathbf{w}^t)=0$, where $\mathbf{w}^t$ is the weight vector in the $t$-th iteration, and then the weight is updated along the direction of projected gradient $-\mathbf{P}\nabla_{\mathbf{w}^t}\mathcal{L}$. From $\mathbf{e}^T(\mathbf{w}-\mathbf{w}^t)=0$, we have $\mathbf{e}^T\mathbf{w}^{t+1}=\mathbf{e}^T\mathbf{w}^t=...=\mathbf{e}^T\mathbf{w}^0$, i.e., $\mathbf{e}^T\mathbf{w}$ is a constant during training. Mathematically, the latent objective function w.r.t. one weight vector $\mathbf{w}$ can be written as follows: $$\begin{aligned}
\min_{\mathbf{w}}\mathcal{L}(\mathbf{w}), \ \ \ \ s.t. \ \ \ \ \mathbf{e}^T(\mathbf{w}-\mathbf{w}^0)=0
\end{aligned}\label{latent_fun}
\vspace{-3mm}$$ Clearly, this is a constrained optimization problem on weight vector $\mathbf{w}$. It regularizes the solution space of $\mathbf{w}$, reducing the possibility of over-fitting on training data. As a result, GC can improve the generalization capability of trained DNN models, especially when the number of training samples is limited.
It is noted that WS [@qiao2019weight] uses a constraint $\mathbf{e}^T\mathbf{w}=0$ for weight optimization. It reparameterizes weights to meet this constraint. However, this constraint largely limits its practical applications because the initialized weight may not satisfy this constraint. For example, a pretrained DNN on ImageNet usually cannot meet $\mathbf{e}^T\mathbf{w}^0=0$ for its initialized weight vectors. If we use WS to fine-tune this DNN, the advantages of pretrained models will disappear. Therefore, we have to retrain the DNN on ImageNet with WS before we fine-tune it. This is very cumbersome. Fortunately the weight constraint of GC in Eq. (\[latent\_fun\]) fits any initialization of weight, e.g., ImageNet pretrained initialization, because it involves the initialized weight $\mathbf{w}^0$ into the constraint so that $\mathbf{e}^T(\mathbf{w}^0-\mathbf{w}^0)=0$ is always true. This greatly extends the applications of GC.
**Output feature space regularization:** For SGD based algorithms, we have $\mathbf{w}^{t+1}=\mathbf{w}^{t}-\alpha^t\mathbf{P}\nabla_{\mathbf{w}^t}\mathcal{L}$. It can be derived that $\mathbf{w}^{t}=\mathbf{w}^{0}-\mathbf{P}\sum_{i=0}^{t-1}\alpha^{(i)}\nabla_{\mathbf{w}^{(i)}}\mathcal{L}$. For any input feature vector $\mathbf{x}$, we have the following theorem:
**Theorem 4.1:**\[Theorem2\] *Suppose that SGD (or SGDM) with GC is used to update the weight vector $\mathbf{w}$, for any input feature vectors $\mathbf{x}$ and $\mathbf{x}+\gamma\mathbf{1}$, we have $$\begin{aligned}
({\mathbf{w}^{t}})^T\mathbf{x}-({\mathbf{w}^{t}})^T(\mathbf{x}+\gamma\mathbf{1})=\gamma\mathbf{1}^T\mathbf{w}^0
\end{aligned}
\vspace{-2mm}$$ where $\mathbf{w}^0$ is the initial weight vector and $\gamma$ is a scalar.*
Please find the proof in the **Appendix**. Theorem \[Theorem2\] indicates that a constant intensity change (i.e., $\gamma\mathbf{1}$) of an input feature causes a change of output activation; interestingly, this change is only related to $\gamma$ and $\mathbf{1}^T\mathbf{w}^0$ but not the current weight vector $\mathbf{w}^t$. $\mathbf{1}^T\mathbf{w}^0$ is the scaled mean of the initial weight vector $\mathbf{w}^0$. In particular, if the mean of $\mathbf{w}^0$ is close to zero, then the output activation is not sensitive to the intensity change of input features, and the output feature space becomes more robust to training sample variations.
Indeed, the mean of $\mathbf{w}^0$ is very close to zero by the commonly used weight initialization strategies, such as Xavier initialization [@glorot2010understanding], Kaiming initialization [@he2015delving] and even ImageNet pre-trained weight initialization. Fig. \[F:weight\_init\] shows the absolute value (log scale) of the mean of weight vectors for Conv layers in ResNet50 with Kaiming normal initialization and ImageNet pre-trained weight initialization. We can see that the mean values of most weight vectors are very small and close to zero (less than $e^{-7}$). This ensures that if we train the DNN model with GC, the output features will not be sensitive to the variation of the intensity of input features. This property regularizes the output feature space and boosts the generalization performance of DNN training.
![The absolute value (log scale) of the mean of weight vectors for convolution layers in ResNet50. The $x$-axis is the weight vector index. We plot the mean value of different convolution layers from left to right with the order from sallow to deep layers. Kaiming normal initialization [@he2015delving] (top) and ImageNet pre-trained weight initialization (bottom) are employed here. We can see that the mean values are usually very small (less than $e^{-7}$) for most of the weight vectors. []{data-label="F:weight_init"}](cweightinitc.png){width="57.00000%"}
Accelerating Training Process
-----------------------------
**Optimization landscape smoothing:** It has been shown in [@Santurkar2018How; @qiao2019weight] that both BN and WS smooth the optimization landscape. Although BN and WS operate on activations and weights, they implicitly constrict the gradient of weights, making the gradient of weight more predictive and stable for fast training. Specifically, BN and WS use the gradient magnitude $||\nabla f(\mathbf{x})||_2$ to capture the Lipschitzness of function $f(\mathbf{x})$. For the loss and its gradients, $f(\mathbf{x})$ will be $\mathcal{L}$ and $\nabla_{\mathbf{w}}\mathcal{L}$, respectively, and $\mathbf{x}$ will be $\mathbf{w}$. The upper bounds of $||\nabla_{\mathbf{w}}\mathcal{L}||_2$ and $||\nabla^2_{\mathbf{w}}\mathcal{L}||_2$ ($\nabla^2_{\mathbf{w}}\mathcal{L}$ is the Hessian matrix of $\mathbf{w}$) have been given in [@Santurkar2018How; @qiao2019weight] to illustrate the optimization landscape smoothing property of BN and WS. Similar conclusion can be made for our proposed GC by comparing the Lipschitzness of original loss function $\mathcal{L}(\mathbf{w})$ with the constrained loss function in Eq. (\[latent\_fun\]) and the Lipschitzness of their gradients. We have the following theorem:
**Theorem 4.2:**\[Theorem1\] *Suppose $\nabla_{\mathbf{w}}\mathcal{L}$ is the gradient of loss function $\mathcal{L}$ w.r.t. weight vector $\mathbf{w}$. With the $\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})$ defined in Eq.(\[GZ\_matrix\]), we have the following conclusion for the loss function and its gradient, respectively: $$\left\{
\begin{aligned}
&||\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})||_2\leq||\nabla_{\mathbf{w}}\mathcal{L}||_2,\\
&||\nabla_{\mathbf{w}}\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})||_2\leq||\nabla^2_{\mathbf{w}}\mathcal{L}||_2.
\end{aligned}
\right.
\vspace{-4mm}$$*
The proof of Theorem \[Theorem1\] can be found in the **Appendix**. Theorem \[Theorem1\] shows that for the loss function $\mathcal{L}$ and its gradient $\nabla_{\mathbf{w}}\mathcal{L}$, the constrained loss function in Eq. (\[latent\_fun\]) by GC leads to a better Lipschitzness than the original loss function so that the optimization landscape becomes smoother. This means that GC has similar advantages to BN and WS on accelerating training. A good Lipschitzness on gradient implies that the gradients used in training are more predictive and well-behaved so that the optimization landscape can be smoother for faster and more effective training.
![The $L_2$ norm (log scale) and max value (log scale) of gradient matrix or tensor vs. iterations. ResNet50 trained on CIFAR100 is used as the DNN model here. The left two sub-figures show the results on the first Conv layer and the right two show the FC layer. The red points represent the results of training without GC and the blue points represent the results with GC. We can see that GC largely reduces the $L_2$ norm and max value of gradient. []{data-label="F_grad_clip"}](gradclipc.png){width="100.00000%"}
**Gradient explosion suppression:** Another benefit of GC for DNN training is that GC can avoid gradient explosion and make training more stable. This property is similar to gradient clipping [@pascanu2012understanding; @pascanu2013difficulty; @kim2016accurate; @abadi2016deep]. Too large gradients will make the weights change abruptly during training so that the loss may severely oscillate and hard to converge. It has been shown that gradient clipping can suppress large gradient so that the training can be more stable and faster [@pascanu2012understanding; @pascanu2013difficulty]. There are two popular gradient clipping approaches: element-wise value clipping [@pascanu2012understanding; @kim2016accurate] and norm clipping [@pascanu2013difficulty; @abadi2016deep], which apply thresholding to element-wise value and gradient norm to gradient matrix, respectively. In order to investigate the influence of GC on clipping gradient, in Fig. \[F\_grad\_clip\] we plot the max value and $L_2$ norm of gradient matrix of the first convolutional layer and the fully-connected layer in ResNet50 (trained on CIFAR100) with and without GC. It can be seen that both the max value and the $L_2$ norm of the gradient matrix become smaller by using GC in training. This is in accordance to our conclusion in Theorem \[Theorem1\] that GC can make training process smoother and faster.
Experimental Results
====================
Setup of Experiments
--------------------
Extensive experiments are performed to validate the effectiveness of GC. To make the results as comprehensive and clear as possible, we arrange the experiments as follows:
- We start from experiments on the Mini-ImageNet dataset [@vinyals2016matching] to demonstrate that GC can accelerate the DNN training process and improve the model generalization performance. We also evaluate the combinations of GC with BN and WS to show that GC can improve them for DNN optimization.
- We then use the CIFAR100 dataset [@krizhevsky2009learning] to evaluate GC with various DNN optimizers (e.g., SGDM, Adam, Adagrad), various DNN architectures (e.g., ResNet, DenseNet, VGG), and and different hyper-parameters.
- We then perform experiments on ImageNet [@russakovsky2015imagenet] to demonstrate that GC also works well on large scale image classification, and show that GC can also work well with normalization methods other than BN, such as GN.
- We consequently perform experiments on four fine-grained image classification datasets (FGVC Aircraft [@maji2013fine], Stanford Cars [@krause20133d], Stanford Dogs [@khosla2011novel] and CUB-200-2011 [@wah2011caltech]) to show that GC can be directly adopted to fine-tune the pre-trained DNN models and improve them.
- At last, we perform experiments on the COCO dataset [@lin2014microsoft] to show that GC also works well on other tasks such as objection detection and segmentation.
GC can be applied to either Conv layer or FC layer, or both of them. In all of our following experiments, if not specified, we always apply GC to both Conv and FC layers. Except for Section \[exp\_cifar\] where we embed GC into different DNN optimizers for test, in all other sections we embed GC into SGDM for experiments, and the momentum is set to 0.9. All experiments are conducted under the Pytorch 1.3 framework and the GPUs are NVIDIA Tesla P100.
We would like to stress that no additional hyper-parameter is introduced in our GC method. Only one line of code is needed to embed GC into the existing optimizers, while keeping all other settings remain unchanged. We compare the performances of DNN models trained with and without GC to validate the effectiveness of GC.
![Training loss (left) and testing accuracy (right) curves vs. training epoch on the Mini-ImageNet. The ResNet50 is used as the DNN model. The compared optimization techniques include BN, BN+GC, BN+WS and BN+WS+GC.[]{data-label="F_miniimagenet"}](minilargebn.png){width="72.00000%"}
Results on Mini-Imagenet
------------------------
Mini-ImageNet [@vinyals2016matching] is a subset of the ImageNet dataset [@russakovsky2015imagenet], which was originally proposed for few shot learning. We use the train/test splits provided by [@ravi2016optimization; @iscen2019label]. It consists of 100 classes and each class has 500 images for training and 100 images for testing. The image resolution is $84\times 84$. We resize the images into $224\times 224$, which is the standard ImageNet training input size. The DNN we used here is ResNet50, which is trained on 4 GPUs with batch size 128. Other settings are the same as training ImageNet. We repeat the experiments for 10 times and report the average results over the 10 runs.
BN, WS and GC operate on activations, weights and gradients, respectively, and they can be used together to train DNNs. Actually, it is necessary to normalize the feature space by methods such as BN; otherwise, the model is hard to be well trained. Therefore, we evaluate four combinations here: BN, BN+GC, BN+WS and BN+WS+GC. The optimizer is SGDM with momentum 0.9. Fig. $\ref{F_miniimagenet}$ presents the training loss and testing accuracy curves of these four combinations. Compared with BN, the training loss of BN+GC decreases much faster and the testing accuracy increases more rapidly. For both BN and BN+WS, GC can further speed up their training speed. Moreover, we can see that BN+GC achieves the highest testing accuracy, validating that GC can accelerate training and enhance the generalization performance simultaneously.
Experiments on CIFAR100 {#exp_cifar}
-----------------------
Model R18 R101 X29 V11 D121
-------- -------------------- -------------------- -------------------- -------------------- --------------------
w/o GC 76.87$\pm$0.26 78.82$\pm$ 0.42 79.70$\pm$0.30 70.94$\pm$ 0.34 79.31$\pm$0.33
w/ GC **78.82$\pm$0.31** **80.21$\pm$0.31** **80.53$\pm$0.33** **71.69$\pm$0.37** **79.68$\pm$0.40**
: Testing accuracies of different DNN models on CIFAR100
\[T:dnn\_models\] \[table1\]
Algorithm SGDM Adam Adagrad SGDW AdamW
----------- -------------------- -------------------- -------------------- -------------------- --------------------
w/o GC 78.23$\pm$0.42 71.64$\pm$0.56 70.34 $\pm$0.31 74.02$\pm$0.27 74.12$\pm$0.42
w/ GC **79.14$\pm$0.33** **72.80$\pm$0.62** **71.58$\pm$0.37** **76.82$\pm$0.29** **75.07$\pm$0.37**
: Testing accuracies of different optimizers on CIFAR100
\[table1\]
CIFAR100 [@krizhevsky2009learning] consists of 50K training images and 10K testing images from 100 classes. The size of input image is $32\times32$. Since the image resolution is small, we found that applying GC to the Conv layer is good enough on this dataset. All DNN models are trained for 200 epochs using one GPU with batch size 128. The experiments are repeated for 10 times and the results are reported in mean $\pm$ std format.
**Different networks:** We testify GC on different DNN architectures, including ResNet18 (R18), ResNet101 (R101) [@he2016deep], ResNeXt29 4x64d (X29) [@xie2017aggregated], VGG11 (V11) [@simonyan2014very] and DenseNet121 (D121) [@huang2017densely]. SGDM is used as the network optimizer. The weight decay is set to 0.0005. The initial learning rate is $0.1$ and it is multiplied by $0.1$ for every 60 epochs. Table \[T:dnn\_models\] shows the testing accuracies of these DNNs. It can be seen that the performance of all DNNs is improved by GC, which verifies that GC is a general optimization technique for different DNN architectures.
**Different optimizers:** We embed GC into different DNN optimizers, including SGDM [@qian1999momentum], Adagrad [@duchi2011adaptive], Adam [@kingma2014adam], SGDW and AdamW [@loshchilov2017decoupled], to test their performance. SGDW and AdamW optimizers directly apply weight decay on weight, instead of using $L_2$ weight decay regularization. Weight decay is set to $0.001$ for SGDW and AdamW, and $0.0005$ for other optimizers. The initial learning rate is set to $0.1$, $0.01$ and $0.001$ for SGDM/SGDW, Adagrad, Adam/AdamW, respectively, and the learning rate is multiplied by $0.1$ for every 60 epochs. The other hyper-parameters are set by their default settings on Pytorch. The DNN model used here is ResNet50. Table \[table1\] shows the testing accuracies. It can be seen that GC boosts the generalization performance of all the five optimizers. It is also found that adaptive learning rate based algorithms Adagrad and Adam have poor generalization performance on CIFAR100, while GC can improve their performance by $>0.9\%$.
**Different hyper-parameter settings:** In order to illustrate that GC can achieve consistent improvement with different hyper-parameters, we present the results of GC with different settings of weight decay and learning rates on the CIFAR100 dataset. ResNet50 is used as the backbone. Table \[T:dnn\_weightdecay\] shows the testing accuracies with different settings of weight decay, including 0, $1e^{-4}$, $2e^{-4}$, $5e^{-4}$ and $1e^{-3}$. The optimizer is SGDM with learning rate $0.1$. It can be seen that the performance of weight decay is consistently improved by GC. Table \[T:dnn\_Lr\] shows the testing accuracies with different learning rates for SGDM and Adam. For SGDM, the learning rates are $0.05$, $0.1$ and $0.2$, and for Adam, the learning rates are $0.0005$, $0.001$ and $0.0015$. The weight decay is set to $5e^{-4}$. Other settings are the same as those in the manuscript. We can see that GC consistently improves the performance.
Weight decay 0 $1e^{-4}$ $2e^{-4}$ $5e^{-4}$ $1e^{-3}$
-------------- -------------------- -------------------- -------------------- -------------------- -------------------- --
w/o GC 71.62$\pm$0.31 73.91$\pm$0.35 75.57$\pm$0.33 78.23$\pm$0.42 77.43$\pm$0.30
w/ GC **72.83$\pm$0.29** **76.56$\pm$0.31** **77.62$\pm$0.37** **79.14$\pm$0.33** **78.10$\pm$0.36**
: Testing accuracies of different weight decay on CIFAR100 with ResNet50.
\[T:dnn\_weightdecay\]
Algorithm SGDM SGDM SGDM Adam Adam Adam
--------------- -------------------- -------------------- -------------------- -------------------- -------------------- --------------------
Learning rate 0.05 0.1 0.2 0.0005 0.001 0.0015
w/o GC 76.81$\pm$0.27 78.23$\pm$0.42 76.53$\pm$0.32 73.88$\pm$0.46 71.64$\pm$0.56 **70.63$\pm$0.44**
w/ GC **78.12$\pm$0.33** **79.14$\pm$0.33** **77.71$\pm$0.35** **74.32$\pm$0.55** **72.80$\pm$0.62** **71.22$\pm$0.49**
: Testing accuracies of different learning rates on CIFAR100 with ResNet50 for SGDM and Adam.
\[T:dnn\_Lr\]
Datesets R50BN R50GN R101BN R101GN
---------- ----------- ----------- ----------- ----------- --
w/o GC 23.71 24.50 22.37 23.34
w/ GC **23.21** **23.53** **21.82** **22.14**
: Top-1 error rates on ImageNet w/o GC and w/ GC.
\[table1\_imagenet\]
![Training error (left) and validation error (right) curves vs. training epoch on ImageNet. The DNN model is ResNet50 with GN.[]{data-label="F_imagenet"}](imagnetr50gn2.png){width="75.00000%"}
Results on ImageNet
-------------------
We then evaluate GC on the large-scale ImageNet dataset [@russakovsky2015imagenet] which consists of $1.28$ million images for training and 50K images for validation from 1000 categories. We use the common training settings and embed GC to SGDM on Conv layer. The ResNet50 and ResNet101 are used as the backbone networks. For the former, we use 4 GPUs with batch size 64 per GPU, and for the latter, we use 8 GPUs with batch size 32 per GPU.
We evaluate four models here: ResNet50 with BN (R50BN), ResNet50 with GN (R50GN), ResNet101 with BN (R101BN) and ResNet101 with GN (R101GN). Table \[table1\_imagenet\] shows the final Top-1 errors of these four DNN models trained with and without GC. We can see that GC can improve the performance by $0.5\%\sim 1.2\%$ on ImageNet. Fig. \[F\_imagenet\] plots the training and validation error curves of ResNet50 (GN is used for feature normalization). We can see that GC can largely speed up the training with GN.
Datasets \#Category \#Training \#Testing
--------------- ------------ ------------ -----------
FGVC Aircraft 100 6,667 3,333
Stanford Cars 196 8,144 8,041
Stanford Dogs 120 12,000 8,580
CUB-200-2011 200 5,994 5,794
: The statistics of fine-grained datasets used in this paper.
\[T\_Fine\_grained\]
Datesets FGVC Aircraft Stanford Cars Stanford Dogs CUB-200-2011
---------- -------------------- -------------------- -------------------- -------------------- --
w/o GC 86.62$\pm$0.31 88.66$\pm$0.22 76.16$\pm$0.25 82.07$\pm$0.26
w/ GC **87.77$\pm$0.27** **90.03$\pm$0.26** **78.23$\pm$0.24** **83.40$\pm$0.30**
: Testing accuracies on the four fine-grained image classification datasets.
\[T\_fine-grained2\]
Results on Fine-grained Image Classification
--------------------------------------------
In order to show that GC can also work with the pre-trained models, we conduct experiments on four challenging fine-grained image classification datasets, including FGVC Aircraft [@maji2013fine], Stanford Cars [@krause20133d], Stanford Dogs [@khosla2011novel] and CUB-200-2011 [@wah2011caltech]. The detailed statistics of these four datasets are summarized in Table \[T\_Fine\_grained\]. We use the official pre-trained ResNet50 provided by Pytorch as the baseline DNN for all these four datasets. The original images are resized into $512\times512$ and we crop the center region with $448\times448$ as input for both training and testing. The models are pre-trained on ImageNet. We use SGDM with momentum of $0.9$ to fine-tune ResNet50 for 100 epochs on 4 GPUs with batch size 256. The initial learning rate is $0.1$ for the last FC layer and 0.01 for all pre-trained Conv layers. The learning rate is multiplied by $0.1$ at the 50th and 80th epochs. Please note that our goal is to validate the effectiveness of GC but not to push state-of-the-art results, so we only use simple training tricks. We repeat the experiments for 10 times and report the result in mean $\pm$ std format.
![Training accuracy (solid line) and testing accuracy (dotted line) curves vs. training epoch on four fine-grained image classification datasets. []{data-label="fine_grid2_c"}](finegridc.png){width="100.00000%"}
Fig. \[fine\_grid2\_c\] shows the training and testing accuracies of SGDM and SGDM+GC for the first 40 epochs on the four fine-grained image classification datasets. Table \[T\_fine-grained2\] shows the final testing accuracies. We can see that both the training and testing accuracies of SGDM are improved by GC. For the final classification accuracy, GC improves SGDM by $1.1\%\sim2.1\%$ on these four datasets. This sufficiently demonstrates the effectiveness of GC on fine-tuning pre-trained models.
Method Backbone $\text{AP}$ $\text{AP}_{.5}$ $\text{AP}_{.75}$ Backbone $\text{AP}$ $\text{AP}_{.5}$ $\text{AP}_{.75}$
-------- ---------- ------------- ------------------ ------------------- ------------ ------------- ------------------ -------------------
w/o GC R50 36.4 58.4 39.1 X101-32x4d 40.1 62.0 43.8
w/ GC R50 37.0 59.0 40.2 X101-32x4d 40.7 62.7 43.9
w/o GC R101 38.5 60.3 41.6 X101-64x4d 41.3 63.3 45.2
w/ GC R101 38.9 60.8 42.2 X101-64x4d 41.6 63.8 45.4
: Detection results on COCO by using Faster-RCNN and FPN with various backbone models.
\[Tfaster\]
Method Backbone $\text{AP}^b$ $\text{AP}^b_{.5}$ $\text{AP}^b_{.75}$ $\text{AP}^m$ $\text{AP}^m_{.5}$ $\text{AP}^m_{.75}$ Backbone $\text{AP}^b$ $\text{AP}^b_{.5}$ $\text{AP}^b_{.75}$ $\text{AP}^m$ $\text{AP}^m_{.5}$ $\text{AP}^m_{.75}$
-------- ------------ --------------- -------------------- --------------------- --------------- -------------------- --------------------- ------------ --------------- -------------------- --------------------- --------------- -------------------- ---------------------
w/o GC R50 37.4 59.0 40.6 34.1 55.5 36.1 R50 (4c1f) 37.5 58.2 41.0 33.9 55.0 36.1
w/ GC R50 37.9 59.6 41.2 34.7 56.1 37.0 R50 (4c1f) 38.4 59.5 41.8 34.6 55.9 36.7
w/o GC R101 39.4 60.9 43.3 35.9 57.7 38.4 R101GN 41.1 61.7 44.9 36.9 58.7 39.3
w/ GC R101 40.0 61.5 43.7 36.2 58.1 38.7 R101GN 41.7 62.3 45.3 37.4 59.3 40.3
w/o GC X101-32x4d 41.1 62.8 45.0 37.1 59.4 39.8 R50GN+WS 40.0 60.7 43.6 36.1 57.8 38.6
w/ GC X101-32x4d 41.6 63.1 45.5 37.4 59.8 39.9 R50GN+WS 40.6 61.3 43.9 36.6 58.2 39.1
w/o GC X101-64x4d 42.1 63.8 46.3 38.0 60.6 40.9
w/ GC X101-64x4d 42.8 64.5 46.8 38.4 61.0 41.1
: Detection and segmentation results on COCO by using Mask-RCNN and FPN with various backbone models.
\[Tmask\]
Objection Detection and Segmentation
------------------------------------
Finally, we evaluate GC on object detection and segmentation tasks to show that GC can also be applied to more tasks beyond image classification. The models are pre-trained on ImageNet. The training batch size for object detection and segmentation is usually very small (e.g., 1 or 2) because of the high resolution of input image. Therefore, the BN layer is usually frozen [@he2016deep] and the benefits from BN cannot be enjoyed during training. One alternative is to use GN instead. The models are trained on COCO $train2017$ dataset (118K images) and evaluated on COCO $val2017$ dataset (40K images) [@lin2014microsoft]. COCO dataset can be used for multiple tasks, including image classification, object detection, semantic segmentation and instance segmentation.
We use the MMDetection [@chen2019mmdetection] toolbox, which contains comprehensive models on object detection and segmentation tasks, as the detection framework. The official implementations and settings are used for all experiments. All the pre-trained models are provided from their official websites, and we fine-tune them on COCO $train2017$ set with 8 GPUs and 2 images per GPU. The optimizers are SGDM and SGDM+GC. The backbone networks include ResNet50 (R50), ResNet101 (R101), ResNeXt101-32x4d (X101-32x4d), ResNeXt101-64x4d (X101-32x4d). The Feature Pyramid Network (FPN) [@lin2017feature] is also used. The learning rate schedule is $1{X}$ for both Faster R-CNN [@ren2015faster] and Mask R-CNN [@he2017mask], except R50 with GN and R101 with GN, which use $2{X}$ learning rate schedule.
Tabel \[Tfaster\] shows the Average Precision (AP) results of Faster R-CNN. We can see that all the backbone networks trained with GC can achieve a performance gain about $0.3\%\sim0.6\%$ on object detection. Tabel \[Tmask\] presents the Average Precision for bounding box ($\text{AP}^b$) and instance segmentation ($\text{AP}^m$). It can be seen that the $\text{AP}^b$ increases by $0.5\%\sim0.9\%$ for object detection task and the $\text{AP}^m$ increases by $0.3\%\sim0.7\%$ for instance segmentation task. Moreover, we find that if 4conv1fc bounding box head, like R50 (4c1f), is used, the performance can increase more by GC. And GC can also boost the performance of GN (see R101GN) and improve the performance of WS (see R50GN+WS). Overall, we can see that GC boosts the generalization performance of all evaluated models. This demonstrates that it is a simple yet effective optimization technique, which is general to many tasks beyond image classification.
Conclusions
===========
How to efficiently and effectively optimize a DNN is one of the key issues in deep learning research. Previous methods such as batch normalization (BN) and weight standardization (WS) mostly operate on network activations or weights to improve DNN training. We proposed a different approach which operates directly on gradients. Specifically, we removed the mean from the gradient vectors and centralized them to have zero mean. The so-called Gradient Centralization (GC) method demonstrated several desired properties of deep network optimization. We showed that GC actually improves the loss function with a constraint on weight vectors, which regularizes both weight space and output feature space. We also showed that this constrained loss function has better Lipschitzness than the original one so that it has a smoother optimization landscape. Comprehensive experiments were performed and the results demonstrated that GC can be well applied to different tasks with different optimizers and network architectures, improving their training efficiency and generalization performance.
Appendix {#appendix .unnumbered}
========
A1. Proof of Theorem 4.1 {#a1.-proof-of-theorem-4.1 .unnumbered}
------------------------
*Proof*. First we show below a simple property of $\mathbf{P}$: $$\mathbf{1}^T\mathbf{P}=\mathbf{1}^T(\mathbf{I}-\mathbf{e}\mathbf{e}^{T})=\mathbf{1}^T-\frac{1}{M}\mathbf{1}^T\mathbf{1}\mathbf{1}^T=\mathbf{0}^T,\vspace{-3mm}$$ where $M$ is the dimension of $\mathbf{e}$. For each SGD step with GC, we have: $$\mathbf{w}^{t+1}=\mathbf{w}^{t}-\alpha^t\mathbf{P}\nabla_{\mathbf{w}^t}\mathcal{L}.\vspace{-3mm}$$ It can be easily derived that: $$\mathbf{w}^{t}=\mathbf{w}^{0}-\mathbf{P}\sum_{i=0}^{t-1}\alpha^{(i)}\nabla_{\mathbf{w}^{(i)}}\mathcal{L},\vspace{-3mm}$$ where $t$ is the number of iterations. Then for the output activations of $\mathbf{x}$ and $\mathbf{x}+\gamma\mathbf{1}$, there is $$\begin{aligned}
({\mathbf{w}^{t}})^T\mathbf{x}-({\mathbf{w}^{t}})^T(\mathbf{x}+\gamma\mathbf{1})&=\gamma \mathbf{1}^T{\mathbf{w}^{t}}\\
&=\gamma \mathbf{1}^T(\mathbf{w}^{0}-\mathbf{P}\sum_{i=0}^{t-1}\alpha^{(i)}\nabla_{\mathbf{w}^{(i)}}\mathcal{L})\\
&=\gamma \mathbf{1}^T\mathbf{w}^{0}-\gamma\mathbf{1}^T\mathbf{P}\sum_{i=0}^{t-1}\alpha^{(i)}\nabla_{\mathbf{w}^{(i)}}\mathcal{L}\\
&=\gamma \mathbf{1}^T\mathbf{w}^{0}.
\end{aligned}\label{11}
\vspace{-2mm}$$ Therefore, $$\begin{aligned}
({\mathbf{w}^{t}})^T\mathbf{x}-({\mathbf{w}^{t}})^T(\mathbf{x}+\gamma\mathbf{1})=\gamma\mathbf{1}^T\mathbf{w}^0.
\end{aligned}\label{12}
\vspace{-1mm}$$
For SGD with momentum, the conclusion is the same, because we can obtain a term $\gamma\mathbf{1}^T\mathbf{P}\sum_{i=0}^{t-1}\alpha^{(i)}\mathbf{m}^{i}$ in the third row of Eq.(\[11\]), where $\mathbf{m}^{i}$ is the momentum in the $i$th iteration, and this term is also equal to zero.
The proof is completed. $\blacksquare$
A2. Proof of Theorem 4.2 {#a2.-proof-of-theorem-4.2 .unnumbered}
------------------------
*Proof*. Because $\mathbf{e}$ is a unit vector, there is $\mathbf{e}^T\mathbf{e}=1$. We can easily prove that: $$\begin{aligned}
\mathbf{P}^T\mathbf{P}&=(\mathbf{I}-\mathbf{e}\mathbf{e}^{T})^T(\mathbf{I}-\mathbf{e}\mathbf{e}^{T})\\
&=\mathbf{I}-2\mathbf{e}\mathbf{e}^{T}+\mathbf{e}\mathbf{e}^{T}\mathbf{e}\mathbf{e}^{T}\\
&=\mathbf{I}-\mathbf{e}\mathbf{e}^{T}\\
&=\mathbf{P}.
\end{aligned}
\vspace{-2mm}$$ Then for $\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})$, we have: $$\begin{aligned}
||\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})||_2^2&=\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})^T\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})\\
&=(\mathbf{P}\nabla_{\mathbf{w}}\mathcal{L})^T(\mathbf{P}\nabla_{\mathbf{w}}\mathcal{L})\\
&=\nabla_{\mathbf{w}}\mathcal{L}^T\mathbf{P}^T\mathbf{P}\nabla_{\mathbf{w}}\mathcal{L}\\
&=\nabla_{\mathbf{w}}\mathcal{L}^T\mathbf{P}\nabla_{\mathbf{w}}\mathcal{L}\\
&=\nabla_{\mathbf{w}}\mathcal{L}^T(\mathbf{I}-\mathbf{e}\mathbf{e}^{T})\nabla_{\mathbf{w}}\mathcal{L}\\
&=\nabla_{\mathbf{w}}\mathcal{L}^T\nabla_{\mathbf{w}}\mathcal{L}- \nabla_{\mathbf{w}}\mathcal{L}^T\mathbf{e}\mathbf{e}^{T}\nabla_{\mathbf{w}}\mathcal{L}\\
&=||\nabla_{\mathbf{w}}\mathcal{L}||_2^2-||\mathbf{e}^{T}\nabla_{\mathbf{w}}\mathcal{L}||_2^2\\
&\leq ||\nabla_{\mathbf{w}}\mathcal{L}||_2^2.
\end{aligned}
\vspace{-2mm}$$ For $\nabla_{\mathbf{w}}\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})$, we also have $$\begin{aligned}
||\nabla\Phi_{GC}(\nabla_{\mathbf{w}}\mathcal{L})||_2^2&=||\mathbf{P}\nabla^2_{\mathbf{w}}\mathcal{L}||_2^2\\
&=\nabla^2_{\mathbf{w}}\mathcal{L}^T\mathbf{P}^T\mathbf{P}\nabla^2_{\mathbf{w}}\mathcal{L}\\
&=\nabla^2_{\mathbf{w}}\mathcal{L}^T\mathbf{P}\nabla^2_{\mathbf{w}}\mathcal{L}\\
&=||\nabla^2_{\mathbf{w}}\mathcal{L}||_2^2-||\mathbf{e}^{T}\nabla^2_{\mathbf{w}}\mathcal{L}||_2^2\\
&\leq ||\nabla^2_{\mathbf{w}}\mathcal{L}||_2^2.
\end{aligned}
\vspace{-2mm}$$
The proof is completed. $\blacksquare$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the predictions of a model for proton-proton total cross-section at LHC. It takes into account both hard partonic processes and soft gluon emission effects to describe the proper high energy behavior and to respect the Froissart bound.'
author:
- 'A. Achilli$^1$[^1], R. Godbole$^2$, A.Grau$^3$, G. Pancheri$^4$, Y.N. Srivastava$^1$'
title: |
MPI08\
QCD Mini-jet contribution to the total cross section
---
Introduction
============
A reliable prediction of the total proton-proton cross section is fundamental to know which will be the underlying activity at the LHC and for new discoveries in physics from the LHC data. In this article, we shall describe a model [@lastPLB] [@ourmodel] for the hadronic total cross section based on QCD minijet formalism. The model includes a resummation of soft gluon radiation which is necessary to tame the fast high-energy rise typical of a purely perturbative minijet model. It is called the BN model from the Bloch and Nordsiek discussion of the infrared catastrophe in QED. In the first section, results are presented concerning the behavior of the QCD minijet cross section. It will then be explained how this term is included into an eikonal formalism where infrared soft gluon emission effects are added. The last section is devoted to the link between the total cross-section asymptotic high energy behavior predicted by our model and the model parameters. This relation also shows that our prediction is in agreement with the limit imposed by the Froissart bound.
Mini-jet cross section
======================
Hard processes involving high-energy partonic collisions drive the rise of the total cross section [@cline]. These jet-producing collisions are typical perturbative processes and we can describe them through the usual QCD expression: $$\sigma^{AB}_{\rm jet} (s,p_{tmin})=
\int_{p_{tmin}}^{\sqrt{s/2}} d p_t \int_{4
p_t^2/s}^1 d x_1 \int_{4 p_t^2/(x_1 s)}^1 d x_2 \times
\sum_{i,j,k,l}
f_{i|A}(x_1,p_t^2) f_{j|B}(x_2,p_t^2)
\frac { d \hat{\sigma}_{ij}^{ kl}(\hat{s})} {d p_t},
\label{eq1}$$ with $A,B = p, \bar p$. This expression depends on the parameter $p_{tmin}$ which represents the minimum transverse momentum of the scattered partons for which one allows a perturbative QCD treatment. Its value is usually around $\approx 1-2$ GeV and it distinguishes hard processes (that are processes for which a perturbative approach is used) from the soft ones that dominate at low energy, typically for $\sqrt{s}\le 10\div 20 \ GeV$, i.e, well before the cross-section starts rising. The Minijet expression also depends on the DGLAP evoluted Partonic Densities Functions $f_{i|A}$ for which there exist in the literature different LO parameterizations(GRV, MRST, CTEQ [@densities]). We obtain an asymptotic growth of $\sigma_{jet}$ with energy as a power of $s$. As shown in figure \[fig:1\], the value of the exponent depends on the PDF used and one has $$\ \sigma_{jet}^{GRV}\approx s^{0.4} \ \ \ \
\sigma_{jet}^{MRST}\approx s^{0.3} \ \ \ \
\sigma_{jet}^{CTEQ}\approx s^{0.3}.$$ This result can be derived by considering the relevant contribution to the integral in (\[eq1\]) in the $\sqrt{s}>>p_{tmin}$ limit. In this limit, the major contribution comes from the small fractions of momentum carried by the colliding gluons with $x_{1,2}<<1$. In this limit we know that the relevant PDF’s behave approximately like powers of the momentum fraction $x^{-J}$ with $J \sim 1.3$ [@Lomatch]. From the previous consideration and noting that $\frac { d \hat{\sigma}_{ij}^{ kl}(\hat{s})} {d
p_t} \propto \frac{1}{p_t ^3}$ we obtain from (\[eq1\]) the following asymptotic high-energy expression for $\sigma_{jet}$:
$$\sigma _{jet} \propto \frac{1}{{p_{t\min }^2 }}\left[
{\frac{s}{{4p_{t\min }^2 }}} \right]^{J - 1}.$$
The dominant term is just a power of $s$ and the estimate obtained for the exponent $\epsilon=J-1\sim0.3$ is in agreement with our previous results. We now need to understand how to incorporate into a model for the total cross section this very fast rise at very high energy, which is present in the perturbative regime.
![minijet cross section for different input parton densities.[]{data-label="fig:1"}](sigjet_115_band.eps){height="3.9in"}
Firstly it is important to note that $\sigma_{jet}$ is an inclusive cross section and therefore contains in itself a multiplicity factor, linked to the average number $< n >$ of partonic collisions that take place during the hadronic scattering. We can approximate the energy driving term at high energy [@durand] $< n >$ as $$< n > \approx \sigma_{jet} \cdot A ,
\label{mult}$$ where $A$ is a function representing the overlap between the two hadrons.
Now we can derive an expression for the total cross section as a function of $<n>$. Assuming that the number of partonic collisions follows a Poisson distribution, since each interaction is independent from the other, the probability of having $k$ partonic collisions is: $$P(k, < n > ) = \frac{{ < n > ^k e^{ - < n > } }}{{k!}}.$$ The average number of partonic collisions should depend on the energy and on the impact parameter $b$ relative to the hadronic process $< n > \equiv < n(b,s) >$. From the previous expression it is possible to obtain the inelastic hadronic cross section: $$\sigma _{inelastic} = \int {d^2 b\sum\limits_{k = 1} {P(k, < n(b,s) > )} } = \int {d^2 b\left[ {1 - e^{ - < n(b,s) > } }
\right]} ,$$ which is the usual eikonal expression if we consider the link between $< n(b,s) >$ and the eikonal $\chi (b,s)$: $$< n(b,s) > = 2{\mathop{\rm Im}\nolimits} \chi (b,s).$$
Eikonal model
=============
The eikonal representation allows to implement multiple parton scattering and to restore a finite size of the interaction. Neglecting the real part of the eikonal function, an acceptable approximation in the high energy limit, the expression for the total cross section is $$\sigma _{tot} = 2\int {d^2 b} \left[ {1 - e^{ - n(b,s)/2} }
\right] .$$ The average number of partonic collisions receives contributions both from hard and soft physics processes and we write it in the form $$n(b,s) = n_{soft} (b,s) + n_{hard} (b,s) ,$$ where the soft term parameterizes the contribution of all the processes for which the partons scatter with $p_t<p_{tmin}$. It is the only relevant term at low-energy and it establishes the overall normalization, while the hard term is responsible for the high-energy rise. From (\[mult\]), we approximate this term with $$n_{hard} (b,s) = A(b,s)\sigma _{jet} (s) , \\
\label{nhard}$$ where the minijet cross section drives the rise due to the increase of the number of partonic collisions with the energy and $A(b,s)$ is the overlap function which depends on the (energy dependent) spatial distribution of partons inside the colliding hadrons. In some older models [@durand] a simpler factorized expression for $n(b,s)$ was used, with the overlap function depending only on $b$. However, when up-to-date realistic parton densities are used, such impact parameter distributions, inspired by constant hadronic form factors, led to an excessive rise of $\sigma_{tot}$ with the energy. In our BN model we include an $s$-dependence in the overlap function that has to tame the strong growth due to the fast asymptotic rise of $\sigma_{jet}$ [@ourmodel].
We identify soft gluon emissions from the colliding partons as the physical effect responsible for the attenuation of the rise of the total cross section. These emissions influence matter distribution inside of the hadrons, hence changing the overlap function. They break collinearity between the colliding partons, diminishing the efficiency of the scattering process. The number of soft emissions increases with the energy and this makes their contribution important, also at very high energy. The calculation of this effect uses a semiclassical approach based on a Block-Nordsieck inspired formalism [@ddt; @*pp; @*oldkt], the basic assumption of this technique is that all emissions are independent from each other, so the number of gluons emitted follows a Poisson distribution. Thereof one obtains a distribution of the colliding partons as function of the transverse momentum of the soft gluons emitted in the collision, i.e. $$d^2P({\bf K_\perp})=d^2{\bf K_\perp} {{1} \over{(2\pi)^2}}\int
d^2 {\bf b}\ e^{i{\bf K_\perp\cdot b} -h( b )}\ ,$$ the factor $h(b)$ is given by
$$h(b) = \int {d^3 } n_g (k)[1 - e^{ - ik_ \bot \cdot b} ] = \int
{\frac{{d^3 k}}{{2k_0 }}} \sum\limits_{m,n = colors} {|j^{\mu ,m}
(k)j_{\mu ,n} (k)|} [1 - e^{ - ik_ \bot \cdot b} ],$$ where $d^3 n_g(k)$ is the distribution for single gluon emission in a scattering process and it is linked to the QCD current $j^{\mu}$ responsible for emission.
We have proposed to obtain the overlap function as the Fourier transform of the previous expression of the soft gluon transverse momentum resummed distribution, namely to put $$A_{BN}(b,s)=N \int d^2{\bf K_{\perp}}\ e^{-i{\bf K_\perp\cdot b}}
{{d^2P({\bf K_\perp})}\over{d^2 {\bf K_\perp}}}={{e^{-h( b,q_{max})}}\over
{\int d^2{\bf b} \ e^{-h( b,q_{max})}
}} ,$$ with $$h( b,q_{max}) = \frac{16}{3}\int_0^{q_{max} }
{{ \alpha_s(k_t^2) }\over{\pi}}{{d
k_t}\over{k_t}}\log{{2q_{\max}}\over{k_t}}[1-J_0(k_tb)]
\label{hbq} ,$$ this integral is performed up to a maximum value which is linked to the maximum transverse momentum allowed by the kinematics for a single gluon emitted, $q_{max}$ [@greco]. In principle, this parameter and the overlap function should be calculated for each partonic sub-process, but in the partial factorization of Eq.(\[nhard\]) we use an average value of $q_{max}$ obtained considering all the sub-processes that can happen for a given energy of the main hadronic process[@ourmodel]: $$q_{\max } (s) = \sqrt {\frac{s} {2}}\, \frac{{\sum\limits_{i,j}
{\int {\frac{{dx_1 }} {{x_1 }}\int {\frac{{dx_2 }} {{x_2 }}
\int_{z_{min}}^1 {dz f_i (x_1) f_j (x_2) \sqrt {x_1 x_2 } (1 - z)}
} } } }} {{\sum\limits_{i,j} {\int {\frac{{dx_1 }} {{x_1 }}\int
{\frac{{dx_2}} {{x_2}} \int_{z_{min}}^1 {dz} f_i (x_1)f_j (x_2) }
} } }} ,
\label{qmax}$$ with $z_{min}=4p_{tmin}^2/(s x_1 x_2)$. Notice that consistency of the calculation requires that the PDF’s used in Eq. (\[qmax\]) be the same as those used in $\sigma_{jet}$. In are presented our results for $q_{max}$ as function of $\sqrt{s}$ using $p_{tmin}=1.15$ GeV.
![Results for $q_{max}$ using different input parton densities with $p_{tmin}=1.15$ GeV.[]{data-label="fig:2"}](qmax.eps){height="3.9in"}
The integral in (\[hbq\]) has another relevant feature, it extends down to zero momentum values, and to calculate it we have to take an expression of $\alpha_s$ different from the perturbative QCD expression which is singular and not integrable in (\[hbq\]). We use a phenomenological expression[@npb382], which coincides with the usual QCD limit for large $k_t$, and is singular but integrable for $k_t \to 0$: $$\alpha_s(k_t^2)={{12 \pi}\over{33-2N_f}} {{p}
\over{\ln[1+p({{k_t}\over{\Lambda}})^{2p}]}} .$$ This expression for $\alpha_s$ is inspired by the Richardson expression for a linear confining potential [@Richardson], and we find for the parameter $p$ that
- $p<1$ to have a convergent integral (unlike the case of the Richardson potential where $p=1$)
- $p>1/2$ for the correct analyticity in the momentum transfer variable.
[@lastPLB] shows our predictions, obtained for the total cross-section using a set of phenomenological values for $p_{tmin}$ and $p$, and varying the parton densities. We also make a comparison with data and other current models.
![Results from our total cross-section model (for different parton densities) compared with data [@data] and with other models [@bibmodels].[]{data-label="fig:3"}](figure3aug2007.eps){height="3.9in"}
Restoration of Froissart Bound
==============================
The Froissart Martin Bound [@froissmar] states that $\sigma_{tot}$ cannot rise faster than a function which is proportional to $log^2(s)$. In order to see that in our model this bound is respected, we approximate our total cross section at very large energies as $$\sigma _{tot} \approx 2\pi \int {db^2 \left[ {1 - e^{ - n_{hard} (b,s)/2} } \right]}
\label{sigtothard} ,$$ with $ n_{hard} (b,s) \approx \sigma _{jet} (s)A_{hard} (b,s)$. We then take for $\sigma_{jet}$ the asymptotic high energy expression: $$\sigma_{jet}=\sigma _1 \left( {\frac{s}{{GeV^2 }}}
\right)^\varepsilon ,$$ with $\sigma_1=$constant and $\epsilon \sim 0.3-0.4$. Being $
A_{hard} (b,s) \propto e^{ - h(b,s)}$, we can consider in (\[hbq\]) the infrared limit $k_t \to 0$ where the integral receives the dominant contribution. In this limit we have $$\alpha _s (k_t^2 ) \approx \left( {\frac{\Lambda }{{k_t }}}
\right)^{2p} ,$$ apart from logarithmic terms. Then, with $ h(b,s) \propto (b\bar \Lambda )^{2p}$ [@ourmodel] (again apart from logarithmic terms), we have $$A_{hard} (b) \propto e^{ - (b\bar \Lambda )^{2p} },$$ and from this expression $$n_{hard} = 2C(s)e^{ - (b\bar \Lambda )^{2p} } ,$$ with $C(s) = \frac{A_0 \sigma _1}{2} \left( {\frac{s}{{GeV^2 }}}
\right)^\varepsilon$. The very high energy limit of Eq. (\[sigtothard\]) then gives $$\sigma _{tot} \approx 2\pi \int_0^\infty {db^2 [1 - e^{ - C(s)e^{ - (b\bar \Lambda )^{2p} } } ] \to \left[ {\varepsilon \ln \left( {\frac{s}{{GeV^2 }}} \right)} \right]^{1/p}
} .$$ The asymptotic growth of $\sigma_{tot}$ in our model depends on the parameter $\epsilon$ which fixes the asymptotic rise of the minijet cross section, and on $p$ which modulates the infrared behavior of $\alpha_s$. Notice that $1/2<p<1$ and thus this approximated result links the restoration of the Froissart bound in our model with the infrared behavior of $\alpha_s$. We can now understand why a knowledge of the confining phase of the strong interaction is necessary if we want to restore the finite size of the hadronic interaction.
[99]{} A. Achilli, R. Hegde, R. M. Godbole, A. Grau, G. Pancheri and Y. Srivastava, Phys. Lett. B [**659**]{} (2008) 137 \[arXiv:0708.3626 \[hep-ph\]\]. R. M. Godbole, A. Grau, G. Pancheri and Y. N. Srivastava, Phys. Rev. D [**72**]{} (2005) 076001 \[arXiv:hep-ph/0408355\]; A. Grau, G. Pancheri and Y. N. Srivastava, Phys. Rev. D [**60**]{} (1999) 114020 \[arXiv:hep-ph/9905228\]. D. Cline, F. Halzen and J. Luthe, Phys. Rev. Lett. [**31**]{} (1973) 491. M. Gluck, E. Reya, and A. Vogt, Z. Phys. [**C53**]{} (1992) 127–134; Z. Phys. [**C67**]{} (1995) 433–448; Eur. Phys. J. [**C 5**]{} (1998) 461–470; A. D. Martin, R. G. Roberts, W. J. Stirling, and R. S. Thorne, Phys. Lett. [**B531**]{} (2002) 216–224; H.L. Lai , J. Botts , J. Huston , J.G. Morfin , J.F. Owens , Jian-wei Qiu, W.K. Tung, H. Weerts, Phys.Rev. [**D51**]{} 4763-4782,1995. S. Lomatch, F. I. Olness and J. C. Collins, Nucl. Phys. B [**317**]{} (1989) 617. L. Durand and H. Pi, Phys. Rev. Lett. 58 (1987) 303. Y. L. Dokshitzer, D. I. D’Yakonov and S. I. Troyan, Phys. Lett. [**79B**]{} (1978) 269. G. Parisi and R. Petronzio, Nucl. Phys. B [**154**]{} (1979) 427. G. Pancheri-Srivastava and Y.N. Srivastava, Phys.Rev. **D15** (1977) 2915. P. Chiappetta and M. Greco, Nucl. Phys. B [**221**]{} (1983) 269. A. Corsetti, A. Grau, G. Pancheri and Y. N. Srivastava, Phys. Lett. B [**382**]{} (1996) 282 \[arXiv:hep-ph/9605314\]. J. L. Richardson, Phys. Lett. B [**82**]{} (1979) 272. For total cross section data see:\
W.-M. Yao [*et al.*]{} [**PDG**]{}, J. Phys. G. [**33**]{} (2006) 1; G. Arnison [*et al.*]{}, [**UA1**]{} Collaboration, Phys. Lett. [**128B**]{} (1983) 336; R. Battiston [*et al.*]{} [**UA4**]{} Collaboration, Phys. Lett. [**B117**]{} (1982) 126; C. Augier [*et al.*]{} [**UA4/2**]{} Collaboration, Phys. Lett. [**B344**]{} (1995) 451; M. Bozzo [*et al.*]{} [**UA4**]{} Collaboration, Phys. Lett. [**147B**]{} (1984) 392; G.J. Alner [*et al.*]{} [**UA5**]{} Collaboration, Z. Phys. [**C32**]{} (1986) 153; N. Amos [*et. al.*]{}, [**E710**]{} Collaboration, Phys. Rev. Lett. [**68**]{} (1992) 2433–2436; C. Avila [*et. al.*]{}, [**E811**]{} Collaboration, Phys. Lett. [**B445**]{} (1999) 419–422; F. Abe [*et. al.*]{}, [**CDF**]{} Collaboration, Phys. Rev. [**D50**]{} (1994) 5550–5561. For total cross section models see:\
M. M. Block and F. Halzen, Phys. Rev. D [**73**]{} (2006) 054022 \[arXiv:hep-ph/0510238\]; E. G. S. Luna and M. J. Menon, arXiv:hep-ph/0105076; J. R. Cudell and O. V. Selyugin, arXiv:hep-ph/0612046; A. Donnachie and P. V. Landshoff, Phys. Lett. B [**296**]{} (1992) 227 \[arXiv:hep-ph/9209205\]; A. Donnachie and P. V. Landshoff, Phys. Lett. B [**595**]{} (2004) 393 \[arXiv:hep-ph/0402081\] M. Froissart, Phys.Rev. [**123**]{} (1961) 1053; A. Martin, Phys. Rev. [**129**]{} (1963) 1432.
[^1]: speaker
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Polarizations of electromagnetic waves from distant galaxies are known to be correlated with the source orientations. These quantities have been used to search for signals of cosmological birefringence. We review and classify transformation properties of the polarization and source orientation observables. The classifications give a firm foundation to certain practices which have sprung up informally in the literature. Transformations under parity play a central role, showing that parity violation in emission or in the subsequent propagation is an observable phenomenon. We also discuss statistical measures, correlations and distributions which transform properly and which can be used for systematic data analysis.'
---
psfig.tex 0.3in /\#1[0=\#1-0to0[*/*]{}]{}
[**Covariant Symmetry Classifications for Observables of Cosmological Birefringence**]{}
0.25in [John P. Ralston]{}
[*Department of Physics and Astronomy, Kansas University, Lawrence, KS-66045, USA\
[email protected]*]{}
0.25in [*and*]{} 0.25in
[Pankaj Jain]{}
[*Physics Department, I.I.T. Kanpur, India 208016\
[email protected]*]{}
Introduction
============
Galaxies at cosmological distances are known to emit polarized electromagnetic radiation according to certain statistical regularities. The processes range from linear polarization from radio galaxies due to synchrotron radiation, to centro-symmetric polarizations sometimes seen in optical measurements of QSO’s. The observables can include the percentage of polarization, a linear polarization angle, a galaxy symmetry axis, and the redshift of the galaxy, as well as the angular position coordinates of the galaxy on the dome of the sky. One typical correlation is an observed alignment of the bisector of a radio galaxy’s long axis and linearly polarized radiation in the “FR1" class of objects. This has been interpreted as probing the source’s magnetic field in conventional models. While this effect is in the context of conventional physics, other polarization correlations can test extremely small effects that might indicate new physics–sensitive to “cosmological birefringence" – that might occur in the subsequent propagation. This has led to several interesting studies, from seminal early work on propagation in parity-violating electro-weak media by Leo Stodolsky [@Leo], to the discussion of “optical activity of the universe" by Gabriel Karl and collaborators [@Karl], to recent studies of possible cosmological anisotropy in propagation [@NR; @JR].
Galaxy and polarization axes are labeled by angles, typically measured in “degrees East of North" in a spherical -polar system of angular coordinates. The galaxy major axis projected into the tangent plane is measured with an angle $\psi$. A linearly polarized electric field orientation in the same coordinates is assigned polarization angle $\chi$ . The angles, however, do not represent [*vectors*]{} in the rigorous sense, because a rotation of $\pi$ radians returns an axis to its original orientation. These observables are described by elements of the real projective group, which is not very familiar in physics. In this paper we will investigate and then automate proper transformation properties of such observables, and list quantities which are invariant under different symmetries. This provides a firm foundation for certain traditions and informal uses of variables which have sprung up in the literature. Rather interestingly, the split-up of the three dimensional orthogonal group into disconnected subgroups of different parity also occurs for observables connected with polarization. One then has quantities which are even or odd under parity, which extends the list of symmetries one can test using polarization data.
The physical motivation for this is as follows. Consider a linear polarization correlated with an emission axis. From “symmetry", statistics would naively be expected to show equal offsets of polarization axis relative to galaxy axis for either a positive or a negative sense of rotation. The particular “symmetry" assumed here is [*parity*]{} symmetry. While the emergence of parity as an observable may be surprising, it should be intuitively clear that observation of the converse would allow one to extract a definition of right–versus –left handed senses of rotation from a data set, breaking parity symmetry. An even more obvious violation of parity symmetry would be preference for one kind of circular polarization over another. While one may not expect an object the size of a galaxy to break parity symmetry in its emission characteristics, the possibility constitutes a valid observable, subject to classification and study. Another more interesting possibility, moreover, is to assume that objects do respect parity symmetry in a statistical sense upon emission, while the intervening medium can be tested for parity violation to an extremely fine degree. Many models predict just such effects [@sikivie; @ni]. We will proceed by setting up the transformation properties, listing covariant and invariant quantities, and illustrating the remarks by some statistical quantities which embody the ideas for data analysis.
**Setup**
=========
Let us discuss transformation properties. We are not concerned with Lorentz boosts, nor the usual Lorentz covariant characterization of radiation in terms of the energy momentum tensor, etc. Our focus is the “little group" of angular variables in the rest frame of the observer, and the peculiar fact that the angles ${\chi}$ and ${\psi}$ actually observed give the orientation of “sticks". A rotation of just $\pi$ radians about the axis of propagation makes no observable effect. Due to this unusual transformation property, both the electric field and the galaxy axis are conveniently represented by certain tensors.
Let $\hat p$ be the radial unit vector pointing toward a particular source. Construct a local Cartesian frame with its $z$–axis oriented along $\hat p$ (Fig 1). The 2–dimensional subspace orthogonal to $\hat p$ is a plane tangent to the unit sphere of the sky. Any consistent orientation of the local $\hat x $ and $\hat y$ axes spanning the tangent plane can be used; one can always take these to be in the directions of increasing azimuthal and polar angles ($\hat
\phi$ and $\hat \theta$) relative to a global North pole (hence the astronomer’s “East of North"). The electric field $(E_x, E_y, 0)$ is transverse to $\hat
p$ and conventionally represented by a 2 component complex doublet, a convention hearkening back to optics [@mandle]. In the circular polarization basis, $|E>={1\/\sqrt{2}}(E_x+iE_y,E_x-iE_y)$. The doublet notation is a hybrid and the transformation properties need examination.
First, in averaging over many cycles of the wave, the observer does not actually measure the electric field. Instead what can be measured is an Hermitian density matrix [@mandle] $E_i E_j^* =J_{ij} =(J_{unpolz}+J_{polz})_{ij}$. In our local coordinates, the non–zero entries of $J$ are a $2 \times 2$ block in the upper left hand corner of a $3 \times 3$ matrix. If the coordinate system is rotated, this tensor will transform in the usual way, because $E_i$ and $E_j$ transform like 3-vectors.
The unpolarized part $J_{unpolz}$ is proportional to “${\bf 1}$" with a proportionality constant fixed by the total power. The polarization density matrix $J_{polz}$ is defined [@mandle] to be the rest of the matrix, “as if" it were a pure state. Thus $det(J_{polz})=0$. This matrix is conventionally expanded in Pauli matrices:
$$J_{polz}={1\over 2} P_{max} (1 +\hat\chi\cdot\vec \sigma) ={1\over
2}P_{max}(1 +\slash\chi\ )$$ where $P_{max}$ is the degree of polarization, $\hat\chi$ is a unit–3 vector, and $\vec\sigma$ are the Pauli matrices. Equivalent are the “Poincarè" coordinates $\vartheta_p$ and $\chi$ (for describing elliptical and circular polarization [@mandle]): $$\slash\chi=\pmatrix{\cos\vartheta_p&\sin\vartheta_p\; e^{2i\chi}\cr
\sin\vartheta_p\; e^{-2i\chi}& -\cos\vartheta_p\cr}$$ We have introduced the common “slash" notation for contraction of a vector with the Pauli matrix. We can also use $\slash\chi$ for the $3 \times 3$ matrix when no confusion between the two is possible. The polarization parameters have now been organized in a form convenient for transformations.
Consider the important case of a rotation of tangent plane coordinates about the propagation axis, that is the local z-axis, $|E>\rightarrow U| E>$, with $U$ $3\times 3$ and in $SO(3)$. $U$ takes the block form $U=(u_\phi, 1)$ where $u_\phi$ is $2\times 2$ and unitary, in fact an element of SO(2). Of course $\slash \chi \rightarrow u_\phi\slash\chi u^\dagger_\phi$. This is the same rule as the rotations in the spin 1/2 representation, but in this case operating on vector components, not spinors. It follows that the parameters $\hat \chi$ transform with the angle of rotation [*doubled*]{}. As a check, a short calculation shows that a linear polarization oriented at angle $\chi$ relative to the local $x$–axis gives $\hat\chi=(\cos 2\chi,\; \sin 2\chi, 0)$ in the local basis. The factor of “2" is just right to account for the periodicity of the observable plane of the electric field under rotations by ${\pi}$ (as opposed to ${2\pi}$). This supports the informal use of $\hat\chi$ in literature ranging from biology [@batschelet; @fisher] to astronomy [@radio; @astron], where the “2" is inserted intuitively and to make things come out right while making it clear that $\hat\chi$ is not a true vector.
The result is elucidated by the following general argument. Recall the familiar decomposition of angular momentum ($\vec j$) states for the product $ | E >
\times <E | ={\bf 1 \times 1}= {\bf 2 + 1+ 0}$. The states are labeled by $(m, j)$ representing the eigenvalues of $j_z$ and $\vec j^2$, respectively. However, since $E$ is transverse, all tensor products made from $m=0$ are absent. From the addition rule for $m$, these absent representations are the $m= 0, m=\pm 1$ states of $j=2$, and the $m=\pm 1$ states of $j=1$. This eliminates 5 of 9 possible combinations, leaving 4 total which are $ (2, 2),
(-2,
2), (0,1), (0, 0)$. The 4 possibilities are the 4 Stokes parameters, with the total power $\vec E \cdot \vec E^*$ being the singlet $(0,0)$. The object $\hat\chi$, a deceptively vector–like position on the Poincarè sphere, is made from the peculiar combination $(2, 2), (-2, 2), (0,1)$. Making a short calculation in the helicity basis, $i P_{max} \cos(\vartheta_p) =1/2 Tr [ \sigma_z J] =
i(E_+E^*_+
- E_-E^*_-) = (\vec E \times\vec E^*)_z$ is the $(0,1)$ component, showing that the ellipticity transforms like a 3-vector attached to the photon’s direction of motion.
A similar treatment is needed for a galaxy axis. An observed galaxy axis is the projection of the physical major axis (a signless eigenvector of a $3 \times 3$ tensor of intensity distribution, say) onto the tangent plane. We can again define in a helicity basis a 2 component $|\psi>=(\cos\psi+i\sin\psi,
\cos\psi-i\sin\psi)/\sqrt{2}$. Again $|\psi>$ is not a good variable, because the observable is actually projective, and we must identify $|\psi>=-|\psi>$. The associated $2\times 2$ matrix, $|\psi><\psi|={1\over 2}+{1\over 2}\slash
{\bf \psi}$, is part of a tensor embedded in a $3\times 3$ matrix. In our coordinates it can be expanded in Pauli matrices $$|\psi><\psi|=1/2\pmatrix{1&e^{2i\psi}
\cr h.c.& 1 \cr} ={1\over
2}{\bf 1}+{1\over 2}\slash{\bf \psi}$$ The matrix $\slash{\bf \psi}$ transforms like $u_\phi{\bf\slash \psi}
u_\phi^\dagger$ and defines a 3 component thing, $\hat\psi=(\cos 2\psi, \sin 2\psi, 0)$. Note again the doubling of angles in $\hat\psi$, representing the information that $|\psi>$ and $-|\psi>$ are now identified, just as informally practiced in the literature. (It is interesting that if one could deduce information on the “pitch" of a galaxy, then $\vec\psi$ could point out of the plane: such an axis might be “elliptically polarized", otherwise $\hat\psi$ is in the local tangent plane.)
We also have the position on the sky $\hat p$, which is a true vector and [*not*]{} to transform with doubled angles.
Finally we should classify the objects under parity: by definition we have that $\hat p \rightarrow -\hat p$. $\hat\psi$ and $\hat\chi$ are only a little more work: since $\hat\psi_j = 1/2 Tr[ \slash \psi \sigma_j]$, with $\slash \psi$ invariant, then $\hat\psi$ and $\hat\chi$ are invariant under parity (although transforming in a complicated way under the full $O(3)$).
Now one can transform away from the special coordinate system, using general $3
\times 3$ elements of $SO(3)$, with the three objects $\slash \chi$, $\slash
\psi$, and $\hat p$ all transforming properly.
Covariants and Invariants
=========================
We now turn to invariants one can make from the matrices $\slash \psi, \slash
\chi$ and $\hat p$. A “local" quantity will be one made from a single source, or (if possible) different sources at the same location on the dome of the sky; a “non-local" quantity anything made from sources at different locations.
[*Local quantities:*]{} Since for any galaxy $i$ a coordinate system exists where all the matrix elements are in the upper left, then we have the covariant identity for such $3\times 3$ matrices $$\slash A\slash B =
A\cdot B\ 1 + i {\slash C } ; C= A\times B$$ The “1" of course means the 2 x 2 unit matrix on the upper left, covariantly written as $\delta_{ij} -\hat p_i\hat p_j$. As a consequence of this identity an obvious invariant is reduced to a simpler form: $$s_1 = {1\over
2}Tr[\slash\chi\slash{\psi}]={1\over 2}Tr[\slash{\psi} \slash \chi] =
\hat\chi\cdot\hat\psi.$$ It follows that $s_1 = \cos(2(\chi-\psi))$, which is clearly invariant under the local rotation $\chi \rightarrow \chi + \delta,
\psi
\rightarrow \psi + \delta$. It also follows that $s_1$ is even under parity.
Another useful quantity is the anti-symmetric $3 \times 3$ commutator $[\slash
\chi, \slash \psi ]$. This is dual to a pseudo-vector: $$\vec A_i = 1/2 Tr [ \epsilon_i \slash \chi \slash{\bf \psi}]$$ where $\epsilon_i$ is the completely anti-symmetric matrix with $j, k$ elements $\epsilon_i^{jk}$. In our local coordinate frame, $\vec A_i$ points in the direction of $\hat p$ and is proportional to $\sin( 2 (\chi-\psi))$. The sign of proportionality depends on the right–handed convention for angles. This remains true however the coordinate system is rotated. The epsilon-tensor is even under parity, making it clear that $\vec A$ is even and therefore a pseudo-vector by construction. Since any quantity which is odd in $(\chi-\psi)$ reverses when a positive “sense" of rotation is reversed to a negative one, such quantities are [*parity–odd*]{} on general grounds.
This is not the only odd-parity observable. The helicity $h$ of the wave is defined in a Lorentz-covariant manner as the projection of its spin along its direction of propagation. This is a pseudo-scalar: it must therefore be equal (up to a constant) to the product $\hat\chi\cdot\hat p$. The other invariants which can be made by contracting $\hat p$ and the matrices are trivially zero. Consulting the identity (1) above, there are 4 real-valued quantities in the products of the 2 matrices, which have now been classified into one scalar and one pseudo-vector, exhausting the possibilities for local bilinears of the two tensors. One can go further, and add another unit vector $\hat\lambda$ to the problem. Such a vector is needed to quantify asymmetries of angular distribution. Then one can make a scalar $s_2$ and a pseudoscalar $p_1$:
$$s_2= \hat p\cdot\hat\lambda = \cos( \theta);\; p_1= \vec
A\cdot\hat\lambda=\sin( 2 (\chi-\psi)) \cos( \theta)$$ where $\theta$ is the angle between $\hat\lambda$ and the position of the source on the dome of the sky. The pseudovector combination $\hat p \times \hat\lambda$ can also be considered, which makes the usual unit vector $\hat \theta$ sitting in the tangent plane, transforming like $(\pm 1, 1)$. By the $j_z$ addition rule for angular momentum, this cannot be combined with the $(2,2), (-2,2), (0,1) $ representations available from $\slash \chi$ and $\slash \psi$ to make new invariants. Using $\hat p$, which transforms like $(0,1)$, we can take the $(0,1)$ part of $\slash\chi$ and make a pseudo-scalar, but this is the helicity $h$ already discussed.
Continuing in this way, products of higher order can always be reduced to sums of smaller dimensional representations, much like the usual decomposition of angular momentum. Some care is needed, however, because the polarization variables are made from incomplete representations because the fields are transverse. One generally then has fewer invariants than straightforward counting using rotation group methods might indicate.
[*Non-Local Invariants:*]{} Combining different sources leads to some interesting quantities at low order. There is for example the familiar angle $\theta_{ij}$ between galaxy $i$ and galaxy $j$ on the sky, given in terms of the true scalar $cos(\theta_{ij}) =\hat p_i\cdot\hat p_j$.
More interesting are symmetric and antisymmetric combinations of $\slash\chi_i$ and $\slash\chi_j$, or $\slash\psi_j$. It is straightforward to treat these by standard methods: thus the symmetric combinations $$Tr[ \slash\chi_i
\slash\chi_j] ;\; \; Tr[ \slash\chi_i \slash{\psi}_j] ;\; \; Tr[
\slash{\psi}_i
\slash\chi_j] ;$$ are scalars; while the antisymmetric combinations or commutators $$[\slash\chi_i , \slash\chi_j] ;\; \; [\slash\chi_i ,
\slash\psi_j] ;\; \; [\slash\psi_i , \slash{\psi_j}] ;$$ are dual to pseudo-vectors.
The quantities above are involved in questions of angular coherence: for example, if one wants to smooth the angular distribution of sources at different locations, while it would be a bit sloppy to add the parameters such as $\hat\chi$, in practice we are aware only of averaging polarization parameters over single sources of small angular dimensions, in a quasi-local way, so that this should not cause a problem. However if one were making a more ambitious study –for example examining the degree of coherent polarization over larger regions of the sky–then more formal care would be needed. One can, for example, create correlation functions of the matrices in a consistent basis and evaluated at different angular positions, which are then reduced to scalars by taking traces. Another interesting quantity is the $h_i h_j$, or helicity-helicity correlation function, and obvious generalizations removing the means. This quantity does not need a galaxy axis for its evaluation, and might probe parity-violating effects of the medium evaluated as a function of angular scales. This may be useful for optical polarizations , for example, which often are not associated with axial structure. An application of interest is the cosmic micro-wave background, for which polarization measurements can be expected in the future.
[* Distributions*]{}
The classifications are useful in constructing statistical distributions or correlations which can be used to quantify observations. We will use a bracket ( $< >$ ) symbol to denote an expectation value in a normalized distribution.
Consider the problem of quantifying the correlations between $\chi$ and $\psi$ mentioned in the Introduction. From our results, one would naturally assume the distribution to be a function of the rotationally invariant quantity $\beta=\chi-\psi$. Given that $\chi$ and $\psi$ are defined up to multiples of $\pi$, the difference $\beta$ ranges in the most general case over $2 \pi$, and not $\pi$, as sometimes assumed [@CF] in the literature. This is seen very simply by making sketches of some trial distributions. Analytically one can expand the distribution $f( \chi, \psi)$ in Fourier series for integer $n, m$: $$f( \chi, \psi) = \sum f_{nm} \exp ( i 2n \chi+ i 2 m \psi) =\sum f_{nm}
\exp
\left[ i (n - m) ( \chi- \psi) + i (n+ m) ( \chi+\psi)\right]$$ Reality of $f$ must also be imposed; it prescribes the negative integer values of $f_{nm}$, but does not restrict whether $n \pm m$ is even or odd. The coefficients $f_{nm}$ for odd values of $n - m$ make $\beta=\chi-\psi$ periodic on the interval of $2\pi$, as claimed. Restricting non-zero coefficients to even values of $n - m$ can be motivated by extra assumptions. One sufficient condition is that the distribution obeys overall rotational symmetry, giving $f_{nm}$ going like $\delta( n+m) $. Such an assumption might seem very general but in fact it is not. It is an interesting fact of optics that even a perfectly transmitting (unitary) medium can treat different polarizations dissimilarly, leading to a non-trivial distribution of $\chi+\psi$ as a signal. The same mechanism can create an anisotropic distribution of linear polarizations along a pencil through a medium from an random uncorrelated distributed set of linear polarization emitters. To simplify the analysis we will assume here, however, that the distribution of $\beta$ is periodic on the interval of $\pi$.
To make invariant distributions from invariants, this leaves us with the quantities constructed earlier, namely $s_1=\cos(2\beta)$ and $p_1=\sin(2
\beta)$. A simple, and indeed well-known distribution that follows is the von Mises form [@batschelet; @fisher; @mardia]: $f_{vM}(\beta) = const.\
\exp(
k \cos(2\beta)) = const. \exp( k s_1 ) $. This distribution has often been used in likelihood tests, but unfortunately without discussion of parity symmetry. By construction, the $vM$ distribution embodies a physical assumption that the twist of one angle relative to another has no preferred parity, which may be unsuitable in some cases. The “shifted" von Mises distribution is similar: $f_{\rm shifted-vM}
(\beta) = {\rm const.} \exp(k \cos(2(\beta- \alpha))$. The parameter $\alpha$ shifts the origin of $\beta$ and might seem to be free. However if parity symmetry is assumed, then $\alpha$ is quite restricted. The distribution is a function of $\cos(2\beta- 2\alpha)=\cos(2\beta) \cos(2 \alpha) +
\sin(2\beta)\sin(2 \alpha)$. Parity symmetry requires $f(\beta)=f(-\beta)$, yielding $\sin(2\alpha)=0$, or $\alpha=0, \pi/2$. Parity symmetry, then, is sufficient to predict that marginal distributions of linear polarizations tend to be oriented either along galaxy major axes, or perpendicular. This has been a misunderstood point, because of assertions that only the single choice of angular origin relative to the perpendicular is sensible. For any statistic which is covariant under a shift of origin, however, the matter is irrelevant, making no difference. Transforming $\beta\rightarrow\beta - \pi/2$ is equivalent to transforming $k\rightarrow -k$ in the $vM$ or shifted $vM$ distributions, showing that either choice is equally well described automatically. Similarly, $\sin(2
\beta)$ is odd in both $\beta$ and $\beta- \pi/2$, a pseudoscalar no matter how angles are measured. Another example of the issue is underscored by the paper of Loredo et al.[@loredo] using the shifted $vM$ distribution in a detailed likelihood analysis. The paper responded to an odd-parity statistic used in Ref. [@NR], but exclusively used an even parity shifted $vM$ model in its analysis. In replacing the original statistic by one with opposite transformation properties, conclusions were drawn on a false basis. This illustrates the problem that can occur when using functions from a different symmetry class than the idea being tested.
There is in fact a long history of mix-ups from lumping together observables of different parity, which has led to interesting contention in the literature. Birch [@birch] in 1982 empirically observed a [*pseudovector*]{} correlation in radio polarizations, odd in the difference $(\chi-\psi)$. Kendall and Young ($KY$) made a model distribution to explore this[@KY], choosing $vM$ for the [*null*]{} distribution of $\beta$. This is an implicit assumption of parity conservation in the null, which is perfectly physical (but which should be stated explicitly). $KY$ then set up a conditional correlation function $C_{KY}( \beta, \hat p)=exp( \mu \hat\lambda\cdot\hat p \sin(2\beta))$ which involved odd-parity to test Birch’s pseudovector result. Unfortunately Kendall and Young did not explain the reason for their ansatz, which in retrospect was entirely appropriate. The overall $KY$ distribution finally can be written in a nicely compact form $$f_{KY}(\chi, \psi, \hat p) =exp( k\cos( 2\beta)
+\mu\hat \lambda\cdot\hat p \sin(2\beta)) =\exp( k s_1 + \mu p_1)$$ where $s_1$ and $p_1$ are our scalar and pseudoscalars from the first section, making it clear that $\mu$ is a parity-violation anisotropic correlation parameter. Likelihood analysis, for example, can be used in a perfectly objective way to see whether any parity–violating effects might be present or not in a data set, and entirely separate from the need to quantify the marginal distribution of $\beta$. It is interesting that Kendall and Young’s data analysis [@KY] along such lines then indicated a high statistical significance for the parity -violating effect Birch had observed; more recent work in this regard can also be cited [@JR]. However, other independent studies which used invariant correlations lumping together different parity seemed to contradict the result, a topic to which we now turn.
[* Correlations*]{}
One of the best-known statistical correlation tests comes from an influential paper by Jupp and Mardia[@JM]. Their prescription for correlation between 2 angular quantities involves mapping them into ”vectors" $v, w$, that is covariantly transforming elements, of which $\hat\chi$ and $\hat\psi$ are examples we have already seen. To test for correlations, $JM$ use canonical $p \times q$ correlation matrices $$\Sigma^{vw}_{ij} = <
(v_i-<v_i>)( w_j-<w_j> )>$$ which behave under the separate transformations of $v, w$ as the indices would indicate. Then the $JM$ correlation test is to calculate $$n\rho^2_{v, w} = n\ Tr[\Sigma^{vw} (\Sigma^{ww})^{-1}
\Sigma^{wv}(\Sigma^{vv})^{-1} ],$$ with matrix products indicated, and where $n$ is the number of data points in the sample. The step of dividing by auto-correlation matrices is used to make the quantity scale-invariant. One easily finds that $n\rho^2_{vw}=0 $ if $v$ and $w$ are independent, and the distribution of fluctuations in an uncorrelated null distribution has also be obtained[@JM]. Thus $n\rho^2_{v, w} $ has served for many years as a useful simple test for independence.
Unfortunately Jupp and Mardia, in discussing their correlation test, did not discuss parity and other discrete symmetries. In the problem at hand, to test whether the difference $\beta$ is correlated with sky position, one might consider the $JM$ correlation of a “natural vector“ $v=(\cos(2\beta),\sin
(2\beta))$ and $w= \hat p$. This particular combination $v$ was used to test Birch’s correlation by Bietenholz and Kronberg[@BK]. Recall, however, the even-odd rule of parity for those quantities even-or odd in $\beta$. It is clear that the above ”natural" vectorlike combination $v$ needlessly mixes two quantities which are of opposite parity. The two pieces are also separately invariant under rotations, and should not be combined into a vector. Ironically, history shows that the mixed parity combination was used while at the same time citing that Birch had a correlation of pseudovector character [@BK]; without recognizing the parity properties of Kendall and Young’s procedure, both their and Birch’s results were then rejected. From our analysis it is sufficient to use $n\rho^2_{v_1, w} $ and $n\rho^2_{v_2, w} $, which allows separate tests of scalar or pseudoscalar kind. One can, of course, also work directly with scalar and pseudoscalar measures such as $<s_1>$ and $<p_1> $ to explore certain features of data, and create any number of statistics, once the proper transformation properties have been respected.
Summary
=======
We have classified several combinations of the polarization and angular correlation observables under parity and angular momentum. The features of local observables, and their simple appearance in classic distributions and correlation analysis, should be helpful to those interested in the area. While not exploring very far into the non-local quantities, they appear to offer many possibilities for interesting studies. The possibility of exploring parity violation when there are no obvious emission axis variables, as in the case of upcoming polarized cosmic -microwave background observations from ground-based and satellite facilities, seems intriguing and bears further investigation.
ACKNOWLEDGMENTS: This work was supported in part under DOE grant number 85ER40214, by the University of Kansas General Research Fund, and the [*Kansas Institute for Theoretical and Computational Science/ K\*STAR*]{} program.
L. Stodolsky, [*Phys. Rev. Lett.*]{} [**34**]{}, 110, 1975; Erratum-ibid. [**34**]{}, 508, 1975. Gabriel Karl, [*Can. J. Phys.*]{} [**54**]{}, 568, 1976; J.N. Clarke, G. Karl, and P.J.S. Watson, [*Can. J. Phys.*]{} [**60**]{}, 1561, 1982. B. Nodland and J. P. Ralston, [*Physical Review Letters*]{} [**78**]{}, 3043 (1997). P. Jain and J. P. Ralston, Kansas preprint 98/07, astro-ph/9803164. R. D. Peccei and H. Quinn, [*Phys. Rev. Lett.*]{} [**38**]{}, 1440 (1977); S. Weinberg, [*Phys. Rev. Lett.*]{} [**40**]{}, 223 (1978); F. Wilczek, [*Phys. Rev. Lett.*]{} [**40**]{}, 279 (1978); P. Sikivie, [*Phys. Lett.*]{} [**B137**]{}, 353 (1984); D. Harari and P. Sikivie, [*Phys. Lett.*]{} [**B289**]{}, 67 (1992). W.-T. Ni, [*Phys. Rev. Lett.*]{} [**38**]{}, 301 (1977), C. Wolf, Phys. Lett. [**A132**]{}, 151,(1988). M. Mandle and E. Wolf, [*Coherence and Quantum Optics*]{}, Cambridge University Press, 1985; L. D. Landau, E. M. Lifshitz, and L.P. Piatevskii, [*Electrodynamics of Continuous Media*]{}, 2nd Edition (Pergamon Press 1984). 1995). E. Batschelet, [*Circular Statistics in Biology*]{}, (London: Academic Press, 1981). N. I. Fisher, [*Statistics of Circular Data*]{}, (Cambridge, 1993). K. Rohlfs and T. L. Wilson, [*Tools of Radio Astronomy*]{}, 2nd Edition (Springer Verlag, 1996). S. M. Carroll and G. B. Field, [*Phys. Rev. Lett.*]{} [**79**]{}, 2934 (1997); S. M. Carroll, G. B. Field and R. Jackiw, [*Phys. Rev.*]{} [**D41**]{}, 1231 (1990) K. V. Mardia, [*Statistics of Directional Data*]{}, (London: Academic Press, 1972). T. J. Loredo, E. E. Flanagan, and I. M. Wasserman, astro-ph/9706258 (1997). P. Birch, [*Nature*]{} [**298**]{}, 451 (1982). D. G. Kendall and A. G. Young, [*Mon. Not. Roy. Ast. Soc.*]{} [**207**]{}, 637, (1984). P. E. Jupp and K. V. Mardia, [*Biometrika*]{} [**67**]{}, 163 (1980). M. F. Bietenholz and P. P. Kronberg, [*Astrophys J.* ]{} [**287**]{}, L1-L2 (1984).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We find the inconsistency of dimensional reduction and naive dimensional regularization in their applications to Chern-Simons type gauge theories. Further we adopt a consistent dimensional regularization to investigate the quantum correction to non-Abelian Chern-Simons term coupled with fermionic matter. Contrary to previous results, we find that not only the Chern-Simons coefficient receives quantum correction from spinor fields, but the spinor field also gets a finite quantum correction.'
address: |
High Energy Physics Division, Department of Physics\
and\
Helsinki Institute of Physics,\
P.O. Box 9 (Siltavuorenpenger 20 C), FIN-00014 Helsinki, Finland
author:
- 'M. Chaichian and W. F. Chen'
title: ' Inconsistency of Naive Dimensional Regularizations and Quantum Correction to Non-Abelian Chern-Simons-Matter Theory Revisited'
---
Introduction
============
There has been a considerable amount of popularity in perturbative Chern-Simons-type theory due to the particular feature of the finite renormalization of its coupling constant. Almost all the old regularization schemes $\cite{ref1,ref2,ref3,ref4}$ and even some newly developed ones $\cite{ref5,ref6}$ have been applied. The existence of the antisymmetric tensor in Chern-Simons term makes the implement of the regularization method much more non-trivial. Specially the one-loop quantum correction to a general three-dimensional field theory is very delicate $\cite{ref7}$. Different regularization methods can easily produce ambiguity in the quantum corrections. In particular, it appears that some old regularization schemes can bring non-physical quantum corrections$\cite{ref8,ref9}$. This makes one be cautious about the use of some regularization methods. In this paper we show that dimensional reduction and naive dimensional regularization are inconsistent when they are applied to Chern-Simons type theories. Indeed, when we use the consistent dimensional regularization to re-calculate the one-loop quantum correction for one typical example, Chern-Simons term coupled with spinor field, we obtain a result different from the previous ones$\cite{ref10}$.
Regulating Chern-Simons theory by dimensional reduction means evaluating all the antisymmetric tensor algebra in three dimensions but performing the loop momentum integration in $n$-dimension$\cite{ref10,ref11,ref12}$. The concrete definition in three dimensions is as follows: $$\begin{aligned}
&& {\epsilon}^{\mu\nu\rho}{\epsilon}_{\alpha\beta\gamma}
={\delta}^{\mu}_{[~\alpha}{\delta}^{\nu}_{~\beta}\delta^{\rho}_{~\gamma]},~
{\delta}_{~\mu}^{\mu}=3, ~\mu,\nu,\cdots=0,1,2,\nonumber\\
&& \delta^{\mu}_{~\nu}=\tilde{\delta}^{\mu}_{~\nu}+\hat{\delta}^{\mu}_{~\nu},
~\tilde{\delta}^{\mu}_{~\nu}\tilde{\delta}^{\nu}_{~\rho}
=\tilde{\delta}^{\mu}_{~\rho},~
\hat{\delta}^{\mu}_{~\nu}\hat{\delta}^{\nu}_{~\rho}
=\hat{\delta}^{\mu}_{~\rho},
~\tilde{\delta}^{\mu}_{~\nu}\hat{\delta}^{\nu}_{~\rho}=0,\nonumber\\
&&k_{\mu}\tilde{\delta}^{\mu}_{~\nu}=k_{\mu},
~k_{\mu}\hat{\delta}^{\mu}_{~\nu}=0; ~\tilde{\delta}^{\mu}_{~\mu}=n,
~\hat{\delta}^{\mu}_{~\mu}=3-n.
\label{eq1}\end{aligned}$$ The inconsistency of this regularization method in four-dimensional supersymmetric field theories had already been found by its inventor$\cite{ref11}$. For three-dimensional case, it can also be easily shown that this regularization method is not consistent. From Eq.(\[eq1\]), we have $$\begin{aligned}
\tilde{\epsilon}^{\mu\nu\rho}\tilde{\epsilon}_{\mu\nu\rho}&=&
\tilde{\delta}^{\mu}_{[~\mu}\tilde{\delta}^{\nu}_{~\nu}
\tilde{\delta}^{\rho}_{~\rho]}
=n^3-3n^2+2n=n(n-1) (n-2);\nonumber\\
\hat{\epsilon}^{\mu\nu\rho}\hat{\epsilon}_{\mu\nu\rho}&=&
\hat{\delta}^{\mu}_{[~\mu}\hat{\delta}^{\nu}_{~\nu}\hat{\delta}^{\rho}_{~\rho]}
=(3-n)^3-3(3-n)^2+2(3-n)=(3-n)(2-n)(1-n).
\label{eq2}\end{aligned}$$ So we can obtain $$\begin{aligned}
0&=&\left(\tilde{\epsilon}^{\mu\nu\rho}\hat{\epsilon}_{\alpha\beta\gamma}\right)
\left(\tilde{\epsilon}_{\mu\nu\rho}\hat{\epsilon}^{\alpha\beta\gamma}\right)
=\left(\tilde{\epsilon}^{\mu\nu\rho}\tilde{\epsilon}_{\mu\nu\rho}\right)
\left(\hat{\epsilon}_{\alpha\beta\gamma}\hat{\epsilon}^{\alpha\beta\gamma}
\right)\nonumber\\
&=&n(3-n)(n-1)^2(n-2)^2.
\label{eq3}\end{aligned}$$ Therefore, it is only valid for $n=0,1,2,3$, and thus it is not the analytic dimensional continuation as that required by dimensional regularization.
As for the naive dimensional regularization, it defines the antisymmetric tensor algebra to satisfy$\cite{ref10}$ $$\begin{aligned}
{\epsilon}_{\mu\sigma\eta}{\epsilon}^{\mu\lambda\tau}
=({\delta}_{~\sigma}^{\lambda}
{\delta}_{~\eta}^{\tau}-{\delta}_{~\sigma}^{\tau}{\delta}_{~\eta}^{\lambda})
{\Gamma}(n-1),
~~{\delta}_{~\sigma}^{\sigma}=n.
\label{eq4}\end{aligned}$$ We can show that this definition in essence makes the theory defined in three dimensions, but not in $n$-dimension as it should be. This can be seen from the following simple algebraic manipulations. Consider the quantity ${\epsilon}_{\mu\sigma\eta}{\epsilon}^{\mu\lambda\tau}
{\epsilon}_{\alpha\lambda\tau}$. On one hand, it equals to $$\begin{aligned}
({\epsilon}_{\mu\sigma\eta}{\epsilon}^{\mu\lambda\tau})
{\epsilon}_{\alpha\lambda\tau}
&=&{\Gamma}(n-1)({\delta}_{~\sigma}^{\lambda}
{\delta}_{~\eta}^{\tau}-{\delta}_{~\sigma}^{\tau}
{\delta}_{~\eta}^{\lambda}){\epsilon}_
{\alpha\lambda\tau}\nonumber\\
&=&{\Gamma}(n-1)({\epsilon}_{\alpha\sigma\eta}
-{\epsilon}_{\alpha\eta\sigma})
=2{\Gamma}(n-1){\epsilon}_{\alpha\sigma\eta};
\label{eq5}\end{aligned}$$ on the other hand, we have $$\begin{aligned}
{\epsilon}_{\mu\sigma\eta}({\epsilon}^{\mu\lambda\tau}
{\epsilon}_{\alpha\lambda\tau})
&=&{\epsilon}_{\mu\sigma\eta}{\Gamma}(n-1)
({\delta}_{~\mu}^{\alpha}{\delta}_{~\lambda}
^{\lambda}-{\delta}_{~\lambda}^{\mu}{\delta}_{~\alpha}^{\lambda})\nonumber\\
&=&{\Gamma}(n-1)(n-1){\epsilon}_{\alpha\sigma\eta}.
\label{eq6}\end{aligned}$$ Comparing Eq.(\[eq5\]) with Eq.(\[eq6\]), we can see that only $n=3$, otherwise ${\Gamma}(n-1)=0$. Thus the naive dimensional regularization also does not make the theory well defined.
This motivates us to re-consider Chern-Simons type theory in the dimensional regularization schemes proposed by ’t Hooft and Veltman$\cite{ref13}$. To our knowledge, up to now this is the only dimensional continuation scheme compatible with gauge symmetry to deal with ${\gamma}_5$ and the similar problems. In the following we use this consistent dimensional regularization to investigate the one-loop quantum corrections of non-Abelian Chern-Simons term coupled to spinor field, the classical action of which in Minkowski space is $$\begin{aligned}
S={\int}d^3x\left[{\epsilon}^{\mu\nu\rho}
\left(\frac{1}{2}A_{\mu}^a{\partial}_{\nu}A_{\rho}^a
+\frac{1}{3!}gf^{abc}A_{\mu}^aA_{\nu}^bA_{\rho}^c\right)+
\bar{\psi}(i/\hspace{-2mm}{\partial}-m
+g/\hspace{-2.5mm}A^aT^a){\psi}\right],
\label{eq7}\end{aligned}$$ where ${\psi}$ belongs to the fundamental representation of gauge group, and for simplicity we only consider the one-flavour case; $\gamma_{\mu}$ ($\mu =0,1,2$) are usually chosen as follows$\cite{ref14}$, $$\begin{aligned}
&&{\gamma}^0={\sigma}_2,~{\gamma}^1=i{\sigma}_3,~{\gamma}^3=i{\sigma}_1;
\nonumber\\
&&{\gamma}_{\mu}{\gamma}_{\nu}=
g_{\mu\nu}-i{\epsilon}_{\mu\nu\rho}{\gamma}^{\rho},
~g_{\mu\nu}=\mbox{diag}(1,-1,-1).
\label{eq8}\end{aligned}$$ This model has become revived in recent years owing to its possible physical application to condensed matter theory. The coefficient of Chern-Simons term (called statistical parameter) plays a crucial role in transmuting the spin and the statistics of the anyon particles. The quantum corrections up to two-loop for this model was investigated in dimensional reduction and naive dimensional regularization$\cite{ref10}$. It was found that there exist no quantum corrections and all the renormalization constant are identically equal to one.
For the case of Chern-Simons typical theory, the dimensional continuation proposed by ’t Hooft and Veltman, which had been been explicitly written down in Refs.$\cite{ref5,ref18}$, is as follows, $$\begin{aligned}
&&{\epsilon}^{\mu_1\mu_2\mu_3}{\epsilon}_{\nu_1\nu_2\nu_3}
={\sum}_{{\pi}{\in}P_3}\mbox{sgn}(\pi)\Pi_{i=1}^3
\tilde{\delta}^{\mu_i}_{~\nu_{\pi(i)}},
~g_{\mu\nu}=\tilde{g}_{\mu\nu}{\oplus}\hat{g}_{\mu\nu},
~p_{\mu}=\tilde{p}_{\mu}{\oplus}\hat{p}_{\mu},\nonumber\\[2mm]
&& {\epsilon}^{\mu\nu\rho}\hat{g}_{\rho\alpha}=0,
~ {\epsilon}^{\mu\nu\rho}\hat{p}_{\rho}=0,
~\tilde{\delta}_{~\mu}^{\mu}=3, ~\hat{\delta}_{~\mu}^{\mu}=n-3~.\end{aligned}$$ Then $n$ is continued to a complex variable to regulate the theory. However, the price that must be paid for this consistent definition, as that pointed out in Ref.$\cite{ref4}$, is first performing higher covariant derivative regularization. Otherwise this dimensional continuation will lead to linear dependence of Chern-Simons kinetic operator even after gauge-fixing. As usual, we choose the simplest higher covariant derivative term, the Yang-Mills Lagrangian, $$\begin{aligned}
-\frac{1}{4M}F_{\mu\nu}^aF^{a\mu\nu},
~~~F_{\mu\nu}^a={\partial}_{\mu}A_{\nu}^a -
{\partial}_{\nu}A_{\mu}^a+gf^{abc}A_{\mu}^bA_{\nu}^c.
\label{eq10}\end{aligned}$$ The ghost and gauge-fixing terms have the well known form in the covariant gauge, $$\begin{aligned}
S_{\rm ghost}+S_{\rm g.f.}=\int d^3x \left[-\partial^{\mu}\bar{c}^a
\left(\partial_{\mu}c^a+gf^{abc}A_{\mu}^bc^c\right)
-\frac{1}{2\alpha}(\partial^{\mu}A_{\mu}^a)^2\right].\end{aligned}$$ There still exists another difficulty, that is, this dimensional prescription in fact defines the $n$-dimensional ${\epsilon}^{\mu\nu\rho}$ effective only in three dimensions This makes the regulated theory possess the $SO(3){\otimes}SO(n-3)$ covariance rather than $SO(n)$, the regulated propagator will not only take very complicated form, but it is also not $SO(n)$ covariant. This will make the loop integration very difficult to carry out. However, thanks to $\cite{ref4}$, one can prove that the propagator of gauge field can be decomposed into two parts: one part is composed of evanescent quantity, which has no contribution to the loop integration in the limit $n{\longrightarrow}3$; then one can make use of the second part as an effective propagator $$\begin{aligned}
G^{ab}_{\mu\nu}(p)=-
\frac{i M}{p^2(p^2-M^2)}
\left(iM{\epsilon}_{\mu\nu\rho}p^{\rho}+p^2g_{\mu\nu}-p_{\mu}p_{\nu}\right)
\label{eq11}\end{aligned}$$ in order to perform calculation. Note that we have chosen the Landau gauge ($\alpha=0$). The other Feynman rules are listed as below:
- Fermion propagator $$\begin{aligned}
S(p)=i\frac{/\hspace{-2mm}p+m}{p^2-m^2}\delta_{ij};
\label{eq12}\end{aligned}$$
- Quark-Gluon vertex $$\begin{aligned}
ig{\gamma}_{\mu}T^a_{ij}(2\pi)^3 {\delta}^{(3)}(p+q+r).
\label{eq13}\end{aligned}$$
In Sect.II, using the consistent dimensional continuation, we re-consider some of the new one-loop two-point Green functions such as the fermionic self-energy, the ghost self-energy and the contribution to vacuum polarization tensor from the fermionic loop. Sect.III is about the one-loop three-point functions such as the fermion-gluon vertex, the ghost-gluon vertex and so on. Since it is quite complicated to straightforwardly calculate the fermion-gluon vertex in a non-Abelian gauge theory, we make use of the Slavnov-Taylor identity between fermion-gluon vertex and the composite ghost-fermion vertex to facilitate the calculation. In Sect.IV, we define the finite renormalization of the coupling constant with mass-shell renormalization convention and we show that the result is different from that presented in the literature previously$\cite{ref3}$. Finally, in Sect.V we emphasize our conclusions and discuss the justification for our results. For clarity and completeness, a derivation of the needed Slavnov-Taylor identities from BRST symmetry is presented in Appendix.
One-loop Two-Point Function
===========================
Contribution to Vacuum Polarization Tensor from Fermionic Loop
--------------------------------------------------------------
The contributions to vacuum polarization tensor from the self-interaction of gauge fields and the ghost loop have been shown in many works$\cite{ref3,ref4,ref19}$. Here we only consider the extra contribution from the spinor field.
The relevant Feynman diagram is shown in Fig.1 and the amplitude is $$\begin{aligned}
i{\Pi}_{\mu\nu}^{(f)ab}&=&-
g^2\mbox{Tr}(T^aT^b){\int}\frac{d^nk}{(2\pi)^n}
\frac{\mbox{Tr}\left\{{\gamma}_{\nu}[/\hspace{-2mm} k+/\hspace{-2mm} p+m]
{\gamma}_{\mu}
(/\hspace{-2mm} k+m]\right\}}
{(k^2-m^2)[(p+k)^2-m^2]}\nonumber\\[2mm]
&=& -\frac{1}{2}g^2{\delta}^{ab}
{\int}\frac{d^nk}{(2\pi)^n}\frac{-im {\epsilon}_{\mu\nu\rho}p^{\rho}+
2k_{\mu}k_{\nu}+k_{\mu}p_{\nu}+k_{\nu}p_{\mu}-g_{\mu\nu}[k{\cdot}(k+p)-m^2]}
{(k^2-m^2)[(p+k)^2-m^2]},
\label{eq14}\end{aligned}$$ where we choose the normalization of group factor as, $$\begin{aligned}
\mbox{Tr}(T^aT^b)=\frac{1}{2}\delta^{ab}.\end{aligned}$$ Calculation (after taking the limit $n{\longrightarrow}3$) gives $$\begin{aligned}
i{\Pi}_{\mu\nu}^{(f)ab}&=&\frac{ig^2}{16\pi}{\delta}^{ab}\left\{
i{\epsilon}_{\mu\nu\rho}p^{\rho}\frac{m}{p}
\ln\left[\frac{1+p/(2m)}{1-p/(2m)}\right]
\right.\nonumber\\[2mm]
&-&\left.\left(p^2g_{\mu\nu}-p_{\mu}p_{\nu}\right)\frac{1}{m}
\left[-\frac{m^2}{p^2}+\left(\frac{1}{4}\frac{m}{p}+\frac{m^3}{p^3}\right)
\ln\left(\frac{1+p/(2m)}{1-p/(2m)}\right)\right] \right\},
\label{eq15}\end{aligned}$$ where $p{\equiv}|p|$. Using the expansion near $p=0$, $$\begin{aligned}
\ln\left[\frac{1+p/(2m)}{1-p/(2m)}\right]
=\frac{p}{m}+\frac{1}{12}\frac{p^3}{m^3}
+\frac{1}{80}\frac{p^5}{m^5}+{\cdots},
\label{eq16}\end{aligned}$$ we have $$\begin{aligned}
{\Pi}_{\mu\nu}^{(f)ab}(0)=\frac{g^2}{8\pi}{\delta}^{ab}\left[
i{\epsilon}_{\mu\nu\rho}p^{\rho}-\frac{1}{3m}(p^2g_{\mu\nu}-p_{\mu}p_{\nu})
\right].
\label{eq17}\end{aligned}$$ Combining Eq.(\[eq15\]) with the contributions to polarization tensor from gluon and ghost loops$\cite{ref3,ref4,ref19}$, $$\begin{aligned}
{\Pi}_{\mu\nu}^{({\rm gl})ab}(p)+{\Pi}_{\mu\nu}^{({\rm gh})ab}(p)
=-\frac{7}{3}\frac{g^2}{4\pi}C_V\delta^{ab}\epsilon_{\mu\nu\rho}p^{\rho},\end{aligned}$$ and choosing the renormalization condition on gluon mass-shell [^1] $p=0$, $$\begin{aligned}
\Pi^{abR}_{\mu\nu}(0)=0,\end{aligned}$$ we can define the gluon wave function renormalization constant, $$\begin{aligned}
Z_A=1-\frac{g^2}{4\pi}\left(\frac{7}{3}C_V+\frac{1}{2}\right).\end{aligned}$$
Self-energy for Spinor Field
----------------------------
Let us consider fermionic self-energy. Its Feynman diagram is shown in Fig.2 and the amplitude is read as $$\begin{aligned}
-i\Sigma (p,M)&=&-g^2(T^aT^a)M{\int}\frac{d^nk}{(2\pi)^n}
\frac{{\gamma}_{\nu}[/\hspace{-2mm} p+/\hspace{-2mm} k+m]{\gamma}_{\mu}
[iM{\epsilon}_{\mu\nu\rho} k^{\rho}
+k^2g_{\mu\nu}-k_{\mu}k_{\nu}]}{[(k+p)^2-m^2]k^2(k^2-M^2)}
\nonumber \\[2mm]
&=&-2Mg^2C_2(R){\bf 1} \int\frac{d^nk}{(2\pi)^n}
\frac{(M-/\hspace{-2mm}k)k{\cdot}(k+p)
+m(k^2-\gamma_\alpha k^{\alpha}M)}{[(k+p)^2-m^2]k^2(k^2-M^2)},
\label{eq18}\end{aligned}$$ where ${\bf 1}$ is the unit matrix in colour space, and we have used Eq.(\[eq8\]) and the identity $$\begin{aligned}
\int \frac{d^nk}{(2\pi)^n}\frac{{\epsilon}_{\mu\nu\rho} p^{\nu}k^{\rho}}
{[(k+p)^2-m^2]k^2(k^2-M^2)}=0.
\label{eq19} \end{aligned}$$ Using the decomposition $$\begin{aligned}
\frac{1}{k^2(k^2-M^2)}=\frac{1}{M^2}\left(
\frac{1}{k^2-M^2}-\frac{1}{k^2}\right),
\label{eq20}\end{aligned}$$ and $$\begin{aligned}
k{\cdot}p&=&\frac{1}{2}\{[(k+p)^2-m^2]-k^2-(p^2-m^2)\}\nonumber\\[2mm]
&=&\frac{1}{2}\{[(k+p)^2-m^2]-(k^2-M^2)-(p^2-m^2-M^2)\},
\label{eq21}\end{aligned}$$ we can write Eq.(\[eq18\]) as follows: $$\begin{aligned}
-i\Sigma (p,M)&=&-2Mg^2C_2(R){\bf 1} \int\frac{d^nk}{(2\pi)^n}\left\{
\left(\frac{p^2-m^2}{2M^2}-\frac{1}{2}-\frac{m}{M}\right)
\frac{/\hspace{-2mm} k-M}{(k^2-M^2) [(k+p)^2-m^2]}\right.\nonumber\\
&+&\frac{1}{2M}\frac{1}{k^2-M^2}
+\frac{p^2-m^2}{2M}\frac{1}{k^2[(k+p)^2-m^2]}\nonumber\\
&-&\left. \left(\frac{p^2-m^2}{2M^2}
-\frac{m}{M}\right)\frac{/\hspace{-2mm}k}{k^2[(k+p)^2-m^2]}\right\}.
\label{eq22}\end{aligned}$$ The standard integration gives $$\begin{aligned}
-i\Sigma (p,M)&=&-\frac{i}{4\pi}g^2C_2(R){\bf 1}M\left\{1+\frac{p^2-m^2}{Mp}
\ln\left(\frac{1+p/m}{1-p/m}\right)\right.\nonumber\\
&-&\gamma_{\mu}p^{\mu}\left(\frac{p^2-m^2}{2M^2}-\frac{m}{M}
\right)\left[\frac{m}{p^2}+\left(\frac{1}{p}-\frac{m^2}{p^3}\right)
\ln\left(\frac{1+p/m}{1-p/m}\right)\right]\nonumber\\[2mm]
&+&\frac{/\hspace{-2mm}p}{p^2}\left(\frac{p^2-m^2}{2M^2}-\frac{1}{2}
-\frac{m}{M}\right)\left[M-m-\frac{p^2-m^2+M^2}{2p}
\ln\left(\frac{1+(p+m)/M}{1-(p+m)/M}\right) \right]\nonumber\\
&+&\left.\left(\frac{m}{p}+\frac{M}{2p}-\frac{p^2-m^2}{2Mp}\right)
\ln\left(\frac{1+(p+m)/M}{1-(p+m)/M}\right)\right\}.
\label{eq23}\end{aligned}$$ After taking the large-$M$ limit, we obtain the quark self-energy $$\begin{aligned}
-i\Sigma (p)&=&-\frac{i}{4\pi}g^2C_2(R){\bf 1}\left\{2M+m
+\frac{p^2-m^2}{p}\ln\left(\frac{1+p/m}{1-p/m}\right)\right.\nonumber\\
&-&\left./\hspace{-2mm}p
\left[\frac{m^2}{p^2}+\left(\frac{m}{p}-\frac{m^3}{p^3}\right)
\ln\left(\frac{1+p/m}{1-p/m}\right)-\frac{2}{3}\right]\right\}.
\label{eq24}\end{aligned}$$ As usual, this quark self-energy can be written in the form of quark mass expansion, $$\begin{aligned}
\Sigma (p)&=&\frac{1}{2\pi}g^2C_2(R){\bf 1}\left(M+\frac{m}{3}\right)
+\frac{1}{4\pi}g^2C_2(R){\bf 1}\frac{5}{3}(/\hspace{-2mm} p-m)
\nonumber\\
&+&\frac{1}{4\pi}g^2C_2(R){\bf 1}\left\{2m
+\frac{p^2-m^2}{p}\ln\left(\frac{1+p/m}{1-p/m}\right)\right.\nonumber\\
&-&\left./\hspace{-2mm}p
\left[1+\frac{m^2}{p^2}+\left(\frac{m}{p}-\frac{m^3}{p^3}\right)
\ln\left(\frac{1+p/m}{1-p/m}\right)\right]\right\}\nonumber\\
&=&\delta m{\bf 1} -(Z_\psi^{-1}-1)(/\hspace{-2mm}p-m){\bf 1}
+Z_\psi^{-1}\Sigma_R (p).
\label{eq25}\end{aligned}$$ Thus in the quark mass-shell renormalization scheme, we have the renormalization constants and the radiative correction of quark self-energy as: $$\begin{aligned}
m_{\rm ph}&=&m-\delta m=m-\frac{g^2}{2\pi}C_2(R)(M+m);\nonumber\\
Z_\psi&=&1+\frac{5}{3}\frac{g^2}{4\pi}C_2(R);\nonumber\\
\Sigma_R (p)&=&\frac{1}{4\pi}g^2C_2(R){\bf 1}\left\{2m_{\rm ph}
+\frac{p^2-m^2_{\rm ph}}{p}
\ln\left(\frac{1+p/m_{\rm ph}}{1-p/m_{\rm ph}}\right)
\right.\nonumber\\
&-&\left./\hspace{-2mm}p
\left[1+\frac{m^2_{\rm ph}}{p^2}+\left(\frac{m_{\rm ph}}{p}
-\frac{m^3_{\rm ph}}{p^3}\right)
\ln\left(\frac{1+p/m_{\rm ph}}{1-p/m_{\rm ph}}\right)\right]\right\}.
\label{eq26}\end{aligned}$$
Self-energy for Ghost Field
---------------------------
The self-energy for ghost field had been explicitly shown in $\cite{ref19}$ in a different method. Here for completeness and later use, we re-calculate it in terms of consistent dimensional regularization (Fig.3), $$\begin{aligned}
i\Sigma_g^{(1)ab}(p)p^2&=&
\lim_{M\to\infty}g^2 C_V\delta^{ab}\int\frac{d^nk}{(2\pi)^n}
\frac{M}{k^2(k^2-m^2) (k+p)^2}\left[k^2p^2-(k.p)^2\right]\nonumber\\
&=&\lim_{M\to\infty}g^2 C_V\delta^{ab}\int\frac{d^nk}{(2\pi)^n}\left[
\frac{Mp^2}{(k^2-M^2)(k+p)^2}-\frac{1}{M}\frac{(k\cdot p)^2}{(k^2-M^2)(k+p)^2}
\right] \nonumber\\
&=&\lim_{M\to\infty}
g^2C_V\delta^{ab}ip^2\frac{1}{8\pi}\left[\frac{1}{2}+\frac{1}{2}
\frac{M^2}{p^2}-\frac{1}{4} \frac{M^3}{p^3}\left(1-\frac{p^2}{M^2}\right)^2
\ln\left(\frac{1+p/M}{1-p/M}\right)\right]\nonumber\\
&=&g^2C_V\delta^{ab}\frac{i}{4\pi}p^2\frac{2}{3}.
\label{eq27}\end{aligned}$$ Consequently, one can define the wave function renormalization constant for ghost field, $$\begin{aligned}
Z_c=1+\frac{g^2}{4\pi}\frac{2}{3}C_V.
\label{eq27n}\end{aligned}$$
One-loop Three-point function
=============================
One-loop On-shell Quantum Correction to Fermion-Gluon Vertex
------------------------------------------------------------
Let us see the one-loop quantum correction to quark-gluon vertex, which receives contributions from two Feynman diagrams (Fig.4). The first diagram is quite simple and can be calculated analytically. However the calculation for the second digram is quite complicated since it contains one three-gluon vertex and two gauge field propagators. Thus we shall make use of the Slavnov-Taylor identity to convert the calculation of fermion-gluon vertex into that of composite fermion-ghost vertices, whose amplitude can be easily calculated. The detailed derivation of this identity and its one-loop form are listed in Appendix.
From Eq.(\[eqa13\]), we can see that to calculate the quark-gluon vertex, three parts need to be considered. The first part is associated with the ghost field self-energy, which can be easily obtained from Eq.(\[eq27\]), $$\begin{aligned}
\Sigma_g^{(1)}(p)=\frac{g^2}{4\pi}\frac{2}{3}C_V.
\label{eq28}\end{aligned}$$
Now we turn to the second part, which is connected with the quark self-energy. To calculate its contribution to the on-shell quark-gluon vertex, we should first pull out the factor $(q-p)_{\mu}$ and then put it on mass-shell. From Eq.(\[eq24\]), we have $$\begin{aligned}
&&-i\left[\Sigma^{(1)}(q)-\Sigma^{(1)}(p)\right]\nonumber\\
&& =-2g^2C_2(R)M\left\{\left(m+\frac{M}{2}\right)\int\frac{d^3k}{(2\pi)^3}
\frac{1}{k^2-M^2}\left[\frac{1}{(k+q)^2-m^2}
-\frac{1}{(k+p)^2-m^2}\right]\right.\nonumber\\
&-&\left(\frac{m}{M}+\frac{1}{2}\right)\int\frac{d^nk}{(2\pi)^n}
\frac{1}{(k^2-M^2)}\left[\frac{/\hspace{-2mm}k}{(k+q)^2-m^2}
-\frac{/\hspace{-2mm}k}{(k+p)^2-m^2}\right]\nonumber\\
&+&\frac{m}{M}\int\frac{d^nk}{(2\pi)^n}
\left[\frac{/\hspace{-2mm}k}{k^2\left[(k+q)^2-m^2\right]}
-\frac{/\hspace{-2mm}k}{k^2\left[(k+p)^2-m^2\right]}\right]\nonumber\\
&-&\frac{q^2-m^2}{2M}\int\frac{d^3k}{(2\pi)^3}
\frac{1}{(k^2-M^2)\left[(k+q)^2-m^2\right]}
+\frac{p^2-m^2}{2M}\int\frac{d^3k}{(2\pi)^3}
\frac{1}{(k^2-M^2)\left[(k+p)^2-m^2\right]} \nonumber\\
&+&\frac{q^2-m^2}{2M^2}\int\frac{d^3k}{(2\pi)^3}
\frac{/\hspace{-2mm}k}{(k^2-M^2)\left[(k+q)^2-m^2\right]}
-\frac{p^2-m^2}{2M^2}\int\frac{d^3k}{(2\pi)^3}
\frac{/\hspace{-2mm}k}{(k^2-M^2)\left[(k+p)^2-m^2\right]}
\nonumber\\
&+&\left.\frac{q^2-m^2}{2M}\int\frac{d^3k}{(2\pi)^3}
\frac{1}{k^2\left[(k+q)^2-m^2\right]}
-\frac{p^2-m^2}{2M}\int\frac{d^3k}{(2\pi)^3}
\frac{1}{k^2\left[(k+p)^2-m^2\right]}\right\} \nonumber\\
&=&2(q-p)^{\mu}g^2C_2(R)M\left\{\left(m+\frac{M}{2}\right)
\int\frac{d^3k}{(2\pi)^3}\frac{2k_{\mu}}{(k^2-M^2)\left[(k+q)^2-m^2\right]
\left[(k+p)^2-m^2\right]}\right.\nonumber\\
&-&\left(\frac{m}{M}+\frac{1}{2}\right)\int\frac{d^3k}{(2\pi)^3}
\frac{2/\hspace{-2mm}kk_{\mu}}{(k^2-M^2)\left[(k+q)^2-m^2\right]
\left[(k+p)^2-m^2\right]}\nonumber\\
&+&\frac{m}{M}\int\frac{d^3k}{(2\pi)^3}
\frac{2/\hspace{-2mm}kk_{\mu}}{k^2\left[(k+q)^2-m^2\right]
\left[(k+p)^2-m^2\right]} \nonumber\\
&-&\frac{q_{\mu}+p_{\mu}}{2M}\int\frac{d^3k}{(2\pi)^3}
\frac{M^2}{k^2(k^2-M^2)\left[(k+q)^2-m^2\right]}\nonumber\\
&-&\frac{q_{\mu}+p_{\mu}}{2M^2}\int\frac{d^3k}{(2\pi)^3}
\frac{/\hspace{-2mm}k}{(k^2-M^2)\left[(k+q)^2-m^2\right]}\nonumber\\
&-&\frac{p^2-m^2}{2M}
\int\frac{d^3k}{(2\pi)^3}\frac{2M^2k_{\mu}}{k^2 (k^2-M^2)
\left[(k+q)^2-m^2\right]
\left[(k+p)^2-m^2\right]}\nonumber\\
&+&\left.\frac{p^2-m^2}{2M^2}\int\frac{d^3k}{(2\pi)^3}
\frac{2/\hspace{-2mm}kk_{\mu}}{(k^2-M^2)\left[(k+q)^2-m^2\right]
\left[(k+p)^2-m^2\right]} \right\},
\label{eq29}\end{aligned}$$ where we have thrown away the vanishing terms in the large-$M$ limit.
As above, to compute the terms in Eq.(\[eq29\]), we cannot take the large-$M$ limit directly. So we still make use of above decomposition, and then we have $$\begin{aligned}
&&-i\left[\Sigma^{(1)}(q)-\Sigma^{(1)}(p)\right]
= 2(q-p)^{\mu}g^2C_2(R)\left\{-\int\frac{d^nk}{(2\pi)^n}
\frac{k_{\mu}}{(k^2+2k{\cdot}p)(k^2+2k{\cdot}q)}\right.\nonumber\\
&-&\left(m+\frac{M}{2}\right)
\int\frac{d^nk}{(2\pi)^n}\frac{/\hspace{-2mm}kk_{\mu}}{(k^2-M^2)(k^2-m^2)^2}
+m\int\frac{d^nk}{(2\pi)^n}
\frac{/\hspace{-2mm}kk_{\mu}}{k^2(k^2+2k{\cdot}p)(k^2+2k{\cdot}q)}
\nonumber\\
&-&\frac{q_{\mu}+p_{\mu}}{2M}\int \frac{d^nk}{(2\pi)^n}
\frac{/\hspace{-2mm}k}{(k^2-M^2)(k^2+2k{\cdot}q)}
+\left.\frac{p^2-m^2}{M}\int \frac{d^nk}{(2\pi)^n}
\frac{/\hspace{-2mm}kk_{\mu}}{(k^2-M^2)(k^2-m^2)^2}
\right\}\nonumber\\
&=&2(q-p)^{\mu}g^2C_2(R)\left\{\frac{i}{16\pi}(p_{\mu}+q_{\mu})
\frac{1}{r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)+
\frac{i}{4\pi}\frac{\gamma_{\mu}}{3}\right.\nonumber\\
&&+\left.
\frac{i}{8\pi}\left[\frac{m}{r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)
\gamma_{\mu}
-\frac{p_{\mu}+q_{\mu}}{m}\frac{2}{4-r^2/m^2}\right]\right\}\nonumber\\
&=&(q-p)^{\mu}\frac{i}{2\pi}g^2C_2(R)\left\{\frac{p_{\mu}+q_{\mu}}{m}\left[
\frac{m}{4r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)-
\frac{1}{4-r^2/m^2}\right]\right.\nonumber\\
&&\left.+\gamma_{\mu}\left[\frac{1}{3}
+\frac{m}{2r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)
\right] \right\},
\label{eq30} \end{aligned}$$ where $r_{\mu}=q_{\mu}-p_{\mu}$.
Now we consider the contribution from the one-loop on-shell composite ghost-fermion vertex. The corresponding Feynman diagrams are shown in Fig.5 and the amplitude we need reads as follows: $$\begin{aligned}
i\gamma^{(1)} &{\equiv}&
g^2\bar{u}(q)\left[\gamma^{(1)a}(p,q,r)(/\hspace{-2mm}q-m)-
(/\hspace{-2mm}p-m)\gamma^{(1)a}(p,q,r)\right]u(p)\nonumber\\
&=&(q-p)^{\mu}\frac{1}{4}g^2C_VT^a \bar{u}(q)\left[\int\frac{d^nk}{(2\pi)^n}
\frac{(/\hspace{-2mm}k+/\hspace{-2mm}p+m)(\gamma_{\mu}/\hspace{-2mm}k
-/\hspace{-2mm}k\gamma_{\mu})}{k^2\left[(k+p)^2-m^2\right]
\left[k-r\right]^2}(/\hspace{-2mm}q-m)\right.\nonumber\\
&& +\left.\int\frac{d^nk}{(2\pi)^n}(/\hspace{-2mm}p-m)
\frac{(/\hspace{-2mm}k\gamma_{\mu}-\gamma_{\mu}/\hspace{-2mm}k)(/\hspace{-2mm}k
+/\hspace{-2mm}q+m)}{(l+r)^2l^2\left[(l+p)^2-m^2\right]}\right]u(p)\nonumber\\
&=&(q-p)^{\mu}\frac{1}{2}g^2C_VT^a \bar{u}(q)\left\{\int\frac{d^nk}{(2\pi)^n}
\left[\frac{/\hspace{-2mm}kk_{\mu}
+(/\hspace{-2mm}p+m)(k_{\mu}-/\hspace{-2mm}k\gamma_{\mu})}
{k^2\left[(k+p)^2-m^2\right](l-r)^2}\right.\right.\nonumber\\
&& \left.-\frac{\gamma_{\mu}}{\left[(k+p)^2-m^2\right]
(l-r)^2}\right](/\hspace{-2mm}q-m)\nonumber\\
&&+\left.\int\frac{d^nk}{(2\pi)^n}(/\hspace{-2mm}p-m)
\left[\frac{/\hspace{-2mm}kk_{\mu}+(k_{\mu}-\gamma_{\mu}/\hspace{-2mm}k)
(\gamma_{\mu}/\hspace{-2mm}q+m)}{(l+r)^2l^2\left[(l+q)^2-m^2\right]}-
\frac{\gamma_{\mu}}{\left[(k+q)^2-m^2\right]
(l+r)^2}\right]\right\}u(p)\nonumber\\
&=&(q-p)^{\mu}\left[\gamma^{(1)}_{\mu}+\gamma^{(2)}_{\mu}
+\gamma^{(3)}_{\mu}+\gamma^{(4)}_{\mu}\right],
\label{eq31}\end{aligned}$$ where we have taken the large-$M$ limit and used the mass-shell condition. Correspondingly, we have $$\begin{aligned}
\gamma^{(1)}_{\mu}&=&-\frac{g^2C_VT^a}{2}\int\frac{d^nk}{(2\pi)^n}
\left[\frac{2(q_{\mu}-m\gamma_{\mu})}{(k^2+2k\cdot p)(k-r)^2}
+\frac{2(p_{\mu}-m\gamma_{\mu})}{(k^2+2k\cdot q)(k+r)^2}\right]\nonumber\\
&=&g^2C_VT^a\frac{i}{8\pi}\left[(m\gamma_{\mu}-q_{\mu})\frac{1}{q}
\ln\left(\frac{1+q/m}{1-q/m}\right)|_{q^2=m^2}\right.\nonumber\\
&&\left.+(m\gamma_{\mu}-p_{\mu})\frac{1}{p}
\ln\left(\frac{1+p/m}{1-p/m}\right)|_{p^2=m^2}\right]\nonumber\\
&=&g^2C_VT^a\frac{i}{8\pi}\left[\frac{p_{\mu}+q_{\mu}}{m}-2\gamma_{\mu}\right]
\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}\right);
\label{eq32}\end{aligned}$$ $$\begin{aligned}
\gamma^{(2)}_{\mu}&=&\frac{g^2C_VT^a}{2}\int\frac{d^nk}{(2\pi)^n}
\left[\frac{2q{\cdot}kk_{\mu}}{k^2(k^2+2p\cdot k)(k-r)^2}
+\frac{2p{\cdot}kk_{\mu}}{k^2(k^2+2q\cdot k)(k+r)^2}\right.\nonumber\\
&&+\left.\frac{k_{\mu}(2p\cdot q-2 m^2)}{k^2(k^2+2k\cdot p)(k-r)^2}+
\frac{k_{\mu}(2p\cdot q-2 m^2)}{k^2(k^2+2k\cdot q)(k+r)^2}\right]\nonumber\\
&=&\frac{g^2C_VT^a}{2}\int\frac{d^nk}{(2\pi)^n}\left[\frac{k_{\mu}}{k^2(k-r)^2}
-\frac{k_{\mu}}{k^2(k^2+2k\cdot p)}+\frac{r^2k_{\mu}}{k^2(k^2+2k\cdot p)(k-r)^2}\right.
\nonumber\\
&&+\frac{k_{\mu}}{k^2(k+r)^2}
-\frac{k_{\mu}}{k^2(k^2+2k\cdot q)}+\frac{r^2k_{\mu}}{k^2(k^2+2k\cdot q)(k+r)^2}\nonumber\\
&&+\left.\frac{k_{\mu}(2p\cdot q-2 m^2)}{k^2(k^2+2k\cdot p)(k-r)^2}+
\frac{k_{\mu}(2p\cdot q-2 m^2)}{k^2(k^2+2k\cdot q)(k+r)^2}\right]\nonumber\\
&=&-\frac{g^2C_VT^a}{2}\int\frac{d^nk}{(2\pi)^n}\left[\frac{k_{\mu}}{k^2(k^2+2k\cdot p)}
+\frac{k_{\mu}}{k^2(k^2+2k\cdot q)}\right]\nonumber\\
&=&\frac{i}{16\pi}C_VT^ag^2\frac{p_{\mu}+q_{\mu}}{m};
\label{eq33}\end{aligned}$$ $$\begin{aligned}
\gamma^{(3)}_{\mu}&=&g^2C_VT^a m\int\frac{d^nk}{(2\pi)^n}\left[
\frac{/\hspace{-2mm}kk_{\mu}}{k^2(k^2+2p\cdot k)(k-r)^2}
+\frac{/\hspace{-2mm}kk_{\mu}}{k^2(k^2+2q\cdot k)(k+r)^2}\right]|_{p^2=q^2=m^2}
\nonumber\\
&=&g^2C_VT^a\frac{i}{32\pi}\left[m\gamma_{\mu}-\frac{p_{\mu}+q_{\mu}}{2}\right]
\left[\frac{1}{q}
\ln\left(\frac{1+q/m}{1-q/m}\right)|_{q^2=m^2}\right.\nonumber\\
&&\left.+
\frac{1}{p}\ln\left(\frac{1+p/m}{1-p/m}\right)|_{p^2=m^2}\right]
+\frac{i}{16\pi}\left(\gamma_{\mu}-\frac{3(p_{\mu}+q_{\mu})}{2}\right)
\nonumber\\
&=&-g^2C_VT^a\frac{i}{16\pi}\left[\frac{p_{\mu}+q_{\mu}}{2m}-\gamma_{\mu}\right]
\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}\right)
+\frac{i}{16\pi}\left[\gamma_{\mu}-\frac{3(p_{\mu}+q_{\mu})}{2}\right],
\label{eq33i}\end{aligned}$$ where $\epsilon_{\rm IR}=3-n$ and $\mu$ is the artificial parameter with mass dimension. One notices that there is an IR pole term, which is induced purely by the mass-shell condition. During the above procedure, we have used the identities $$\begin{aligned}
2 p\cdot k&=&(k^2+2 p\cdot k)-k^2,\nonumber\\
2k\cdot r &=& (k+r)^2-k^2-r^2=k^2+r^2-(k-r)^2,
\label{eq34}\end{aligned}$$ and the mass-shell condition $$\begin{aligned}
2p\cdot q-2m^2=p^2+q^2-(q-p)^2-2m^2=-r^2.
\label{eq35}\end{aligned}$$
As for $\gamma^{(4)}_{\mu}$, it is a little more complicated, namely $$\begin{aligned}
\gamma^{(4)}_{\mu}&=&-\frac{1}{2}g^2C_VT^a\int\frac{d^nk}{(2\pi)^n}
\left[ \frac{4p\cdot k q_{\mu}-4q\cdot k p_{\mu}+2(p\cdot q-m^2)/\hspace{-2mm}k
\gamma_{\mu} }{k^2(k^2+2k\cdot p)(l-r)^2}\right.\nonumber\\
&&+\frac{4mp_{\mu}/\hspace{-2mm}k-4mk\cdot p\gamma_{\mu}}
{k^2(k^2+2k\cdot p)(l-r)^2}
+\frac{4q\cdot kp_{\mu}-4k\cdot p q_{\mu}+2(p\cdot q-m^2)\gamma_{\mu}
/\hspace{-2mm}k}{k^2 (k^2+2k\cdot q)(l+r)^2}\nonumber\\
&&+\left.\frac{4mq_{\mu}/\hspace{-2mm}k-4mk\cdot q\gamma_{\mu}}
{k^2 (k^2+2k\cdot q)(l+r)^2}\right].
\label{eq36}\end{aligned}$$ However, from Eq.(\[eq36\]) one can see that we only need two Feynman integrals (on-shell), $$\begin{aligned}
\int\frac{d^nk}{(2\pi)^n}\frac{k_{\mu}}{k^2(k^2+2k\cdot p)(k-r)^2}|_{p^2=m^2}&=&
Ap_{\mu}+Br_{\mu};
\nonumber\\
\int\frac{d^nk}{(2\pi)^n}\frac{k_{\mu}}{k^2(k^2+2k\cdot q)(k+r)^2}|_{q^2=m^2}
&=&Aq_{\mu}-Br_{\mu},
\label{eq37}\end{aligned}$$ where $A$ and $B$ are the form factors needed to be determined. It is easy to obtain that $$\begin{aligned}
A&=&\frac{1}{2(p^2-q^2)}\left\{\int\frac{d^nk}{(2\pi)^n}\left[
\frac{1}{\left[(k+p)^2-m^2\right](k+r)^2}
-\frac{1}{\left[(k+q)^2-m^2\right](k+r)^2}\right]|_{p^2=q^2=m^2}\right.
\nonumber\\
&&+\left.\int\frac{d^nk}{(2\pi)^n}\left[\frac{q^2-m^2}
{k^2\left[(k+q)^2-m^2\right](k+r)^2}
- \frac{p^2-m^2}
{k^2\left[(k+p)^2-m^2\right](k-r)^2}\right]|_{p^2=q^2=m^2}\right\}
\nonumber\\
&=&\frac{i}{16\pi}\frac{1}{m(p^2-q^2)}\left[\frac{1}{q}
\ln\left(\frac{1+q/m}{1-q/m}\right)
-\frac{1}{p}\ln\left(\frac{1+p/m}{1-p/m}\right)\right]|_{p^2=q^2=m^2}
-\frac{I}{2}\nonumber\\
&=&\frac{i}{16\pi}\frac{1}{(p^2-q^2)m^{1+\epsilon}}\int^1_0 dx \left[
\frac{1}{\left[x-q^2/m^2 x (1-x)\right]^{(1+\epsilon)/2} }|_{p^2=m^2}
\right. \nonumber\\
&&\left.-\frac{1}{\left[x-p^2/m^2 x (1-x)\right]^{(1+\epsilon)/2}}
\right]|_{q^2=m^2}-\frac{I}{2}
\nonumber\\
&=&\frac{i}{16\pi} \frac{1+\epsilon}{2}\frac{1}{m^{3+\epsilon}}
\int^1_0 dx x (1-x)\nonumber\\
&&\times \int_0^1 dy\frac{1}{\left\{\left[x-q^2/m^2 x (1-x)\right]
+(q^2-p^2)/m^2 x (1-x) y\right\}^{(3+\epsilon)/2}}|_{p^2=q^2=m^2}-\frac{I}{2}\nonumber\\
&=&\frac{i}{16\pi} \frac{1+\epsilon}{2}\frac{1}{m^{3+\epsilon}}
\int^1_0 dx (1-x) x^{-2-\epsilon}-\frac{I}{2}
=\frac{i}{32\pi}\frac{1}{m^3}\left(\frac{1}{\epsilon_{IR}}
+\ln\frac{\mu}{m}\right)-\frac{I}{2},
\label{eq38}\end{aligned}$$ where we have used the relation $$\begin{aligned}
\frac{1}{a^{(1+\epsilon)/2}}-\frac{1}{b^{(1+\epsilon)/2}}
=\frac{1+\epsilon}{2}\int_0^1 dy \frac{b-a}{\left[a+(b-a)y
\right]^{(3+\epsilon)/2}},
\label{eq39}\end{aligned}$$ and $$\begin{aligned}
I&{\equiv}&\int\frac{d^nk}{(2\pi)^n} \frac{1}
{k^2\left[(k+q)^2-m^2\right](k+r)^2}|_{q^2=m^2}\nonumber\\
&=&\int\frac{d^nk}{(2\pi)^n} \frac{1}
{k^2\left[(k+p)^2-m^2\right](k-r)^2}|_{p^2=m^2}\nonumber\\
&=& \frac{i}{4\pi^2}\Gamma\left(\frac{3-\epsilon}{2}\right)
\Gamma\left(\frac{3+\epsilon}{2}\right)\frac{1}{m^3}
\left(\frac{\mu}{m}\right)^{\epsilon}\nonumber\\
&\times& \int_0^1dx\int_0^1dy
\frac{y}{\left[(1-x)^2y^2-k^2/m^2 x y (1-y)\right]^{(3+\epsilon)/2}}.\end{aligned}$$ Thus we obtain $$\begin{aligned}
\gamma^{(4)}_{\mu}&=&-\frac{1}{2}g^2C_VT^aA\left[(4m^2-2 r^2)(p_{\mu}+q_{\mu})+
(2mr^2-8m^3)\gamma_{\mu}\right]\nonumber\\
&=&-\frac{i}{32\pi}g^2C_VT^a\left(\frac{1}{\epsilon_{\rm IR}}+
\ln\frac{\mu}{m}-16m^3I\right)
\left[\left(2- \frac{r^2}{m^2}\right)\frac{p_{\mu}+q_{\mu}}{m}+
\left(\frac{r^2}{m^2}-4\right)\gamma_{\mu}\right].
\label{eq40}\end{aligned}$$ and hence $$\begin{aligned}
i\gamma^{(1)}&=&(q-p)^{\mu}g^2C_VT^a\frac{i}{8\pi}
\left\{\frac{p_{\mu}+q_{\mu}}{m}\left[
\frac{1}{4}\left(\frac{1}{\epsilon_{\rm IR}}
\ln\frac{\mu}{m}\right)\left(1+\frac{r^2}{m^2}\right)
-\frac{1}{4}+4m^3I\left(2-\frac{r^2}{m^2}\right)\right]\right.\nonumber\\
&&\left. -\gamma_{\mu}\left[\frac{1}{2}\left(\frac{1}{\epsilon_{\rm IR}}
+\ln\frac{\mu}{m}\right)\left(1+\frac{r^2}{2m^2}\right)-\frac{1}{2}
+4m^3I\left(\frac{r^2}{m^2}-4\right)\right]\right\}.
\label{eq42}\end{aligned}$$
From Eqs.(\[eqa13\]), (\[eq28\]), (\[eq30\]) and (\[eq42\]), we finally obtain the on-shell one-loop fermion-gluon vertex, $$\begin{aligned}
\Gamma_{\mu}^{(1)a}(r)&=&-\frac{g^2}{4\pi}T^a\left\{\gamma_{\mu}
\left[-C_V\left(\frac{5}{12}
+\frac{1}{4}\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}\right)
\left(1+\frac{r^2}{2m^2}\right)\right.\right.\right.\nonumber\\
&& \left.\left. -2m^3I\left(\frac{r^2}{m^2}-4\right)\right)
+C_2(R)\left(\frac{2}{3}+\frac{m}{r}
\ln\frac{1+r/(2m)}{1-r/(2m)}\right)\right]
\nonumber\\
&&+\frac{p_{\mu}+q_{\mu}}{m}\left[C_2(R)
\left(\frac{m}{2r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)
-\frac{2}{4-r^2/m^2}
-\frac{1}{2}\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}\right)\right)
\right. \nonumber\\
&&\left.\left.+C_V\left(
\frac{1}{8}\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}\right)
\left(1+\frac{r^2}{m^2}\right)
-\frac{1}{8}+2m^3I\left(2-\frac{r^2}{m^2}\right)\right)\right]\right\}
\nonumber\\
&=&-\frac{g^2}{4\pi}T^a\gamma_{\mu}\left\{C_V\left[-\frac{2}{3}+
\frac{1}{8}\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}-16m^3I\right)
\frac{r^2}{m^2}\right]\right.
\nonumber\\
&&\left.+C_2(R)\left[\frac{2}{3}
+\frac{3}{2}\frac{m}{r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)
-\frac{2}{4-r^2/m^2} \right]\right\}\nonumber\\
&&+\frac{g^2}{4\pi}T^a\frac{i\epsilon_{\mu\nu\rho}r^{\nu}\gamma^{\rho}}{m}
\left\{C_2(R)
\left[\frac{m}{2r}\ln\left(\frac{1+r/(2m)}{1-r/(2m)}\right)-\frac{2}{4-r^2/m^2}
\right]\right.\nonumber\\
&&\left.
+C_V\left[\frac{1}{8}\left(\frac{1}{\epsilon_{\rm IR}}+\ln\frac{\mu}{m}\right)
\left(1+\frac{r^2}{m^2}\right)
-\frac{1}{8}+2m^3I\left(2-\frac{r^2}{m^2}\right)\right]\right\}\nonumber\\
&=&\gamma_{\mu}T^af_1(r)+iT^a\epsilon_{\mu\nu\rho}r^{\nu}\gamma^{\rho}f_2(r),
\label{eq43}\end{aligned}$$ where we have used the three-dimensional analogue of the Gordon identity: $$\begin{aligned}
{\gamma}_{\mu}=\frac{1}{2m}\left[(p_{\mu}+q_{\mu})
+i{\epsilon}_{\mu\nu\lambda}r^{\nu}
{\gamma}^{\lambda}\right].
\label{eq44}\end{aligned}$$ We define the vertex at the renormalization point $r=0$ as that done in $\cite{ref20}$, $$\begin{aligned}
&&\Gamma_{\mu}^{a(R)}(0)=0;\nonumber\\
&& \Gamma^{a(R)}_{\mu}(r)=T^a{\gamma}_{\mu}(Z_3^{-1}-1)+Z_3^{-1}
{\Gamma}^{a(R)}_{\mu}(r),
\label{eq45}\end{aligned}$$ then the quark-gluon vertex renormalization constant is $$\begin{aligned}
Z_{3}^{-1}&=&1+f_1(0),
\nonumber\\[2mm]
Z_3&=&1+\frac{g^2}{4\pi}\left[-\frac{2}{3}C_V+C_2(R)\frac{5}{3}\right].
\label{eq46}\end{aligned}$$ One can notice that actually $f_2(0)$ does not vanish. This will induce a non-minimal (colour) magnetic moment interaction between two three-dimensional quarks, which may play a certain role in the application of this model to condensed matter physics. The similar result had also been obtained in Abelian case$\cite{ref21}$.
One-loop Ghost-Gluon Vertex
----------------------------
For discussing the renormalization of coupling constant, we shall have a brief look at the one-loop ghost-gluon vertex (Fig.6), whose value had been predicted in Ref.$\cite{ref18}$ from the general result in Landau gauge$\cite{ref2}$ and was explicitly calculated in Ref.$\cite{ref18}$. It is obvious that after taking the large-$M$ limit, the amplitude indeed vanishes. Therefore we can always define the gluon-ghost vertex renormalization constant as, $$\begin{aligned}
Z_2=1.
\label{eq47}\end{aligned}$$
Finite Renormalization
======================
Now let us consider the renormalization of the coupling constant. There are two ways to implement this: one way is using the Slavnov-Taylor identities and the known one-loop results to determine the local part of the one-loop quantum effective action$\cite{ref4,ref6}$; another way, which we shall adopt in the following, is to use the relations among various coupling constants imposed by the Slavnov-Taylor identities to determine the finite renormalization of the coupling constant$\cite{ref3,ref19}$. Since the on-shell renormalization is compatible with the Slavnov-Taylor identity and the renormalized coupling constant is unique, we can write the local effective action in terms of the renormalized fields in the following two forms, $$\begin{aligned}
S&=&\int d^3x\left[\frac{1}{2}Z_A\epsilon^{\mu\nu\rho}A_{\mu}^a
\partial_{\nu}A_{\rho}^a+Z_c\bar{c}^a\partial^{\mu}\partial_{\mu}c^a
+Z_{\psi}\bar{\psi}(i/\hspace{-2mm}\partial-Z_mm_{\rm ph})\psi\right.\nonumber\\
&+&\left.\frac{1}{3!}Z_1gf^{abc}\epsilon^{\mu\nu\rho}
A_{\mu}^aA_{\nu}^bA_{\rho}^c-Z_2gf^{abc}\partial^{\mu}\bar{c}^aA_{\mu}^bc^c
+Z_3g\bar{\psi}/\hspace{-2.5mm}A\psi\right]\nonumber\\
&=&\int d^3x\left[\frac{1}{2}Z_A\epsilon^{\mu\nu\rho}A_{\mu}^a
\partial_{\nu}A_{\rho}^a+Z_c\bar{c}^a\partial^{\mu}\partial_{\mu}c^a
+Z_{\psi}\bar{\psi}(i/\hspace{-2mm}\partial-Z_mm_{\rm ph})\psi\right.\nonumber\\
&+&\left.\frac{1}{3!}g_BZ_A^{3/2}\epsilon^{\mu\nu\rho}
A_{\mu}^aA_{\nu}^bA_{\rho}^c-g_BZ_cZ_A^{1/2}
f^{abc}\partial^{\mu}\bar{c}^aA_{\mu}^bc^c
+g_BZ_\psi Z_A^{1/2}\bar{\psi}/\hspace{-2.5mm}A\psi\right],
\label{eq48}\end{aligned}$$ where the correspondence between the renormalized quantities and the bare ones is defined as usual, $$\begin{aligned}
&&A^a_{B\mu}=Z_A^{1/2} A_{\mu}^a,~~ c^a_B=Z_c^{1/2} c^a,~~
\bar{c}^a_B=Z_c^{1/2} \bar{c}^a,~~
\psi_B=Z_\psi^{1/2}\psi,\nonumber\\
&&\bar{\psi}_B=Z_\psi^{1/2}\bar{\psi}, ~~m=m_{\rm ph}+\delta m=Z_mm_{\rm ph}.
\label{eq49}\end{aligned}$$ Eq.(\[eq48\]) gives $$\begin{aligned}
g=g_BZ_1^{-1}Z_A^{3/2}&=&g_BZ_2^{-1}Z_cZ_A^{1/2}=g_BZ_3^{-1}Z_\psi Z_A^{1/2};
\nonumber\\
\frac{Z_A}{Z_1}&=&\frac{Z_c}{Z_2}=\frac{Z_\psi}{Z_3}.
\label{eq50}\end{aligned}$$ From Eqs.(\[eq26\]), (\[eq27n\]), (\[eq46\]) and (\[eq47\]), we can see that the relation $Z_c/Z_2=Z_{\psi}/Z_3$ is indeed satisfied. Using Eq.(\[eq50\]), we obtain $$\begin{aligned}
Z_1=1-\frac{g^2}{4\pi}\left(3C_V+\frac{1}{2}\right).
\label{eq51}\end{aligned}$$
Summary and Discussion
======================
We have found the inconsistency of dimensional reduction and naive dimensional regularization when they are applied to Chern-Simons type theories. Further we use the consistent dimensional continuation to re-investigate the one-loop quantum correction of Chern-Simons term coupled with spinor fields. As it is pointed out in Ref.$\cite{ref4}$, the practice of consistent dimensional regularization requires the introduction of the higher covariant derivative term like Yang-Mills term, since this special prescription of dimensional continuation results in the linear dependence of the $n$-dimensional kinetic operator, even though the gauge fixing has been performed. Therefore the regularization we adopt in essence consists of higher covariant derivative regularization combining with consistent dimensional continuation.
With this regularization prescription, we have calculated all the one-loop two-point amplitudes and have given the analytical result of one-loop on-shell quark-gluon vertex with aid of the Slavnov-Taylor identity. In the mass-shell renormalization convention, we have found that not only the coupling constant receives an extra finite renormalization from the fermionic loop, but the fermionic matter also has a finite renormalization. This is different from the result given in Ref.$\cite{ref3}$, where it was shown that all the renormalization constants are defined as $Z_i=1$.
Of course, purely from the viewpoint of renormalization, our results do not contradict the ones of Ref.$\cite{ref3}$ since a difference in a finite renormalization can always be explained as a different choice of renormalization convention. However, since Chern-Simons type theory is finite at one-loop level, the $\beta$-function and the anomalous dimensions of all the fields vanish identically, and we have no objects like renormalization group equation to show the renormalization convention independence. As pointed out in Ref.$\cite{ref23}$, the only criterion for the equivalence among different renormalization conditions is that all the regularization schemes preserving the fundamental symmetry such as gauge invariance should give the same gauge invariant radiative corrections. Therefore, we interpret these differences as the inconsistency of naive dimensional regularization.
The financial support of the Academy of Finland under the Project No. 37599 is greatly acknowledged. We would like to thank V.Ya. Fainberg for useful discussions.
Slavnov-Taylor Identity from BRST Symmetry
==========================================
The BRST transformation reads as follows: $$\begin{aligned}
\delta A_{\mu}^a&=&-D_{\mu}^{ab}c^b,~\delta c^a=\frac{g}{2}f^{abc}c^bc^c,~
\delta\bar{c}^a=-\frac{1}{\alpha}\partial^{\mu}A_{\mu}^a,\nonumber\\
\delta \psi&=&-igT^ac^a\psi,~\delta\bar{\psi}=ig\bar{\psi}T^ac^a,
\label{eqa1}\end{aligned}$$ which is nilpotent, $$\begin{aligned}
\delta^2=0.
\label{eqa2}\end{aligned}$$ The generating functional with all the external source terms is $$\begin{aligned}
Z&{\equiv}&Z[J_{\mu}^a,\eta,\bar{\eta},\bar{K},K,\bar{L},L,u^a_{\mu},v^a]
\nonumber\\
&=&\int {\cal D}X\exp\left\{i\left[S+\int d^3x \left(J^{\mu a}A_{\mu}^a+
\bar{\psi}\eta+\bar{\eta}\psi + \bar{K}^ac^a+\bar{c}^aK^a\right)\right.\right.
\nonumber\\
&&\left.\left. +\int d^3x \left(-\bar{L}gT^ac^a\psi+\bar{\psi}gT^ac^aL
-u^a_{\mu}D^{\mu ab}c^b+v^a\frac{g}{2}f^{abc}c^bc^c\right)\right]\right\}.
\label{eqa3}\end{aligned}$$ The BRST invariance of the generating functional lead to the following general Ward identity: $$\begin{aligned}
&&\delta Z=\int {\cal D}X \left\{\int d^3u \left[J_{\mu}^a \delta A_{\mu}^a
+ \delta \bar{\psi}\eta-\bar{\eta}
\delta \psi+\delta\bar{c}^aK^a-\bar{K}^a\delta c^a\right]\right.\nonumber\\
&& \times \left.\exp\left[iS+i\int (\mbox{the source term})\right]\right\} =0,
\nonumber\\
&& \int d^3 u\left[J_{\mu}^a\frac{\delta}{\delta u^a_{\mu}}-i\bar{\eta}
\frac{\delta}{\delta \bar{L}}-i\frac{\delta}{\delta L}\eta-\bar{K}^a
\frac{\delta}{\delta v^a}-\frac{1}{\alpha}\left(\partial_{\mu}
\frac{\delta}{\delta J^a_{\mu}}\right)K^a\right]Z=0.
\label{eqa4}\end{aligned}$$ It can be directly written out the Ward identities for the generating functional of the connected Green functions due to the linearity of the functional differential operator in Eq.(\[eqa4\]) $$\begin{aligned}
\int d^3 u\left[J_{\mu}^a\frac{\delta}{\delta u^a_{\mu}}-i\bar{\eta}
\frac{\delta}{\delta \bar{L}}-i\frac{\delta}{\delta L}\eta-\bar{K}^a
\frac{\delta}{\delta v^a}-\frac{1}{\alpha}\left(\partial_{\mu}
\frac{\delta}{\delta J^a_{\mu}}\right)K^a\right]W=0,
\label{eqa5}\end{aligned}$$ where $Z=\exp[iW]$. Acting $\delta/\delta \bar{\eta}(x)$, $\delta/\delta {\eta}(y)$ and $\delta/\delta K^a(z)$ on above identity and then the external sources to zero, we obtain the Ward identity containing the quark-gluon vertex, $$\begin{aligned}
&&\left[\frac{1}{\alpha}\frac{\delta}{\delta \bar{\eta}(x)}
\frac{\delta}{\delta {\eta}(y)}\partial_{\mu}\frac{\delta}{\delta J^a_{\mu}(z)}
+i\frac{\delta}{\delta {\eta}(y)}\frac{\delta}{\delta K^a(z)}
\frac{\delta}{\delta \bar{L}(x)}\right.\nonumber\\
&& \left.+i\frac{\delta}{\delta \bar{\eta}(x)}\frac{\delta}{\delta K^a(z)}
\frac{\delta}{\delta L(y)}\right]W|_{\mbox{\rm all the external sources $=0$}}
=0,
\nonumber\\
&&\frac{1}{\alpha}\frac{\partial}{\partial z_{\mu}}\langle\psi(x)\bar{\psi}(y)
A_{\mu}^a(z)\rangle_C
+igT^b\langle\bar{\psi}(y)\bar{c}^a(z)c^b(x)\psi (x)\rangle_C\nonumber\\
&&-igT^b\langle\psi(x)\bar{c}^a(z)\bar{\psi}(y)c^b(y)\rangle_C =0,
\label{eqa6}\end{aligned}$$ where the subscript “ C ” means the connected part of the Green functions. Decomposing the above Green functions into 1PI part, we get $$\begin{aligned}
&&\frac{1}{\alpha}\frac{\partial}{\partial z_{\mu}}
\int d^3u d^3v d^3 w \,iD^{aa'}_{\mu\nu}(z-w) iS(x-u) g \Gamma^{a'}(u,v,w)
iS(v-y)\nonumber\\
&&+ig\int d^3u d^3v\left[\gamma^{a'}(x,u,v)iS(u-y)
-iS(x-u)\gamma^{a'}(u,y,v)\right] iD^{a'a}(v-z)=0,
\label{eqa7}\end{aligned}$$ where $\Gamma^{a}(u,v,w)$ is the 1PI part of the fermion-gluon vertex function, $\gamma^{a}(x,u,v)$ and $\gamma^{a}(u,y,v)$ are the composite ghost-gluon vertex functions. After Fourier transformation, we obtain $$\begin{aligned}
&&\frac{1}{\alpha}r^{\mu}D^{aa'}_{\mu\nu}(r)S(p)\Gamma^{\nu a'}(p,q,r)S(q)
+\gamma^{a'}(p,q,r)D^{aa'}(r)S(q)\nonumber\\
&&-S(p)\gamma^{a'}(p,q,r) D^{a'a}(r)=0,
\label{eqa8}\end{aligned}$$ where $r_{\mu}=q_{\mu}-p_{\mu}$. Considering the fact that the longitudinal part of gauge field receives no quantum correction, i.e. $$\begin{aligned}
r^{\mu}D_{\mu\nu}^{ab}(r)
=r^{\mu}D_{\mu\nu}^{ab(0)}(r)=-\alpha \frac{r_{\nu}}{r^2}
\delta^{ab},
\label{eqa9} \end{aligned}$$ and using the general form of the full ghost propagator $$\begin{aligned}
D^{ab}(r)=-\frac{i\delta^{ab}}{r^2\left[1+\Sigma_g(r^2)\right]},
\label{eqa10}\end{aligned}$$ we get the required Ward identity $$\begin{aligned}
r^{\mu}\Gamma^a_{\mu}(p,q,r)\left[1+\Sigma_g(r^2)\right]
=\gamma^{a}(p,q,r)S^{-1}(q)-S^{-1}(p)\gamma^{a}(p,q,r).
\label{eqa11}\end{aligned}$$ Expanding the above identity up to one-loop order and using $$\begin{aligned}
&& \Gamma^a_{\mu}(p,q,r)=\gamma_{\mu}T^a+g^2\Gamma_{\mu}^{(1)a}+{\cal O}(g^4),~~
\Sigma_g(r^2)=g^2\Sigma_g^{(1)}(r^2)+{\cal O}(g^4),~~\nonumber\\
&& S^{-1}(p)=/\hspace{-2mm}p-m-g^2 \Sigma^{(1)}(p)+{\cal O}(g^4) ,~~
\gamma^a(p,q,r)=T^a+g^2\gamma^{(1)a}(p,q,r)+{\cal O}(g^4),
\label{eqa12}\end{aligned}$$ we obtain the desired one-loop Slavnov-Taylor identity: $$\begin{aligned}
(q^{\mu}-p^{\mu})\Gamma_{\mu}^{(1)a}(p,q,r)&=&
-(/\hspace{-2mm}q-/\hspace{-2mm}p)T^a
\Sigma_g^{(1)}(r^2)-T^a\left[\Sigma^{(1)}(q)-\Sigma^{(1)}(p)\right]
\nonumber\\
&+&g^2\left[\gamma^{(1)a}(p,q,r)(/\hspace{-2mm}q-m)-
(/\hspace{-2mm}p-m)\gamma^{(1)a}(p,q,r)\right].
\label{eqa13}\end{aligned}$$
E. Witten, Comm. Math. Phys. [**121**]{} (1989) 351; D. Birmingham, M. Rakowsky and G. Thompson, Phys. Lett. [**B251**]{}(1990) 121.
E. Guadagnini, M. Martellini and M. Mintchev, Phys. Lett. [**B227**]{} (1989) 111; L. Alvarez-Gaumé, J.M.F. Labastida and A.V. Ramallo, Nucl.Phys. [**B334**]{} (1990) 103.
W. Chen, G.W. Semenoff and Y.S. Wu, Mod. Phy. Lett. [**A5**]{} (1990) 1833; Phys. Rev. [**D46**]{} (1992) 5521.
C.P. Martin, Phys. Lett. [**B241**]{} (1990) 513; G. Giavarini, C.P. Martin and F. Ruiz Ruiz. Nucl. Phys. [**B381**]{} (1992), 222.
M. Asorey and F. Falceto, Phys. Lett. [**B241**]{} (1990) 31.
W.F. Chen and Z.Y. Zhu, J. Phys. [**A27**]{} (1994), 1781; W.F. Chen, H.C. Lee and Z.Y. Zhu, Phys. Rev. [**D55**]{} (1997) 3664.
E.R. Speer, J. Math. Phy. [**15**]{} (1974) 1.
C.P. Martin and F. Ruiz Ruiz, Nucl. Phys. [**B436**]{} (1995) 545; J.H. Leon, C.P. Martin and F. Ruiz Ruiz, Phys. Lett. [**B355**]{} (1995) 531.
M. Asorey and F. Falceto, Phys. Rev. [**D54**]{} (1996) 5290; T.D. Bakeev and A.A. Slavnov, Mod. Phys. Lett. [**A11**]{} (1996) 1539.
W. Chen, Phys. Lett. [**B251**]{} (1990) 415. W. Chen, G.W. Semenoff and Y.S. Wu, Phys. Rev. D44 (1991) R1625; Phys. Rev. [**D46**]{} (1992) 5521.
W. Siegel, Phys. Lett. [**B84**]{} (1979) 193; Phys. Lett. [**B94**]{} (1980) 37.
P. West, [*Introduction to Supersymmetry and Supergravity*]{}, World Scientific Publishing (1986).
G. ’t. Hooft and M. Veltman, Nucl. Phys. [**B44**]{} (1972) 189.
S. Deser, R. Jackiw and S. Templeton, Ann. Phys. (N.Y) [**140**]{} (1982) 372; I. Affleck, J. Harvey and E. Witten, Nucl. Phys. [**B206**]{} (1982) 413.
G.W. Semenoff and P. Sodano, Phys. Rev. Lett. [**57**]{} (1986) 1195; A.M. Polyakov, Mod. Phys. Lett. [**A3**]{}(1988) 325; J. Dunne, R. Jackiw and C. Trugenberger, Ann. Phys. (N.Y.) [**194**]{} (1989) 197. G.W. Semenoff and N. Weiss, Phys. Lett. [**B250**]{} (1990) 117.
Breitenlohner and Maison, Comm. Math. Phys. [**52**]{} (1977) 11.
R. Pisarski and S.Rao, Phys. Rev. [**D32**]{} (1985) 2081.
C. Itzykson and J.B. Zuber, [*Quantum Field Theory*]{}, McGraw-Hill Inc. (1980).
I.I. Kogan and G.W. Semenoff, Nucl. Phys. [**B368**]{} (1992) 718; G. Gat and R. Ray, Phys. Lett. [**B340**]{} (1994) 162; M. Chaichian, W.F. Chen and V. Ya. Fainberg, [hep-th/9706068]{}, Eur. Phys. J. [**C2**]{} (1998).
J.C. Taylor, Nucl. Phys. [**B33**]{} (1971) 436.
A. Blasi and R. Collina, Nucl. Phys. [**B345**]{} (1990) 472; F. Delduc, O. Piguet, C. Lucchesi and S.P. Sorella, Nucl. Phys. [**B346**]{} (1990) 313; A. Blasi, N. Maggiore and S.P. Sorella, Phys. Lett. [**B285**]{} (1992) 54.
G. Giavarini, C.P. Martin and F. Ruiz Ruiz, [*BRS Invariance, Supersymmetry and the Shift of the Bare Parameter in Chern-Simons Theory*]{}, talk given at 19th International Colloquium on Group Theoretical Methods in Physics, Salamanca, Spain (1992) (unpublished).
FEYNMAN
(25000,8000) (14000,5000) (0,0)\[2\]=12600 =5000 by -1 by -1 by by (,)\[2\](,)\[2\]=15400 =5000 by by (,)\[2\](,)\[2\]
(8000,8000) (0,0)\[2000\] (,) (,)\[2000\] (,)\[7000\]
(10000,5000) (0,0)\[2\] (,) (,)\[2\] (,)\[4\]
(40000,20000) (9000,11000)\[4\] (,)\[6400\] (,)\[3000\] (,) (2000,3000)[$p$]{} (,)\[6400\] (,)\[3000\] (,) (15000,3000)[$q$]{} (,)\[8\] (8000,200)[$(a)$]{}
(31000,11500)\[4\] (,)\[3000\] (,) (24000,3000)[$p$]{} (,)\[9600\] (,)\[3000\] (,) (37000,3000)[$q$]{} (30000,200)[$(b)$]{}
(40000,20000) (9000,10000)\[6000\] (8600,10100)[$\bigotimes$]{} (,)\[3\] (,)\[3000\] (,) (2000,5000)[$p$]{} (,)\[8\] (,)\[2\] (,) (15000,5000)[$r=q-p$]{} (8000,2000)[$\gamma^a(p,q,r)$]{}
(32000,10000)\[3\] (31600,10100)[$\bigotimes$]{} (,)\[6000\] (,)\[2\] (,) (23000,5000)[$r=q-p$]{} (,)\[8\] (,)\[3000\] (,) (38000,5000)[$q$]{} (30000,2000)[$\gamma^a(p,q,r)$]{}
(40000,15000) (9000,11000)\[3\] (,)\[3\] (,)\[2\] (,)\[3\] (,)\[2\] (,)\[8\] (8000,200)[$(a)$]{}
(31000,11200)\[3\] (,)\[2\] (,)\[3\] (,)\[2\] (30000,200)[$(b)$]{}
[^1]: For pure Chern-Simons theory, this renormalization scheme is equivalent to taking the large mass limit $M{\rightarrow}{\infty}$. However here, owing to the mass parameter $m$, they are not equivalent.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Á. Kóspál'
- 'P. Ábrahám'
- 'Ch. Westhues'
- 'M. Haas'
date: 'Received date; accepted date'
title: 'Brightness variations of the FUor-type eruptive star V346 Nor[^1]'
---
Introduction
============
FU Orionis-type stars (FUors) are low-mass pre-main sequence objects characterized by 4-6magnitude optical outbursts due to temporarily enhanced accretion from the circumstellar disk to the star [@hk96]. Following the outburst of the first such object, FU Ori in 1937, now more than two dozen FUors and FUor candidates are known [@audard2014]. V346 Nor was discovered by @elias1980 as a source within a few arcseconds of the HH 57 nebulosity, the latter being a faint, compact H$\alpha$ emitting knot [@schwartz1977], located in the Sa 187 molecular cloud within the Norma 1 association, at a distance of 700pc (@reipurth1981, see also Fig. \[fig:composit\]). A few years later @graham1983 reported the appearance of a star-like source at the northeastern tip of HH 57, probably coinciding with the source in @elias1980. They mentioned that the star was not visible in 1976 [@schwartz1977], but a diffuse patch is clearly discernible in the blue plates of the ESO/SERC Sky Survey, obtained in April–June 1975 [@holmberg1974; @reipurth1981]. Therefore, V346 Nor transformed from a faint diffuse nebula to a bright point-source some time between 1976 and 1980. Based on this and on the spectroscopic properties of the star, @reipurth1983a suggested that V346 Nor was undergoing a FUor-type outburst. @frogel1983 presented photometry from 2.2 to 20$\,\mu$m and remarked that the colors of the object are similar to those of FU Ori and V1057 Cyg.
![V346 Nor (black [*plus*]{} sign in the center) and its surroundings in a $JHK_{\rm S}$ color composit image. The observations were taken within the VISTA Variables in The Via Lactea (VVV) Survey on March 15, 2010. North is up and east is to the left. The displayed area is 2$\farcm$5${\times}$2$\farcm$5 in size. The dark lane across the image is due to extinction by the Sa 187 molecular cloud.[]{data-label="fig:composit"}](composit.ps){width="7.5cm"}
Subsequent near-infrared (near-IR) photometry indicated that V346 Nor was gradually brightening in the $K$ band, reaching a broad maximum between about 1990 and 2000 (@frogel1983 [@reipurth1983b; @reipurth1985; @kh1991; @molinari1993; @prusti1993; @reipurth1997; @abraham2004], see also Fig. \[fig:light\]). @kh1991 presented the broad-band optical-IR spectral energy distribution (SED) of V346 Nor, while @weintraub1991 published submillimeter and millimeter photometry for it. Both groups concluded that the object is surrounded by a significant amount of circumstellar material, in the form of an actively accreting disk and a flattened envelope. Recently, @kraus2016 reported a dramatic brightness decrease of V346 Nor and a subsequent brightening. They interpret these results as a 2-3 orders of magnitude drop in the accretion rate, followed by the onset of a new outburst. In order to better understand this spectacular event, and to follow up the evolution of the system, we present new near-IR observations of V346 Nor, and re-evaluate the data taken by the VISTA telescope. We analyze the brightness and color variations observed in V346 Nor and compare the results with similar fading events of highly accreting young stellar objects from the literature.
Observations, data reduction, and photometry
============================================
We observed V346 Nor with the NaCo adaptive optics instrument on the UT4 of European Southern Observatory’s Very Large Telescope (VLT) at Cerro Paranal, Chile, on April 10/11, 2008, as part of project 381.C-0241 (PI: Á. Kóspál). The weather conditions were good and the typical optical seeing was around 1$''$. We obtained $J$, $H$, and $K_{\rm S}$-band images with the N20C80 dichroic and the 13maspixel$^{-1}$ scale camera. We observed 2MASS J16323308$-$4457314 as a photometric standard with the same instrumental setup as the science target. This star is only 2$'$ away from V346 Nor and has similar 2MASS magnitudes ($J$=10.288mag, $H$=8.395mag, $K_{\rm S}$=7.242mag). Our NaCo images are displayed in Fig. \[fig:naco\]. We obtained aperture photometry for both V346 Nor and the photometric standard using an aperture radius of 2$''$ and sky annulus between 2$\farcs$6 and 3$\farcs$9. This large aperture was chosen to include all the flux of V346 Nor, which appears slightly extended in our NaCo images (see below), therefore, our photometry can be compared to earlier unresolved photometry from the literature. The obtained instrumental magnitudes were converted to standard magnitudes using the 2MASS values for the photometric standard. We also downloaded archival $J$ and $K_{\rm S}$-band NaCo observations from June 12, 2003, and reduced and extracted photometry from them in a similar way. The obtained brightnesses of V346 Nor are presented in Tab. \[tab:phot\].
![[*Top:*]{} VLT/NaCo $J$, $H$, and $K_{\rm S}$ images of V346 Nor from 2008. The color scale is logarithmic. [*Bottom:*]{} Radial intensity profiles of V346 Nor (solid curves) and of another, fainter star visible in the field of view (dashed curves) measured in our NaCo $J$, $H$, and $K_{\rm S}$ images. The uncertainty of the brightness profile of the fainter star is indicated by error bars, while the uncertainty of V346 Nor’s profile is less than the curve thickness.[]{data-label="fig:naco"}](map4b.ps "fig:"){width="8.5cm"} ![[*Top:*]{} VLT/NaCo $J$, $H$, and $K_{\rm S}$ images of V346 Nor from 2008. The color scale is logarithmic. [*Bottom:*]{} Radial intensity profiles of V346 Nor (solid curves) and of another, fainter star visible in the field of view (dashed curves) measured in our NaCo $J$, $H$, and $K_{\rm S}$ images. The uncertainty of the brightness profile of the fainter star is indicated by error bars, while the uncertainty of V346 Nor’s profile is less than the curve thickness.[]{data-label="fig:naco"}](radprof3.ps "fig:"){width="8.5cm"}
Our group observed the area around V346 Nor with the InfraRed Imaging System (IRIS) at Bochum Observatory near Cerro Armazones. IRIS is a 80cm telescope equipped with a 1k$\times$1k infrared camera. The system provides a resolution of 0.74$''$/pixel and a field-of-view of 13$'{\times}$13$'$. Data were taken between June 26, 2010 and July 1, 2010, as well as between August 22 and 25, 2016, in the $J$, $H$ and $K_{\rm S}$ bands. Individual frames were combined to eliminate the sky signal and correct for flatfield differences. V346 Nor was not visible in the $J$ band, but was detected in all $H$ and $K_{\rm S}$ frames. All $J$-band images obtained on the same night were combined into mosaics, and 3$\sigma$ upper limits were determined. For the other filters, we performed aperture photometry using the same aperture and sky annulus sizes as for the NaCo images. For the photometric calibration, we used a set of about 50 2MASS stars with quality flag ‘A’ to determine the offset between the instrumental and the 2MASS magnitudes. We found that no color term was needed for this transformation. The uncertainty of the final photometry is the quadratic sum of the formal uncertainty of the aperture photometry and the photometric calibration. The resulting $J$ upper limits and $HK_{\rm S}$ magnitudes are presented in Tab. \[tab:phot\].
We observed V346 Nor with the SMARTS 1.3m telescope at Cerro Tololo on June 7 and August 9, 2016. The telescope is equipped with the ANDICAM instrument, which provides simultaneous optical and IR images. The CCD for the ANDICAM is a Fairchild 447 2k$\times$2k chip, which we used with 2$\times$2 binning, resulting in a binned pixel scale of 0.371$''$/pixel, and a field of view of about 6$'\times$6$'$. The IR Array for the ANDICAM is a Rockwell 1k$\times$1k HgCdTe “Hawaii” Array, also used with 2$\times$2 binning, with 0.276$''$/pixel binned scale and 2$\farcm$4$\times$2$\farcm$4 field of view. We used the Johnson-Kron-Cousins $VRI$ optical and CIT/CTIO $JHK$ IR filters. A 5-point dithering was done to enable bad pixel removal and sky subtraction in the IR images. Bias and flat correction for the optical images were done by the Yale SMARTS team. Although HH 57 is faintly visible in our $V$ and $R$ images, V346 Nor itself is not detected in the optical. We used magnitude values from the UCAC4 catalog [@zacharias2013] to calibrate the images and determined 3$\sigma$ upper limits of $V\,{>}\,$20.8mag, $R\,{>}\,$19.6mag, and $I\,{>}\,$18.8mag for V346 Nor. In the near-IR regime, V346 Nor was detected in all three bands, and we performed photometry the same way as described for the IRIS images. The resulting $JHK$ magnitudes for V346 Nor are presented in Tab. \[tab:phot\].
V346 Nor was covered as part of the VISTA Variables in the Via Lactea Survey, an ESO public survey using the VISTA 4.1m telescope and the VIRCAM near-IR camera [@minniti2010]. We downloaded all VIRCAM images from this survey covering V346 Nor. To obtain photometry that can be compared with our NaCO, IRIS, and SMARTS data, we performed our own flux extraction with the same aperture as described above. At some epochs, our values differ from those published by @kraus2016 for the same measurements. The reason is probably a different treatment of the known non-linearity of the VIRCAM detectors for bright sources [e.g., @saito2012]. The correction we applied is described in details in Appendix \[sec:appendix\_a\]. The $J$ and $H$ photometry, as well as the $Ks$-band results after the non-linearity correction are also given in Tab. \[tab:phot\].
Results
=======
The top part of Figure \[fig:naco\] shows our NaCo $JHK_{\rm S}$ images from 2008 of V346 Nor, while the bottom panel displays the normalized, azimuthally averaged radial brightness distributions of V346 Nor, and another, fainter star visible in the field of view at a distance of 5$\farcs$7, position angle of 13$^{\circ}$ east of north. Assuming that this nearby faint star is a point source, the comparison of the brightness profiles show that V346 Nor is extended in the $J$ and $H$ bands, while it is consistent with a point source in the $K_{\rm S}$ band. Deconvolved sizes using Gaussian deconvolution are 0$\farcs$041$\pm$0$\farcs$016 (29$\pm$11au) in $J$, 0$\farcs$037$\pm$0$\farcs$010 (25$\pm$7au) in $H$, and we can give an upper limit of 0$\farcs$02 (13au) for the $K_{\rm S}$-band size. The size of the near-IR emitting region in circumstellar disks is typically only a few au, therefore, the emission seen in the NaCo images has to be scattered light. The $J$ and $H$ images are slightly asymmetric with the northern part slightly more extended than the southern part. No such asymmetry is evident in the $K_{\rm S}$ image.
![Near-IR light curves of V346 Nor. Data points are from @elias1980 [@frogel1983; @graham1985; @reipurth1983a; @reipurth1983b; @reipurth1985; @kh1991; @molinari1993; @prusti1993; @reipurth1997; @quanz2007; @connelley2008], the 2MASS, DENIS, AllWISE and NEOWISE catalogs [@cutri2003; @cutri2013; @cutri2015], and this work. Downward arrows indicate upper limits.[]{data-label="fig:light"}](light.ps "fig:"){width="8.5cm"} ![Near-IR light curves of V346 Nor. Data points are from @elias1980 [@frogel1983; @graham1985; @reipurth1983a; @reipurth1983b; @reipurth1985; @kh1991; @molinari1993; @prusti1993; @reipurth1997; @quanz2007; @connelley2008], the 2MASS, DENIS, AllWISE and NEOWISE catalogs [@cutri2003; @cutri2013; @cutri2015], and this work. Downward arrows indicate upper limits.[]{data-label="fig:light"}](light_zoom.ps "fig:"){width="8.5cm"}
Figure \[fig:light\] shows the near-IR light curves of V346 Nor. The first few data points indicate a brightening between 1979 and 1983, followed by a relatively constant period until about 1988. Afterwards, the $K$ and $L$-band data show a gradual brightening until 1992, already reported in @abraham2004. The $J$ and $H$ light curves were rather flat, with small, $<$1mag brightness variations. After 2003-2004 the source significantly faded, especially at short wavelengths and in 2008 displayed similar brightness although slightly redder color than the first $JHK$ photometry in 1979.
Some time around 2008, V346 Nor started a dramatic fading, and reached a minimum around late 2010-early 2011 (Fig. \[fig:light\], bottom). Afterwards, the source quickly brightened by $\Delta{}K\,=\,$2.4mag in about 3 years, and another $\Delta{}K\,=\,$0.6mag in the following 3 years, indicating a slower brightening rate. As of 2016 August, V346 Nor has not yet reached the brightness level it displayed in 1980–2000. This is well visible in the near-IR light curves, but our optical upper limits also supports this, as the star would have been visible in our $VRI$ images had it been as bright as between 1980–2000. The deep minimum was also visible in the WISE 3.4$\,\mu$m photometry, although with smaller amplitude.
The lower panel of Fig. \[fig:light\] shows that the minimum in the $K$-band has a parabola-like light curve shape. By fitting and subtracting a second-order polinomial from the photometry between 2010 and 2014, the obtained residuals are on the order of 0.2mag, and suggest a possible periodic modulation. We calculated a Lomb-Scargle periodogram for the residuals, and found a tentative 58$\pm$2day period in the data with a false alarm probability of 3$\times$10$^{-3}$ (see also Fig. \[fig:periodogram\] in Appendix \[sec:appendix\_b\]). Similar periodicities in the light curves were already found in, e.g., V1647 Ori, where it was explained by an orbiting dust cloud [@acosta2007], and in V960Mon, where is was explained by a putative close companion [@hackstein2015].
![Near-IR color-magnitude diagram for V346 Nor. The reddening path is marked with dashed line [@cardelli1989]. The solid gray curves indicate our reddened accretion disk model fits (see details in text).[]{data-label="fig:tcd"}](tcd2b.ps){width="8.5cm"}
Discussion and Conclusions
==========================
The color-magnitude diagram in Fig. \[fig:tcd\] shows that all data points until 2008 form an approximately linear strip, while the later measurements deviate from this trend, suggesting different physical mechanisms for the brightness and color changes before and after 2008. In order to reproduce the observations, in each epoch we fitted the $JHK_{\rm S}$ data points by a steady, optically thick, geometrically thin, viscous accretion disk, with radially constant mass accretion rate. Such disk models were successfully proposed and used to reproduce the SEDs of FUors by @hk96, @zhu2007, and @kospal2016. We calculated the disk’s SED by integrating the fluxes of concentric annuli between the stellar radius and R$_{\rm
out}$, assuming blackbody emission. We reddened the model fluxes by different $A_{V}$ values, using the extinction law from @savage1979 with $R_V$=3.1. Fixing the outer radius of the accretion disk (R$_{\rm out}$ = 2au, the exact value has a negligible effect on the near-IR fluxes), we have only two free parameters, the product of the stellar mass and the accretion rate $M\dot{M}$, and the extinction $A_{V}$. We fixed the stellar mass and radius to typical low-mass YSO values of 1$\,M_{\odot}$ and 3.0$\,R_{\odot}$ (the resulting accretion rate is inversely proportional to the adopted stellar mass). The inclination of the V346 Nor disk is not known, thus we adopted 2/$\pi\,{\approx}\,$40$^{\circ}$, the mean expected value if the disk is randomly oriented. The fitting procedure was performed with $\chi^2$ minimization. More extreme extinction laws (up to $R_V$=5.3, @cardelli1989) would change the fitted $A_{V}$ and $\dot{M}$ values by less than 15%, which is within the formal uncertainty of our fitting procedure.
The bluest points in Fig. \[fig:tcd\] correspond to measurements obtained in 1983-88 and 2003, when the system exhibited approximately the same brightness and color. We found that these points can be well fitted by our disk model with $\dot{M}$ = 1.0$\times10^{-5}\,M_{\odot}$yr$^{-1}$ and line-of-sight reddening of $A_V$=6.7mag. We obtained similarly good fits for the 1979 and 2008 SEDs with $\dot{M}$ = 2.1$\times10^{-5}\,M_{\odot}$yr$^{-1}$ and $\dot{M}$ = 4.5$\times10^{-5}\,M_{\odot}$yr$^{-1}$, respectively. The reddening, however, was significantly higher, 16.7mag in 1979 and 21.5mag in 2008. We simulated the time evolution of the system by computing a sequence of models of gradually changing $A_V$ from 21.5 to 6.7 mag and $\dot{M}$ from 4.5$\times10^{-5}$ to 1.0$\times10^{-5}\,M_{\odot}$yr$^{-1}$. The resulting line is plotted in Fig. \[fig:tcd\]. For most epochs, our disk model fits the SED shape well with typical formal uncertainties of 1-2mag in $A_V$ and 10-30% in $\dot{M}$. After the minimum in 2010, the shape of the SEDs can be reproduced less well with the accretion disk model, resulting in formal uncertainties of 6-8mag in $A_V$, and up to a factor of 6 in $\dot{M}$.
The data points obtained between 1989 and 1999 are situated above the model line. These SEDs can also be fitted with our disk model, but with higher $\dot{M}$ values than at any time before. $A_V$ was between 12.1 and 19.2 mag in this period. In particular, we found that the accretion rate showed a maximum in 1992 January, with $\dot{M}$=9.8$\times10^{-5}$ and $A_V$=16.8 mag. Here, we again computed a sequence of models with gradually changing $A_V$ from 19.2 to 12.1 mag and $\dot{M}$ from 9.8$\times10^{-5}$ to 3.5$\times10^{-5}\,M_{\odot}$yr$^{-1}$, also plotted in Fig. \[fig:tcd\].
@kraus2016 suggested that the minimum around 2010 is related to a large drop in the accretion rate. Using our accretion disk model, we found that by keeping a constant $A_V$=21.5mag between 2008 and 2010, and adding an accretion disk model of gradually increasing $\dot{M}$ to the SED measured in 2010 we can reach the 2008 data point (Fig. \[fig:tcd\]). In the minimum, the measured fluxes constrain the model accretion rate below 4$\times10^{-7}\,M_{\odot}$yr$^{-1}$, thus the change in accretion rate was at least a factor of 100 or more, in agreement with the findings of @kraus2016. In our modeling, the scattered light component, indicated by our NaCo observations in 2008, was not included, since its consistent treatment would be beyond the scope of this Letter.
Our results demonstrated that while the rapid fading in 2010 was an accretion event, the flux evolution beforehand was due to a correlated change in extinction and accretion rate together. That is, increasing accretion rate is accompanied by growing extinction towards the source. V346 Nor is similar to a group of highly variable young stellar objects whose flux changes are due to a combined effect of changing accretion rate and variable circumstellar extinction. Such objects are, e.g., H$\alpha$11, PV Cep, V1647 Ori, and V899 Mon [@kun2011b; @kun2011; @mosoni2013; @ninan2015]. V346 Nor also resembles the young eruptive star V2492 Cyg in several aspects: the pre-outburst position of V2492 Cyg in the near-IR color-color diagram is close to the point of the 2010 minimum of V346 Nor, it also underwent a large accretion change at the beginning of its outburst, and in the high state, the line-of-sight extinction is continuously varying [@kospal2011; @kospal2013; @hillenbrand2013]. Therefore, the observed variability in both V346 Nor and V2492 Cyg are governed by a combination of changing accretion and extinction.
The minimum of V346 Nor in 2010 was immediately followed by the re-brightening of the source. As of 2016, the source is on its way back to its 2008 state in the color-magnitude diagram (Fig. \[fig:tcd\]). This suggests that the brightening is governed by the same process that was responsible for the fading, namely changing accretion at a constant high extinction. Indeed, our simple accretion disk model can reproduce the $JHK_{\rm S}$ fluxes measured in 2016 by assuming $\dot{M} =$ 3.3$\times10^{-6}\,M_{\odot}$yr$^{-1}$, and $A_V$ = 19.8mag. The brightening of the source is still ongoing, and our latest observations suggest that after a relatively constant period after 2013, V346 Nor might have entered again a fast brightening phase.
In the classical FUor outburst models, the eruption ends when the inner disk completely depletes, and a new eruption can start only when the disk material is replenished, typically in several thousand or ten thousand years [@bell1994]. Therefore, the relatively short minimum of V346 Nor is unlikely to signal the end of the large FUor outburst in the classical sense. It was more likely a temporary halt of the accretion onto the star. A similar phenomenon was observed in V899 Mon by @ninan2015, and in V1647 Ori, where the 2004–2006 outburst finished and then restarted a few years later. The physical mechanism of this temporary stop is not known yet. @ninan2015 speculated about several possible explanations, with the constraint that these processes should be able to return the system to the pre-fading state within a few years. The detailed understanding of the evolution of the V346 Nor outburst will bring us closer to the understanding of these enigmatic sudden fadings of eruptive stars.
The authors thank the referee for his/her useful comments. This work was supported by the Momentum grant of the MTA CSFK Lendület Disk Research Group, and by the NKFIH research fund OTKA101393.
[41]{} natexlab\#1[\#1]{}
, P., [K[ó]{}sp[á]{}l]{}, [Á]{}., [Csizmadia]{}, S., [et al.]{} 2004, , 428, 89
, J. A., [Kun]{}, M., [[Á]{}brah[á]{}m]{}, P., [et al.]{} 2007, , 133, 2020
, M., [[Á]{}brah[á]{}m]{}, P., [Dunham]{}, M. M., [et al.]{} 2014, Protostars and Planets VI, 387
, K. R. & [Lin]{}, D. N. C. 1994, , 427, 987
, J. A., [Clayton]{}, G. C., & [Mathis]{}, J. S. 1989, , 345, 245
, M. S., [Reipurth]{}, B., & [Tokunaga]{}, A. T. 2008, , 135, 2496
, R. M. 2013, VizieR Online Data Catalog, 2328
, R. M., [Mainzer]{}, A., [Conrow]{}, T., [et al.]{} 2015, [Explanatory Supplement to the NEOWISE Data Release Products]{}, Tech. rep.
, R. M., [Skrutskie]{}, M. F., [van Dyk]{}, S., [et al.]{} 2003, VizieR Online Data Catalog, 2246
, J. H. 1980, , 241, 728
, J. A. & [Graham]{}, J. A. 1983, , 3792
, J. A. 1983, , 3785
, J. A. & [Frogel]{}, J. A. 1985, , 289, 331
, M., [Haas]{}, M., [K[ó]{}sp[á]{}l]{}, [Á]{}., [et al.]{} 2015, , 582, L12
, L. & [Kenyon]{}, S. J. 1996, , 34, 207
, L. A., [Miller]{}, A. A., [Covey]{}, K. R., [et al.]{} 2013, , 145, 59
, E. B., [Lauberts]{}, A., [Schuster]{}, H.-E., & [West]{}, R. M. 1974, , 18, 463
, S. J. & [Hartmann]{}, L. W. 1991, , 383, 664
, [Á]{}., [[Á]{}brah[á]{}m]{}, P., [Acosta-Pulido]{}, J. A., [et al.]{} 2013, , 551, A62
, [Á]{}., [[Á]{}brah[á]{}m]{}, P., [Acosta-Pulido]{}, J. A., [et al.]{} 2011, , 527, A133
, [Á]{}., [[Á]{}brah[á]{}m]{}, P., [Acosta-Pulido]{}, J. A., [et al.]{} 2016, ArXiv e-prints
, S., [Caratti o Garatti]{}, A., [Garcia-Lopez]{}, R., [et al.]{} 2016, , 462, L61
, M., [Szegedi-Elek]{}, E., [Mo[ó]{}r]{}, A., [et al.]{} 2011, , 733, L8
, M., [Szegedi-Elek]{}, E., [Mo[ó]{}r]{}, A., [et al.]{} 2011, , 413, 2689
, D., [Lucas]{}, P. W., [Emerson]{}, J. P., [et al.]{} 2010, , 15, 433
, S., [Liseau]{}, R., [Lorenzetti]{}, D., & [Graham]{}, J. 1993, , 5727
, L., [Sipos]{}, N., [[Á]{}brah[á]{}m]{}, P., [et al.]{} 2013, , 552, A62
, J. P., [Ojha]{}, D. K., [Baug]{}, T., [et al.]{} 2015, , 815, 4
, T., [Bontekoe]{}, T. R., [Chiar]{}, J. E., [Kester]{}, D. J. M., & [Whittet]{}, D. C. B. 1993, , 279, 163
, S. P., [Henning]{}, T., [Bouwman]{}, J., [et al.]{} 2007, , 668, 359
, B. 1981, , 44, 379
, B. 1985, , 143, 435
, B. & [Krautter]{}, J. 1983, , 3823
, B., [Olberg]{}, M., [Gredel]{}, R., & [Booth]{}, R. S. 1997, , 327, 1164
, B. & [Wamsteker]{}, W. 1983, , 119, 14
, R. K., [Hempel]{}, M., [Minniti]{}, D., [et al.]{} 2012, , 537, A107
, B. D. & [Mathis]{}, J. S. 1979, , 17, 73
, R. D. 1977, , 35, 161
, D. A., [Sandell]{}, G., & [Duncan]{}, W. D. 1991, , 382, 270
, N., [Finch]{}, C. T., [Girard]{}, T. M., [et al.]{} 2013, , 145, 44
, Z., [Hartmann]{}, L., [Calvet]{}, N., [et al.]{} 2007, , 669, 483
Near-IR photometry of V346 Nor {#sec:appendix_a}
==============================
The comparison of our aperture photometry with the 2MASS catalog revealed the known detector issue that stars brighter than about 11mag in the $K_{\rm S}$ band enter the non-linear regime of the VIRCAM detector [@saito2012]. While close to its minimum V346 Nor was below this limit, after about May 2013 it became brighter than 11mag. In order to correct for the underestimation of the signal, we plotted the offsets between the instrumental and the 2MASS magnitudes for all stars with quality flag ‘A’ in the image as a function of the instrumental magnitude. We fitted the distribution of points with a first or second order polynomial, and determined the offset valid for V346 Nor from this fit (for an example, see Fig. \[fig:nonlinearity1\]). The necessary correction due to the non-linearity was typically in the 0.1–0.2 mag range, with a few higher values up to 0.4–0.6mag. Fig. \[fig:nonlinearity2\] shows the $K_{\rm S}$-band light curve without and with the non-linearity correction, demonstrating that our correction significantly reduced the scatter of the data points obtained close in time.
![Demonstration of the non-linearity correction applied for the photometry of V346 Nor.[]{data-label="fig:nonlinearity1"}](fig_app_example.ps){width="8.5cm"}
![VISTA/VIRCAM photometry of V346 Nor without (green plus signs) and with (red circles) correction for non-linearity. Asterisks indicate our IRIS photometry. The black solid curve is a parabola fitted to the data points to remove the long-term trend before the period analysis (see Sec. \[sec:appendix\_b\]).[]{data-label="fig:nonlinearity2"}](fig_app_lightcurve2.ps){width="8.5cm"}
Period analysis {#sec:appendix_b}
===============
![[*Top:*]{} Lomb-Scargle periodogram of the light curve of V346 Nor after removing a parabolic trend, as illustrated in Fig. \[fig:nonlinearity2\]. The highest peak corresponds to a period of 58 days. The dashed lines show the powers corresponding to false alarm probabilities of 10$^{-3}$ and 10$^{-2}$. [*Bottom:*]{} Phase-folded light curve showing the data points after removing the parabolic trend and folded with a period of 58 days.[]{data-label="fig:periodogram"}](fig_app_LS.ps){width="8.5cm"}
Date JD$\,{-}\,$2400000 $J$ $H$ $K_{\rm S}$ Telescope
------------ -------------------- ---------------- ---------------- ---------------- ------------- --
2003-06-12 52802.86 9.81$\pm$0.20 … 7.48$\pm$0.20 VLT
2008-04-10 54567.78 12.63$\pm$0.16 10.11$\pm$0.09 8.16$\pm$0.14 VLT
2010-05-15 55271.34 17.65$\pm$0.04 15.09$\pm$0.09 12.27$\pm$0.07 VISTA
2010-06-26 55373.52 … … 12.69$\pm$0.09 IRIS
2010-06-27 55374.51 $>$17.37 … … IRIS
2010-06-28 55375.51 … … 12.73$\pm$0.09 IRIS
2010-07-01 55378.51 $>$16.98 15.86$\pm$0.22 12.66$\pm$0.09 IRIS
2010-08-14 55423.01 … … 12.57$\pm$0.04 VISTA
2011-05-12 55694.34 … … 12.38$\pm$0.03 VISTA
2011-05-13 55695.40 … … 12.41$\pm$0.03 VISTA
2011-05-15 55697.28 … … 12.41$\pm$0.04 VISTA
2011-05-16 55698.34 … … 12.45$\pm$0.04 VISTA
2011-05-29 55711.30 … … 12.65$\pm$0.04 VISTA
2011-06-08 55721.26 … … 12.68$\pm$0.04 VISTA
2011-08-17 55791.04 … … 12.40$\pm$0.04 VISTA
2011-08-22 55796.06 … … 12.40$\pm$0.04 VISTA
2011-08-31 55805.02 … … 12.43$\pm$0.03 VISTA
2011-09-01 55806.03 … … 12.45$\pm$0.04 VISTA
2011-09-19 55824.06 … … 12.58$\pm$0.04 VISTA
2011-10-02 55837.00 … … 12.58$\pm$0.03 VISTA
2012-06-04 56083.04 … … 11.86$\pm$0.04 VISTA
2012-06-04 56083.16 … … 11.86$\pm$0.04 VISTA
2012-06-20 56098.98 … … 11.75$\pm$0.04 VISTA
2012-06-20 56099.01 … … 11.75$\pm$0.04 VISTA
2012-06-20 56099.06 … … 11.76$\pm$0.04 VISTA
2012-06-26 56105.02 … … 11.69$\pm$0.05 VISTA
2012-06-26 56105.09 … … 11.68$\pm$0.04 VISTA
2012-06-29 56108.18 … … 11.74$\pm$0.04 VISTA
2013-05-26 56439.37 … … 10.61$\pm$0.04 VISTA
2013-06-26 56470.26 … … 10.65$\pm$0.04 VISTA
2013-06-27 56471.16 … … 10.67$\pm$0.04 VISTA
2013-06-29 56473.17 … … 10.67$\pm$0.03 VISTA
2013-06-30 56474.13 … … 10.73$\pm$0.04 VISTA
2013-07-01 56475.17 … … 10.72$\pm$0.04 VISTA
2013-07-13 56487.16 … … 10.56$\pm$0.04 VISTA
2013-07-17 56490.97 … … 10.51$\pm$0.04 VISTA
2013-07-20 56494.01 … … 10.49$\pm$0.03 VISTA
2013-07-20 56494.07 … … 10.48$\pm$0.04 VISTA
2013-07-20 56494.12 … … 10.47$\pm$0.03 VISTA
2013-07-20 56494.16 … … 10.48$\pm$0.03 VISTA
2013-07-20 56494.21 … … 10.48$\pm$0.04 VISTA
2013-07-21 56495.01 … … 10.44$\pm$0.03 VISTA
2013-07-21 56495.08 … … 10.47$\pm$0.03 VISTA
2013-07-21 56495.14 … … 10.46$\pm$0.04 VISTA
2013-07-21 56495.20 … … 10.45$\pm$0.04 VISTA
2013-07-24 56498.14 … … 10.44$\pm$0.04 VISTA
2013-07-27 56501.09 … … 10.46$\pm$0.03 VISTA
2013-07-30 56504.00 … … 10.51$\pm$0.04 VISTA
2016-06-07 57546.56 15.07$\pm$0.03 12.67$\pm$0.02 10.22$\pm$0.03 SMARTS 1.3m
2016-08-10 57610.52 14.44$\pm$0.03 12.17$\pm$0.02 9.75$\pm$0.04 SMARTS 1.3m
2016-08-22 57622.51 … … 9.90$\pm$0.02 IRIS
2016-08-25 57625.51 14.52$\pm$0.02 12.26$\pm$0.02 9.92$\pm$0.02 IRIS
[^1]: Based on observations collected at the European Organisation for Astronomical Research in the Southern Hemisphere under ESO programmes 71.C-0526(A), 179.B-2002, and 381.C-0241(A).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recent neutron scattering measurements reveal spin and charge ordering in the half-doped nickelate, La$_{3/2}$ Sr$_{1/2}$ NiO$_4$. Many of the features of the magnetic excitations have been explained in terms of the spin waves of diagonal stripes with weak single-ion anisotropy. However, an optical mode dispersing away from the $(\pi,\pi)$ point was not captured by this theory. We show here that this apparent optical mode is a natural consequence of stripe twinning in a diagonal stripe pattern with a magnetic coupling structure which is two-fold symmetric, [*i.e.*]{} one possessing the same spatial rotational symmetry as the ground state.'
author:
- 'D. X. Yao$^1$ and E. W. Carlson$^2$'
date: 'September 12, 2006'
title: 'Spin-wave dispersion in half-doped La$_{3/2}$Sr$_{1/2}$ NiO$_4$'
---
Strongly correlated electronic systems often exhibit some evidence of charge, spin, or orbital order, or some combination thereof. The interplay between these degrees of freedom can lead to a wide variety of novel phases. The nickelates in particular show a wide range of doping in which both charge and spin order coexist in stripe patterns. The less-ordered stripe structures in some families of cuprates have been widely studied for their possible connection to high temperature superconductivity.
Recent experiments on doped La$_{2-x}$ Sr$_{x}$ NiO$_4$ have shown clear evidence of static diagonal charge and spin stripe order.[@boothroyd03a; @boothroyd05] Spin wave theory has been successful at describing much of the behavior in these spin-ordered systems.[@erica04; @yaocarlson06b; @tranquada; @tranquada03; @kruger03; @freeman05a] We consider here the recent experiments by Freeman [*et al.*]{} on the spin dynamics of half-doped La$_{3/2}$ Sr$_{1/2}$NiO$_4$ using inelastic neutron scattering.[@freeman05a] In this material, the spins are in a diagonal stripe phase, where stripes run $45^o$ from the Ni-O bond direction, and the charged domain walls are only 2 lattice constants apart. The charge density modulation can either be considered as densely packed stripes, or as a checkerboard, for domain walls centered on the Ni sites, since in that case the two ways of describing the charge pattern are indistinguishable at this filling. However, if the domain walls are centered on oxygen sites ([*i.e.*]{} for bond-centered stripes), the charge checkerboard description is not applicable.
In this paper, we are interested in the extra magnetic mode dispersing away from the antiferromagnetic wavevector $Q_{\rm AF} = (0.5,0.5)$ above $50$meV in Fig. 3 of Ref. . One explanation put forth by the authors is that diagonal discommensurations in the spin order may be able to account for this extra scattering mode. We show here that the mode could also be due to asymmetry in the spin coupling constants, in which case the “extra mode” is really an extension of the acoustic band, made visible due to stripe twinning.
The observed ordering vector is $Q=(0.275,0.275)$, which is close to a commensurate stripe value of $Q=(0.25,0.25)$. The slight deviation from commensurability is believed to be due to the discommensurations described above. Since we are interested in describing relatively high energy effects, in what follows we neglect the small incommensurability, and consider commensurate diagonal stripe structures of spacing $p=2$. We consider two patterns in this paper: site-centered stripes as shown in Fig. \[lattice\](a), and the corresponding bond-centered stripes shown in Fig. \[lattice\](b). To model these two systems, we use a suitably parametrized Heisenberg model on a square lattice, $$%H= \frac{1}{2}\sum_{\left< i,j\right>} J_{i,j} \mathbf{S}_{i} \cdot \mathbf{S}_{j}
H= \frac{1}{2}\sum_{i,j} J_{i,j} \mathbf{S}_{i} \cdot \mathbf{S}_{j}
%+ K_c\sum_{i} (S_i^z)^2 %EC removed.
\label{model}$$ where the indices $i$ and $j$ run over all sites, and the couplings $J_{i,j}$ are illustrated in Fig. \[lattice\]. For diagonal, site-centered stripes of spacing $p=2$ (DS2), there is no need for nearest-neighbor coupling, and so we set $J_a = 0$. The straight-line coupling $J_b$ across the domain wall is antiferromagnetic. The diagonal coupling $J_c$ across the domain wall is also antiferromagnetic, but the diagonal coupling $J_d$ parallel to the stripes we take to be ferromagnetic, $J_d <0$, as explained below. In the diagonal bond-centered case (DB2), the nearest neighbor coupling $J_a >0$ is finite and antiferromagnetic. We also include the ferromagnetic coupling $J_b <0$ across the domain wall. Since we are interested in describing high energy effects, we neglect the very weak single-ion anisotropy term, which splits the $2$-fold degenerate acoustic band at low energy, with one mode remaining gapless at the IC point $Q_{\rm IC} = (0.25,0.25)$, and the other mode developing a small gap.[@freeman05a]
For the case of site-centered stripes in the absence of the diagonal couplings, [*i.e.*]{} for $J_c=J_d=0$, the spin system reduces to two interpenetrating antiferromagnets, with two separate Néel vectors but identical Néel ordering temperatures. Any weak diagonal coupling is sufficient to establish a unique relative direction between the two Néel vectors, and the ground state becomes the stripe structure shown in Fig. \[lattice\](a). The number of reciprocal lattice vectors is also decreased by a factor of $2$ in the presence of the diagonal couplings $J_c \ne 0$ or $J_d \ne 0$, as can be seen in the bandstructure of Fig. \[plot3d\]. In either case, independent of the value of $J_c$ and $J_d$, although the antiferromagnetic point $Q_{\rm AF}$ is a magnetic reciprocal lattice vector and therefore must have a spin wave cone dispersing out of it, there is no net antiferromagnetism in the system, so that weight is forbidden at zero frequency at $Q_{\rm AF}$. The cone that emanates out of $Q_{\rm AF}$ gains finite weight as energy is increased, but remains faint at low energies.
Another key feature of nonzero couplings $J_c$ and $J_d$ for site-centered stripes is the symmetry of the spin wave structure, as shown in Fig. \[plot3d\]. In the limit where $J_c=J_d=0$, the spin wave dispersion is symmetric under $90^o$ rotations, as shown in Fig. \[plot3d\](a). However, when either $J_c$ or $J_d$ or both are nonzero, the symmetry is broken, and the spin wave structure is now only symmetric under $180^o$ rotations, as shown in Fig. \[plot3d\](b). This means that for any nonzero $J_c$ or $J_d$, the spin wave velocity of the acoustic mode dispersing out of $Q_{\rm AF}=(0.5,0.5)$ is different parallel and perpendicular to the stripe direction. In the presence of stripe twins, the two velocities will appear as two branches in plots of $\omega$ [*vs.*]{} $\vec{k}$, as shown in Fig. \[fig:DS2\].
The magnon dispersion from Eq. (\[model\]) can be solved analytically, and for DS2 we find that $$\omega(k_x,k_y)=2\sqrt{A^2-B^2},
% EC K_c \omega(k_x,k_y)=2\sqrt{A^2-B^2 + K_c (A\pm B) },$$ where $$\begin{aligned}
A&=&2J_b+J_c-J_d+J_d\cos(k_x-k_y) \nonumber \\
B&=&J_b\cos(2k_x)+J_b\cos(2k_y)+J_c\cos(k_x+k_y).\end{aligned}$$ There are two different spin wave velocities for the cones emanating from the IC peak $Q_{\rm IC} = (0.25,0.25)$ and symmetry-related points, one corresponding to spin wave velocities perpendicular to the direction of the domain walls ([*i.e.*]{} perpendicular to the stripes), and the other parallel to the direction of the domain walls, $$\begin{aligned}
v_{\perp}&=&4(2J_b+J_c) \nonumber \\
v_{\parallel}&=&4\sqrt{(2J_b+J_c)(2J_b-J_d)}~. %yao: easier to understand\end{aligned}$$
In Fig. \[fig:DS2\], we show the expected dispersions and scattering intensities for DS2. Plots are shown for twinned stripes, summing the contributions parallel and perpendicular to the stripe direction, [*i.e.*]{} along the $(k_x,-k_x)$ and $(k_x,k_x)$ directions, respectively. Because $J_c$ and $J_d$ are nonzero, there is an apparent optical mode. Note that it is not a true optical mode, since in this configuration there are only two spins per unit cell, leading to only one (twofold degenerate) band. As in the bond-centered case, weight is forbidden at low energy at the antiferromagnetic point $Q_{\rm AF}$, so that the spin-wave cone emanating from this magnetic reciprocal lattice vector has vanishing weight as $\omega \rightarrow 0$. In Fig. \[fig:DS2\](a), we have set $J_c = J_b$ and $J_d = -0.5J_b$. In Fig. \[fig:DS2\](b), we use $J_c = 2J_b$ with $J_d = -0.5J_b$. Notice that in panel (a) of the figure, the apparent optical mode is flat.
We have chosen the coupling constants with the following in mind: The “acoustic” branch peaks at $\omega(3 \pi/4, 3 \pi/4) = 4 J_b + 2 J_c$. The apparent “optical” mode peaks at $\omega(\pi/2, 3 \pi/2) = 4 \sqrt{(2J_b-J_d)(J_c-J_d)}$. The data indicate that the apparent optical mode is higher in energy than the top of the “acoustic” part: $\omega(\pi/2, 3 \pi/2) > \omega(3 \pi/4, 3 \pi/4)$, which implies that $$J_d \le (1/2)(2 + J_c - \sqrt{2}\sqrt{4 + J_c^2})
%~~{\rm Constraint~}1
\label{eqn:constraint1}$$ when $J_b = 1$. However, the extra mode above $50$meV is not too high in energy, so parameters need to be chosen so as to satisfy this constraint, but remain close to the equality.
We also require the apparent “optical” branch to be concave, since there is no evidence of a dip in the extra mode. This requirement gives $${\partial^2 \over \partial k_x^2} \omega (k_x, -k_x) =
{4 (J_d - 2)(J_c - 2 -2J_d) \over \sqrt{-(J_c - J_d)(J_d -2)}}\le0~,$$ resulting in the constraint that $$J_d \le {1 \over 2}J_c - 1~.
\label{eqn:constraint2}$$
As long as the second constraint, Eqn. \[eqn:constraint2\], is satisfied in the range $0\le J_c \le 2 J_b$, then the first constraint, Eqn. \[eqn:constraint1\], is also satisfied. To describe the data, then, we find that we need a sizeable $J_c$, on the order of $J_b$. This makes $J_c$ significantly larger than that reported at the lower doping $x=1/3$, where diagonal site-centered stripes of spacing $p=3$ (DS3) were used to explain the data successfully.[@boothroyd05] Although the fits in Ref. were good for $J_c =0$ and were not significantly improved by letting $J_c$ increase to $J_c \approx 0.5 J_b$, the data were still well described using a nonzero $J_c$. We find that the data at doping $x=1/3$ are also well described by taking $J_c$ to be as large as $J_b$ or even $2 J_b$ as in our Fig. \[fig:DS2\].
We also find that we need $J_d<0$, [*i.e.*]{} the diagonal coupling parallel to the stripes needs to be ferromagnetic, in order to describe the data. If one considers the diagonal spin couplings $J_c$ and $J_d$ to be derived from, [*e.g.*]{}, a perturbative treatment of a single-band Hubbard or three-band Emery model on a square Ni-O lattice, we expect $J_c = J_d$. Given that the spin ground state breaks the $4$-fold rotational symmetry of the square lattice to only $2$-fold rotational symmetry, any finite spin-lattice coupling results in the two diagonal directions being inequivalent, and leads to $J_c \ne J_d$. For weak rotational symmetry breaking of the square lattice in the perturbative regime, one expects $J_c = J_o + \epsilon$ with $J_d = J_o - \epsilon$ where $\epsilon$ is small compared to $J_o$, so that the anisotropy between the two diagonal coupling directions is weak as well. (This preserves the spin ground state of Fig. \[lattice\](a).) We find, however, that this regime of the coupling constants leads to the apparent optical mode being too low in energy to capture the data. This may indicate that the materials are far from the perturbative limit of the single-band Hubbard or three-band Emery model.
Fig. \[cut.ds2\] shows constant energy cuts for DS2 corresponding to the parameters in Fig. \[fig:DS2\](b). An important feature of this configuration is that although the antiferromagnetic point $Q_{\rm AF} = (0.5,0.5)$ is a magnetic reciprocal lattice vector, zero-frequency weight is forbidden there by symmetry, since the stripes have no net Néel vector at $Q_{\rm AF}$. Combined with the fact that there is no optical band, the DS2 configuration [*can never have spectral weight at the antiferromagnetic point*]{} $Q_{\rm AF}$, even at finite frequency. Notice that as energy is increased in Fig. \[cut.ds2\], a faint spin wave cone emerges from $Q_{\rm AF}$ in a light ring of scattering, but none of the constant energy plots show any weight at $Q_{\rm AF}$. This is consistent with the constant energy plots for La$_{3/2}$Sr$_{1/2}$ NiO$_4$ shown in Ref. . By contrast, the corresponding bond-centered configuration (DB2) shown in Fig. \[lattice\](b) has an optical band which displays a saddlepoint at $Q_{\rm AF}$, and rather large scattering intensity at finite frequency at $Q_{\rm AF}$ as a result. (See Fig. \[fig:DB2\] of this paper, as well as Fig. 8 of our previous paper, Ref. .)
While diagonal, site-centered stripes of spacing $p=2$ (DS2) are able to account for the behavior of the apparent optical mode observed to disperse away from $Q_{\rm AF}$ in Fig. 3 of Freeman [*et al.*]{}[@freeman05a], there are two other high energy features which this model has not captured. One is the asymmetry in intensity observed above $30$meV in the spin wave cones emanating from the main IC peaks. The other is a mode in the $31-39$meV range propagating away from $(h,k)$ structural reciprocal lattice points. These (as well as the apparent optical mode) have been attributed to discommensurations in the magnetic order.[@freeman05a]
In Fig. \[fig:DB2\], we show the expected dispersions and intensities for the diagonal, [*bond-centered*]{} stripes (DB2) shown in Fig. \[lattice\](b). This configuration has a true optical mode. Fig. \[fig:DB2\](a) shows weak coupling across the charged domain walls, with $J_b = -0.1 J_a$, and Fig. \[fig:DB2\](b) has somewhat stronger coupling across the domain walls, with $J_b = -0.5 J_a$. Note that in the bond-centered case, couplings across the domain walls are ferromagnetic. Results are shown for twinned stripe patterns, summing the contribution parallel and perpendicular to the stripe direction, [*i.e.*]{} along the $(k_x,-k_x)$ and $(k_x,k_x)$ directions, respectively. Although there is a reciprocal lattice vector at $Q_{\rm AF}$, weight is forbidden there at zero frequency, since the Néel vector switches sign across the domain walls. We have reported the analytic form of the spin wave dispersion in this case in a previous publication.[@erica04]
Because this spin configuration is only $180^o$ symmetric, the spin-wave velocity emanating from $Q_{\rm AF}$ is different parallel and perpendicular to the stripe direction. However, the branch emanating from $Q_{\rm AF}$ in the direction parallel to stripes has so little weight as to be effectively invisible in the plots. This configuration displays a true optical mode because there are four spins in the unit cell. The optical mode has a [*saddlepoint*]{} at $Q_{\rm AF}$ and finite energy, with increased weight at the saddlepoint. For weak coupling across the domain walls ($|J_b|<|J_a|$), the optical mode always displays significant weight at $Q_{\rm AF}$ at finite frequency. This is not supported by the data[@freeman05a], which at the energies measured display no scattering at $Q_{\rm AF}$ and finite frequency. This likely indicates that the domain walls are not bond-centered in this material, but are probably site-centered.
In conclusion, we have used linear spin-wave theory to describe the magnetic excitations recently observed in neutron scattering[@freeman05a] on La$_{3/2}$Sr$_{1/2}$ NiO$_4$. Many features of the data were captured in the spin wave analysis of Ref. . Other features, including an apparent optical mode dispersing away from $Q_{\rm AF}$ above $50$meV, were attributed to discommensurations in the spin order. We show here that the apparent optical mode may also be captured in linear spin wave theory by using a spin coupling configuration that preserves the symmetry of the spin ground state. Namely, we have shown that diagonal, site-centered stripes of spacing $p=2$ capture this mode when the pattern of couplings is $2$-fold symmetric. This is because the $2$-fold symmetric coupling pattern gives rise to two different spin-wave velocities ([*i.e.*]{} $v_{\parallel} \ne v_{\perp}$) emanating from the antiferromagnetic point $Q_{\rm AF}=(0.5,0.5)$. For twinned samples, the two velocities are simultaneously visible, and the higher velocity mode $v_{\parallel}$ parallel to the stripes is responsible for the “extra” scattering above $50$meV. Furthermore, this configuration is forbidden to display scattering at the antiferromagnetic point $Q_{\rm AF}$, whereas bond-centered stripes have a true optical mode with significant weight at $Q_{\rm AF}$, which is not supported by the data. We therefore conclude that the magnetic excitations observed in Ref. are consistent with site-centered stripes, but not with bond-centered stripes..
It is a pleasure to thank D. K. Campbell and A. T. Boothroyd for helpful discussions. This work was supported by Boston University (DXY), and by the Purdue Research Foundation (EWC). EWC is a Cottrell Scholar of Research Corporation.
[8]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , , , , ****, ().
, , , , , , , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , , , ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Every GRB model where the progenitor is assumed to be a highly relativistic hadronic jet whose electron-pair secondaries are feeding the jet’s engine, necessarily (except for very fine-tuned cases) leads to a high average neutrino over photon radiant exposure (radiance) ratio well above unity, though the present observed average IceCube neutrino radiance is at most comparable to the gamma in the GRB one. Therefore no hadronic GRB, fireball or hadronic thin precessing jet, escaping exploding star in tunneled beam, can fit the actual observations. A new model is shown here, based on a purely electronic progenitor jet, fed by neutrons stripped from a neutron star (NS) by tidal forces of a black hole or NS companion, it may overcome these limitations. Such thin precessing spinning jets explain unsolved puzzles such as the existence of the X-ray precursor in many GRBs. The present pure electron jet model, disentangling gamma and (absent) neutrinos, explains naturally why there is no gamma GRB correlates with any simultaneous TeV IceCube astrophysical neutrinos. A thin persistent electronic beaming, born in an empty compact binary system has the ability to offer the answer for a sudden engine (the thin jet) whose output may be comparable, off axis, to $10^{44}$–$10^{47}$ erg s$^{-1}$. The jet power is fed by a stripped neutron mass skin by tidal forces. The consequent jet blazing to us on axis occurs within the inner jet cone beammed by a spiral charged ring at highest apparent output. In rare cases, the NS, while being stripped by the BH companion, will suddenly become unstable and it will explode and shine during the GRB afterglow, with an (apparent) late SN-like event birth. Primitive SN outer chemical mass shells, should be retro illuminated by such a NS explosion, re-brightening the relic nuclei as in a SN-like spectral line signature. To disentangle SN from NS explosion we note that only radiative shining due to Cobalt and Nichel decay, present in most SN, will be absent in present NS explosion. Recent IceCube-160731A $\nu_\mu$ event with absent X-$\gamma$ traces confirm the present model.'
author:
- Daniele Fargion
- Pietro Oliva
title: 'Solving the missing GRB neutrino and GRB-SN puzzles'
---
Introduction {#sec:intro}
============
Gamma ray burst (GRB) physics represents today a half-century (1967–2016) unsolved puzzle which brings together a long list of unanswered questions related to the many faces a GRB can show. The main popular fireball model and its modern variations are always doomed to fail in front of a key lethal unanswered question: how we do explain the existence of tiny X-ray precursors (present in hundreds of GRBs) seconds or minutes before the huge apparent gamma explosion? No fireball nor any one-shot fountain model even try to face this reality or seems to be comfortable with the existence of precursors. Maybe the time has come to embrace a change.
One of the most important puzzles to recall is: how is it possible that a huge GRB (apparently isotropic) power $P_{\rm GRB}\sim10^{53}$ erg s$^{-1}$ can sometimes coexist (see i.e., ) with a late correlated supernova (SN) event of the typical order of $P_{\rm SN}\sim10^{44}$ erg s$^{-1}$, a power billion times weaker? Indeed, this question represents only the tail of a long chain of mysteries surrounding the nature of GRBs. First of all, because of the fast millisecond-second scale of GRB variability, how could any corresponding compact source emit at MeV energies any apparent spherical GRB luminosity $P_{\rm GRB}\gtrsim10^{51}\div10^{53}$ erg s$^{-1}$ several orders of magnitude above Eddington limit for such objects $(\sim10^{38}$ erg s$^{-1})$? In such a model photon scattering will lead to the birth of electron pairs so dense and opaque that they will definitively screen off and shield the GRB self prompt compact spherical explosion. Moreover no GRB show just a single bang (as in a SN), on the contrary the most of them show a sequence of peaks in gamma.
\[fig:06\]
The early (1980–2000) “fireball” model \[[@Cavallo01071978; @1986ApJ...308L..47G; @1986ApJ...308L..43P; @1992MNRAS.258P..41R; @1994ApJ...430L..93R; @1993ApJ...418L...5P; @1997ApJ...489L..33W; @1997ApJ...489L..37S; @1997ApJ...478L...9V; @1999ApJ...517L.113C]\] tried to explain that the sea of electron pairs from a GRB will spread out and dilute in a sphere, the so-called fireball, hence cooling the photons in an adiabatic expansion from MeV to keV energies. The model then foresaw that when the pair-sea shell would have become sufficiently diluted and transparent, these keV photons (ejected and scattered by these ultrarelativistic electron pairs) would reach us boosted at MeV energies like the ones observed in GRBs. Since the Beppo-SAX identification and discovery of the high cosmic redshift of some GRBs with extremely high luminosity this simple isotropic model depicting “spherical” GRBs failed, mostly because of the observed highest GRB integrated energy $(E_{\rm GRB}\gtrsim10^{54}\,\mathrm{erg})$ which is comparable or larger than the same source budget allowable energy mass, a mass derived and constrained by the object’s Schwarzschild radius (fixed or constrained by its variability). Clearly, such an energy budget paradox could not be solved by an increase of the GRB mass and its Schwarzschild radius because of the subsequent increase of the variability time scale in disagreement with the observed fast ms GRB timescales.
Subsequently in 2000, most authors abandoned the spherical fireball model and turned to a mildly beamed jet-explosive fountain model with a $\Delta\Omega/\Omega\sim10^{-3}$ ratio [@1999ApJ...517L.109S; @2000ApJ...529..146E; @2000AIPC..526..514M] while the inner (random) variability (peaks and sudden re-brightening) of the GRB luminosity was explained assuming that the fountain jet would hit relic shells of matter around (but external) the GRB, where shock waves revived the GRB luminosity. Unluckily for fireball believers, this ad hoc model was and still is not able to explain the multi-peak structure of some GRBs: to face this variability and to keep alive the fireball model several authors considered the far external relic shells of the exploding GRB star as the additional onion-like screens where, by scattering of the expanding shock waves, the explosive luminosity re-brightens several times. Obviously this process, fireball defenders said, must open the fireball fountain jet into an increasingly spread out spherical explosion with a more and more diluted luminosity. Several GRBs on the contrary proved an opposite growing peak luminosity trace. Moreover each onion shell in such models must be diluted enough to transfer outside the GRB shock wave but not too diluted for being transparent to the scattering: such a fine-tuned GRB dressing for fireball is purely ad hoc and unexplained. In particular the fireball one shot model is totally incapable of describing and justifying the early X-ray precursor [@Fargion:2001xf; @2001foap.conf..347F] present in a significant fraction $(\sim 7\%\div15\%)$ of GRB curves up to date. These earliest bright X-ray flares may hold a million times the SN luminosity even several minutes (ten minutes for GRB 06124) before the main (billion times brighter) harder GRB event. Moreover, the wide beaming of the fountain $(\Delta\theta\sim10^\circ\div15^\circ)$ is assumed ad hoc and the single-shot model cannot describe some observed long life and “day after” re-brightening GRBs, nor the several week X-ray afterglows. Moreover the fireball model is unable to justify the apparent “conspiracy” that makes GRB more and more (in apparent) brightest power at larger and larger redshift, in a spread of apparent luminosity of nearly a factor a billion discussed below: a beaming factor of just a thousand as in fireball model, cannot explain more than a thousand in luminosity range variability. On the contrary a thinner precessing jet whose solid angle is a million or billion times smaller, may embrace a million or a billion luminosity variability. The same plot play a role in making (apparent) harder and harder the GRB spectra with the more and more distance (and red-shift). Naturally, we are observing a statistical geometry evolution that allows the most distant and richest sample to have the most aligned and thinner jets pointing towards us, while the nearer (smaller cosmic volume) and rarer GRB are usually off-axis and they shine with low fluxes.
An anti-Copernican GRB Luminosity evolution?
============================================
Among the contradictions of all GRB one-shot models stand the apparent conspiracy of GRB luminosity around us: nearby (lowest red-shift) GRBs show on average a peak luminosity and a soft energy spectra versus the much brighter and harder luminosity of far away (large redshift) GRB events. The conjure or the apparent luminosity evolution is so fast that it suggest that we (in our local Universe) are at the center of the Universe. There is not any ad hoc luminosity evolution that may explain such a sudden ($z\gtrsim0.01$) growth in spectra and luminosity evolution. This result is manifest in Fig. \[fig:06\], and it calls for an explanation. A wide fountain and a marginal beaming as in a fireball model cannot explain such a factor of a billion in luminosity spread; a very thin beaming (as will be discussed below within a millionth or less of steradian solid angle) spinning and precessing jet has a characteristic angle linked to the peak electron Lorentz factor above thousands value: the inverse of the solid angle and the apparent luminosity grows as large as the square of the Lorentz factor (of highest energetic electrons). Of course also a hierarchic cannibal event between binary compact objects may play a role, showing new rare powerful jet with wider distances and volumes. However as it is well known binary (Schwarzschild or Kerr neutral) BH merging systems are ejecting only gravitational waves (GW). Therefore only (or mainly) neutron star merging, as discussed below, in BH-NS or in NS-NS systems are a guaranteed source of electromagnetic radiation and the NS are a well-bounded amount of mass-energy. Therefore even if the GRB event is fed by NS-NS or NS-BH binary merging, even for large and large BH, the outgoing energy budget in GRB is nearly fixed and bounded by the NS mass. The huge luminosity variability is due to the very thin beaming geometry associated to tens-hundred GeV electron pair jets, not to any hierarchic growth of objects.
 
\[fig:01\]
Precessing and spinning of thin decaying $\gamma$ jet {#sec:prec_jet}
=====================================================
In order to overcome the GRBs puzzles we proposed since 1994 [@1994dsu..conf...88F; @Fargion1995269] a model to describe both GRBs (and/or SGRs) based on the blazing of a very thin $\gamma$ beamed jet $(\Delta\theta\sim0.1^\circ\div0.02^\circ)$, $\Delta\Omega/\Omega\lesssim 10^{-6}\div10^{-8}$ whose birth was associated to tens GeV electron pairs showering via inverse Compton scattering (ICE) into MeV-GeV photons . Our precessing-spinning $\gamma$ jet was assumed fed at a low power (fitting today SGR or AX-PSRs) in our galaxy $(P_{\rm SGR}\sim10^{38}\,\mathrm{erg}\,\mathrm{s}^{-1})$ or, since 1998 , also at highest power as large as a SN powering and beamed jet for cosmic GRBs $(P_{\rm GRB}\sim\gamma_e^2\mathrm{P}_{\mathrm{SN}}\simeq10^{50}\div10^{54}\,\mathrm{erg}\,\mathrm{s}^{-1})$.
![Protons follow their ring trajectory while in $\beta$-decay forming a net charged current and a huge aligned magnetic field $\mathbf{B}_{p}$. The evaporating electrons are easily captured and aligned along $\mathbf{B}_{p}$; their crowding at the North and the South Poles create a huge electrostatic gradient that makes a powerful linear active accelerator: an electronic jet arises and ejects electrons and/or electron pairs by bremsstrahlung as well as photons (by inverse Compton scattering and synchrotron radiation); the thin spinning and (by tidal gravity forces) precessing jet, drives a collinear $\gamma$ jet making a blazing dance by its geometry beaming [@2006ChJAS...6a.342F; @Fargion:2006we]. Once on axis, we are dazzled and we call it a GRB event.](f2a.eps)
\[fig:02\] Late GRB jet power, decaying with a power law $\approx t^{-1}$, may shine as an nearby exhausted soft gamma repeater (SGR) jet source where the output power is correlated with a thousand year time delay with the early GRB and present SGR output. The geometrical spinning and precessing of the thin GRB-SGR jet naturally explain the huge GRB variability and the quasi-periodic behaviors found in well recorded SGR events. In the present model discussed below, the feeding of stripped matter of a NS by a black hole (BH) or a NS companion, is shining energy: indeed stripped neutrons and protons condense into a charged spiral ring that is paying the energetic output budget to eject a thin collimated, spinning and precessing electron jet, at $10^{44}-10^{47}$ erg s$^{-1}$ output; moreover the bending geometry of the electron jet (by bending magnetic fields of the accreting ring and the BH spin) and its consequent beamed variability, explain the huge and fast GRB-SGR luminosity. The fact that the neutron by NS star stripped matter and its decayed protons will follow the spiral geodesic around the BH or NS cannibal companion (while the electrons and neutrinos will not) will lead to a charged ring and a sudden collimating magnetic field. This decaying neutron-$e^+e^-$ pairs-proton ring, which is also pulsating, can shrink the magnetic lines and it can force the electrons trapped in the poles into an ultra-relativistic jet which will later create the observable gamma jet. This novel electronic model is able to avoid the pion progenitor and the overcrowded neutrino tails foreseen in all hadronic GRB models explaining GRB-$\nu$ absence.
GRB with SN event
-----------------
In some occasion such an electronic jet model formed around the BH, or heaviest NS companion, may also lead to an explosion of the relic stripped NS binary, which is now unstable because of the spoiled and stolen external weights.
Because of the extremely beamed angle $(\Delta\Omega/\Omega\sim10^{-6}\div10^{-8})$ these apparent luminosity, if seen in-axis by the observer, would shine apparently as bright as a $\tilde{P}_{\rm SGR}\sim10^{44}\div10^{46}\,\mathrm{erg}\;\mathrm{s}^{-1}$ while $\tilde{P}_{\rm GRB}\sim10^{50}\div10^{54}\,\mathrm{erg}\;\mathrm{s}^{-1}$. The lifetime of the jet has been assumed not to be a one-shot event (as the fireball model does). On the contrary our thin precessing and spinning jet has a characteristic decay life about $t^{\rm GRB}_{\mathrm{decay}}\simeq\left(t/t_{0}\right)^{-1}$, where $t_{0}\simeq 3\times10^4$ s. This half-a-day timescale was chosen to connect, by a time decay law $P\sim \left(t/t_{0}\right)^{-1}$ the highest GRB output to late, thousand years later, less powerful relic, almost steady (Galactic as SS433) Soft Gamma Repeaters, SGRs.
Despite being able to explain even the X-ray precursor (by a peripherals skimming shine of the jet to the Earth, before the main jet blazes as a GRB) and the late GRB re-brightening through simple geometry beaming, the precessing jet model unifying GRB and SGRs was (and it is) often underestimated or un-noticed since twenty years.
Hadronic jet feeding a fireball lepton–$\gamma$ jet {#sec:fireball}
---------------------------------------------------
The fountain-fireball model was — and is — based on shock interacting shells of hadrons (UHECR at PeVs$\div$EeV, protons and nuclei) leading to neutral pions $(\pi^0\to2\gamma)$ as well as to charged ones $(\pi^\pm)$ whose final decay results in electron pairs, the ones that later will shine in $\gamma$ in the GRB and a rich tail of neutrinos $(\nu_e, \nu_\mu, \bar{\nu}_e, \bar{\nu}_\mu)$ as well. There is also the possibility to feed pions by UHE nucleons and nuclei interacting with photons in flight. Also more violent charmed hadronic reactions lead to prompt secondaries as the ones above. In this context the most popular fireball model foresees a comparable trace of $\gamma$ luminosity under the form of GRBs with respect to a neutrino radiance, as they were just secondaries of charged pions in decay in vacuum space. Nevertheless, we repeat, GRB occur in dense stella shells in fireball model. Naturally, because of the photon-photon interaction and/or IR-tens TeV opacity most of highest TeVs photons degrade and decay into MeV$\div$GeV ones (directly at their source or along their cosmic flight). This is not the case for tens TeVs or PeVs complementary neutrinos that may reach us unabsorbed showing (in this popular and ideal fountain-fireball model) the same radiance imprint of the partially absorbed gamma observed in GRBs. As we shall comment, the transparent pion decay in flight, in fireballs, is a wishful chain of events, mostly very unrealistic because most of the onion shell barrier encountered by the fireball jet will be (mainly at the inner core) opaque to photons but not to neutrinos. Photons will fed the kinetic energy of the barrier shells while UHE neutrinos will escape with nearly no losses. If the inner star core shells are opaque even to the neutrinos then only the rare interacting UHE neutrinos, making UHE penetrating muons at the external edges, may feed the GRB with electromagnetic secondaries, while most of the primary neutrinos will export to us much more energy than gamma in GRB anyway. In conclusion the ratio gamma - neutrinos comparable to the unity is a “chimera".
The so-called Waxman-Bahcall (WB) limit or bound [@1999PhRvD..59b3002W], which connects ten EeV cosmic ray (CR) radiance ($\Phi_{\rm CR}\sim$10 eV cm$^{-2}$ s$^{-1}$ sr$^{-1}$) with average cosmic GRBs one $(\Phi_{\rm GRB}\sim\Phi_{\rm CR})$, constrains the expected cosmic tens TeV$\div$PeV GRB neutrinos (GRB$\nu$s) at similar GRB energy radiance. Indeed, the expected WB neutrino signal didn’t arise with any correlated GRB yet, or it might be rarely $(\sim1\%)$ arose as a possible precursor. The absence of any prompt GRB–$\nu$ correlation represents a remarkable failure of any one-shot fireball version, even the most beamed one. No room for one-shoot GRB neutrino and gamma event [@2014arXiv1408.0227F]. Furthermore, any hypothetical dark or hidden population of GRB should not be considered, for this would call for a higher and higher ratio $(\Phi^\nu_{\rm GRB}/\Phi^\gamma_{\rm GRB}\gg1)$ while the observations are telling us $(\Phi^\nu_{\rm GRB}/\Phi^\gamma_{\rm GRB}\sim1)$ [@2012Natur.484..351A; @2016arXiv160106484I].
 
\[fig:03\]
In our thinner precessing jet we might solve the huge apparent GRB power spread puzzle in a first approach because of the ultra-relativistic beaming and the consequent thin beaming angle: the higher the energy, the thinner the jet cone and thus the rarer the blazing, which of course explains why we have observed (at tens to hundred keV) thousands of GRBs, a few hundred GRBs at a MeV to tens MeV, a few dozen at a hundred MeV to GeV energies and only few rare events at a hundred GeV, the beming explaining their rarety. The precessing jet model can also shine in an almost cyclic fashion (like SGRs) and might blaze partially as a rare precursor, ruling out the mysterious 10%$\div$20% GRB events with precursors. In principle a thin relativistic beaming may explain that TeV neutrinos are so beamed that their shining inside the wider X-$\gamma$ cones happens very rarely. Furthermore, this requires a prompt $\nu$ detection with a fast follow up in X-$\gamma$ range. The first attempts (see next section), have failed.
An hadronic or electronic precessing jet?
-----------------------------------------
We admit that our precessing $\gamma$ jet was originally based on hadronic-UHECR primaries, leading to PeVs $\mu^+\mu^-$ secondaries whose decay in flight were able to escape and survive the eventual opaque stellar mass layer and photosphere of a SN explosion [@2005NCimC..28..809F]. In addition, the same $\mu^+\mu^-$ shined in $\nu$, $\bar{\nu}$ at higher and higher than unity ratio respect to photons; this applies for the following reasons: if GRB’s $\gamma$ are made by relativistic electrons radiation and if the GRB jet are originated by UHECR hadrons inside the collapsing star, than only a small fraction of the UHECR energy radiance is able to escape the matter barrier in the form of secondary final $\gamma$ constituting the GRB. Most of the hadron jet energy is dispersed and wasted inside the baryonic shell kinetic energy and its temperature along the jet shock wave propagation. The basic huge absorption of any electromagnetic traces respect to neutrino ones is a severe argument against any hadronic GRB origination. Present low (or missing) neutrino records in IceCube with respect to same observed gamma radiance in nearly a thousand GRB probe it [@2016ApJ...824..115A].
\[fig:03b\]
The absence of $\gamma$-X signal from IceCube-160731A {#sec:lack}
-----------------------------------------------------
The very recent prompt search of an electromagnetic trace by an astrophysical candidate event IceCube-160731A have proven the embarrassing absence of any optical (H.E.S.S.), X-$\gamma$ (Swift, Fermi, Agile) correlated signal (see [@report]). A first and rough estimate of downgoing energetic $\nu_\mu$ neutrino at $\sim100$ TeV in a km$^3$ IceCube has a pure probability to interact ($\sim10^{-4}$). therefore, the consequent expected IceCube-160731A released energy is $\sim10^{18}$ eV over a km$^2$ area, or $10^8$ eV/cm$^2$ energy fluence. On the contrary the electromagnetic bounds in Swift (as well as in Fermi and other detectors) are as low as 1 eV/cm$^2$ energy fluence. The main consequence is that even no a part of a million is related in $\gamma$ respect to $\nu$ signal. Therefore, there are no hadronic jets in GRBs (or worse no clear understanding of IceCube astropysical $\nu$ nature, [@2015arXiv151208794F]).
Cosmic rays and hadronic jet surviving analogy {#sec:ana}
----------------------------------------------
To depict the analogy in a more clear way let’s recall the CR metamorphosis along their flight inside the Earth atmosphere, which is a ten meters water equivalent (w.e.) screen: at ground level only a small amount of the CR energy is observable under the form of electromagnetic secondaries ($e^\pm$, $\gamma$). Most of the surviving electromagnetic traces are indeed $\mu^+\mu^-$, whose energy radiance is already suppressed by two orders of magnitude with respect to the primary GeV $p$ (nuclei) at the top of the atmosphere. Most of the relic energy is lost as heat and as kinetic energy spread by CR showering in air. A large fraction of the surviving CR trace is represented by the atmospheric neutrinos at a hundred MeV that exceed by 3$\div$4 orders of magnitude the corresponding MeV $\gamma$ component arriving at sea level, although in very special fine-tuned cases of EeV airshowers we can find a great electromagnetic component comparable to the $\nu$ one on the ground. In general the surviving atmospheric neutrino secondary tail exceeds by many orders of magnitude the corresponding electromagnetic component (mainly muons) while crossing the hadron barrier along the jet propagation.
To be more quantitative let’s recall the ratio between $\nu$ and the electromagnetic tail of atmospheric CR both on the ground and in deep kilometer-underground detectors as well as across the Earth (for neutrino event rates in different scenarios see i.e. [@2012ApJ...758....3F]). Atmospheric muons or $e^\pm$, $\mu^\pm$ from $\nu_{\mu, e}$, $\bar{\nu}_{\mu, e}$ are the observable electromagnetic traces in the last case: $\Phi_{\rm CR}/\Phi_\nu\simeq\Phi_{\rm CR}/\Phi_{\mu^+\mu^-}\gtrsim10^2$ on the ground; $\Phi_{\rm CR}/\Phi_{\mu^+\mu^-}\gtrsim10^8$ in underground detectors; $\Phi_{\rm CR}/\Phi_{\mu^+\mu^-}\gtrsim10^{14}$ in case of up-going signals [@AAFargion02; @2004ApJ...613.1285F]. The corresponding shields are namely 10 m w.e., 2 km w.e. and $10^5$ km w.e. In general the ratio between $\Phi_{\rm CR}/\Phi_\gamma$ is related to the ratio between the baryon barrier size $D_b$, the propagating lepton $\mu^+\mu^-$ distance $l_\mu$ and the interacting and propagating $\nu_\mu$, $\bar{\nu}_\mu\to\mu^+$, $\mu^-$. In summary, the ratio $\Phi_{\rm CR}/\Phi_{\mu^+\mu^-}$ is related to the surviving muons and the propagating distance: $\Phi_{\rm CR}/\Phi_{\mu^+\mu^-}\simeq e^{-D_b/l_\mu}$ and for largest baryon barrier ($D_b\gg12$ km) the muons arise by the appearance of high energy atmospheric neutrinos interacting with matter.
The lowest ratio (in first approximation) between a survived neutrino over a gamma average GRB radiance (assuming a dozen km size rock shell along the hadronic jet trajectory) maybe estimated assuming (as for IceCube) a primary prompt 30 TeV neutrinos whose most penetrating secondaries (the muons) escape as well after tens km rock they are shining outside the shell as muon first and later on as electron pairs and gamma: $\Phi_{\nu}/\Phi_{\mu^+\mu^-}\simeq l_{\nu}/l_\mu$, above ten thousand. In conclusion the minimal ratio of neutrino over gamma radiance should be around ten thousand and not one, if GRB are hadronic in primary nature.
Where is the gamma radiance lost?
---------------------------------
If in the hadronic GRB jet, a large fraction of the gamma output is lost in opaque shells, one may wonder that this is impossible because the energy conservation is lost. Indeed in the sun the radiation is both in photons and in neutrinos. Why should it not be the same in GRBs? The reason is that the solar photons are in late thermal equilibrium stage while GRB photons are out of equilibrium. Therefore where did the gamma energy fade (with respect to the neutrino one)? We believe that in any hadronic GRB (outside of a fine-tuned case where the external shell are just transparent ad hoc) a large part of the gamma energy should be absorbed by the baryon matter while being scattered and-or being absorbed, accelerating the shell masses in the form of kinetic shells. The explosive kinetic shell masses as well as part of the survived cosmic rays (escaped along the jet) might contain the primary hadron and gamma energy, while the neutrino component (born at the same inner sources) will suffer a negligible depletion, surviving with higher energy fluency. In conclusion, once again, neutrino radiance should be much larger than gamma one in the most general hadronic jet model crossing star shells. However, the data show a comparable or a minor neutrino radiance with respect to the gamma ones. This is the main need for a pure electronic jet in GRBs.
Binary BH-NS feeding accretion disk and powering $\gamma$ jet {#sec:BH-NS}
=============================================================
In the light of this absence of GRB-$\nu$ (\[sec:lack\]), we are forced to consider a new engine process able to avoid any pion decay chain. The most natural one is a binary system in empty space made by a neutron star (NS) and a black hole (BH) in an elliptical trajectory with each other. At a nearby encounter, as depicted in Fig. \[fig:01\], the NS may suddenly lose a fragment of its mass because of tidal forces at relativistic Roche limit [@1973ApJ...185...43F]. These neutrons are led within tens of minutes toward the last extended boundary ($r\sim3R_{\textrm{Schwarzschild}}$) of the BH while the decay $n\to p+e^-+\bar{\nu}_e$ takes place. The electrons will then escape at low MeV energy, leading to a poor spherical (hard to detect) signal, while the protons which don’t gain too much energy, nor relevant momentum in the decay, will proceed in its geodetic spiraling in a disk-ring around the BH; the ring will be therefore a positive charged ring. The almost relativistic electrons in the meantime will spread themselves in a nearly spherical fashion. The neutron-proton coherent spiraling around the BH will then define a net positive charged current in a ring that is not compensated by a relativistic electronic component of the decay. This induces a huge axial magnetic field $\mathbf{B}_p$ proton-induced which is represented in Fig. \[fig:02\]; the magnetic lines force the electrons to concentrate themselves toward the BH accretion disk’s poles (let’s call them north and south according to the magnetic field polarity). The electrons will then be forced and squeezed by a powerful charged pump that accelerate the $e^-$ in a jet at highest energies well above the starting MeV ones. Within such a dense relativistic electron beam flow, because of self-electron Compton scattering, inverse Compton and pair production, collinear pairs $e^+e^-$ and $\gamma$ will arise resulting in a final $\gamma$ jet. The BH spin and the ring spin will interact and precess among themselves.
![Unstable NS explodes in a spherical SN-like event, observable days or weeks after first GRB blaze. Shells of energy of the supernova embrace the same BH jet. The asymmetric binary BH is suddenly without a companion and it is launched tangentially with a high speed kick (see [@2007ARep...51..308B]) in a fast flight holding alive its ring and its jet. The latest stages of the BH fed jet may shine as a SGR. The model NS-BH maybe dressed in a similar NS-NS evolution where the final relic is a spinning NS jet; this version may fit the SGRs or AX PSRs relics observed in our own galaxy.](f4.eps)
\[fig:04\]
In the proton disk, meanwhile, for the accumulated charged asymmetry, some of the external circuiting protons will start to escape at equatorial disk edges (see Fig. \[fig:02\]). Clearly, the extreme collimation of the pairs $e^+e^-$ and $\gamma$ avoids the Eddington opacity that normally occurs for spherical luminosity and the huge dense NS mass feeding the proton ring represents a very powerful engine ($\dot{m}_{\rm NS}\simeq10^{-6}\div10^{-5}\,\mathrm{M}_\odot\,\mathrm{s}^{-1}$). This mass loss, then, powers the BH accretion disk and the jet, whose blazing toward the Earth is perceived as a GRB. After a few days or months the NS is doomed; its strip for the benefit of the BH ring may lead to instabilities (see Fig. \[fig:03\]) and the reason for that is simple: a very minimal NS mass $(m^{\rm NS}_{min}\lesssim0.2\,\mathrm{M}_\odot)$ may become too light to hold together nuclei (see ) and its surface gravity weight becomes unable to compensate the nuclear chemical repulsion potential (as happens in a normal NS). Neutrons from the surface would then start to decay and escape making the degenerate system totally unstable in a matter of tens of seconds or few minutes (Fig. \[fig:03\]). This would lead to a sudden spherical explosion appearing from Earth as a SN event (Fig. \[fig:03b\]). However, it is not trivial to tell if the critical minimal neutron star mass could release much more or much less energy than of a canonical SN. The energy potential budget for a NS collapsing in a normal SN accounts for around 10% of the object rest mass ($\sim\mathrm{M}_\odot$). Therefore, an apparent SN-like event like the one celebrated SN-GRB related to GRB 980425 may be attributed to such a simple process of a minimal NS explosion without any correlated beamed neutrino and with a few days (or a week, [@2007ApJ...658L...5M]) delay with respect to the main GRB blaze. Naturally, the shining of the spherical NS explosion may heat and excite the external surrounding (original SN shell from where NS itself or BH is formed) shell leading to spectroscopic emission and absorption lines that may mimic the SN explosion. On the contrary, the Ni and/or the Co radioactive decay mode are not naturally born (therefore there might be a remarkable imprint to be discussed elsewhere that might distinguish the SN from the NS-like explosions).
We like to stress that this electromagnetic pump accelerator mechanism does not require any hadron parental engine, any consequent muons or energetic neutrinos, explaining the the observed absence of ICECUBE neutrino radiance (larger than the photon one) and-or the missing GRB-$\nu$ correlation.
Bimodal Short and long GRB {#sec:slGRB}
--------------------------
There are also natural corollary consequences of the proposed model: we can find a similar tale for a NS-SN binary collapse where one of the two NS “eats” and “strips” matter from the companion NS leading to a similar story-board. Because such a NS-NS binary systems are among the narrow ones then we imagine that also their characteristic blazing times are sharper leading to more short duration GRBs. Therefore these shorter GRBs may populate the short events, whose duration is below 2 s.
Larger sized BH-NS binaries, like the very recent candidate in the LIGO-VIRGO gravitational wave detection [@PhysRevLett.116.061102], system may imply a wider family of NS-BH with BH masses as large as $10\div100\div1000\,\mathrm{M}_\odot$. These are possibly the longer duration GRBs whose characteristic time is longer than 2 s. The infrequent and sporadic presence of largest BH makes rarer and rarer the longest GRB events explaining the rarest long life GRB , thousand second long. Also late GRBs whose early explosion has not been in axis but whose late precessing jet is pointing (as a young SGRs) to us at a still high output, may appear as a short GRBs mostly at nearer cosmic distances (respect peaked GRB luminosity).
Conclusions {#sec:end}
===========
If the SGRB and LGRB are explained by NS-NS (SGRB) and NS-BH (LGRB) models, then the main puzzle of the apparent over-Eddington luminosity is simply solved by high collimated beaming. The tidal ring-jet perturbation and the spinning of the BH versus the disk makes the jet spin and precessing as well as blaze in the observed almost chaotic way (see Fig. \[fig:04\]). The absence of longest events, almost comparable with largest optically violent variable quasar 3C 279 gamma flare is simply related with the rarity of super-massive BH (as the AGNs) respect lighter tens-of-hundreds or thousands solar masses. The coexistence of a SN-like event (for a quick review see i.e. ) is solved by light tidal NS sudden evaporation and consequent explosion. The absence of TeV neutrinos correlated with GRBs is guaranteed by the absence of any hadronic accelerator as well as leptonic neutrino tails in GRB. The thinner precessing jet moreover still explains the statistics we see, i.e. in Fig. \[fig:06\].
The model consistence is based on the geometrical evolution of a thin persistent jet whose acceptance today, after twenty years, is becoming more and more obvious. We admit that for a long time we also assumed that such thin jets were powered by hadronic engine (muons) and later on by their electron pairs fed by muons, [@2009astro2010S..75F; @2010MmSAI..81..440F; @2012MmSAI..83..312F], but the absence of a $\nu$-$\gamma$ correlation and in particular the paucity of $\Phi_{\rm GRB}^\nu$ with respect to $\Phi_{\rm GRB}^\gamma$ forced us to the present “neutron striptease” jet-SN model, made by pure electron jets, mostly or totally free of hadronic engines.
We are in debt to Prof. Maxim Khlopov for deep and useful comments, to Prof. B. Mele, Prof. P. Lipari, Prof. G. Salmé, Prof. R. Jatzen for their discussions and comments. This paper is dedicated to the memory of Ameglio Fargion, born on 2 June 1913, died on 22 April 2011.
Fishbone, L. G. 1973, , 185, 43
Cavallo, G., Rees, M. J. 1978 , 183, 359–365
Goodman, J. 1986 , 308, L47–L50
Paczynski, B. 1986 , 308, L43–L46
Rees, M. J., & M[é]{}sz[á]{}ros, P. 1992, , 258, 41P
Paczynski, B., & Rhoads, J. E. 1993, , 418, L5
Rees, M. J., & M[é]{}sz[á]{}ros, P. 1994, , 430, L93
Fargion, D. 1994, The Dark Side of the Universe, World Scientific, 88-97
Fargion, D., & Salis, A. 1995, Nuclear Physics B Proceedings Supplements, 43, 269 - 273
Waxman, E. 1997, , 489, L33
Sari, R. 1997, , 489, L37
Vietri, M. 1997, , 478, L9
Piro, L., & BeppoSAX Team 1997, The Transparent Universe, 382, 179
Feroci, M., Frontera, F., Costa, E., et al. 1997, , 3114, 186
Fargion, D., Konoplich, R. V., & Salis, A. 1997, Zeitschrift fur Physik C Particles and Fields, 74, 571
Fargion, D., & Salis, A. 1998, Physics Uspekhi, 41, 823
Iwamoto, K., Mazzali, P. A., Nomoto, K., et al. 1998, , 395, 672
Feroci, M., Antonelli, L. A., Guainazzi, M., et al. 1998, , 332, L29
Sumiyoshi, K., Yamada, S., Suzuki, H., & Hillebrandt, W. 1998, , 334, 159
Waxman, E., & Bahcall, J. 1999, , 59, 023002
Sari, R., & Piran, T. 1999, , 517, L109
Cen, R. 1999, , 517, L113
Fargion, D. 1999, , 138, 507, First published as INFN.1215;13.07.98 preprint
M[é]{}sz[á]{}ros, P. 2000, Gamma-ray Bursts, 5th Huntsville Symposium, 526, 514
Eichler, D., & Levinson, A. 2000, , 529, 146
Fargion, D. 2000, 2nd Workshop on Gamma Ray Bursts in the Afterglow Era Rome, Italy, 315–317
Fargion, D. 2001, Frontier Objects in Astrophysics and Particle Physics, 73 eds. Giovannelli, F. et al., 347–360
Fargion, D. 2002, , 570, 909
Fargion, D., De Sanctis Lucentini, P. G., De Santis, M., & Grossi, M. 2004, , 613, 1285
Fargion, D., & Grossi, M. 2005, Nuovo Cimento C Geophysics Space Physics C, 28, 809
Fargion, D., & Grossi, M. 2006, Chinese Journal of Astronomy and Astrophysics Supplement, 6, 342
Fargion, D., Lanciano, O., & Oliva, P. 2006, 4th Workshop on Science with the New Generation of High Energy Gamma-Ray Experiments (SciNeGHE 2006) Portoferraio, Isola d’Elba, Italy
Woosley, S. E., & Bloom, J. S. 2006, , 44, 507
Gal-Yam, A., Fox, D. B., Price, P. A., et al. 2006, , 444, 1053
Maeda, K., Kawabata, K., Tanaka, M., et al. 2007, , 658, L5
Bogomazov, A. I., Lipunov, V. M., & Tutukov, A. V. 2007, Astronomy Reports, 51, 308
Fargion, D., D’Armiento, D., Oliva, P., & Manniti, F. 2009, Frontier Objects in Astrophysics and Particle Physics, 351
Fargion, D., & D’Armiento, D. 2009, astro2010: The Astronomy and Astrophysics Decadal Survey, 2010,
Fargion, D., & D’Armiento, D. 2010, , 81, 440
Pierre Auger Collaboration, Abreu, P., Aglietta, M., et al. 2012, Astroparticle Physics, 35, 354
Fargion, D., D’Armiento, D., Desiati, P., & Paggi, P. 2012, , 758, 3
Fargion, D. 2012, , 83, 312
Bersier, D. 2012, arXiv:1206.6979
Abbasi, R., Abdou, Y., Abu-Zayyad, T., et al. 2012, , 484, 351
Melandri, A., Pian, E., D’Elia, V., et al. 2014, , 567, A29
Fargion, D. 2014, Frascati Phys. Ser. 58, 263
Fargion, D., Oliva, P., & Ucci, G. 2015, Proceedings of Science 26-31-May-2014, 028
Abbott, B. P. et al. 2016, , 116, 061102
IceCube Collaboration, Aartsen, M. G., Abraham, K., et al. 2016, arXiv:1601.06484
Aartsen, M. G., Abraham, K., Ackermann, M., et al. 2016, , 824, 115
Swift: <http://www.astronomerstelegram.org/?read=9294>, Fermi: <http://www.astronomerstelegram.org/?read=9303> H.E.S.S.: <http://www.astronomerstelegram.org/?read=9301>, AGILE: <http://www.astronomerstelegram.org/?read=9295>
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Noam D. Elkies and Matthias Schütt'
title: Modular forms and K3 surfaces
---
**MSC(2000):** 14J28; 11F11, 11F23, 11G40, 14G10.
Introduction {#s:intro}
============
The question of modularity for algebraic varieties over ${\mathbb{Q}_{}}$ has been studied in great detail in recent years. Historically, it began with an observation by Eichler who noticed the modularity of an elliptic curve over ${\mathbb{Q}_{}}$ of conductor $11$. Shimura then proved that every Hecke eigenform of weight $2$ is associated to an abelian variety over ${\mathbb{Q}_{}}$. In the case of rational eigenvalues, the corresponding variety is an elliptic curve.
Conversely, the Tanyiama–Shimura–Weil conjecture stated that every elliptic curve over ${\mathbb{Q}_{}}$ is modular. The celebrated proof of this conjecture by Wiles et al. [@Wi; @TW; @BCDT] not only implies Fermat’s Last Theorem, but also catalyzed many further developments in this area, notably the recent great progress on Serre’s conjecture.
By now, we know modularity for several classes of varieties. For instance, Dieulefait and Manoharmayum [@DM] prove modularity of rigid Calabi–Yau threefolds over ${\mathbb{Q}_{}}$ under a mild condition. There is also work by Livné [@L] on modularity for two-dimensional motives with complex multiplication (CM) that we use here, citing it as Theorem \[Thm:mod\].
On the other hand, the problem of geometric realizations is harder for Hecke eigenforms of weight greater than two. Deligne [@Deligne] gives a geometric construction of $\ell$-adic Galois representations for Hecke eigenforms. However, the varieties involved vary greatly with the level. In this sense, his construction is not as uniform as one might wish (cf. Remark \[Rem:EMS\]).
This paper gives the first case of higher weight where we can realize all known Hecke eigenforms with rational eigenvalues in a single class of varieties:
\[thm\] Assume the extended Riemann Hypothesis (ERH) for odd real Dirichlet characters. Then every Hecke eigenform of weight $3$ with rational eigenvalues is associated to a K3 surface over ${\mathbb{Q}_{}}$.
This result answers a question asked independently by Mazur and van Straten. It builds on the classification of CM forms with rational coefficients by the second author which we recall in section \[s:CM\]. That section also explains the dependence of Theorem \[thm\] on the ERH. Section \[s:K3\] recalls the notion of singular K3 surfaces and Livné’s modularity result. We review the relevant known examples and obstructions in sections \[s:Ex\] and \[s:obs\]. Our main technique for proving Theorem \[thm\] is constructing one-dimensional families of K3 surfaces and searching for singular specializations over ${\mathbb{Q}_{}}$. This is explained in section \[s:tech\] and exhibited in detail for one particular family in section \[s:fam\]. The paper concludes with the remaining surfaces needed to prove Theorem \[thm\].
Singular K3 surfaces {#s:K3}
====================
A K3 surface is a smooth, projective, simply connected surface $X$ with trivial canonical bundle $\omega_X={\mathcal{O}}_X$. The most prominent examples are smooth quartics in ${\mathbb{P}}^3$ and Kummer surfaces. Later we will work with elliptic K3 surfaces.
Throughout this paper, modularity will refer to classical modular forms (cf. sect. \[s:CM\]). This classical kind of modularity is a very special property of a variety; a general K3 surface over ${\mathbb{Q}_{}}$ cannot be modular for several reasons (cf. the discussion before Theorem \[Thm:mod\]), though the Langlands Program predicts a correspondence with some automorphic forms.
K3 surfaces and their moduli have been studied in great detail. We will come back to these questions in section \[s:tech\]. The only complex K3 surfaces that can be classically modular are those that have no moduli at all. In terms of the Picard number $\rho(X) = \text{rk}{\mathop{\rm NS}\nolimits}(X)$, the condition that $X$ have no moduli is that $$\rho(X) = 20,$$ the maximum in characteristic zero. K3 surfaces with Picard number $20$ are often referred to as *singular K3 surfaces*. The terminology is reflected in the structure (cf. sect. \[s:Ex\]) which relates any singular K3 surface to a product of two isogenous elliptic curves with complex multiplication (CM), thus with *singular moduli*.
Our results will often be stated in terms of the *discriminant* $d=d(X)$ of a singular K3 surface $X$, i.e. the discriminant of the intersection form on the lattice, which is the group endowed with the cup-product pairing: $$d=d(X)={\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(X)).$$
\[Ex:Fermat\] The Fermat quartic in ${\mathbb{P}}^3$, $$S=\{(x_0 : x_1 : x_2 : x_3) \in {\mathbb{P}}^3 \mid x_0^4+x_1^4+x_2^4+x_3^4=0\}$$ is a singular K3 surface. There are $48$ obvious lines on $S$, all defined over the 8th cyclotomic field ${\mathbb{Q}_{}}(e^{2\pi i/8})$. Pjateckiĭ-Šapiro and Šafarevič [@PSS] proved that the lines generate ${\mathop{\rm NS}\nolimits}(S)$, with ${\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(S)) = d = -64$. However, the proof depended on a claim by Demjanenko whose argument turned out to be incorrect. The proof was later independently completed by Cassels [@Cassels] and Inose [@Inose-Kummer].
For any smooth projective surface $X$ over ${\mathbb{C}^{}}$, we define the *transcendental lattice* $T(X)$ as the following sublattice of $H^2(X,{\mathbb{Z}})$: $$T(X)={\mathop{\rm NS}\nolimits}(X)^\bot \subset H^2(X,{\mathbb{Z}}).$$ When $X$ is a K3 surface, $H^2(X,{\mathbb{Z}})$ has rank $22$ and signature $(3,19)$. Hence $T(X)$ has rank $22-\rho(X)$ and signature $(2,20-\rho(X))$.
If $X$ is a singular K3 surface, then $T(X)$ has rank $2$, and is even, positive-definite and equipped with an orientation [@SI §4]. Using the intersection form, we will identify the transcendental lattice with a $2\times 2$ matrix $$T(X) \leftrightarrow \begin{pmatrix} 2a & b\\
b & 2c\end{pmatrix}$$ with integral coefficients and discriminant $d=b^2-4ac<0$. Applying an ${\mathop{\rm SL}\nolimits}_2({\mathbb{Z}})$ change of basis, we will always use a reduced representative with $-a < b\leq a\leq c$. We will also use the shorthand $[2a,b,2c]$ for $T(X)$.
In consequence, if the singular K3 surface $X$ is defined over some number field $L$, then $T(X)$ gives rise to a two-dimensional Galois representation $\varrho$ over $L$. Over some extension of the ground field, Shioda and Inose associated a Hecke character to this Galois representation [@SI]. Livné then proved modularity over ${\mathbb{Q}_{}}$ as an application of a more general result concerning motives:
\[Thm:mod\] Every singular K3 surface over ${\mathbb{Q}_{}}$ is modular: there is a Hecke eigenform with system of eigenvalues $\{c_p\}$, such that $$\text{trace $\varrho($Frob}_p) = c_p \;\;\; \text{for almost all } p.$$
The corresponding Hecke eigenform $f$ has weight $3$ and CM by the imaginary quadratic field $K={\mathbb{Q}_{}}(\sqrt{d})$, where $d={\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(X))<0$ is the discriminant of $X$. The Hecke eigenvalues of $f$ are all rational integers.
For instance, the Hecke eigenform corresponding to the Fermat quartic has CM by ${\mathbb{Q}_{}}(\sqrt{-1})$. For the model $S$ in Example \[Ex:Fermat\], the newform has level $16$, as given in [@S-CM Table 1].
CM newforms {#s:CM}
===========
Modular forms for congruence subgroups come with weight, level and nebentypus character that appear in the transformation law. For fixed invariants, the modular forms constitute a module over the Hecke algebra. On cusp forms, the Hecke operators can be diagonalized simultaneously. The set of eigenvalues as in Theorem \[Thm:mod\] determines a unique primitive normalized eigenform, a so-called *newform*.
By the transformation law, a modular form $f$ always has a Fourier expansion $$f=f(\tau)=\sum_{n\geq 0} a_n q^n,\;\;\;q=e^{2\pi i \tau}.$$ Here a newform is characterized by the property that the Mellin transform $$L(f,s) = \sum_{n\geq 0} a_n n^{-s}$$ has an Euler product. In particular, the Fourier coefficients $a_p$ are multiplicative and equal the Hecke eigenvalues $c_p$. Throughout this paper, we will use the terms Fourier coefficients and Hecke eigenvalues for a newform interchangeably.
We are interested in newforms of weight $3$ with rational eigenvalues. Generally, odd weight is special in the sense that the nebentypus character is necessarily nontrivial. In fact, by a result of Ribet [@R], a newform $f$ of odd weight with real Hecke eigenvalues has CM by its nebentypus character. In particular, $f$ comes from a Hecke character.
CM newforms with rational coefficients have been classified by the second author in [@S-CM]. The analysis of the associated Hecke characters revealed the following structure:
\[Thm:CM\] For fixed weight $k+1$, there is a bijective correspondence $$\left\{\begin{matrix} \text{CM newforms of weight}\; k+1\\
\text{with rational eigenvalues}\\ \text{up to twisting}
\end{matrix}\right\}
\stackrel{1:1}{\longleftrightarrow}\left\{\begin{matrix}
\text{Imaginary-quadratic fields $K$}\\\text{with class group }
Cl(K)\subseteq ({\mathbb{Z}}/k{\mathbb{Z}})^g\\ \text{for some}\;\, g\in{\mathbb{N}_{}}\end{matrix}\right\}$$
Unless $K={\mathbb{Q}_{}}(\sqrt{-1})$ or ${\mathbb{Q}_{}}(\sqrt{-3})$, twisting refers to modifying the Fourier coefficients by a quadratic Dirichlet character $\chi$ (since otherwise the rationality of Fourier coefficients is not preserved): $$\begin{aligned}
\label{eq:twist}
f\otimes \chi = \sum_{n\geq 1} a_n \chi(n) q^n.\end{aligned}$$ Any twist of a newform is again a newform, although the level and the nebentypus character will differ in general. If $K={\mathbb{Q}_{}}(\sqrt{-1})$ or ${\mathbb{Q}_{}}(\sqrt{-3})$, then we can also twist the associated Hecke character by a biquadratic resp. cubic character. For geometric equivalents of twisting, see (\[eq:Kummer\]) and the following explanations.
For fixed weight, we will refer to the Hecke eigenforms from Theorem \[Thm:CM\] as newforms of class number $h(K)$. In this terminology, it should be understood that the newforms have rational Hecke eigenvalues.
Conjecturally, the number of imaginary quadratic fields on the right hand side in Theorem \[Thm:CM\] is always finite. The case of $k=1$ was proven by Heilbronn and made explicit by Heilbronn and Linfoot [@Heil]. The exponents $k=2, 3$ are due to Weinberger [@Wb] (the latter also proven by Boyd and Kisilevsky [@BK]). Recently, Heath-Brown proved finiteness for $k=5$, $2^n$, or $3\cdot 2^n$ [@H-B].
Most of these results are not effective. However, for $k=2$ (and $k=1$), our knowledge goes much further. There are 65 imaginary quadratic fields whose class groups are at most two-torsion. Their class numbers go up to $16$ for $K={\mathbb{Q}_{}}(\sqrt{-1365})$. Many of these fields were already known to Euler through his search for idoneal numbers. We list them by discriminants $d_K$ and class numbers $h(d_K)$:
$$\begin{array}{c|l}
\hline
h(d_K) & d_K\\[0.2ex]
\hline
\hline
1 & -3, -4, -7, -8, -11, -19, -43, -67, -163\\
\hline
2 & -15, -20, -24, -35, -40, -51, -52, -88, -91,\\
& -115, -123, -148, -187, -232, -267, -403, -427\\
\hline
4 & -84, -120, -132, -168, -195, -228, -280, -312,\\
& -340, -372, -408, -435, -483, -520, -532, -555,\\
& -595, -627, -708, -715, -760, -795, -1012, -1435\\
\hline
8 & -420, -660, -840, -1092, -1155, -1320, -1380,\\
& -1428, -1540, -1848, -1995, -3003, -3315\\
\hline
16 & -5460\\
\hline
\end{array}$$
In [@Wb], Weinberger proved that there is at most one further imaginary quadratic field with class group exponent $2$. Assuming the absence of Siegel–Landau zeros for the of odd real Dirichlet characters, the known list is complete. This condition would be a consequence of the extended Riemann hypothesis (ERH) for odd real Dirichlet characters. Thus the (stronger) assumption in Theorem \[thm\].
This paper solves the geometric realization problem in weight 3. By Theorem \[Thm:CM\], our primary task is to find a singular K3 surface over ${\mathbb{Q}_{}}$ for each of the 65 newforms (up to twisting) corresponding to imaginary quadratic fields with class group exponent $2$. As we will work with elliptic surfaces, we need not worry about twisting (cf. the discussion around (\[eq:twist\]) and (\[eq:Kummer\])).
\[Rem:EMS\] There is a canonical way to realize newforms of weight 3, via the elliptic modular surface (or universal elliptic curve) for the corresponding congruence subgroup $\Gamma_1(N)$ [@ShEMS]. However, this correspondence is not uniform in several respects. First, the complex surfaces involved vary greatly with $N$. Even if one considers a newform and a twist as in (\[eq:twist\]), the corresponding surfaces will in general not be . Moreover, the surfaces will in general have more than one associated newform (i.e. $p_g = h^{2,0} > 1$).
There are two classes of surfaces which might come to mind first when thinking of modular forms: abelian surfaces and K3 surfaces — both generalizations of elliptic curves to dimension two. They will be discussed in the next section.
Singular abelian surfaces and Kummer surfaces {#s:Ex}
=============================================
Shioda and Inose derived a canonical way to produce a singular K3 surface of given isomorphism class (i.e. of given transcendental lattice) [@SI]. The main object in their construction is the Kummer surface Km$(A)$ of the abelian surface $A$ with the given transcendental lattice. Hence we shall briefly discuss singular abelian surfaces over ${\mathbb{Q}_{}}$.
A complex abelian surface $A$ has $H^2(A, {\mathbb{Z}})$ of rank $6$. The surface is called singular if $\rho(A)=4$, that is, if its transcendental lattice $T(A)$ has rank $2$. By a result of Shioda and Mitani [@SM], every singular abelian surface is isomorphic to a product of two isogenous elliptic curves $E, E'$ with CM. These can be given explicitly in terms of the transcendental lattice $T(A)$.
If $E$ and $E'$ are defined over ${\mathbb{Q}_{}}$, we deduce that the product $A=E\times E'$ realizes a newform of weight 3 with CM by the corresponding imaginary quadratic field of class number $1$.
For singular abelian surfaces over ${\mathbb{Q}_{}}$, there is one further possible construction: If $K$ has class number $2$, consider an elliptic curve $E$ with CM by the ring of integers ${\mathcal{O}}_K$ in $K$. Then $E$ is defined over a quadratic extension of ${\mathbb{Q}_{}}$. Let $A$ be the Weil restriction of $E$ to ${\mathbb{Q}_{}}$. By [@SM], $A$ is a singular abelian surface with transcendental lattice corresponding to the non-principal class in $Cl(K)$. Since $E$ is associated to some Hecke character, $A$ geometrically realizes a newform of weight 3 with CM by $K$.
It follows that we can realize all newforms of weight 3 and class number $1$ or $2$ in singular abelian surfaces over ${\mathbb{Q}_{}}$. We cannot pursue this approach any further because there is an obstruction coming from the cohomological structure of singular abelian surfaces:
\[Lem:ab\] Let $A$ be a singular abelian surface of discriminant $d$. Assume that $A$ is defined over a number field $L$. Let $H(d)$ denote the ring class field. Then $$[L.H(d):L] \mid 4.$$
*Proof:* Consider the compatible system of $\lambda$-adic Galois representations on $H^1(A)$. By [@SM], $A$ is isomorphic over some extension of $L$ to the product of two isogenous CM elliptic curves. The eigenvalues of Frobenius on $H^1(A)$ therefore lie in the ring class field $H(d)$. By class field theory, the extension $(H(d).L)/L$ has degree dividing the dimension of the Galois representation on $H^1(A)$, which is $4$. [$\Box$]{}
By the lemma, we can only define singular abelian surfaces of class number $1$ and $2$ over ${\mathbb{Q}_{}}$. In order to realize the newforms of weight 3 and greater class number geometrically, we therefore turn to singular K3 surfaces. This has two advantages: On the one hand, the above constructions for class number $1$ and $2$ carry over to the corresponding Kummer surfaces (cf. Lemma \[Lem:Kummer\]). On the other, the descent obstructions are much milder for singular K3 surfaces. This will be studied in the next section.
We now turn to Kummer surfaces. Here we consider the quotient of an abelian surface $A$ by its involution $-1$. The minimal resolution of the resulting 16 singularities of type $A_1$ is a K3 surface. One multiplies the intersection form on $T(A)$ by $2$ to obtain $T(\text{Km}(X))=T(A)[2]$. For example, the Fermat surface (Example \[Ex:Fermat\]) is ${{\,\overline{\!{\mathbb{Q}_{}}\!}\,}}$-isomorphic to the Kummer surface of the product of elliptic curves with CM by ${\mathbb{Z}}[\sqrt{-1}]$ and ${\mathbb{Z}}[2\sqrt{-1}]$.
\[Lem:Kummer\] Any newform of weight 3 and class number $1$ or $2$ is associated to a Kummer surface over ${\mathbb{Q}_{}}$.
*Proof:* The Kummer quotient is defined over the ground field of the abelian surface (cf. (\[eq:Kum\])). For each imaginary quadratic field $K$ of class number $1$ or $2$, we have found a singular abelian surface over ${\mathbb{Q}_{}}$. Therefore we obtain a Kummer surface over ${\mathbb{Q}_{}}$ for some newform with CM by $K$. We now address the issue of twisting.
We will use elliptic fibrations. Let $E, E'$ be elliptic curves. Denote the $j$-invariants by $j, j'$. Inose exhibited an explicit elliptic fibration on Km$(E\times E')$ in [@Inose]. By [@S-CM proof of Prop. 4.1], this elliptic fibration can be defined over $L={\mathbb{Q}_{}}(j+j', j\cdot j')$: $$\begin{aligned}
\label{eq:Kum}
\text{Km}(E\times E'):\;\;\; y^2 = x^3 + B(t)\,x+C(t),\;\;\; B, C\in L[t].\end{aligned}$$
In the present situation, the abelian surface $A$ that we start with is isomorphic to the product of two isogenous CM elliptic curves of class number $1$ or $2$. Then $L={\mathbb{Q}_{}}$ because $j,j'$ are equal for class number $1$ and quadratic conjugate for class number $2$. Let $f$ denote the associated newform. Then the ${\mathbb{Q}_{}}(\sqrt{d})$-isomorphic fibration $$\begin{aligned}
\label{eq:Kummer}
\text{Km}(A):\;\; d\,y^2 = x^3 + B(t) x + C(t), \;\;\; d\in{\mathbb{Q}_{}}^*\end{aligned}$$ realizes the twist of $f$ by the quadratic Dirichlet character $\left(\frac d{\cdot}\right)$ as in (\[eq:twist\]). This leaves cubic and biquadratic twists in case of $K={\mathbb{Q}_{}}(\sqrt{-3}), {\mathbb{Q}_{}}(\sqrt{-1})$. Here we use the fact that for the elliptic curves with $j$-invariant, $j=0$ resp. $1728$, the Kummer surface admits an automorphism of order $6$ resp. $4$. In fact, in the first case, Inose’s elliptic fibration (\[eq:Kum\]) is isotrivial ($B\equiv 0$). Hence every fiber admits an automorphism of order $6$, so we can apply cubic twists fiberwise. In the second case we could also argue directly with the Fermat quartic (Example \[Ex:Fermat\]), since it is Kummer. [$\Box$]{}
By Lemma \[Lem:Kummer\], we are left to find singular K3 surfaces over ${\mathbb{Q}_{}}$ for all newforms of weight 3 with class number 4, 8 or 16. Up to twisting, there are 38 such forms.
Obstructions for singular K3 surfaces {#s:obs}
=====================================
Before continuing our search for singular K3 surfaces over ${\mathbb{Q}_{}}$, we discuss obstructions to the field of definition. These are much milder than for singular abelian surfaces (cf. Lemma \[Lem:ab\]), notably because $H^1(X)$ is trivial for a K3 surface $X$. Moreover, not every singular K3 surface is Kummer although the relation with Kummer surfaces is very close.
Shioda and Inose showed in [@SI] how to produce a singular K3 surface with given transcendental lattice. Every singular K3 surface $X$ admits a Nikulin involution such that the quotient is Kummer. The resulting picture is often referred to as * structure*. We sketch it in the following figure. Here $A$ and $X$ are chosen with the same transcendental lattice $T(X)=T(A)$, and $T($Km$(A))=T(A)[2]$. $$\begin{array}{ccccc}
A &&&& X\\
& \searrow && \swarrow &\\
&& \text{Km}(A) &&
\end{array}$$
The construction is exhibited over some finite extension of the field of definition of $A$; the elliptic fibration (\[eq:Kum\]) on the Kummer surface is a base change from $X$. Hence the question is when we can descend $X$ to ${\mathbb{Q}_{}}$. We now discuss the known obstructions.
The first obstruction comes from lattice theory. It involves the genus of a lattice and was first studied by Shimada [@Shimada] in the case of fundamental discriminant. The second author then proved the general case in [@S-fields].
\[Thm:genus\] Let $X$ be a singular K3 surface over some number field. Let $d$ denote the discriminant of $X$ and $K={\mathbb{Q}_{}}(\sqrt{d})$. Then $$\{T(X^\sigma) \mid \sigma\in{\mathop{\rm Aut}\nolimits}({\mathbb{C}^{}}/K)\} = \text{genus of } T(X).$$
If $X$ is defined over ${\mathbb{Q}_{}}$, the genus of $T(X)$ consists of a single class. This implies that $Cl(K)$ is at most two-torsion. Independently this follows from Theorems \[Thm:mod\] and \[Thm:CM\].
The second obstruction is related to class field theory. It essentially says that even if a singular K3 surface descends to some number field $L$, it still carries the structure of the ring class field $H(d)$ through the Galois action on the group. This property was first noted by the first author in [@Elkies]. The second author gave an alternative proof in [@S-NS].
\[Thm:NS\] Let $X$ be a singular K3 surface of discriminant $d$. Let $H(d)$ be the ring class field for $d$. Let $L$ be a number field such that ${\mathop{\rm NS}\nolimits}(X)$ is generated by divisors over $L$. Then $H(d)\subseteq L(\sqrt{d})$.
In consequence, as $h(d)$ increases, singular K3 surfaces become more and more complicated. In particular the number of singular K3 surfaces over ${\mathbb{Q}_{}}$ up to ${\mathbb{C}^{}}$-isomorphism is finite. This result is originally due to Šafarevič [@Shafa]. Hence, in our attempt to find singular K3 surfaces for all newforms of weight three with rational Fourier coefficients, we are searching within a finite set.
We now resume our search for singular K3 surfaces over ${\mathbb{Q}_{}}$ for the newforms of weight 3 and class numbers 4, 8 and 16. Here we mention another class of singular K3 surfaces which has been investigated before: *extremal elliptic K3 surfaces*. These are singular K3 surfaces admitting an elliptic fibration with finite group of sections. Up to the torsion sections, such surfaces are determined by the configuration of singular fibers. Extremal elliptic K3 fibrations were classified by Shimada and Zhang in [@SZ]. They are finite in number.
Many explicit defining equations have been obtained by the second author in [@S-Rocky] and by Beukers and Montanus in [@BM]. In addition to previous newforms of class number $1$ and $2$, they realize ten discriminants of class number $4$ and $8$. The next table lists the discriminants and one possible configuration of singular fibers and group for each newform. In each case the surface is semistable, that is, all reducible fibers are of type $I_n$ for some $n\geq 1$; so we simplify the notation by listing only the indices $n$. For completeness, we also give the transcendental lattice $T_X$ in the shorthand notation $[2a,b,2c]$ for the intersection form.
$$\begin{array}{|c|c|c|c|}
\hline
\text{discriminant} & \text{configuration} & MW & T_X\\
\hline
\hline
-84 & [1,2,2,2,3,14] & {\mathbb{Z}}/2{\mathbb{Z}}& [2,0,42]\\
\hline
-120 & [1,2,2,2,5,12] & {\mathbb{Z}}/2{\mathbb{Z}}& [6,0,20]\\
\hline
-4\cdot 132 & [1,1,3,4,4,11] & (0) & [24, 12, 28]\\
\hline
-168 & [1,1,1,3,4,14] & (0) & [4, 0, 42]\\
\hline
-195 & [1,1,1,3,5,13] & (0) & [6, 3, 34]\\
\hline
-280 & [1,1,1,4,7,10] & (0) & [2, 0, 140]\\
\hline
-312 & [1,1,2,3,4,13] & (0) & [6, 0, 52]\\
\hline
\hline
-4\cdot 420 & [1,3,4,4,5,7] & (0) & [24, 12, 76]\\
\hline
-660 & [1,2,2,3,5,11] & (0) & [2, 0, 330]\\
\hline
-840 & [1,1,4,5,6,7] & (0) & [12, 0, 70]\\
\hline
\end{array}$$
By Theorem \[Thm:NS\], the Galois group of the ring class field $H(d)$ acts nontrivially on the group ${\mathop{\rm NS}\nolimits}(X)$, i.e. on the reducible fibers. A nontrivial action is possible when a fiber has type $I_n$ for $n\geq 3$, allowing an involution of the components that preserves the identity component and incidence relation, or when several fibers of the same type lie over points on the base curve ${\mathbb{P}}^1$ that are Galois conjugates. Later in this paper we again need a term for points on ${\mathbb{P}}^1$ lying under reducible fibers of an elliptic surface; we call them “cusps”$\!$, consistent with the special case of the universal elliptic curve over the modular curve $X(N)$ or $X_1(N)$.
In terms of the associated newforms, the discriminants in the above table exhaust the extremal elliptic K3 surfaces over ${\mathbb{Q}_{}}$. Note that since we consider elliptic surfaces, we can again twist as in (\[eq:Kummer\]). In the next section we will explain our main techniques to exhibit singular K3 surfaces for the remaining 28 imaginary quadratic fields.
The main techniques {#s:tech}
===================
We search for singular K3 surfaces over ${\mathbb{Q}_{}}$ with particular discriminants. The main idea is to take advantage of the moduli theory of complex K3 surfaces. Any one-dimensional family of K3 surfaces of Picard number $\rho\geq 19$ has infinitely many specializations with $\rho=20$. We will only search for the specializations over ${\mathbb{Q}_{}}$ (which are finite in number by Šafarevič’s result [@Shafa]). Because of Theorem \[Thm:NS\], these surfaces and therefore the families involved will be very special. Hence one of the key steps will be to construct suitable families. This will be achieved in the next section. Here we explain how we find the specializations of a given family.
Given a one-dimensional family of K3 surfaces $X_\lambda$ over ${\mathbb{Q}_{}}$ satisfying $\rho(X_\lambda) \geq 19$ for all $\lambda$, there is an easily checked necessary condition that must be satisfied by any member $X$ with $\rho(X)=20$. The condition is based on the Lefschetz fixed point formula at a good prime $p$, formulated in terms of $\ell$-adic étale cohomology $H_{\acute{e}t}^i(\bar X, {{\mathbb{Q}_{}}}_\ell)$ for some prime $\ell\neq p$. Here we work with the base change $\bar X$ of the reduction of $X \bmod p$ to an algebraic closure of ${\mathbb{F}}_p$. For simplicity, we will just write $H^i(X)$ for $H_{\acute{e}t}^i(\bar X, {{\mathbb{Q}_{}}}_\ell)$ in the following.
The cohomology groups $H^i(X)$ are equipped with an induced action of the geometric Frobenius morphism Frob$_p$. The set $X({\mathbb{F}}_p)$ of ${\mathbb{F}}_p$-rational points on $X$ is exactly the fixed set of Frob$_p$. For a K3 surface $X/{\mathbb{Q}_{}}$ with good reduction at $p$, the Lefschetz fixed point formula simplifies to $$\begin{aligned}
\label{eq:Lefschetz}
\# X({\mathbb{F}}_p) = 1 + \text{trace Frob}_p^*(H^2(X)) + p^2.\end{aligned}$$ Because $\rho(X)\geq 19$, we can predict 19 of the 22 eigenvalues of Frob$_p^*$ on $H^2(X)$. Since the absolute Galois group operates through a permutation on the algebraic cycles, all these eigenvalues have the form $\zeta\cdot p$ for some roots of unity $\zeta$. By the Weil conjectures, one further eigenvalue has the form $\pm p$, and the remaining two eigenvalues are algebraic integers $\alpha_p,\beta_p$ of absolute value $p$. In particular, the unordered pair $(\alpha_p, \beta_p)$ is determined by (\[eq:Lefschetz\]) and the sign of the other eigenvalue $\pm p$. In general, $\beta_p=\bar\alpha_p$; the only exception is the supersingular case where $\alpha_p+\beta_p=0$.
If the specialization $X$ at some $\lambda_0\in{\mathbb{Q}_{}}$ is a singular K3 surface, then it is modular by Theorem \[Thm:mod\]. Hence, for the right choice of sign in the other eigenvalue, $$\begin{aligned}
\label{eq:match}
\alpha_p+\beta_p=a_p\end{aligned}$$ where $a_p$ is the Fourier coefficient of the corresponding newform $f$ of weight 3. In particular, both $\alpha_p$ and $\beta_p$ lie in the imaginary quadratic extension $K$ associated to $f$. By Theorem \[Thm:CM\], $K$ has class group exponent $2$. Moreover, $K$ remains fixed when $p$ varies. This gives a criterion to either rule out $\lambda_0$ or collect evidence for $\rho(X)=20$.
As it stands, our condition for $\rho=20$ is necessary but not sufficient. To search for the CM-specializations, we will use the condition in a different, almost opposite approach: we search for good parameters mod $p$ and try to lift $\lambda_0$ to ${\mathbb{Q}_{}}$. We use the following algorithm:
\[Alg\] Let $X_\lambda$ be a family of K3 surfaces over ${\mathbb{Q}_{}}$ with Picard number $\rho\geq 19$. Then the following algorithm returns candidate parameters $\lambda_0\in{\mathbb{Q}_{}}$ such that the specialization $X$ at $\lambda_0$ might have $\rho(X)=20$:
(i) Fix one of the 65 weight 3 newforms $f$ up to twisting (that is, one of the imaginary quadratic fields of class group exponent two).
(ii) \[item\] For each of several primes $p=p_i$ ($i=1,\ldots,n$), use (\[eq:Lefschetz\]) to compute $\alpha_p,\beta_p$ for every $\lambda\in{\mathbb{F}}_p$, and find all $\lambda$ such that one choice of sign for the other eigenvalue leads to $\alpha_p, \beta_p$ matching $f$ as in (\[eq:match\]). Even though $f$ varies, we need only compute $\#X_\lambda({\mathbb{F}}_p)$ once for each pair $(p,\lambda)$.
(iii) \[item-3\] For a collection of parameters $\lambda_1\bmod p_1, \hdots, \lambda_n \bmod p_n$ matching the newform $f$, compute a lift $\lambda_0\in{\mathbb{Q}_{}}$ of small height using the Chinese Remainder Theorem and the Euclidean algorithm.
For each newform $f$, the algorithm returns a number of guesses for specializations in ${\mathbb{Q}_{}}$, if any. Often, there will be one $\lambda_0$ among them which looks particularly likely (small height, small primes involved etc.). At this point, we can continue to collect numerical evidence by running the above test for further primes $p$. In the end, however, we want to prove that $\rho(X_{\lambda_0})=20$. Therefore we have to find explicitly an additional divisor on the specialization $X$ at $\lambda_0$. This is where we turn to elliptic surfaces. Until this point the procedure works for any one-dimensional family. For instance, we computed candidates for the singular specializations for the Dwork pencil $$x_0^4+x_1^4+x_2^4+x_3^4=\lambda x_0x_1x_2x_3$$ of deformations of the Fermat quartic from Example \[Ex:Fermat\] (cf. [@ES] for a detailed account). But the only way we know to systematically search for extra divisors on a K3 family $X_\lambda$ uses models of $X_\lambda$ as elliptic surfaces.
The advantage of working with elliptic surfaces is the following. By the formula of Shioda and Tate, their groups are always generated by horizontal and vertical divisors, i.e. by fiber components and sections. Hence for the Picard number to increase in a family, either the singular fibers degenerate further (which happens for only finitely many $\lambda_0 \in {\mathbb{C}^{}}$), or there is an additional section $P$. Then the discriminant $d$ of the specialization can be computed purely in terms of the intersection behavior of $P$ with the singular fibers and the other sections. This is made explicit through the theory of lattices.
In our setting, $d$ is predicted up to a square factor by the newform $f$ and its CM-field $K$ by Theorem \[Thm:mod\] and the explanation following it. This provides us with additional information about the conjectural section $P$, information that is often crucial to the feasibility of a direct computation. The next section illustrates this by the detailed analysis of a particular family of K3 surfaces.
Specializations of a one-dimensional family {#s:fam}
===========================================
We want to search in one-dimensional families of K3 surfaces with Picard number $\rho\geq 19$ for singular specializations over ${\mathbb{Q}_{}}$. In fact, it is a nontrivial task to find such families with interesting specializations in the first place. This difficulty is due to the nontrivial Galois action that the group must admit by Theorem \[Thm:NS\]. In this section, we discuss one particular family in detail.
We start with the following two-dimensional family in extended Weierstrass form with parameters $\lambda\in{\mathbb{P}}^1, \mu\neq 0$: $$\begin{aligned}
X_{\lambda, \mu}:\;\;y^2 & = & x^3 +
(t-\lambda) A \,x^2 + t^2 (t-1) (t-\lambda)^2 B\, x +
t^4 (t-1)^2 (t-\lambda)^3 C,\\
A & = & \frac 1{24} \Bigl( \frac 19 (2 \mu+9)^3 t^3-
(22 \mu-9) (2 \mu-27) t^2-27 (14 \mu-9) t-81 \Bigr),\\
B & = & \mu\, \Bigl(\frac 19 (2 \mu+9)^3 t^2
-2 (10 \mu-9) (2 \mu-9) t-27 (2 \mu-3) \Bigr),\\
C & = & \frac 23 \mu^2 ((2 \mu+9)^3 t-81 (2 \mu-3)^2).\end{aligned}$$ This elliptic surface has discriminant $$\begin{array}{l}
\Delta = 36 \mu^4 t^5 (t-1)^3 (t-\lambda)^6\\
\;\;\;\; ((2 \mu+9)^4 t^3-9 (32 \mu+27) (2 \mu+9)^2 t^2+
81 (308 \mu^2+243-864 \mu) t+729 (4 \mu-9)).
\end{array}$$ It has the following singular fibers: $$\begin{array}{|c||c|c|c|c|c|}
\toprule
\text{cusp} & 0 & 1 & \infty & \lambda & \text{cubic with coefficients in } \mu\\
\midrule
\midrule
\text{fiber} & I_5 & I_3 & I_7 & I_0^* & I_1, I_1, I_1\\
\bottomrule
\end{array}$$ In general, $X_{\lambda, \mu}$ has lattice $${\mathop{\rm NS}\nolimits}(X_{\lambda, \mu}) = U + A_2 + A_4 + A_6 + D_4.$$ Here $U$ denotes the hyperbolic plane (generated by the 0-section $O$ and a general fiber $F$), and $A_i, D_i$ are the root lattices corresponding to the reducible singular fibers. In particular, we deduce that $$\rho(X_{\lambda, \mu})\geq 18.$$ We briefly explain how we found the above family. As in [@E-Shimura], we work with an extended Weierstrass form. Here we can translate $x$ so that all singular fibers have their singularities at $x=y=0$. In the present situation, this gives rise to the family $$\begin{aligned}
y^2 & = & x^3 +
(t-\lambda)\, a_2(t) \,x^2 + t^2\, (t-1)\, (t-\lambda)^2\, a_1(t)\, x +
t^4\, (t-1)^2\, (t-\lambda)^3\, a_0(t).\end{aligned}$$ In general, the $a_i(t)$ are polynomials of degree $\deg(a_i)\leq i+1$. Hence we have ten parameters to choose including $\lambda$, relative to one normalization by rescaling $x, y$. The above extended Weierstrass form guarantees that the fiber types are at least $I_4$, $I_2$, $I_0^*$, $I_2$ at $t=0,1,\lambda,\infty$ respectively. We can easily choose the coefficients of $a_2(t)$ to promote the fiber at $\infty$ to type $I_6$. Then we solve a system of three nonlinear equations in the five coefficients of $a_0(t), a_1(t)$ to derive the family $X_{\lambda, \mu}$. This can be achieved by appropriate combinations of the equations and a suitable choice of normalization.
The family $X_{\lambda, \mu}$ can easily be specialized to a family $X_\lambda$ with $\rho(X_\lambda)\geq 19$ by degenerating the singular fibers, i.e. merging fibers. There are two independent ways to do so. On the one hand, we can match $\lambda$ with one of the cusps. This results in four families. Each of them has several interesting specializations. The case where we merge $I_0^*$ with a $I_1$-fiber will be taken up in the next section (Example \[Ex:1\]). On the other hand, we can merge one of the $I_1$ fibers with another singular fiber using $\mu$. We now discuss one particular case.
We merge two fibers of type $I_1$ by setting $\mu=\frac{243}{10}$. We obtain the one-dimensional family $X_\lambda$ with the following singular fibers: $$\begin{array}{|c||c|c|c|c|c|c|}
\toprule
\text{cusp} & 0 & 1 & \frac{35}{32} & -\frac 5{1024} & \infty & \lambda\\
\midrule
\midrule
\text{fiber} & I_5 & I_3 & I_2 & I_1 & I_7 & I_0^*\\
\bottomrule
\end{array}$$ The general member $X_\lambda$ has lattice $${\mathop{\rm NS}\nolimits}(X_\lambda) = U + A_1 + A_2 + A_4 + A_6 + D_4,$$ so $\rho(X_\lambda)\geq 19$. The Galois action on the group is encoded in the fields where the singular fibers with at least three components split (depending on $\lambda$): $$\begin{array}{|c||c|c|c|c|}
\toprule
\text{fiber} & I_5 & I_3 & I_7 & I_0^*\\
\midrule
\midrule
\text{splitting field} & {\mathbb{Q}_{}}(\sqrt{6\lambda}) &
{\mathbb{Q}_{}}(\sqrt{10\,(1-\lambda)}) & {\mathbb{Q}_{}}(\sqrt{15}) & {\mathbb{Q}_{}}(f(\lambda))\\
\bottomrule
\end{array}$$ Here ${\mathbb{Q}_{}}(f(\lambda))$ is the splitting field of the cubic polynomial $$\begin{aligned}
f(\lambda) & = & x^3 + \left({\frac {4096}{15}}\,{\lambda}^{3}-115\,\lambda
-146\,{\lambda}^{2}-{\frac {25}{24}}\right)\, x^2\\
&& \; + 6\,{\lambda}^{2} \left( \lambda-1 \right)
\left( 8192\,{\lambda}^{2}-7150\,\lambda-475 \right)\,x
+ 1080\,{\lambda}^{4} \left( \lambda-1 \right) ^{2}
\left( 2048\,\lambda-1805 \right).\end{aligned}$$
There are five obvious specializations with $\rho=20$ where we match $\lambda$ (i.e. the $I_0^*$ fiber) with another cusp. This way, we obtain explicit equations for extremal elliptic K3 surfaces over ${\mathbb{Q}_{}}$ that were not derived in [@BM] or [@S-Rocky]. These fibrations provide alternative realizations of weight 3 newforms for the following discriminants previously realized by semistable extremal elliptic K3 surfaces (cf. Section \[s:Ex\]):
$\lambda$ degeneration in ${\mathop{\rm NS}\nolimits}(X_\lambda)$ disc ${\mathop{\rm NS}\nolimits}(X_\lambda)$ $T(X_\lambda)$
------------------- --------------------------------------------------------- ---------------------------------------------- ----------------
$-\frac{5}{1024}$ $D_4 \rightsquigarrow D_5$ $-840$ \[20, 0, 42\]
$\frac{35}{32}$ $A_1 + D_4 \rightsquigarrow D_6$ $-420$ \[6, 0, 70\]
$1$ $A_2 + D_4 \rightsquigarrow D_7$ $-280$ \[4, 0, 70\]
$0$ $A_4 + D_4 \rightsquigarrow D_9$ $-168$ \[4, 0, 42\]
$\infty$ $A_6 + D_4 \rightsquigarrow D_{11}$ $-120$ \[6, 0, 20\]
Every other specialization with $\rho=20$ requires the existence of an additional section. We ran Algorithm \[Alg\] through the first 30 primes $p$ to find candidate parameters $\lambda_0\in{\mathbb{Q}_{}}$. In step (\[item\]), we restricted to those $p$ which split in the field $K$ corresponding to a given newform $f$. In almost every case, this implied that there was exactly one candidate parameter $\lambda\mod p$ (if there were any at all). Some lifts to ${\mathbb{Q}_{}}$ are given in Table 1. We shall now explain how the structure of $X_\lambda$ as an elliptic surface helps us verify that the K3 surface at such a lift $\lambda_0$ is singular.
Elliptic surfaces {#elliptic-surfaces .unnumbered}
-----------------
Once we know the group of sections of an elliptic surface, we easily compute the discriminant of its lattice using the theory of lattices [@ShMW]. In the case at hand, the specialization $X$ at $\lambda_0$ has a conjectural section $P\neq O$, generating the group (for the general situation cf. sect. \[ss:2\]). By [@ShMW Thm. 8.6] the discriminant is given by the product of the discriminants of the root lattices and the height of $P$: $$\begin{aligned}
\label{eq:disc}
{\mathop{\rm disc}\nolimits}{\mathop{\rm NS}\nolimits}(X) = -2\cdot 3 \cdot 5\cdot 7 \cdot 4 \cdot \hat h(P)
= -840\, \hat h(P).\end{aligned}$$ The height of $P$ is determined by the intersection number $(P.O)$ of $P$ and $O$ in ${\mathop{\rm NS}\nolimits}(X)$ and some correction terms corr$_v(P)$ according to the fiber components which $P$ meets: $$\begin{aligned}
\label{eq:height}
\hat h(P) = 4 + 2\,(P.O) - \sum_v {\mathop{\rm corr}\nolimits}_v(P).\end{aligned}$$ To describe the correction term at a $I_n$ fiber, we number the components cyclically $\Theta_0,\hdots,\Theta_{n-1}$ such that $\Theta_0$ is the identity component (i.e. the component meeting $O$). In this notation, $${\mathop{\rm corr}\nolimits}_v(P)=\begin{cases}
0 &
\text{if $P$ meets the identity component of the fiber at $v$},\\
1 &
\text{if $P$ meets a non-identity component at a $I_0^*$ fiber},\\
\frac{i(n-i)}n &
\text{if $P$ meets the component $\Theta_i$ of a $I_n$ fiber}.
\end{cases}$$
Assume that we have a candidate parameter $\lambda_0$ for a specialization $X$ to realize some newform $f$. Let $K$ denote the imaginary quadratic field corresponding to $f$. Then we have to arrange for a section $P$ such that ${\mathop{\rm NS}\nolimits}(X)$ has discriminant $d<0$ so that ${\mathbb{Q}_{}}(\sqrt{d})=K$. Let us look at one example in detail.
\[Ex:-1540\] Let $f$ be a newform for the field $K={\mathbb{Q}_{}}(\sqrt{-385})$ of discriminant $d=-1540$. Then Algorithm \[Alg\] suggests the candidate parameter $\lambda_0 = 7^2\,11 / 2^9$. We want to find a section $P$ such that disc ${\mathop{\rm NS}\nolimits}(X)=d$.
Consider the formula (\[eq:disc\]) for the discriminant of the elliptic surface with section $P$. Our target discriminant $d=-1540$ requires that we eliminate a factor $6$ from (\[eq:disc\]) while preserving $5$ and $7$. Hence $P$ must meet the $I_5$ and $I_7$ fibers on their identity components, but meet the $I_2$ and $I_3$ fibers on non-identity components. Now we assume that $(P.O)=0$ and that $P$ intersects a non-identity component of the $I_0^*$ fiber. Then (\[eq:height\]) reads $$\hat h(P) = 4 - \frac 12 - \frac 23 -1 = \frac{11}6.$$ Hence (\[eq:disc\]) gives $${\mathop{\rm disc}\nolimits}{\mathop{\rm NS}\nolimits}(X) = -840 \cdot \frac{11}6 = -1540$$ as required. The assumption $(P.O)=0$ implies that $P$ has coordinates $(u, v)$ for polynomials $u, v$ of degree $4$ resp. $6$ in the coordinate $t$ of the base curve. Note that inserting $P$ into the Weierstrass equation of $X$ allows us to express the coefficients of $v$ in terms of those of $u$. For the fibers to be met as prescribed, we must have $$(t-1)\left(t-\frac{35}{32}\right)(t-\lambda_0)^2 \mid v,\;\;\;
(t-1)(t-\lambda_0) \mid u.$$ Moreover, the fibers at $\lambda_0$ and $\frac{35}{32}$ give two more linear relations in the coefficients of $u$. Since $u$ has degree four, there is only one degree of freedom left. Thus the problem is easily solved using a computer algebra system. The solution $P$ and the resulting transcendental lattice can be found in Table 1.
Often we are not lucky enough to arrive at a system of equations that we can solve directly as in Example \[Ex:-1540\]. If we cannot find a direct solution, we apply the following algorithm, essentially a $p$-adic Newton iteration in several variables.
\[Newton\] Given a system $f_1=\hdots=f_n=0$ of algebraically independent polynomial equations over ${\mathbb{Q}_{}}$ in $n$ variables $z_1,\hdots,z_n$. The following procedure tests for a solution over ${\mathbb{Q}_{}}$:
(1) \[Newton-1\] With an exhaustive search, find a solution $(\bar z_1,\hdots,\bar z_n)$ modulo some prime $p$.
(2) \[Newton-2\] Using difference quotients, double the $p$-adic accuracy of the solution a few times.
(3) \[Newton-3\] Compute a lift in ${\mathbb{Q}_{}}$ with the Euclidean algorithm.
(4) \[Newton-4\] Test whether the lift solves the system of equations over ${\mathbb{Q}_{}}$. While it does not, return to step (\[Newton-2\]) to double the precision once more and try (\[Newton-3\]) and (\[Newton-4\]) again.
For step (\[Newton-2\]) to converge, we need some regularity assumptions for the polynomials $f_i$. For instance, it will converge if the coefficients of the $f_i$ are integers and the Jacobian determinant $|\partial(f_1,\ldots,f_n) / \partial(z_1,\ldots,z_n)|$ does not vanish at $(\bar z_1,\hdots,\bar z_n)$.
We next outline some further implementation issues specific to our setting.
Algorithm \[Newton\] does not require that the section itself be defined over ${\mathbb{Q}_{}}$, only its $x$-coordinate. Even if the $y$-coordinate involves a square root of a rational number as a factor, we can always arrange to solve a system of equations over ${\mathbb{Q}_{}}$.
A more delicate point about Algorithm \[Newton\] is the choice of the prime $p$. Here we distinguish whether or not $p$ splits in the fixed imaginary quadratic field $K$.
For a singular K3 surface $X$ over a number field $L$, one can predict the geometric Picard number of the reductions. Namely, let ${\mathfrak{p}}$ denote a prime of $L$ above $p$. If $X$ has good reduction mod ${\mathfrak{p}}$, write $X_{\mathfrak{p}}$ for the reduced K3 surface. Then $$\begin{aligned}
\label{eq:Tate}
\rho(X_{\mathfrak{p}}) =
\begin{cases}
20, & \text{if $p$ splits in $K$},\\
22, & \text{if $p$ is inert or ramified in $K$}.
\end{cases}\end{aligned}$$ This follows from the structure (cf. sect. \[s:obs\]) since the above cases decide exactly whether the elliptic curves $E, E'$ (and the abelian surface $E\times E'$) are supersingular.
In more detail, we apply the Tate conjecture [@Tate-C] to a (conjectural) singular K3 surface $X$ over ${\mathbb{Q}_{}}$. Here we consider elliptic K3 surfaces with section, so the Tate conjecture holds true by [@ASD]. Note that the assumption implies that the primes ramifying in $K$ are always bad, since they divide the level of the associated Hecke eigenform by Theorem \[Thm:mod\].
At the inert primes $p$ in $K$, the Fourier coefficient $a_p$ of the associated newform is zero. The resulting eigenvalues of Frob$_p$ are $p$ and $-p$. Hence the Tate conjecture predicts that the reduction $X_p$ has an additional algebraic cycle over ${\mathbb{F}}_p$ and one more over ${\mathbb{F}}_{p^2}$. On an elliptic surface, these extra cycles would either change the configuration of reducible fibers or appear as extra sections. In the former case we might miss the reduced surface entirely. In the latter case, reduction mod $p$ would increase the rank compared to the rank over ${\mathbb{Q}_{}}$. Depending on the conditions we impose — on $(P.O)$ and the fiber components met — step (\[Newton-1\]) of Algorithm \[Newton\] might then return more than one section. However, only one of these sections would lift to ${\mathbb{Q}_{}}$, so we would have to make the right choice.
Therefore, we always run Algorithm \[Newton\] at a prime $p$ that splits in $K$. Under the assumption that the K3 surface is singular, the Picard number in (\[eq:Tate\]) guarantees that the rank is constant at $20$ upon reduction.
\[Rem:no-guess\] We will apply Algorithm \[Newton\] after guessing the parameter $\lambda_0$ with the help of Algorithm \[Alg\]. Hence there is one further equation $f_0=0$ in $z_1,\hdots,z_n$. After step (\[Newton-4\]), we thus also have to verify that $f_0$ vanishes at the lift.
The same ideas can also be applied without a candidate parameter $\lambda_0$ at hand. In Example \[Ex:-1540\], this would have sufficed as well: in the end, we would have to solve two equations in two variables. However, if we have to apply Algorithm \[Newton\] to find the section explicitly, it is computationally very convenient for step (\[Newton-1\]) to have one parameter fewer.
The following table collects the specializations of $X_\lambda$ where we verified $\rho=20$. The additional section $P$ can be recovered from its $x$-coordinate $u(t)$ by taking $x=u(t)$ in the Weierstrass equation and choosing a root for $v(t)$. We also list the height $\hat h(P)$ of $P$ and the discriminant of ${\mathop{\rm NS}\nolimits}(X)$. The discriminant will be justified below. The computation of the transcendental lattices $T(X)$ will be explained in the next subsection.
$$\begin{array}{c|c|c|c|c}
\toprule
{\scriptstyle\lambda } & {\scriptstyle \text{disc } {\mathop{\rm NS}\nolimits}(X) } & {\scriptstyle P: u(t) } & {\scriptstyle \hat h(P) } & {\scriptstyle T(X) } \\
\midrule
\midrule
{\scriptstyle -\frac 12} & {\scriptstyle -225\cdot 4} &
{\scriptstyle -\frac{3^5}{2^3\cdot 5}\,(24\,t+35)\,(t-1)\,(2\,t+1)} &
{\scriptstyle \frac{15}{14}} & {\scriptstyle [30,0,30]}\\
\midrule
{\scriptstyle\frac 58 } & {\scriptstyle -4\cdot 43 } & {\scriptstyle
-{\frac {2\cdot 3^6}{5}}\,{t}^{2} \left( t-1 \right)
} & {\scriptstyle \frac{43}{210} } & {\scriptstyle [4,2,44]} \\
\midrule
{\scriptstyle\frac{5^3}{2^7} } & {\scriptstyle -4\cdot 67 } & {\scriptstyle
-{\frac {3^6}{2^6\cdot 5}}\, \left( t-1 \right) t \left( 128\,t-105 \right)
} & {\scriptstyle \frac{67}{210} } & {\scriptstyle [4,2,84]} \\
\midrule
{\scriptstyle\frac{5}{2^5} } & {\scriptstyle -88 } & {\scriptstyle
-{\frac {3^6}{2^4\cdot 5}}\, \left( t-1 \right) t \left( 32\,t-5 \right)
} & {\scriptstyle \frac{11}{105} } & {\scriptstyle [2,0,44]} \\
\midrule
{\scriptstyle\frac{5}{2^3\cdot 19^2} } & {\scriptstyle -4\cdot 163 } & {\scriptstyle
-{\frac {2\cdot 3^6\cdot 29^2}{5\cdot 19^2}}\,{t}^{2} \left(t-1 \right)
} & {\scriptstyle \frac{163}{210} } & {\scriptstyle [4,2,164]} \\
\midrule
{\scriptstyle\frac 52 } & {\scriptstyle -228 } & {\scriptstyle -\frac{3^6}{2^3\cdot 5}\,t\,(2\,t-5)\,(8\,t-7)
} & {\scriptstyle \frac{19}{70} } & {\scriptstyle [2,0,114]} \\
\midrule
{\scriptstyle\frac{3^2\cdot 5}{2^5} } & {\scriptstyle -312 } & {\scriptstyle
-{\frac {3^6}{2^4\cdot 5}}\,{t}^{2} \left( 32\,t-45 \right) } & {\scriptstyle \frac{13}{35}
} & {\scriptstyle [6,0,52]} \\
\midrule
{\scriptstyle-\frac 12 } & {\scriptstyle -340 } & {\scriptstyle
-{\frac {3^5}{2^3\cdot 5}}\, \left( t-1 \right)
\left( 2\,t+1 \right) \left( 24\,t+35 \right)
} & {\scriptstyle \frac{17}{42} } & {\scriptstyle [20,10,22]} \\
\midrule
{\scriptstyle\frac{5\cdot 11^2}{2^9} } & {\scriptstyle -372 } & {\scriptstyle
-{\frac {3^6}{2^8\cdot 5^2}}\,{t}^{2} \left( 512\,t-605 \right) } & {\scriptstyle \frac{31}{70}
} & {\scriptstyle [6,0,62]} \\
\midrule
{\scriptstyle\frac{2\cdot 5}{3^2} } & {\scriptstyle -408 } & {\scriptstyle
\frac{3^3}{2^4\cdot 5}\, t\, (9\, t-10) (-96\, t+81)
} & {\scriptstyle \frac{17}{35} } & {\scriptstyle [6,0,68]} \\
\midrule
{\scriptstyle-\frac 5{2^3\cdot 3} } & {\scriptstyle -4\cdot 435 } & {\scriptstyle
\frac 3{2^4\cdot 5}\, (-7776\, t^3+4816\, t^2+1660\, t+175)
} & {\scriptstyle \frac{29}{14} } & {\scriptstyle [20,10,92]} \\
\midrule
{\scriptstyle\frac{3\cdot 5\cdot 7}{2^7} } & {\scriptstyle -4\cdot 483 } & {\scriptstyle
\frac{3^2}{2^8\cdot 5\cdot 7^2}\,(-5242880\, t^2+9202816\, t-3988061)
} & {\scriptstyle \frac{23}{10} } & {\scriptstyle [4,2,484]} \\
\midrule
{\scriptstyle-\frac 5{7^2} } & {\scriptstyle -520 } & {\scriptstyle
-{\frac {3^6}{2^4\cdot 5\cdot 7^2}}\, \left( t-1 \right)
\left( 32\,t+5 \right) \left( 49\,t+5 \right)
} & {\scriptstyle \frac{13}{21} } & {\scriptstyle [20,0,26]} \\
\midrule
{\scriptstyle\frac{5\cdot 7^2}{2^5\cdot 19} } & {\scriptstyle
-532 } & {\scriptstyle -{\frac {2\cdot 3^5}{5^2\cdot 19}}\,
\left( t-1 \right)\, {t}^{2} \left( 608\,t-245 \right)
} & {\scriptstyle \frac {19}{30} } & {\scriptstyle [4,2,134]} \\
\midrule
{\scriptstyle\frac{37}{2^3\cdot 5} } & {\scriptstyle
-4\cdot 555 } & {\scriptstyle \frac{3^6}{2^4\cdot 5^4}\, (-31648\, t^3+86320\, t^2-78300\, t+23625)
} & {\scriptstyle \frac{37}{14} } & {\scriptstyle [4,2,556]} \\
\midrule
{\scriptstyle-\frac{7}{2^7} } & {\scriptstyle
-4\cdot 595 } & {\scriptstyle \frac{3}{2^6\cdot 5^3}\, (t-1)\, (-1048576\, t^3-716160\, t^2-45225\, t-875)
} & {\scriptstyle \frac{17}6 } & {\scriptstyle [20,10,124]} \\
\midrule
{\scriptstyle-\frac 1{2^5} } & {\scriptstyle -660 } & {\scriptstyle
\frac{3^5}{2^{14}\cdot 5}\, (32\, t+ 1) (-3072\, t^2+2592\, t+105)
} & {\scriptstyle \frac{11}{14} } & {\scriptstyle [20,10,38]} \\
\midrule
{\scriptstyle\frac{5\cdot 11^2}{2\cdot 7^2} } & {\scriptstyle -708 } & {\scriptstyle
\frac{3^2}{2^3\cdot 5^2 \cdot 7^3}\, t\, (98\, t-605) (-53176\, t+46585)
} & {\scriptstyle \frac{59}{70} } & {\scriptstyle [6,0,118]} \\
\midrule
{\scriptstyle-\frac{19}{2^5\cdot 7^2} } & {\scriptstyle -760 } & {\scriptstyle
\frac{3^5}{2^4\cdot 5^3 \cdot 7^2}\, (t-1)\, (1568\, t+19) (-75\, t-1)
} & {\scriptstyle \frac{19}{21} } & {\scriptstyle [20,0,38]} \\
\midrule
{\scriptstyle-\frac{1}{2^3\cdot 3\cdot 5}} & {\scriptstyle-4\cdot 795} &
{\scriptstyle -\frac{3}{2^4\cdot 5^4}\,U(t)/(15360\,t+7)^2}
& {\scriptstyle \frac{53}{14}} & {\scriptstyle [20, 10, 164]}\\
\midrule
{\scriptstyle\frac{5^3\cdot 7}{2^9} } & {\scriptstyle -1092 } & {\scriptstyle
\frac{3^6}{2^8\cdot 5^4 \cdot 7^4} \,t^2\, (512\, t-875) (-4096\, t+2765)
} & {\scriptstyle \frac{13}{10} } & {\scriptstyle [6,0,182]} \\
\midrule
{\scriptstyle\frac{3^2\cdot 5\cdot 11}{2^{10}} } & {\scriptstyle -1320 } & {\scriptstyle
\frac{3^6}{2^{30}\cdot 5}\, (1024\, t-495)\, (-2097152\, t^2+5225472\, t-1607445)
} & {\scriptstyle \frac{11}7 } & {\scriptstyle [4,0,330]} \\
\midrule
{\scriptstyle\frac{11^2}{2^5\cdot 3} } & {\scriptstyle -1380 } & {\scriptstyle
\frac{3}{2^{16}\cdot 5}\, (96\, t-121)\, (-4427776\, t^2+9923936\, t-5636785)
} & {\scriptstyle \frac{23}{14} } & {\scriptstyle [6,0,230]} \\
\midrule
{\scriptstyle\frac{3\cdot 5^3}{2^5} } & {\scriptstyle -1428 } & {\scriptstyle
\frac{3^3}{2^4\cdot 5^3\cdot 7^2} \,t\, (32\, t-375) (32\, t^2-99880 \, t+87500)
} & {\scriptstyle \frac{17}{10} } & {\scriptstyle [6,0,238]} \\
\midrule
{\scriptstyle\frac{7^2\cdot 11}{2^9} } & {\scriptstyle -1540 } & {\scriptstyle
-\frac{3}{2^8\cdot 5} \, (512\, t-539)\, (t-1)\, (512000\, t^2-1097257\, t+588245)
} & {\scriptstyle \frac{11}{6} } & {\scriptstyle [22,0,70]} \\
\midrule
{\scriptstyle-\frac{5\cdot 7}{2^{10}\cdot 3^2} } & {\scriptstyle -1848 } & {\scriptstyle
-\frac{3^4}{2^7\cdot 5\cdot 7^3} \, t\, (9216\, t+35)
(75582720\, t^2+631582\, t+1323)
} & {\scriptstyle \frac{11}5 } & {\scriptstyle [42,0,44]} \\
\bottomrule
\end{array}$$
[Table 1:]{} Singular specializations over ${\mathbb{Q}_{}}$ in the family $X_\lambda$
For the largest discriminant $-4\cdot 795$, the section $P$ is not integral. Here $U(t)$, the numerator of $u(t)$, is given by $$\begin{aligned}
U(t) & = & 229323571200000\,t^5-191371714560000\,t^4-9553942361376\,t^3\\
&& \;\; -151103350160\,t^2-953437100\,t-2100875.\end{aligned}$$
The table is not complete. For each of the four discriminants $$d =-4 \cdot 1435,\;\; -4 \cdot 1155,\;\; -4 \cdot 1995,\;\; -5460,$$ we find using Algorithm 10 the lift $$\lambda_0 =
-\frac{5 \cdot 7 \cdot 11^2}{2^3\cdot 41},\;\;
-\frac{7}{2^3},\;\;
\frac{7\cdot 19}{2^7\cdot 5},\;\;
\frac{3 \cdot 5 \cdot 7^2}{2 \cdot 19^2}$$ respectively for which we expect $X_{\lambda_0}$ to be singular with discriminant $d$; but the additional section is too complicated for us to compute easily, even using Algorithm 12. In the next section we realize each of the remaining nine discriminants, including these four, in a singular K3 surface not in this family $X_\lambda$.
\[Lem:NS\] For all K3 surfaces in the table, ${\mathop{\rm NS}\nolimits}(X)$ is generated by fiber components and the sections $O$ and $P$. In particular, the discriminant is as claimed.
*Proof:* The fiber components together with the sections $O$ and $P$ generate a lattice $N$ of rank 20. We must show that $N={\mathop{\rm NS}\nolimits}(X)$. The discriminant $d$ of $N$ is given by (\[eq:disc\]) and (\[eq:height\]). If the index of $N$ in ${\mathop{\rm NS}\nolimits}(X)$ were greater than $1$, then ${\mathop{\rm NS}\nolimits}(X)$ would have discriminant $d$ divided by the square of the index.
For any elliptic K3 surface $X_\lambda$ in our family, the group is torsion-free. Using the formula (\[eq:height\]), we see that $2\hat h(P)$ is a integer for any section $P$, because $(P.O)$ is an integer and each correction term ${\mathop{\rm corr}\nolimits}_v(P)$ is in $\frac12 {\mathbb{Z}}_2$. By (\[eq:disc\]), we derive the general relation $$\begin{aligned}
\label{eq:4}
4 \mid {\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(X_\lambda)).\end{aligned}$$ In each case in the table, the quotient $d/\!{\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(X))$ can therefore only be a square if it equals $1$. The claim follows. [$\Box$]{}
Transcendental lattices {#transcendental-lattices .unnumbered}
-----------------------
The transcendental lattices for the singular K3 surfaces above were computed using lattice theory as developed by Nikulin [@N]. Here we sketch the argument.
Given an even integral lattice $L$, we denote its dual by $L^\vee$. In [@N], Nikulin introduced a quadratic form on the quotient $L^\vee/L$ which he called the discriminant form: $$\begin{aligned}
q_L\0:\;\; L^\vee/L & \to & {\mathbb{Q}_{}}\mod 2{\mathbb{Z}}\\
x & \mapsto & x^2 + 2{\mathbb{Z}}\end{aligned}$$
For each singular K3 surface $X$ in the table, we know the lattice ${\mathop{\rm NS}\nolimits}(X)$ by Lemma \[Lem:NS\]. Hence we can compute its discriminant form.
Let $N$ be an even integral unimodular lattice. Let $L$ be a primitive non-degenerate sublattice and $M=L^\bot$. Then $$q_L\0 = -q_M\0.$$
Since ${\mathop{\rm NS}\nolimits}(X)$ always embeds primitively into $H^2(X,{\mathbb{Z}})$, the theorem provides the discriminant form of the transcendental lattice $T(X)$.
The genus of an even integral lattice is determined by its signature and discriminant form.
We now use the fact that each singular K3 surface $X$ in the table is defined over ${\mathbb{Q}_{}}$. By Theorem \[Thm:genus\], the genus of the transcendental lattice $T(X)$ consists of a single class. Hence the discriminant form of the lattice ${\mathop{\rm NS}\nolimits}(X)$ determines $T(X)$ uniquely via the above two theorems. Thus the computation is completed by verifying that for the given transcendental lattices $q_{T(X)}\0=-q_{{\mathop{\rm NS}\nolimits}(X)}\0$.
The remaining discriminants {#s:rem}
===========================
So far we have matched all but nine newforms of weight 3 from Section \[s:CM\] with singular K3 surfaces over ${\mathbb{Q}_{}}$. For some of the remaining discriminants, we found candidate surfaces in other one-dimensional families of K3 surfaces with $\rho\geq 19$, see Examples \[Ex:1\]–\[Ex:4\]. For the other forms, we used slightly different techniques to derive elliptic K3 surfaces designed for those particular forms. These will be sketched in Examples \[Ex:5\]–\[Ex:8\]. The transcendental lattices are computed using the discriminant form as before.
\[Ex:1\] In the family $X_{\lambda, \mu}$ we choose $\lambda$ to merge fibers of type $I_0^*$ and $I_1$. A general member $X$ of the resulting family has $${\mathop{\rm NS}\nolimits}(X) = U + A_2 + A_4 + A_6 + D_5,$$ so $\rho(X)\geq 19$. For a specialization to be defined over ${\mathbb{Q}_{}}$, we furthermore need a rational cusp at a $I_1$ fiber, i.e. the cubic factor of $\Delta$ encoding the $I_1$ fibers in terms of $\mu$ must have a rational zero. This can be achieved by the following substitution: $$\mu=9\,{\frac { \left( \nu+1 \right) ^{3}}{5\,{\nu}^{3}+15\,{\nu}^{2}-5\,\nu+1}}.$$ Then the rational cusp gives $$\lambda = -{\frac { \left( \nu-3 \right) \left( 5\,{\nu}^{3}+
15\,{\nu}^{2}-5\,\nu+1 \right) }{ \left( 7\,{\nu}^{2}+1 \right) ^{2}}}$$ Algorithm \[Alg\] suggests several singular specializations over ${\mathbb{Q}_{}}$. Here we verify two of them. The corresponding newforms seem to occur in the family $X_\lambda$ from the previous section as well, but there the conjectural sections have double height because of the relation (\[eq:4\]).
[**Discriminant $\mathbf{-1155}$:**]{} Let $\nu=-3/5$. Then there is a section $P$ of height $\hat h(P)=11/4$. Its is given by $$u(t)=-\frac{2\cdot 27}{5^8\cdot 7\cdot 11^5\cdot 13} \,
(242\, t-585)\, (46060586\, t^3+422472710\, t^2+32588325\, t+8292375).$$ This singular K3 surface has discriminant $d=-1155$ and transcendental lattice $T(X)=[6,3,194]$.
[**Discriminant $\mathbf{-1995}$:**]{} Let $\nu=9/35$. Then there is a section $P$ of height $\hat h(P)=19/4$. Its is given by $$u(t) = -\frac{27\cdot 11^3}{2^5\cdot 5^{10}\cdot 7^7\cdot 53} \,
\dfrac{(784\, t-795)\,U(t)}{(8757\, t-9010)^2},$$ where $$\begin{aligned}
U(t) & = & 519278509294553530368\, t^5-2767640394056706623700\, t^4\\
&& {} +5908183745712577772625\, t^3 -6312492415348218806875\, t^2\\
&& {} +3374618170228790821875\, t-721947602876973103125.\end{aligned}$$
This singular K3 surface has discriminant $d=-1995$ and transcendental lattice $T(X)=[46,11,46]$.
For the remaining discriminants, we constructed single families of elliptic K3 surfaces. Here we give only the specializations in extended Weierstrass form $$\begin{aligned}
\label{eq:Weier}
X:\;\;\; y^2 \, = \, x^3 +A\,x^2 + B\,x+C.\end{aligned}$$
\[Ex:2\] We consider an elliptic K3 surface with singular fibers $I_3, I_6, I_{11}$ at $\frac 17, 0, \infty$: $$\begin{aligned}
A & = & \frac {25}{24}}\,{t}^{4}+
{\frac {293}{6}}\,{t}^{3}-{\frac {23645}{16}}\,{t}^{2}
\mbox{}+{\frac {1705}{12}}\,t-{\frac {1331}{384},\\
B & = & - \left( 7\,t-1 \right) {t}^{2} \left( 200\,{t}^{3}
+9276\,{t}^{2}-92442\,t+4477 \right),\\
C & = & 96\, \left( 7\,t-1 \right) ^{2}{t}^{4}
\left( 100\,{t}^{2}+4588\,t-15059 \right).\end{aligned}$$ It has a section $P$ of height $\hat h(P)=19/6$ and $x$-coordinate $$u(t) = -\frac{3}{2\cdot 7^2\cdot 11^5}\, t \,
(52734375\,{t}^{3}+538828125\,{t}^{2}-2025538427\,t+1004475087).$$ The transcendental lattice is $T(X)=[22,11,34]$ with discriminant $d=-627$.
\[Ex:3\] We consider an elliptic K3 surface with singular fibers $I_4, I_5, I_{11}$ at $1, 0, \infty$: $$\begin{aligned}
A & = & -{\frac {19487171}{3808800}}\,{t}^{4}
+{\frac {3674891}{13800}}\,{t}^{3}
\mbox{}-{\frac {247797}{80}}\,{t}^{2}
+{\frac {37743}{8}}\,t+{\frac {23805}{32}},\\
B & = & -\frac 25\,t \left( t-1 \right)
\left( 161051\,{t}^{3}-5251521\,{t}^{2}+16877745\,t+8212725 \right),\\
C & = & -152352\,{t}^{2} \left( t-1 \right) ^{2}
\left( 1331\,{t}^{2}-17526\,t-23805 \right).\end{aligned}$$ It has a section $P$ of height $\hat h(P)=13/4$ and $x$-coordinate $$u(t) = {\frac {15}{11776}}\, \left( t-1 \right)
\left( 44289025\,{t}^{3}-35970275\,{t}^{2}+11995075\,t-1058529 \right).$$ The transcendental lattice is $T(X)=[22,11,38]$ with discriminant $d=-715$.
\[Ex:4\] We consider an elliptic K3 surface with singular fibers $I_5, I_7, I_{8}$ at $1, 0, \infty$: $$\begin{aligned}
A & = & -{\frac {16807}{332928}}\,{t}^{4}-{\frac {490}{2601}}\,{t}^{3}
+{\frac {333135767}{13872}}\,{t}^{2}
\mbox{}-{\frac {275656745}{5202}}\,t+{\frac {603351125}{20808}},\\
B & = & -{t}^{2} \left( t-1 \right)
\left( t+50 \right) \left( 2401\,{t}^{2}-114044\,t+114244 \right),\\
C & = & -83232\,{t}^{4} \left( t-1 \right) ^{2}
\left( 343\,{t}^{2}+436\,t-3380 \right).\end{aligned}$$ It has a section $P$ of height $\hat h(P)=41/8$ and $x$-coordinate $$u(t) = {\frac {51^2}{2\cdot 5^4\cdot 7}}\,
{\frac {U(t)}{ \left( 20003760\,t+208409617 \right) ^{2}}},$$ where $$\begin{aligned}
U(t) & = & 94791757788196875\,{t}^{5}
-13440531435036024375\,{t}^{4}\\
&& {} + 311827388703362736750\,{t}^{3}
-2250368299914898266350\,{t}^{2}\\
&& {} + 5701998864279209056695\,t
-4700672234454567466251 .\end{aligned}$$ The transcendental lattice is $T(X)=[38,3,38]$ with discriminant $d=-1435$.
The four remaining discriminants require more work for two reasons. For three of them ($-1012, -3003, -3315$), the discriminant has two large prime factors. Hence representing one of them by a singular fiber of an elliptic fibration would be too restrictive for the other singular fibers in a one-dimensional family. Instead, we use two-dimensional families to find singular K3 surfaces over ${\mathbb{Q}_{}}$ with these discriminants. This approach will be sketched in the next subsection.
The final discriminant $d=-5460$ illustrates the constraints that we are facing: The Hilbert class field of ${\mathbb{Q}_{}}(\sqrt{d})$ has Galois group $({\mathbb{Z}}/2{\mathbb{Z}})^4$. Hence Theorem \[Thm:NS\] implies that a family of K3 surfaces with a specialization of discriminant $d$ over ${\mathbb{Q}_{}}$ has to be fairly complicated.
In fact, Algorithm \[Alg\] suggests that both one-dimensional families in Section \[s:fam\] and Example \[Ex:1\] admit a specialization with discriminant $d$ over ${\mathbb{Q}_{}}$. However, the conjectural section would have too big height for computations with Algorithm \[Newton\].
In other words, we must find the right balance between two competing constraints: a family of K3 surfaces that is not too complicated to parametrize, but admits the required Galois action; and a specialization with the given discriminant that is not too complicated to allow an explicit verification by Algorithm \[Newton\].
These complexity problems can be circumvented as follows. We let go the first step of guessing the candidate parameter for a singular specialization on a family of K3 surfaces over ${\mathbb{Q}_{}}$ by Algorithm \[Alg\]. Instead we apply Algorithm \[Newton\] directly to a family of elliptic K3 surfaces over ${\mathbb{F}}_p$. To start the algorithm, we need only determine all family members over ${\mathbb{F}}_p$ — i.e. a finite set. By point counting, we can already filter some surfaces with the test from Section \[s:tech\]. After finding an appropriate section on one of these surfaces, we increase the $p$-adic accuracy and finally compute a lift for both the surface and the section. Then we verify as before that the lifted surface has the prescribed configuration of singular fibers and that the section lifts to this surface.
\[Ex:5\] Using the above approach for $p=37$, we found an elliptic K3 surface $X$ with the following singular fibers: $$\begin{array}{|c||c|c|c|c|c|c|}
\toprule
\text{cusp} & \infty & 0 & 1 & \frac{7^3\cdot 11}{23}
& \alpha, \alpha^\sigma & -\frac{1}{125}\\
\midrule
\midrule
\text{fiber} & I_7 & I_5 & I_4 & I_3 & I_2 & I_1\\
\bottomrule
\end{array}$$ Here $\alpha, \alpha^\sigma$ are the roots of the polynomial $7625\,t^2-1367158\,t-57967$. The surface is given in extended Weierstrass form (\[eq:Weier\]) with coefficients $$\begin{aligned}
A & = & 3125\,{t}^{4}-784700\,{t}^{3}
-40778898\,{t}^{2}-18971036\,t-218491,\\
B & = & -2^{20}\cdot 3^5\,{t}^{2} \left( t-1 \right)
\left( 625\,{t}^{3}-151380\,{t}^{2}-1599171\,t+62426 \right),\\
C & = & 2^{38}\cdot 3^{10}\,{t}^{4} \left( t-1 \right) ^{2}
\left( 125\,{t}^{2}-29164\,t-17836 \right).\end{aligned}$$ It has a section $P$ of height $\hat h(P)=13/4$ and $x$-coordinate $$\begin{aligned}
u(t) & = & - \frac{2^{19}}{3^3\cdot 7^4} \, \left( t-1 \right)
\left( 45273407\,{t}^{3}-3678666\,{t}^{2}+168432\,t-5324 \right).\end{aligned}$$ See [@E-5460] for the details of the computation. The transcendental lattice is $T(X)=[42,0,130]$ with discriminant $-5460$.
Two-dimensional families {#ss:2 .unnumbered}
------------------------
For the remaining three discriminants, we work with two-dimensional families of K3 surfaces with $\rho\geq 18$ because of the two large prime factors of the discriminant. We determine a specialization with two independent sections $P, Q$. The discriminant is then computed in terms of the height pairing on the lattice [@ShMW].
On an elliptic K3 surface $X$, the height pairing of two sections $P, Q$ is by [@ShMW Thm. 8.6] $$\langle P, Q\rangle = 2 + (P.O) + (Q.O) - (P.Q)
- \sum_v {\mathop{\rm corr}\nolimits}_v(P,Q).$$ The height pairing involves intersection numbers in ${\mathop{\rm NS}\nolimits}(X)$ and correction terms according to the non-identity fiber components met by the sections. Since we consider only semistable elliptic fibrations at this time, we require only the correction term at a $I_n$ fiber (or root lattice $A_{n-1}$). Using the cyclical numbering of components such that $O$ meets $\Theta_0$, we have $${\mathop{\rm corr}\nolimits}_v(P, Q) = \frac{i\,(n-j)}n,\;\;\;
\text{if $P$ meets $\Theta_i$ and $Q$
meets $\Theta_j$ of a $I_n$ fiber}.$$ If necessary, we interchange $P$ and $Q$ or renumber the components so that $i<j$. In case $P=Q$, the height pairing specializes to $\hat h(P)$ in (\[eq:height\]).
The height pairing endows the group modulo torsion with the structure of a positive-definite lattice, the so-called * lattice* ${\mathop{\rm MWL}\nolimits}(X)$. Note that the discriminant of the lattice need not be integral if there are reducible fibers.
One way to derive the height pairing is via the orthogonal projection in ${\mathop{\rm NS}\nolimits}(X)\otimes{\mathbb{Q}_{}}$ with respect to zero section and fiber components. Hence the discriminant of an elliptic surface $X$ satisfies $$\begin{aligned}
\label{eq:MWL}
(-1)^{\text{rk}({\mathop{\rm MW}\nolimits}(X))} \left|{\mathop{\rm MW}\nolimits}(X)_\text{tor}\right|^2\,{\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(X)) =
{\mathop{\rm disc}\nolimits}({\mathop{\rm MWL}\nolimits}(X)) \cdot \prod_v {\mathop{\rm disc}\nolimits}(F_v).\end{aligned}$$ Here the product runs over all reducible fibers $F_v$ and involves the discriminants of the corresponding root lattices. We are now ready to exhibit the singular K3 surfaces for the remaining three discriminants. Due to the configuration of singular fibers, the groups are always torsion-free.
\[Ex:6\] We consider an elliptic K3 surface $X$ with singular fibers $I_3, I_5, I_{11}$ at $t=3861/28124$, $0$, and $\infty$: $$\begin{aligned}
A & = & 66\, \left( 351384\,{t}^{4}-372196\,{t}^{3}
+113098\,{t}^{2}-13539\,t+594 \right),\\
B & = & {\frac {2^7\cdot 11}{3}}\,t \left( 28124\,t-3861 \right)
\left( 47916\,{t}^{3}-44484\,{t}^{2}+9479\,t-594 \right) ,\\
C & = & 2^{11}\cdot 11\,{t}^{2} \left( 28124\,t-3861 \right) ^{2}
\left( 242\,{t}^{2}-193\,t+22 \right).\end{aligned}$$ The elliptic surface $X$ has sections $P, P^\sigma$ that are conjugate over ${\mathbb{Q}_{}}(\sqrt{23})$. We found them by an extension of Algorithm \[Newton\] using lattice reduction instead of the Euclidean algorithm to recognize algebraic numbers of degree greater than $1$. The sections have height $\hat h(P) = \hat h(P^\sigma) = 38/15$ and pairing $\langle P, P^\sigma\rangle=8/15$. Let $$U(t) = ( 16427508+3425424\, \sqrt{23} ) {t}^{2}
+ (2159201+450164\, \sqrt{23}) t
\mbox{}+293522+61224\, \sqrt{23}.$$ The $x$-coordinates of the sections $P$ and $P^\sigma$ are $$u(t) \, = \, 3\,{\frac { \left( 6- \sqrt{23} \right) ^{2}t
\left( 28124\,t-3861 \right) U(t) }{ \left( 392
\mbox{}+95\, \sqrt{23} \right) ^{2}}}$$ and the Galois conjugate of $u$. By (\[eq:MWL\]), the elliptic surface $X$ has discriminant $$\begin{aligned}
\label{eq:2-dim}
{\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}(X)) = - 3\cdot 5\cdot 11\cdot
\left|\begin{matrix}\langle P,P\rangle & \langle P, P^\sigma\rangle\\ \langle P, P^\sigma\rangle
& \langle P^\sigma, P^\sigma\rangle\end{matrix}\right| = -1012.\end{aligned}$$ We compute the transcendental lattice $T(X)=[22,0,46]$.
For the final two discriminants, we combined two-dimensional families with the mod $p$ approach for Example \[Ex:5\].
\[Ex:7\] We consider an elliptic K3 surface in extended Weierstrass form (\[eq:Weier\]): $$\begin{aligned}
A & = & 2197\,t^4 - 48516\,t^3 + 393636\,t^2 - 1208764\,t + 411600,\\
B & = & 216\,t \, (3\,t-25) \, (10\,t-67) \,
(13351\,t^3-200168\,t^2+834514\,t-411600),\\
C & = & 108^2\,t^2 \, (3\,t-25)^2 \, (10\,t-67)^2 \,
(81133\,t^2-641164\,t+411600).\end{aligned}$$ This surface has the following reducible singular fibers: $$\begin{array}{|c||c|c|c|c|c|}
\toprule
\text{cusp} & \frac{49}{4} & \frac{25}{3} & \frac{67}{10} & 0 & \infty\\
\midrule
\midrule
\text{fiber} & I_2 & I_3 & I_3 & I_6 & I_{7}\\
\bottomrule
\end{array}$$ There are orthogonal sections $P,Q$ of heights $\hat h(P)=11/6$ and $\hat h(Q)=13/6$. Their $x$-coordinates are $$\begin{aligned}
P: & u(t) = 27 \, t \, (3\,t/25-1) \, (31671\,(t/25)^2 - 675\,t - 6700),\\
Q: & u(t) = -\frac{4}{21} \,t \, (5984\,t^3 - 65032\,t^2 + 16442\,t - 949725).\end{aligned}$$ By (\[eq:MWL\]), $X$ has discriminant $-3003$. The transcendental lattice is $T(X)=[6,3,502]$.
\[Ex:8\] We consider an elliptic K3 surface $X$ in extended Weierstrass form (\[eq:Weier\]): $$\begin{aligned}
A & = & 1105 \, (152628125\,t^4-131340300\,t^3
+35302566\,t^2-4215388\,t+574821), \\
B & = & - 3200 \cdot 11^3 \cdot 13 \cdot 17^2 \, t \, (119\,t-26)
\, (741\,t-289)\\
&& \;\;\;\;\; \times\, (1795625\,t^3-1311895\,t^2+317027\,t-33813), \\
C & = & 80^3 \cdot 11^6 \cdot 13^2\cdot 17^2 \, t^2 \,
(119\,t-26)^2 \, (741\,t-289)^2 \, (27625\,t^2-16594\,t+2601).\end{aligned}$$ This surface has the following reducible singular fibers: $$\begin{array}{|c||c|c|c|c|}
\toprule
\text{cusp} & \frac{26}{119} & \frac{289}{741} & 0 & \infty\\
\midrule
\midrule
\text{fiber} &I_3 & I_4 & I_5 & I_{8}\\
\bottomrule
\end{array}$$ There are orthogonal sections $P,Q$ of heights $\hat h(P)=17/8$ and $\hat h(Q)=13/4$. Their $x$-coordinates are $$\begin{aligned}
P: & u(t) = 44^3 \, (88179\,t^2 - 58882\,t + 2415744/221),\\
Q: & u(t) = -\frac{5}{208} \, (741\,t-289)
\, (1960038171\,t^3+503840883\,t^2-379180711\,t+5454297).\end{aligned}$$ By (\[eq:MWL\]), $X$ has discriminant $-3315$. The transcendental lattice is $T(X)=[2,1,1658]$.
**Acknowledgements:** We thank K. Hulek for suggesting this problem. We benefitted greatly from many discussions with him and B. van Geemen. We also gratefully acknowledge partial funding from the National Science Foundation (first author, grant DMS-0501029) and Deutsche Forschungsgemeinschaft (second author, under research grants Schu 2266/2-1 and Schu 2266/2-2). This collaboration began at the Clay Mathematics Institute 2006 Summer School on “Arithmetic Geometry” at the Mathematisches Institut of Georg-August-Universität in Göttingen. Our thanks go to the organizers for the invitation.
[992]{} Artin, M., Swinnerton-Dyer, P.: *The Shafarevich-Tate conjecture for pencils of elliptic curves on $K3$ surfaces*, Invent. Math. [**20**]{} (1973), 249–266.
Beukers, F., Montanus, H.: *Fibrations of K3-surfaces and Belyi-maps*, preprint (2006); additional data available at\
http://www.math.uu.nl/people/beukers/mirandapersson/Dessins.html. Boyd, D. W., Kisilevsky, H.: *On the exponent of the ideal class group of complex quadratic fields*, Proc. Am. Math. Soc. [**31**]{} (1972), 433–436. Breuil, C., Conrad, B., Diamond, F., Taylor, R.: *On the modularity of elliptic curves over ${\mathbb{Q}_{}}$*, J. Amer. Math. Soc. [**14**]{} (2001), 843–939. Cassels, J. W. S.: *A Diophantine equation over a function field*, J. Austral. Math. Soc. [**25**]{}, Ser. A (1978), no. 4, 489–496. Deligne, P.: *Formes modulaires et représentations $\ell$-adiques*, in: Sém. Bourbaki 1968/69, no. 355 (Lect. Notes in Math. [**179**]{}), Springer-Verlag (1971), 139–172. Dieulefait, L., Manoharmayum, J.: *Modularity of rigid Calabi–Yau threefolds over ${\mathbb{Q}_{}}$* in: Yui, N., Lewis, J. D. (eds.): *Calabi–Yau Varieties and Mirror Symmetry* (Toronto 2001), Fields Inst. Comm. [**38**]{}, AMS (2003), 159–166. Elkies, N. D.: *The maximal rank of an elliptic K3 surface over ${\mathbb{Q}_{}}(t)$*, talk at conference on *Birational Automorphisms of Compact Complex Manifolds and Dynamical Systems* at Nagoya University, Aug 28, 2007; *Elliptic curves of high rank over ${\mathbb{Q}_{}}(t)$ and over ${\mathbb{Q}_{}}$*, in preparation. Elkies, N. D.: *Shimura Curve Computations Via K3 Surfaces of Rank at Least 19*, Lect. Notes in Comp. Sci. [**5011**]{} (ANTS VIII, 2008), Springer (2008), 196–211. Elkies, N. D.: *A singular K3 surface $S/{\mathbb{Q}_{}}$ with ${\mathop{\rm disc}\nolimits}({\mathop{\rm NS}\nolimits}_{{\,\overline{\!{\mathbb{Q}_{}}\!}\,}}(S)) = -5460$*, in preparation. Elkies, N. D., Schütt, M.: *K3 families of high Picard rank*, in preparation. Heath-Brown, D. R.: *Imaginary quadratic fields with class group exponent 5*, Forum Math. [**20**]{} (2008), no. 2, 275–283.
Heilbronn, H. A., Linfoot, E. H.: *On the Imaginary Quadratic Corpora of Class-Number One*, Quart. J. Math. [**5**]{} (1934), 293–301. Inose, H.: *On certain Kummer surfaces which can be realized as non-singular quartic surfaces in ${\mathbb{P}}^3$*, J. Fac. Sci. Univ. Tokyo Sect. IA Math. [**23**]{} (1976), no. 3, 545–560.
Inose, H.: *Defining equations of singular $K3$ surfaces and a notion of isogeny*, in: *Proceedings of the International Symposium on Algebraic Geometry* (Kyoto Univ., Kyoto, 1977), Kinokuniya Book Store, Tokyo (1978), 495–502. Livné, R.: *Motivic orthogonal two-dimensional representations of Gal$(\overline{{\mathbb{Q}_{}}}/{\mathbb{Q}_{}})$*, Israel J. of Math. [**92**]{} (1995), 149–156. Nikulin, V. V.: *Integral symmetric bilinear forms and some of their applications*, Math. USSR Izv. [**14**]{}, No. 1 (1980), 103–167. Pjateckiĭ-Šapiro, I. I., Šafarevič, I. R.: *Torelli’s theorem for algebraic surfaces of type K3*, Izv. Akad. Nauk SSSR Ser. Mat. [**35**]{} (1971), 530–572.
Ribet, K.: *Galois representations attached to eigenforms with Nebentypus*, in: Serre, J.-P., Zagier, D. B. (eds.), *Modular Functions of one Variable V* (Bonn 1976), Lect. Notes in Math. [**601**]{}, Springer (1977), 17–52. Šafarevič, I. R: *On the arithmetic of singular $K3$-surfaces*, in: *Algebra and analysis* (Kazan 1994), de Gruyter (1996), 103–108. Schütt, M.: *Elliptic fibrations of some extremal K3 surfaces*, Rocky Mountain Journal of Mathematics [**37**]{}, 2 (2007), 609–652. Schütt, M.: *CM newforms with rational coefficients*, to appear in Ramanujan Journal, preprint (2005), arXiv: math.NT/0511228. Schütt, M.: *Arithmetic of a singular K3 surface*, to appear in Michigan J. of Math., preprint (2006), arXiv: math.AG/0605660. Schütt, M.: *Fields of definition of singular K3 surfaces*, Communications in Number Theory and Physics [**1**]{}, 2 (2007), 307–321. Schütt, M.: *K3 surfaces with Picard rank 20 over ${\mathbb{Q}_{}}$*, preprint (2008), arXiv: 0804.1558. Shimada, I.: *Transcendental lattices and supersingular reduction lattices of a singular $K3$ surface*, preprint (2006), arXiv: math.AG/0611208. Shimada, I., Zhang, D. Q.: *Classification of extremal elliptic $K3$ surfaces and fundamental groups of open $K3$ surfaces*, Nagoya Math. J. [**161**]{} (2001), 23–54. Shioda, T.: *On elliptic modular surfaces*, J. Math. Soc. Japan [**24**]{} (1972), 20–59. Shioda, T.: *On the lattices,* Comm. Math. Univ. St. Pauli, [**39**]{} (1990), 211–240. Shioda, T., Inose, H.: *On Singular $K3$ Surfaces*, in: Baily, W. L. Jr., Shioda, T. (eds.), *Complex analysis and algebraic geometry*, Iwanami Shoten, Tokyo (1977), 119–136. Shioda, T., Mitani, N.: *Singular abelian surfaces and binary quadratic forms*, in: *Classification of algebraic varieties and compact complex manifolds*, Lect. Notes in Math. [**412**]{} (1974), 259–287. Tate, J.: [*Algebraic cycles and poles of zeta functions*]{}, in: [*Arithmetical Algebraic Geometry*]{} (Proc. Conf. Purdue Univ., 1963), 93–110, Harper & Row (1965). Taylor, R., Wiles, A.: *Ring-theoretic properties of certain Hecke algebras*, Ann. Math. [**141**]{} (1995), 553–572. Weinberger, P. J.: *Exponents of the class groups of complex quadratic fields*, Acta Arith. [**22**]{} (1973), 117–124. Wiles, A.: *Modular elliptic curves and Fermat’s Last Theorem*, Ann. Math. (2) [**141**]{}, No. 3 (1995), 443–551.
------------------------- -------------------------------------
Noam D. Elkies Matthias Schütt
Mathematics Department Department of Mathematical Sciences
Harvard University University of Copenhagen
1 Oxford Street Universitetspark 5
Cambridge, MA 02138 2100 Copenhagen
USA Denmark
[email protected] [email protected]
------------------------- -------------------------------------
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The observation of quantum phenomena in macroscopic mechanical oscillators [@Braginsky1992; @Schwab2005] has been a subject of interest since the inception of quantum mechanics. It may provide insights into the quantum-classical boundary, experimental investigation of the theory of quantum measurements [@Braginsky1992; @Bose1999; @Tittonen1999], the origin of mechanical decoherence [@Marshall2003] and generation of non-classical states of motion. Prerequisite to this regime are both preparation of the mechanical oscillator at low phonon occupancy and a measurement sensitivity at the scale of the spread $\Delta x$ of the oscillator’s ground state wavefunction. Over the past decade, it has been widely perceived that the most promising approach to address these two challenges are electro nanomechanical systems [@Schwab2005; @Cleland1998; @Naik2006; @LaHaye2004; @Regal2008; @Teufel2008], which can be cooled with milli-Kelvin scale dilution refrigerators, and feature large $\Delta x\sim 10^{-14}{\ensuremath{\,\mathrm{m}}}$ resolvable with electronic transducers such as a superconducting single-electron transistor [@Naik2006; @LaHaye2004; @Knobel2003], a microwave stripline cavity [@Regal2008] or a quantum interference device [@Etaki2008]. In this manner, thermal occupation as low as 25 quanta [@Naik2006; @Teufel2008] has been measured. Here we approach for the first time the quantum regime with a mechanical oscillator of mesoscopic dimensions–discernible to the bare eye–and 1000-times more massive than the heaviest nano-mechanical oscillators used to date. Imperative to these advances are two key principles of cavity optomechanics [@Kippenberg2008]: Optical interferometric measurement of mechanical displacement at the attometer level [@Arcizet2006; @Schliesser2008b], and the ability to use measurement induced dynamic back-action [@Dykman1978; @Gigan2006; @Arcizet2006a; @Schliesser2006] to achieve resolved sideband laser cooling [@Schliesser2008; @Regal2008] of the mechanical degree of freedom. Using only modest cryogenic pre-cooling to $1.65{\ensuremath{\,\mathrm{K}}}$, preparation of a mechanical oscillator close to its quantum ground state ($63\pm20$ phonons) is demonstrated. Simultaneously, a readout sensitivity that is within a factor of $5.5\pm1.5$ of the standard quantum limit [@Braginsky1992; @Caves1981] is achieved. Taking measurement backaction into account, this represents the closest approach to the Heisenberg uncertainty relation for continuous position measurements yet demonstrated. The reported experiments mark a paradigm shift in the approach to the quantum limit of mechanical oscillators using optical techniques and represent a first step into a new era of experimental investigation which probes the quantum nature of the most tangible harmonic oscillator: a mechanical vibration.'
author:
- 'A. Schliesser'
- 'O. Arcizet'
- 'R. Rivière'
- 'T.J. Kippenberg'
bibliography:
- 'C:/Dokume\\string\~1/Albert/Eigene\\string\~1/Literature/microcavities.bib'
title: 'Resolved-sideband cooling and measurement of a micromechanical oscillator close to the quantum limit'
---
The experimental setting of the present work is a cavity optomechanical system, which parametrically couples optical and mechanical degrees of freedom via radiation pressure. In the present case, toroidal microresonators are employed which exhibit (cf. Fig. 1) strong, inherent opto-mechanical coupling between high-quality factor ($Q>10^8$) optical whispering gallery modes and the mechanical radial breathing mode [@Kippenberg2005] (RBM), featuring high frequency (65 and 122 MHz for the resonators used in this work), and effective masses [@Pinard1999] on the order of 1-10 ng (cf. Fig. 1b). The quality factors of the RBM can reach values up to 80,000 if clamping losses are mitigated by modal engineering [@Anetsberger2008]. To achieve a regime of low mechanical oscillator occupancy we apply laser cooling to a cryogenically pre-cooled micromechanical oscillator with high frequency. Figure 1 shows a schematic of the experiment. A chip with micro-resonators is inserted into a Helium exchange gas cryostat. Piezoelectric actuators enable positioning of a tapered optical fiber used for evanescent coupling with a resolution sufficient to adjust the taper-toroid gap to critical coupling. The total optical loss through the cryogenic environment can reach values below $25\%$. Low pressure (0.1-50 mbar) Helium exchange gas is admitted into the sample chamber, thermalizing the sample with a heat exchanger through which ${}^4$He is pumped from a reservoir of liquid ${}^4$He. An exchange gas temperature of 1.65 K is achieved. Due to the low heat conductivity of glass, and possible light absorption, it is of prime importance to verify the thermalization of the mechanical oscillator to 1.65 K. To this end, we perform noise thermometry using the RBM. A low power ($<2 \mu$W) laser is tuned into resonance with a high-$Q$ optical mode. Fluctuations of the cavity radius–as induced by thermal excitation of the RBM–induce resonance frequency fluctuations of the cavity, which are imprinted as phase fluctuations on the laser light coupling back to the tapered fiber (cf. Figure 1c). A phase-sensitive detection scheme enables measurement of the Lorentzian displacement noise spectrum of the thermal (Brownian) motion of the RBM, characterized by its resonance frequency ${\ensuremath{\Omega_\mathrm{m}}}$, mechanical damping rate ${\ensuremath{\Gamma_\mathrm{m}}}$ and peak displacement amplitude $S_{xx}^{\mathrm{th}}({\ensuremath{\Omega_\mathrm{m}}})$ (cf. Methods summary). Figure 2a shows the resulting mechanical mode temperature as derived via the equipartition theorem from the independently calibrated noise spectra, where phase-sensitive detection was accomplished using the Pound-Drever-Hall technique. Importantly, the temperature of the sample follows the exchange gas, demonstrating that excellent thermalization is achieved, a key prerequisite for the experiments described from here on. For the 62 MHz sample thermalization to 1.65 K entails an initial average occupancy of $\langle n\rangle =k_\mathrm{B} T_\mathrm{RBM}/\hbar
{\ensuremath{\Omega_\mathrm{m}}}\approx 560$, while for the 122 MHz sample a low occupancy of $\langle n \rangle \approx 280$ is attained. Note that despite the modest pre-cooling to 1.65 K these occupancies are identical to those of a 1 MHz nano-mechanical oscillator thermalized to a dilution refrigerator temperature below 20 mK, emphasizing the significant advantage of working with high frequency oscillators.
Measuring the mechanical displacement associated with such a massive oscillator at low occupancies requires high sensitivity, in particular, since the mechanical quality factor of silica is reduced to $\sim2000$ at 1.65 K due to losses originating from phonon coupling to structural defect states [@Arcizet2009a] (note that damping by the exchange gas is negligible, and that the mechanical Q factor improves again at lower temperatures [@Pohl2002]). The required attometer-level sensitivity can (so far) only be achieved with optical transducers. Following our previous work, we employ homodyne spectroscopy [@Schliesser2008b] based on a quantum-noise limited titanium sapphire laser (in both amplitude and phase), which is resonantly coupled to WGM resonance in the vicinity of a wavelength of 780 nm. The laser’s phase shift introduced by the mechanical fluctuations are detected interferometrically, by comparison with a high-power (2-5 mW) optical phase reference (local oscillator). Frequency analysis of this signal yields the thermal noise displacement spectrum $S_{xx}^{\mathrm{th}}({\ensuremath{\Omega}})$, on top of a measurement background. Figure 2b shows data obtained from the RBM of a $55 \mu\mathrm{m}$-diameter micro-resonator cooled to T=2.4 K, or $\langle n \rangle \approx 770$. The background of this measurement is at a level of $~1.5\cdot
10^{-18}{\ensuremath{\,\mathrm{m}}}/\sqrt{\mathrm{Hz}}$, which is only a factor of $5.5\pm 1.5$-times higher than the standard quantum limit (SQL) [@Braginsky1992; @Caves1981], given by $\sqrt{S_{xx}^{\mathrm{SQL}}({\ensuremath{\Omega_\mathrm{m}}})}=\sqrt{\hbar/{m_\text{eff}}{\ensuremath{\Gamma_\mathrm{m}}}{\ensuremath{\Omega_\mathrm{m}}}}$ for a measurement at the mechanical resonance. This proves the counterintuitive notion, that measurements close to the standard quantum limit are possible, in spite of the here used strategy of using high frequency and comparatively massive oscillators (in contrast to nanomechanical systems).
In order to further decrease the number of thermal quanta of the mechanical oscillator we use cooling via radiation pressure dynamical backaction as predicted [@Dykman1978; @Braginsky2002] and recently experimentally demonstrated [@Gigan2006; @Arcizet2006a; @Schliesser2006]. Similar to the atomic physics [@Wineland1979] case, ground state cooling requires accessing the resolved sideband regime [@Schliesser2008; @Wilson-Rae2007; @Marquardt2007], which necessitates the mechanical oscillator frequency to exceed the cavity decay rate (i.e.${\ensuremath{\Omega_\mathrm{m}}}\gg\kappa$). This regime is moreover prerequisite for schemes such as two transducer quantum non-demolition (QND) measurements [@Braginsky1992; @Braginsky1996] or the preparation of a mechanical oscillator in a squeezed state of motion [@Clerk2008a]. Operation in the RSB regime as demonstrated in [@Schliesser2008; @Teufel2008] is accomplished using a cavity with a narrow resonance (5.5 MHz intrinsic decay rate and 9 MHz mode splitting), which is broadened to a $\kappa/2\pi\approx
19{\ensuremath{\,\mathrm{MHz}}}$-wide resonance due to fiber coupling (corresponding to a loaded finesse of $\sim 70,000$). The laser is subsequently tuned to the lower mechanical sideband, i.e. red-detuned by 65.2 MHz, the resonance frequency of this sample’s RBM. For this detuning the circulating power is reduced by a factor of $4{\ensuremath{\Omega_\mathrm{m}}}^2/\kappa^2$. At the same time, the sensitivity to mechanical displacements is slightly reduced. In the ideal case of a highly overcoupled cavity, with unity detection efficiency and no excess noise except for the laser’s intrinsic quantum noise, the imprecision noise spectral density, i.e. the background of the measurement caused by detection shot noise, is given by: $$S_{xx}({\ensuremath{\Omega}})=\frac{\hbar {\ensuremath{\omega}}}{16 g_0^2 \cdot P}
\left(\frac{\Delta^2+(\kappa/2)^2}{\kappa/2}\right)^2
\left(1+{\ensuremath{\Omega}}^2 \frac{{\ensuremath{\Omega}}^2+(\kappa/2)^2-2 \Delta^2}
{(\Delta^2+(\kappa/2)^2)^2+{\ensuremath{\Omega}}^2(\kappa/2)^2}\right)$$ where the opto-mechanical coupling $g_0=d{\ensuremath{\omega}}/dx$, and $d{\ensuremath{\omega}}/dx={\ensuremath{\omega}}/R$ in the present embodiment. Moreover, $R$ is the cavity radius, $P$ the launched input laser power, ${\ensuremath{\omega}}/2\pi$ the optical resonance frequency, $\Delta/2 \pi$ the detuning from the cavity resonance, and ${\ensuremath{\Omega}}/2\pi$ the analysis frequency. In the resolved sideband case ${\ensuremath{\Omega_\mathrm{m}}}\gg\kappa$, this expression simplifies to $S_{xx}({\ensuremath{\Omega_\mathrm{m}}})=\hbar {\ensuremath{\omega}}{\ensuremath{\Omega_\mathrm{m}}}^2/4 g_0^2 P$ at the mechanical resonance frequency, when detuned to the first (upper or lower) mechanical sideband $|\Delta|={\ensuremath{\Omega_\mathrm{m}}}$. This is only a factor of 4 higher than in the resonant readout case $\Delta=0$. We note the interesting result that this expression does not depend on optical finesse in the deeply resolved sideband regime.
As shown in previous work [@Teufel2008; @Schliesser2008], the laser detuned to the lower sideband leads to a significant reduction of the thermal occupation (i. e. cooling), as evidenced by the reduced area underneath the peaks associated with the oscillator’s thermal noise (cf. Fig. 3b). The underlying physical mechanism giving rise to cooling is enhanced anti-Stokes scattering into the cavity mode, whereby each scattering process annihilates a thermal phonon. In the resolved-sideband regime, ground state cooling is possible in principle, and the minimum occupation that can be reached is given by $\langle \tilde n
\rangle\cong\kappa^2/16{\ensuremath{\Omega_\mathrm{m}}}^2\ll1$ [@Wilson-Rae2007; @Marquardt2007]. As the laser cools the resonator out of equilibrium with the thermal bath (at temperature $T$), however, heating through the bath competes with laser cooling and leads to a final occupation of $\langle n_\mathrm{f}\rangle\approx
\frac{{\ensuremath{\Gamma_\mathrm{m}}}}{{\ensuremath{\Gamma_\mathrm{m}}}+\Gamma_\mathrm{cool}}k_\mathrm{B} T/\hbar {\ensuremath{\Omega_\mathrm{m}}}$, where ${\ensuremath{\Gamma_\mathrm{m}}}$ is the intrinsic damping rate and $\Gamma_\mathrm{cool}$ the laser induced cooling rate. In the case of the data shown in Fig. 3, a strong increase in the damping with a concomitant reduction of the thermal occupation can be observed. The highest attained total damping rate is $({\ensuremath{\Gamma_\mathrm{m}}}+\Gamma_\mathrm{cool})/2 \pi=370{\ensuremath{\,\mathrm{kHz}}}$, reached with a launched power of $\sim 0.2{\ensuremath{\,\mathrm{mW}}}$. Evaluation of the calibrated thermal noise spectrum reveals an effective mode temperature of $200\pm60 {\ensuremath{\,\mathrm{mK}}}$, which corresponds to an average occupation as low as $\langle n_\mathrm{f}\rangle=63\pm20$ quanta. This is the lowest reported occupancy for a cavity optomechanical cooling experiment reported to date; slightly lower occupancy, $\langle n_\mathrm{f}\rangle=25$, has only been attained in the context of conventional dilution refrigeration of nanomechanical oscillators [@Naik2006], albeit with a signal-to-background ratio well below unity. Back-action cooling techniques applied to nano-mechanical oscillators at milli-Kelvin temperatures achieve occupation numbers about one order of magnitude higher [@Naik2006; @Teufel2008].
Moreover, a key aspect of the reported experiments is operation in the resolved sideband regime. To illustrate the instrumental role of the RSB regime for accessing low phonon occupation number of mechanical oscillators, we compare the cooling run just described with a further, independent run with a smaller sample (${\ensuremath{\Omega_\mathrm{m}}}/2\pi=122
{\ensuremath{\,\mathrm{MHz}}}$), exhibiting significantly broader linewidth (fiber coupling broadened to $\kappa/2\pi=155{\ensuremath{\,\mathrm{MHz}}}$), as shown in Fig. 4. In this case, a significant deviation from the linear cooling behavior is observed, with higher temperatures measured than expected based on the cooling rate. It is possible to model this deviation by taking an intracavity-power dependent heating of the mechanical mode into account. Heating of the mechanical mode can be caused by both quantum [@Braginsky1992; @Genes2008] or classical fluctuations [@Schliesser2008] of the radiation pressure force. While the latter is ruled out in our experiment by the use of a low-noise laser source (quantum limited in amplitude and phase for the Fourier frequency of interest), the former effect cannot account for the heating due to its much smaller magnitude, as detailed below. Instead, the effect is attributed to heating via laser absorption. Both data sets exhibit heating of ca. 10 K/Watt of circulating optical power. While quantitative modeling would have to take characteristics of the optical mode and heat transfer in the gas-cooled sample into account, we note that similar values of laser induced heating were extracted from studies of the optical bistability at low temperature at a wavelength of 1.5 $\mu$m [@Arcizet2009a]. Indeed, operating in the resolved-sideband regime [@Schliesser2008] allows to mitigate this effect and enables the results demonstrated in this work, as well as future work aiming at achieving ground-state cooling, as it strongly reduces the circulating intracavity power.
It is also interesting to consider the reported measurements from the perspective of the theory of quantum measurement [@Braginsky1992; @Clerk2008]. For linear continuous measurements as those demonstrated here, the total measurement uncertainty arises from two intrinsic sources of noise: measurement imprecision and measurement backaction. Measurement imprecision arises from fluctuations at the output of the measurement device, which are not related to mechanical oscillator motion. In the case of an optical interferometric measurement as reported here, measurement imprecision can be reduced to the shot noise in the detection process (cf. eq. (1)). Measurement backaction, on the other hand, describes the perturbation of the mechanical oscillator by the process of the measurement. For a mechanical oscillator, this occurs in the form of a fluctuating force, characterized by a spectral density $S_{FF}({\ensuremath{\Omega}})$. Quantum mechanics poses strict limits on how small $S_{xx}({\ensuremath{\Omega}})$ and $S_{FF}({\ensuremath{\Omega}})$ can be; the product obeying $\sqrt{S_{xx}({\ensuremath{\Omega}})\cdot S_{FF}({\ensuremath{\Omega}})}\geq \hbar/2$ where an equality can be reached in the ideal case and in the absence of correlations of imprecision and backaction noise [@Braginsky1992]. While classical fluctuations (caused by laser heating or classical laser amplitude and phase noise) can increase the value of the measurement backaction, the lowest possible force noise for a coherent input is given by the quantum backaction (QBA): $$S_{FF}^{\mathrm{qba}}({\ensuremath{\Omega}})=\frac{\hbar}{2{\ensuremath{\omega}}}g_0^2 P_\mathrm{in}
\left(\frac{\kappa^2}{(\kappa/2)^2+\Delta^2}\right)
\left(\frac{1}{(\kappa/2)^2+(\Delta-{\ensuremath{\Omega_\mathrm{m}}})^2}+\frac{1}{(\kappa/2)^2+(\Delta+{\ensuremath{\Omega_\mathrm{m}}})^2}\right)$$ where $S_{FF}^{\mathrm{qba}}({\ensuremath{\Omega_\mathrm{m}}})=2 g_0^2 P_\mathrm{in} \hbar/{\ensuremath{\omega}}{\ensuremath{\Omega_\mathrm{m}}}^2$ in the deeply resolved sideband regime ($|\Delta|={\ensuremath{\Omega_\mathrm{m}}}\gg\kappa$). For the parameters of the measurement with $\langle n_\mathrm{f}\rangle=63\pm20$, we find $\sqrt{S_{FF}^{\mathrm{qba}}({\ensuremath{\Omega_\mathrm{m}}})}\approx 1
{\ensuremath{\,\mathrm{fN}}}/\sqrt{\mathrm{Hz}}$. For the present experiments, the dominant measurement backaction effect is the heating by laser absorption (as observed in Fig. 4), which increases the magnitude of the thermal Langevin force. Indeed, using the effective mechanical susceptibility (at resonance) of the laser-cooled oscillator $|\chi_\mathrm{eff}({\ensuremath{\Omega_\mathrm{m}}})|=({m_\text{eff}}({\ensuremath{\Gamma_\mathrm{m}}}+\Gamma_\mathrm{cool}){\ensuremath{\Omega_\mathrm{m}}})^{-1}$, we can retrieve the magnitude of the *total* force spectral density from the measurement of its displacement spectrum via $S_{FF}^{\mathrm{tot}}({\ensuremath{\Omega_\mathrm{m}}})=S_{xx}^{\mathrm{th}}({\ensuremath{\Omega_\mathrm{m}}})/|\chi_\mathrm{eff}({\ensuremath{\Omega_\mathrm{m}}})|^2$. It amounts to $\sqrt{S_{FF}^{\mathrm{tot}}({\ensuremath{\Omega_\mathrm{m}}})}\approx 8
{\ensuremath{\,\mathrm{fN}}}/\sqrt{\mathrm{Hz}}$ and is dominated by the thermal Langevin force associated with the toroid’s temperature. While only a part of this force is caused by the measurement process via the heating induced temperature increase, we can nonetheless calculate the product $\sqrt{S_{xx}({\ensuremath{\Omega_\mathrm{m}}})\cdot S_{FF}^{\mathrm{tot}}({\ensuremath{\Omega_\mathrm{m}}})}$ as an upper bound, of how closely the ideal case has been approached in these measurements. For this the conservative assumption that the entire force spectral density acting on the mechanical oscillator (including the thermal Langevin force) is caused by measurement backaction is used. Using $\sqrt{S_{xx}({\ensuremath{\Omega}})}\approx
1.4\cdot10^{-18} {\ensuremath{\,\mathrm{m}}}/\sqrt{\mathrm{Hz}}$ we thereby find a product of $\sqrt{S_{xx}({\ensuremath{\Omega_\mathrm{m}}})\cdot S_{FF}^{\mathrm{tot}}({\ensuremath{\Omega_\mathrm{m}}})}$ as low as $230 \cdot \hbar/2$. This represents a factor of 4 improvement compared to the closest approach made with a nano-electromechanical system [@Naik2006; @Clerk2008]. We expect that improvements on our results are readily attainable by operating more deeply in the resolved-sideband regime [@Schliesser2008], in a colder cryogenic environment (such as ${}^3\mathrm{He}$), and with reduced mechanical dissipation by employing e.g. crystalline resonators [@Ilchenko2004]. The properties of the system demonstrated here–a massive oscillator, prepared with very low average occupation, which is coupled to an ultra low loss fiber transport medium is pivotal for a variety of experiments and protocols involving photons and phonons, such as radiation pressure squeezing [@Fabre1994; @Mancini1994], entanglement [@Vitali2007] or QND measurements [@Clerk2008a; @Heidmann1997].
In summary, our result signal a paradigm shift in the access to the quantum regime of mechanical oscillators by demonstrating the clear feasibility to enter this regime using significantly more massive cavity optomechanical [@Kippenberg2008; @Kippenberg2007] systems and using modest pre-cooling; a regime so far only attained with nanomechanical oscillator thermalized to milli-Kelvin scale dilution refridgerator temperatures. As such, the reported experiments-which demonstrate preparation and sensitive readout of mechanical oscillator with few quanta-mark a first step into a new era of experimental investigation of mechanical systems in the quantum regime, which has applications ranging from fundamental predictions of quantum measurement theory tests of quantum mechanics to the generation of non-classical states of motion and study of mechanical decoherence.
![\[f:setup\] Cryogenic cooling and displacement measurements of a micromechanical oscillator. (A to C) A silica microtoroid is held in a 1.65 K-cold ${}^4$He atmosphere. The toroid supports both high-Q optical whispering gallery modes and a mechanical radial breathing mode, which are parametrically coupled to an optical resonance frequency shift induced by mechanical displacement. High-Q optical resonances are identified using a tunable external-cavity diode laser (ECDL). To probe the mechanical oscillator, the light input is switched to a low-noise Ti:Sapphire. A small fraction of the laser beam is sent into the cryostat and couples to the WGM by evanescent coupling from a fiber taper approached to the rim of the toroid. Balanced homodyne measurement of the laser phase as transmitted through the taper is implemented using a Mach-Zehnder fiber interferometer (phase plates and polarizing beam splitters are only schematically indicated). A modulation $x(t)$ of the radius of the cavity (C) induces a modulation of the phase of the light $\phi(t)$ emerging from the cavity. This phase shift is detected by comparison with a phase reference, derived from the same laser in a beam splitter followed by a balanced detector.](setup_v09){width=".8\linewidth"}
![\[f:setup\] Thermalization and probing of a micromechanical oscillator (A) Using noise thermometry, the temperature of the mechanical oscillator is determined as a function of the temperature of the cryostat (exchange gas). The noise temperature of the 62 MHz mode follows the cryostat temperature in a linear manner, down to an occupation of less than 1000 phonons. (B) In spite of the low phonon occupation (770 in the case of this 65.3 MHz oscillator) displacement monitoring with high signal-to-noise ratio is possible using optical techniques (at the attometer level in the present case). The optical power used to interrogate the mechanical oscillator was ca. $3 {\ensuremath{\,\mathrm{\mu W}}}$ (red trace), $10 {\ensuremath{\,\mathrm{\mu W}}}$ (green trace) and $100 {\ensuremath{\,\mathrm{\mu W}}}$ (blue trace). The measurement background is given by shot noise in the detection.](fig2_v02){width="\linewidth"}
![\[f:setup\] Cryogenic precooling and resolved-sideband laser cooling. (A) Combined with pre-cooling in the cryostat to $1.65{\ensuremath{\,\mathrm{K}}}$, resolved-sideband cooling reduces the occupation of the oscillator further to $\langle n_\mathrm{f} \rangle=63\pm 20$ phonons. The inset illustrates the laser (blue line) detuned to the lower mechanical sideband of the cavity spectrum (red line). (B) Displacement noise spectra of the RBM with four different average occupation numbers, together with Lorentzian fit, at room temperature (red curve) after cryogenic cooling to 10K (green curve), and after additional resolved sideband cooling (blue curves). The sharp calibration peak, and a second mode at slightly lower frequency are also shown. ](GF_all_new_v01){width="\linewidth"}
![\[f:setup\]Resolved sideband cooling of (A) a RBM at 65.1 MHz and (B) an RBM at 121.7 MHz frequency. The mode temperature and corresponding phonon occupation are reduced as the detuned laser induces additional damping, and therefore increases the linewidth of the thermal noise spectrum. Full points correspond to mechanical spectra taken with a detuned cooling laser, open points in (A) correspond to measurements with the laser tuned close to the optical resonance. The deviation from a linear cooling behaviour (blue line) indicates a backaction heating effect, which is compatible with heating of the structure originating from residual absorption (red dashed line). This effect is suppressed in the resolved sideband regime (panel (A)).](results_new_v06){width="\linewidth"}
This work was supported by an Independent Max Planck Junior Research Group of the Max Planck Society, the Deutsche Forschungsgemeinschaft (DFG-GSC) and a Marie Curie Excellence Grant, and the FP7 project “MINOS”. O. A. acknowledges funding from a Marie Curie Grant (project QUOM). Thomas Becker is gratefully acknowledged for support with the cryogenic experiments, and Jörg Kotthaus for sample fabrication.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper we present the first baseline results for the task of few-shot learning of discrete embedding vectors for image recognition. Few-shot learning is a highly researched task, commonly leveraged by recognition systems that are resource constrained to train on a small number of images per class. Few-shot systems typically store a continuous embedding vector of each class, posing a risk to privacy where system breaches or insider threats are a concern. Using discrete embedding vectors, we devise a simple cryptographic protocol, which uses one-way hash functions in order to build recognition systems that do not store their users’ embedding vectors directly, thus providing the guarantee of computational pan privacy in a practical and wide-spread setting.'
author:
- |
Roei Gelbhart\
School of Computing and Information Systems\
University of Melbourne\
`[email protected]`\
Benjamin I. P. Rubinstein\
School of Computing and Information Systems\
University of Melbourne\
`[email protected]`\
bibliography:
- 'citations.bib'
title: 'Discrete Few-Shot Learning for Pan Privacy'
---
Introduction
============
In few-shot learning a classifier is trained to learn previously unseen classes given very few instances per class, motivated by humans who successfully few-shot learn a diverse range of tasks. For example, given a single image of a person, we are able to recognise other images of the same person; we can identify new animals given a single image of them; learn new letters in alphabets we’ve never seen before; we can easily identify fingerprints by matching patterns of skin lines to a reference fingerprint, when granted access to a reference database. Few-shot learning is of significant interest to industry. For example, security systems that rely on users’ bio-metrics for recognition can only practically receive a few instances per person upon enrolment. In support of these applications, great strides have been made recently in the field of few-shot learning [@hoffer2015deep; @vinyals2016matching; @PrototypicalCode; @finn2017model; @ravi2016optimization; @sung2018learning; @chen2019closer; @tian2020rethinking]. However in processing sensitive user data, [*e.g.*,]{}for user recognition or access control, data privacy must be addressed. For more information about the security of authentication systems (both password and bio-metric based), we refer the reader to [@van2020computer Chapter 3].
We focus our attention on two questions:
#### Question 1 - Targeted Data Reconstruction.
Given access to the few-shot learning system, can an attacker recreate users’ original data, [*e.g.*,]{}fingerprints, irises, etc.? We note multiple ways in which an attacker could achieve such access: by hacking into the system; via an offline system that stores data at endpoints; or a government warrant that compels the service provider to release confidential user data. The notion of pan privacy introduced by Dwork [*et al.*]{} [@dwork2010pan], refers to algorithms which can maintain privacy even if their inner state is visible to an attacker. Could neural network-based systems be made pan private? This question can be motivated by extant attacks, for example the demonstration of Hitaj [*et al.*]{} [@hitaj2017deep] in which training images of faces can be regenerated from trained models using GANs.
We take inspiration from modern authentication systems which do not store user passwords directly, but instead only a one-way hash [@goldwasser2015lecture] once salted. This ensures that even an administrator of the system cannot gain access to user passwords.
#### Question 2 - Indiscriminate Model Inversion.
Consider a model that is robust to attacks of the type demonstrated in [@hitaj2017deep]: a user’s training fingerprint cannot be recovered *directly* from a learned model. If the output of the model is not protected, however, then outputs of the model— [*e.g.*,]{}a user’s feature vector in the few-shot learning setting—are possibly susceptible to adversarial examples [@biggio2013evasion] as produced by the FGS attack [@goodfellow2014explaining] for instance. Even for cases in which the attacker can not modify directly the data, such as in airports where a security officer might be overseeing data entry, we see from the work of [@athalye2017synthesizing], that it is possible to create 3D objects as adversarial examples. So a fake fingerprint, or even facial accessories, might be enough to fool the net. While different to sensitive original data, valid and sensitive inputs could still be created that result in target outputs. Can we protect the outputs of our models?
**Related Work.** We observe two kinds of related past work on few-shot learning: (i) Works that rely on learning a transformation from the data space to a Euclidean space, and subsequently classifying based on proximity; (ii) Approaches not based on Euclidean embeddings. For (ii), we highlight some impressive results, such is the work of [@finn2017model; @finn2018probabilistic] on meta-learning, in which neural networks are trained on a variety of learning tasks, so as to adapt to new types of learning tasks with only a few SGD steps. Sung [*et al.*]{} [@sung2018learning], train a first model to produce an embedding vector which they concatenate to the embedding vectors of each candidate in the target class. The second model—a relation module—receives these pairs of embedding vectors as input, and predicts similarity. The approach of our work is to hide the embedding vectors using one-way hash functions, and thus render impossible, use of first model outputs as input to a second neural network. For a good survey over these methods, we refer the reader to [@chen2019closer].
Generally, it seems difficult to ensure theoretical privacy guarantees of highly accurate models which retrain over private data, as neural networks are effectively capable of remembering training data. We thus focus only on approaches of type (i), which train the model solely on public data in the training stage, and then once private users’ data is received, the model is unchanged. Such work can be seen in [@hoffer2015deep; @koch2015siamese; @snell2017prototypical; @ye2018deep]. The main idea of such approaches is to train the model to transform vectors from data space to feature space. The loss function (detailed in Section \[sec\_train\_cnn\]) will encourage the model to bring closer instances from same classes, and create a margin between instances from different classes.
To protect data we design a privacy-preserving algorithm (Section \[sec\_hashing\]), which utilises one-way hash functions . One-way hash functions require exact matches in the input of the function for any type of proximity in the output. This requires that the output of the model be discrete. Work on discrete neural networks has been previously explored in separate contexts [@Liu_2016_CVPR; @Lai_2015_CVPR; @NIPS2017_7210]; to the best of our knowledge, ours is the first related to few-shot learning, in which the neural network has never seen a class it tries to hash. Lai [*et al.*]{} [@Lai_2015_CVPR] evaluate precision-recall curves for hamming radius of up to two. However, for few-shot learning, achieving such low radii is extremely difficult—as experimentally demonstrated in Section \[sec\_train\_cnn\]. Even if a model can achieve such low distances, in order to protect the embedding vectors using one-way hash functions, we must achieve perfect matching for the same classes. To address this problem we introduce a probabilistic algorithm we term *Random Coordinate Projection Hashing (RCPH)*, that can utilise models with even large hamming distances.
**Our Contributions** are summarised as: i) We design a private few-shot recognition algorithm using one-way hash functions. We introduce the RCPH algorithm which enables the use of imperfect matching neural networks. ii) We offer first experimental results for discrete few-shot learning. Finally we analyse the accuracy of RCPH coupled with the learned model.
Privacy-Preserving Hashing Algorithm {#sec_hashing}
====================================
Denote by $f(x)$ a neural network which associates vectors from image space $\mathcal{X}$, into a discrete feature space $\mathcal{V}$. In our experiments, we choose $\mathcal{V} = \left \{0,1 \right \}^{1024}$. $f$ is trained to bring closer (in hamming distance) instances of the same class, while separating instances of different classes. In this way, proximity in the image of $f(\cdot)$ may be used to *match* pairs of input instances. If the trained net is able to achieve perfect matches, [*i.e.*,]{}zero hamming distance with high probability only for members of the same classes, then the task of preserving privacy is straightforward as described in Section \[subsec\_Perfect\_matching\_nets\]. However, this condition is extremely difficult to guarantee. Accordingly we present in Section \[subsec\_non\_matching\_nets\] a probabilistic algorithm named *Random Coordinate Projection Hashing (RCPH)*, which w.h.p effectively matches non-zero hamming-distanced vectors, provided there is a large margin between correct instance distance, and incorrect instance distances.
[**One-Way Function** ([@goldwasser2015lecture Definition 2.2]).]{} A one-way function, $h(x)$, is a function for which, for all $x$, $h(x)$ can be computed in probabilistic polynomial time (PPT), but for every PPT algorithm, given $y=h(x)$, the probability of finding any source of $y$, $z$ for which $f(z)=y$ is negligible. Examples of one-way functions can be found in [@goldwasser2015lecture].
Perfect Matching Nets {#subsec_Perfect_matching_nets}
---------------------
Denote by $h(v)$ a one-way hash function, from $\mathcal{V}$ to a hashing space $\mathcal{H}$. A simple privacy-preserving technique is the following: upon enrolment of a new user with input $x_i$ to the system (learning a new class $y_i$), the system saves $h(f(x_i))$, instead of $f(x_i)$. This provides a form of ‘computational privacy’ provided that the original embedded $f(x_i)$ cannot be generated with computational efficiency. Such an approach would therefore provide pan privacy as its internal state would not reveal sensitive data. Upon testing a given instance $x'$, we search for an exact match in $\{h(f(x_1)),\ldots,h(f(x_{i_i}))\}$ for $h(f(x'))$, which likely exists only if the network model $f(\cdot)$ returns exact matches w.h.p.
**Double hashing with ZKP verification.** Consider a setting where the system is deployed locally on many end-point machines. In such a case, for the sake of pan privacy, we should assume that $h(f(x_i))$ is publicly known. While an attacker may not be able to reconstruct $x_i$ or even $f(x_i)$ given $h(f(x_i))$, they might be able to break into another machine using $h(f(x_i))$, posing as someone else (the user who enrolled $x_i$ in the first place). For example, suppose that the protocol is to compute $f(x_i)$ on a local endpoint, and then $h(f(x_i))$ is sent to the server. If an intruder knows $h(f(x_i))$ (assuming it is public), they can directly send $h(f(x_i))$ to the system and login, without knowing $x_i$ or $f(x_i)$. To prevent this from happening, we modify the system, to save $h^{2}(f(x_i))$, as the user’s ID for matching (which will be public), and $h(f(x_i))$ as a secret password. The password will not be saved anywhere, but only a zero-knowledge proof (ZKP) verifier will be stored, such as the one described in [@bellare2002gq; @kiefer2014zero]. Upon accessing the system, the users will identify themselves with $h^{2}(f(x_i))$, and authenticate by proving to the verifier that they know $h(f(x_i))$. ZKP protocols [@goldwasser1996lecture; @goldwasser1989knowledge; @kiefer2014zero] can verify correctness of the password without storing or transmitting any information about the password itself.
Imperfect Matching Nets {#subsec_non_matching_nets}
-----------------------
Training perfect matching nets, that simultaneously achieve high accuracy, is a significant challenge that remains open. Thus, we take a different tact, motivated by the following example that demonstrates why perfect matching is not a necessity.
\[eg:distance\_vector\] Consider the matrix of a net’s distances, as shown in Equation . We describe the net, applied to the Omniglot data set [@lake2011one], in Section \[sec\_train\_cnn\]. Each row of the matrix is the distance of a row-specific query instance, to each of the 10 classes’ anchors—the hashed embeddings that would be enrolled by new users of the system, as described in the previous section. The bold distances represent the distance to the correct class for the query instance. We note that most of our experiments were performed with 20-way tasks, [*i.e.*,]{}with 20 options per query, but for compactness we illustrate just 10 here. $$M_{\mbox{dist}}=
\begin{pmatrix}
\textbf{7} & 35 & 50 & 28 & 34 & 45 & 28 & 31 & 49 & 37 \\
36 & 50 & 47 & \textbf{15} & 37 & 48 & 25 & 46 & 36 & 42 \\
36 & 25 & \textbf{30} & 38 & 26 & 39 & 21 & 30 & 32 & 26
\end{pmatrix}
\label{distance_vector}$$ A system that stores embedding vectors without hashing, can calculate this distance matrix and return a nearest neighbour, which for the first two rows happens to be the correct class. However, a system that stores only hashes cannot. One-way hash functions send close vectors in the domain to arbitrary vectors in the co-domain, that do not preserve distance—an important distinction with *locality-sensitive hashing* [@indyk1998approximate] that serves approximate nearest neighbour search but that does not protect privacy. We can calculate these matrix rows when estimating the successes rate of our algorithm on a test set, but upon system deployment, this matrix is unknown and any effective system must make decisions without it.
**Random Coordinate Projection Hashing (RCPH)** is detailed in Algorithm \[PM\_algorithm\] as the main algorithm of this paper. RCPH has two parameters $p,m$ that determine its accuracy, time and space complexities, and hashing space preservation. For this part of the paper we consider the following as inputs also: the neural network, the calculated anchors for all the classes (described in Section \[sec\_train\_cnn\]), the size of the feature space $n$, and the number of classes $k$.
The algorithm iterates over hashes of partial combinations of bits from the embedding vector of a given input vector $x$, to the same combinations of bits from the anchors of the classes that were calculated at training time. Parameter $0<p\leq1$ is the portion of bits to match such that the number of bits that are randomly chosen at each iteration is $\lfloor p \cdot n \rfloor$. Here $m$ is the maximum number of iterations that the algorithm runs for. If after $m$ iterations no match is found, the algorithm chooses to abstain, thus limiting its time complexity. Each query can lead to one of three outcomes: correct match, wrong match, or abstention. In our analyses we will bound from below the average correct match rate, bound from above the average incorrect match rate, and bound from above the average time complexity.
In the preprocessing stage $m$ random combinations of size $\lfloor p \cdot n \rfloor$ from $\{1,2,...,n\}$ are drawn, $C=\{c_1,c_2,...,c_m\}$, while $m$ different hash functions $H=\{h_1,h_2,...,h_m\}$ are generated. For each anchor of a class that was registered into the system, the algorithm stores $m$ hashes, choosing bits according to $C$. In iteration $i$, we compare a hash of bits $c_i$ from the query vector $f(x)$, denoted by $h_i(f(x)|c_i)$ to the same choice $c_i$ of bits from all the anchors, $\{ h_i(a_1|c_i),h_i(a_2|c_i),..,h_i(a_k|c_i) \}$, where $\{a_1,a_2,...,a_k\}$ denotes the anchor set. That is, we make comparisons of hashed random coordinate projections, as in $h(v|c_i) = \left(\left(h(v)\right)_{c_{i1}},\ldots,\left(h(v)\right)_{c_{i\lfloor p\cdot n\rfloor}}\right)$. For each anchor $a_j$, we pre-compute during preprocessing the set $A_j = \{ h_1(a_j|c_1),h_2(a_j|c_2),..,h_i(a_j|c_m) \}$, caching them all in $A=\{A_1,A_2,...,A_k\}$ for use during algorithm execution. We note that there could be two anchors that have the same partial hash for some $c_i \in C$. If that is the case, it means that either those two anchors are very close in embedding space (usually because of bad data, or a weak neural network), or that we have observed an unlikely collision event. In either case, the potential damage is quickly mitigated through randomising another coordinate projection. We emphasise that *only hashes of partial embedding vectors are stored, thereby maintaining pan privacy.*
RCPH Analysis {#sub_RCPH_Analysis}
-------------
Despite the fact that the distance vector (as illustrated by Example \[eg:distance\_vector\]) of a query is unknown during activation, we can evaluate the accuracy of the algorithm using these vectors during test time. For example, for the given vector which we denote by $v$, and the correct label index, which we denote by $y$, we know that the probability of finding the correct match in an iteration, denoted by $\emph{E}_{c}$, is $$\label{correct}
\Pr( \emph{E}_{c}) =\frac{\binom{n-v(y)}{\left \lfloor pn \right \rfloor}}{\binom{n}{\left \lfloor pn \right \rfloor}}\approx (1-p)^{v(y)}\enspace.$$ The approximation is valid when $n \gg v(y)$, but we do not use it in our code. The probability of having a wrong match in an iteration ($\emph{E}_{w}$) is bounded by the union bound, $$\label{wrong}
\max_{i\neq y} \frac{\binom{n-v(i)}{\left \lfloor pn \right \rfloor}}{\binom{n}{\left \lfloor pn \right \rfloor}}
\leq
\Pr( \emph{E}_{w}) \leq \sum_{i\neq y} \frac{\binom{n-v(i)}{\left \lfloor pn \right \rfloor}}{\binom{n}{\left \lfloor pn \right \rfloor}}\enspace,$$ and the probability of having no match in an iteration, denoted by $\emph{E}_{\emptyset}$ is similarly bounded as $$\label{no_match}
1-\max_i \frac{\binom{n-v(i)}{\left \lfloor pn \right \rfloor}}{\binom{n}{\left \lfloor pn \right \rfloor}}
\geq
\Pr( \emph{E}_{\emptyset}) \geq 1 - \sum_{i} \frac{\binom{n-v(i)}{\left \lfloor pn \right \rfloor}}{\binom{n}{\left \lfloor pn \right \rfloor}}\enspace.$$
We noted before that during the pre-processing stage, we regenerate any random combination that created a collision between two anchors hashes, which should be in a good neural network, very rare. However, this will clearly complicate the analysis, and thus for simplicity, we assume that collisions can happen, and simply bound the probability of a single correct match ($\emph{E}_{sc}$) by $$\label{one_match}
\Pr(\emph{E}_{sc}) \geq \Pr(\emph{E}_{c}) - \Pr( \emph{E}_{\emptyset})\enspace,$$ and conclude that $$\begin{aligned}
\Pr( \text{RCPH Correct}) = \sum_{i=0}^{m-1} \Pr(\emph{E}_{sc}) \cdot \Pr( \emph{E}_{\emptyset})^i \label{algo_success} \\
= \Pr( \emph{E}_{sc}) \frac{1-\Pr( \emph{E}_{\emptyset})^m}{1-\Pr(\emph{E}_{\emptyset})} \nonumber\enspace. \end{aligned}$$ Similarly, the wrong match probability can be bounded from above by $$\label{algo_fail}
\Pr( \text{RCPH False}) = \Pr( \emph{E}_{w}) \frac{1-\Pr(\emph{E}_{\emptyset})^m}{1-\Pr(\emph{E}_{\emptyset})}\enspace.$$
The time complexity of RCPH is bounded by $m$, but given the random nature of the algorithm, we can calculate its average case complexity. For each iteration, if we use a hash table to look for matches, the average time complexity is $O(1)$. The probability of finding any match in one iteration is at least $$\lambda = \max_i \frac{\binom{n-v(i)}{\left \lfloor pn \right \rfloor}}{\binom{n}{\left \lfloor pn \right \rfloor}}\enspace,$$ thus, by the expectation of a geometric random variable, the desired expectation is bounded by, $$\label{time_complexity}
\mathbb{E} T \leq \min (1/\lambda, m)\enspace.$$ According to the law of total expectation we average over the expectations of all query points to obtain an estimate of the average-case time complexity. Using , for each query we can bound from below the success rate shown in Equation , and from above the fail rate and average time complexity, shown in Equations , . Averaging over these values for an entire query set provides performance bounds of the system.
[0.45]{}[c \*[2]{}[>X]{}]{} Accuracy & Fail Rate & Average Complexity\
(lr)[3-3]{}(lr)[1-1]{}(lr)[2-2]{} 0.9999 & 0.0000 & 130.66\
0.2402 & 0.0002 & 10,000\
0.0000 & 0.0045 & 10,000\
[0.45]{}[c \*[2]{}[>X]{}]{} Accuracy & Fail Rate & Average Complexity\
(lr)[3-3]{}(lr)[1-1]{}(lr)[2-2]{} 0.9995 & 0.0000 & 130.66\
0.9345 & 0.0008 & 36,359\
0.0000 & 0.0445 & 100,000\
In Tables \[table1\] and \[table2\] we calculate the performance bounds for the three query points showed in the distance matrix of Example \[eg:distance\_vector\]. The first point is the best one, as we can see that the correct label is within hamming distance 7. As demonstrated in our experiments, this is very much the common case. The second query is more borderline as we can see that the correct label is within distance 15, while the second best is relatively close in 25. We see that for $m=10000$ the accuracy is very poor, but if we are willing to pay with higher time complexity, the accuracy increases, and naturally so does the failure rate—more iterations mean higher chances for matching. We note that the numbers do not add up to one as they are bounds—one upper and one lower bound—while in addition the algorithm may abstain.
Salting {#Salting}
-------
A common practice in authentication systems is password salting. Given a user which has an ID and password, a random vector of bits is either added or concatenated to the password before it is hashed. Each user has their own salt vector, ensuring that users with identical passwords will have distinct hashed passwords after salting. If one uses large salt vectors, brute-force attacks need to be done per user, instead of over an entire password database. This mitigates amortising rainbow-table attacks.
In our case, salting appears infeasible. The embedding vector serves both as the identity and the password of the user, and thus we can not assign different salt values per user. However, if we are willing to relax that requirement, salting becomes a possibility. For example, consider a system in which the user puts in their ID (their name for example), with their fingerprint serving as the password. Another option is using two neural networks in the process; the first will be used for ID recognition, and its data will be public (for example it will use the face of the user, or only a partial fingerprint), and the second will receive confidential data. For the first case we will use standard distance-based few-shot learning, and the data will not be protected, for the second case we use RCPH with salting.
When using salting, if we add a random binary vector of size $n$ to the feature vector, the distance between two different classes’ feature vectors becomes on average $2^{\frac{n}{2}}$, and the probability of misclassification (fail rate) for reasonable $p,m$ values becomes negligible. So can we reduce $p$ (the required matching portion of the feature vector), to increase the accuracy? The answer depends on the quality of the data and the net. The salt is public knowledge, and thus an attacker that tries to break into a specific account, can add the salt as well. Reducing $p$ means that the attacker needs to guess a shorter part of the feature vector. If the attacker possesses another data point which is similar to the user’s, they can brute-force over a small neighbourhood of that data point to hack in. Smaller $p$ means higher vulnerability to attacks. To reduce the vulnerability, we need to improve the net’s separation abilities.
Training Discrete Neural Networks for Few-Shot Learning {#sec_train_cnn}
=======================================================
**Data set.** We used the Omniglot data set [@lake2011one], and the work on prototypical networks of [@snell2017prototypical] as a base line. We leveraged code from an implementation of [@snell2017prototypical] available at [@PrototypicalCode]. Omniglot contains 1623 characters from 50 different alphabets; each character having 20 different instances, each drawn by a different person. We use the same data splits as in [@vinyals2016matching; @snell2017prototypical], which consists of 90 degree rotations (considered to be different classes), which makes for a total of 6492 characters. The train, validation and test sets are of sizes 4112, 688 and 1692 respectively, and do not share characters from the same alphabet.
{width="\linewidth"}
[Convolutional and fully connected blocks with same colours share weights. K in this example is 5, for one-shot learning is 1. ]{} \[model\_figure\]
**Loss function.** We used softmax over the negated distance vector, to calculate the probability of each class. The loss is the negative log-likelihood. The number of classes we compare each query to is a hyper parameter. We chose 40 for the training stage, and for testing, we used 5 and 20 (corresponding to 5-way and 20-way few-shot learning).
**Model and output discretization.** Snell [*et al.*]{} [@snell2017prototypical] introduced the idea of prototypical networks, which compute distances to prototype representations of each class, which are an average in feature space of examples in the support set of the class. The results from the original paper [@snell2017prototypical] are presented together with our results in Table \[final\_results\]. Our model, presented in Figure \[model\_figure\], is very similar to the model in [@snell2017prototypical], but with one major difference. Our goal is to encourage the net to output vectors which are close to binary vectors even before the final rounding, so as to ensure that the rounding of the vector will not significantly damage accuracy. In [@snell2017prototypical], the main strategy for few-shot learning (demonstrated with 5-shot learning), was to average the 5 anchor instances of the class to a single anchor in feature space. If we try to copy their idea, and the net was able to output close to discrete values for the queries, the average of the anchors will likely be far from discrete: it will be discrete only if they are all identical. To solve this problem, we introduce another strongly-connected layer, that receives as input the sum of the anchors, which is equivalent to their average (up to a factor). In order to share the parameters of the anchors’ net and query’s net, we perform a small modification for queries, where we multiply the output of the final convolutional layer, by the number of anchors per class, as seen in Figure \[model\_figure\].
Using strongly-connected layers, and a final sigmoid activation function, we are able to architecturally limit the values of the net output to be in $(0,1)^{1024}$. We refer to our architecture as *Discrete Prototypical Networks (DPN)*. During training, rounding will eliminate the derivative, thus we must train without it. It might be beneficial to regularise the weights of the net such that the output from the sigmoid layer will be close to either 0 or 1, so that the difference between the artificial rounding, and the actual net’s output will be as small as possible. To test this hypothesis, we added a regularization factor to the loss function, very similar to the one introduced in [@Liu_2016_CVPR], $\lambda \left \| f(x) - \lfloor f(x)\rceil \right \|_2^2$.
We trained the model with different $\lambda$ values, and as result, the average of the distance of the net’s output to the rounded output, which we refer to as “the discretization gap", reduced from **$0.09$** without regularization, to **$0.06$** with the largest $\lambda$. However, the accuracy of the model decreased as $\lambda$ grows larger (graphs in the Appendix). Interestingly, the value of the discretization gap even without regularization seems rather small. We examined the gap as function of training time, and notice that even without regularization it decreases from **$0.21$** at start, to **$0.09$** at finish. It seems that the vanishing gradient phenomenon, which is very common for sigmoid activation functions, actually serves our purpose well. When the output of the net is close to 0/1 the gradient of the sigmoid is small, and thus small changes are made to the net.
**Experiments.** We present our final results in Table \[final\_results\]. When dealing with security validation systems, $m$ of magnitude of $10^6$ is very reasonable, as the delay for the user is only upon entrance to the system, and with today’s computing power, it is still well beneath a second. The average complexity, is usually significantly lower. In Figures \[Accuracy\] and \[fail\_rate\] we can see how the accuracy and failure rate behave as a function of $p,m$ for our net. We emphasise that using different training methods or architectures, which regulate the average best hamming distance, can significantly change the location of the best $p$. In this work we focused purely on maximising the accuracy of the net before discretizing. This is not necessarily the best approach, especially if one desires higher $p$ values (for better security).
[l c \*[8]{}[>X]{}]{} & &\
(lr)[2-5]{}(lr)[6-9]{} & Acc. & Fail Rate & Comp-lexity & Best P & Acc. & Fail Rate & Comp-lexity & Best P\
(lr)[2-5]{}(lr)[6-9]{} **Prototypical Networks[@snell2017prototypical]** & 98.8% & - & - & - & 99.7% & - & - &-\
**DPN** & 97.8% & - & - & - & 99.4% & - & - &-\
**DPN+RCPH, $\mathbf{m=10^3}$** & 86.1% & 6.5% & 174& 0.07 & 94.2% & 2.6% & 82 &0.07\
**DPN+RCPH, $\mathbf{m=10^4}$** & 89.6% & 5.2% & 1182 & 0.09 & 95.9% & 1.8% & 547 & 0.09\
**DPN+RCPH, $\mathbf{m=10^5}$** & 93.9% & 3.7% & 3961 & 0.10 & 97.5% & 1.3% & 1924 & 0.10\
**DPN+RCPH, $\mathbf{m=10^6}$** & 94.5% & 3.0% & 42616& 0.12& 97.6% & 1.3% & 17152 &0.12\
[l c \*[8]{}[>X]{}]{} & &\
(lr)[2-5]{}(lr)[6-9]{} & Acc. & Fail Rate & Comp-lexity & Best P & Acc. & Fail Rate & Comp-lexity & Best P\
(lr)[2-5]{}(lr)[6-9]{} **Prototypical Networks[@snell2017prototypical]** & 96.0% & - & - & - & 98.9% & - & - &-\
**DPN** & 93.6% & - & - & - & 98.0% & - & - &-\
**DPN+RCPH, $\mathbf{m=10^3}$** & 68.9% & 18.4% & 270& 0.08 & 84.0% & 8.6% & 148 &0.08\
**DPN+RCPH, $\mathbf{m=10^4}$** & 77.0% & 13.9% & 1857 & 0.10 & 89.0% & 5.4% & 885 & 0.10\
**DPN+RCPH, $\mathbf{m=10^5}$** & 81.7% & 11.0% & 13704 & 0.12 & 92.5% & 4.1% & 5832 & 0.12\
**DPN+RCPH, $\mathbf{m=10^6}$** & 84.1% & 9.6% & 109302& 0.14& 93.4% & 4.0% & 27666 &0.13\
Without privacy - using nearest neighbour.
Computationally secure system. Given m, the ideal p is calculated over the validation set to maximise the accuracy, and then plugged in to the test set analysis.
Future Work
===========
As always for work that present baseline results for a new setting, our paper opens up a range of new directions for future research.
**Better Accuracy.** From Figures \[Accuracy\] and \[fail\_rate\] we see the trade-off between higher accuracy and lower failure rate. Reducing $p$ lowers the probability of falsely entering the wrong person into the system (and as consequence revealing part of their feature vector). At the same time, the accuracy rises as well (it stops rising due in large part to the fact that we have a lower bound on accuracy, not an exact value). If we want to improve the trade-off, we need to improve the net—lower hamming distance for the correct class, and larger distance to the second best.
**Attacks and Defences.** A brute-force attack against our model will require time complexity of $2^{n*p}$, which for our case is roughly $2^{1000*0.1} \approx 10^{30}$. However, an attacker that has access to data points from the same distribution, can investigate the outputs of the net to find correlation between bits, and thus reduce the brute-force search space. We call these “data dependent Attacks", and they pose a serious threat to the security of our algorithm. The attacker can alternatively start from a specific data point, and search for close neighbours in hamming space. Given enough data points, the actual space that needs to be covered is a lot smaller than $10^{30}$, and dependent on the proximity in hamming space of the attacker queried data to the targeted data point.
When dealing with one-way hash functions, the size of the input space determines the time complexity of brute force attacks. For our case, it is a subgroup of $\mathcal{V} = \left \{0,1 \right \}^{1024}$, but it is still not clear which points in $\mathcal{V}$ are reachable by the net from the data distribution, and what is the final probability distribution of the net’s output over the data. If for example, the net utilises only the first half of the bits, while the second half remains zero, or equal to the first half, then the output space is a lot smaller. How can we measure the randomness of the net’s output, and how can we encourage the net to reduce correlation between output bits? We leave these questions for future work.
We note that the randomness of the net’s output is limited not only by the net ability to discover features in the data, but also by the data itself; the dimension of the input data, and the number of actual features that are possible to detect.
Broader Impact {#broader-impact .unnumbered}
==============
Authentication systems based on computer vision and deep learning are prevalent. It is commonplace to use fingerprint or facial recognition to unlock our smartphones, while airport immigration is increasingly relying on bio-metric data to reduce waiting times in passport control. Despite progress in safe storage of user passwords, AI-based systems introduce new vulnerabilities. First, the data itself might be valuable for some user. Second, even the mere output of the net (without data reconstruction) can be used in an adversarial attack to hack into users’ accounts. In our work we present a first approach for a pan-private deep learning recognition system. We expect that research in this new direction will promote improved user privacy—a fundamental human right that should be afforded to citizens even when an authentication system is required for national security—and improved security—through improved safeguarding of credentials. We highlight data dependent attacks as a potential vulnerability and call for researchers to explore extensions that mitigate any risk from such attacks.
**Appendix**
Further Results on Discretization
=================================
In this appendix we explore the effect of training our model with varying $\lambda$ values. The results, which can be seen in Figure \[acc\_reg\], show little influence over the accuracy, and in fact, the more we regularise, the worse the accuracy becomes. Interestingly, we can see from Figure \[case1\], that even in a model without regularisation, that the average distance to the rounded output (the discretization gap), decreases with training in the first $2.5$ epochs (each epoch is $100$ batches, and it reduces to a minimum of $0.0053$ after $243$ batches). It appears that the vanishing gradient problem, which is very common for sigmoid activation functions, actually serves our purpose well. When the output of the net is close to 0/1 the gradient of the sigmoid is small, and thus only small changes are made to the net.
An interesting phenomenon is the slow increase of the discretization gap after $2.5$ epochs. This increase lasts for approximately $30$ epochs, while the accuracy has already reached $0.91$. The net, which is obviously oblivious to the discretization gap, begins to utilise values in $(0,1)$ which are further from the edges.
\[fig:supp\]
\[fig:supp\]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
First-order variational equations are widely used in $N$-body simulations to study how nearby trajectories diverge from one another. These allow for efficient and reliable determinations of chaos indicators such as the Maximal Lyapunov characteristic Exponent (MLE) and the Mean Exponential Growth factor of Nearby Orbits (MEGNO).
In this paper we lay out the theoretical framework to extend the idea of variational equations to higher order. We explicitly derive the differential equations that govern the evolution of second-order variations in the $N$-body problem. Going to second order opens the door to new applications, including optimization algorithms that require the first and second derivatives of the solution, like the classical Newton’s method. Typically, these methods have faster convergence rates than derivative-free methods. Derivatives are also required for Riemann manifold Langevin and Hamiltonian Monte Carlo methods which provide significantly shorter correlation times than standard methods. Such improved optimization methods can be applied to anything from radial-velocity/transit-timing-variation fitting to spacecraft trajectory optimization to asteroid deflection.
We provide an implementation of first and second-order variational equations for the publicly available [[REBOUND]{}]{}integrator package. Our implementation allows the simultaneous integration of any number of first and second-order variational equations with the high-accuracy IAS15 integrator. We also provide routines to generate consistent and accurate initial conditions without the need for finite differencing.
author:
- |
Hanno Rein$^{1,2}$, Daniel Tamayo$^{1,3,4}$\
$^1$ Department of Physical and Environmental Sciences, University of Toronto at Scarborough, Toronto, Ontario M1C 1A4, Canada\
$^2$ Department of Astronomy and Astrophysics, University of Toronto, Toronto, Ontario, M5S 3H4, Canada\
$^3$ Canadian Institute for Theoretical Astrophysics, 60 St. George St, University of Toronto, Toronto, Ontario M5S 3H8, Canada\
$^4$ Centre for Planetary Sciences Fellow
bibliography:
- 'full.bib'
date: 'Submitted: 14 February 2016, Accepted: 10 March 2016.'
title: 'Second-order variational equations for $N$-body simulations'
---
methods: numerical — gravitation — planets and satellites: dynamical evolution and stability
Introduction {#sec:intro}
============
Calculating the orbital motion of planets and predicting the position of planets in the night sky is one of astronomy’s oldest reoccurring tasks. Today this is considered a solved problem, a simple application of Newtonian physics. Typically the dynamical system is solved by numerically integrating forward in time using an $N$-body integrator. Different techniques are available to do this very accurately over both short [see e.g. @ReinSpiegel2015] and long [see e.g. @WisdomHolman1991; @ReinTamayo2015] timescales.
But often the solutions of even very simple dynamical systems are complex, in some cases exhibiting chaos. This means that small perturbations in the initial conditions lead to exponentially diverging solutions at late times. The solar system is one such chaotic dynamical system [@Roy1988; @SussmanWisdome1988; @LaskarGastineau2009]. One way to characterize chaotic systems is to numerically determine the Maximal Lyapunov characteristic Exponent (MLE), which measures the rate of exponential divergence between neighbouring trajectories in phase space.
Calculating how particle trajectories vary with respect to their initial conditions is therefore an important numerical task in modern celestial mechanics. It is also immediately relevant to orbital fitting and optimization. For example, when fitting an $N$-body simulation of a planetary system to data, one might want to calculate the derivative of the $\chi^2$ value with respect to a planet’s initial orbital eccentricity.
The MLE, or more generally any derivative with respect to initial conditions, can be calculated by running a separate $N$-body simulation with *shadow particles*, where the initial conditions of one or more particles are slightly perturbed. Measuring how fast the distance in phase space of the shadow particles with respect to their unperturbed counterparts grows then yields the MLE [@Benettin1976].
However, it is well known that there are problems associated with this shadow-particle method [@Tancredi2001]. On the one hand, we want to start the shadow particles close so that we obtain a *local* measure of the divergence of trajectories, and so that as the paths begin to drift apart, there are several decades over which to characterize the rate of divergence. On the other hand, the closer we put the shadow particles, the more digits we lose to numerical roundoff error. One workaround is to periodically rescale the separation vectors to keep shadow particles nearby their real counterparts [see, e.g., Sec. 9.3.4 of @solarsystemdynamics]. However, we show in Sec. \[sec:finitedifference\] that the use of shadow particles requires problem-dependent fine-tuning and that the problem is exacerbated when computing higher-order derivatives.
Luckily, instead of integrating a separate simulation of shadow particles, one can also use variational equations to measure divergences. Rather than differencing two nearly equal trajectories, one instead derives a new linearized dynamical system that directly evolves the small distance between two initially offset particles. These variational equations are scale-free and circumvent the numerical pitfalls associated with the shadow-particle method [@Tancredi2001].
First-order variational equations have been widely discussed and applied in the literature [e.g., @MikkolaInnanen1999; @Tancredi2001; @Cincotta2003]. In this paper we derive second-order variational equations for the $N$-body problem for the first time. These provide the second derivatives of the solution with respect to the initial conditions. Although mathematically straightforward to calculate, the number of terms and therefore the complexity rises significantly. As we will see below, some terms involve 7 different (summation) indices.
Our work opens up many new opportunities for a variety of applications. Perhaps most importantly, it is now straightforward to implement derivative-based optimization methods. While the first derivatives provide a gradient that yields the [*direction*]{} towards a local minimum on a $\chi^2$ landscape, the second derivatives provide the [*scale*]{} for how far one should move to reach the minimum. This can, among other things, dramatically improve fitting algorithms for radial velocity and transit planet searches, posterior estimation using Markov Chain Monte Carlo (MCMC) methods and even spacecraft trajectory optimization.
We begin in Sec. \[sec:derivation\] with a formal introduction to variational equations that generalizes to higher order. In Sec. \[sec:varnbody\] we specialize to the case we are interested in, the $N$-body problem. For completeness, in Sec. \[sec:derivation1\] we rederive the first-order variational equations for the $N$-body problem. We then go one step further in Sec. \[sec:derivation2\] and derive the second-order variational equations.
We have implemented the second-order variational equations within the [[REBOUND]{}]{}package and make them freely available. [[REBOUND]{}]{}is a very modular $N$-body package written in `C99` and comes with an optional `python` interface. We have abstracted the complexity of higher order variations significantly, and summarize our adopted syntax in Sec. \[sec:implementation\]. Obtaining consistent initial conditions for variational equations in terms of Keplerian orbital elements without relying on finite difference is non-trivial. We have therefore also implemented several convenience methods for this purpose.
In Sec. \[sec:tests\] we demonstrate how second-order variational equations and Newton’s method can be used to fit observational data to a dynamical $N$-body model. Finally, we compare variational and finite-difference methods in Sec. \[sec:finitedifference\], and conclude in Sec. \[sec:conclusions\] by outlining the next steps in using higher order variational equations efficiently in optimization problems and MCMC methods.
Derivation of differential equations {#sec:derivation}
====================================
Variational Equations {#sec:vareq}
---------------------
In this section, we define what we mean by variational equations and introduce our notation. We follow the work of [@MoralesRuiz2007] and start with an analytic differential equation of the form $$\begin{aligned}
\dot x = X(x).\label{eq:de}\end{aligned}$$ In the case that we are interested in later, $x \in \mathbb{R}^{6N}$ encodes the 3 position and 3 velocity coordinates for each particle. $X$ is then a vector field on $\mathbb{R}^{6N}$. The dot represents a time derivative. Given a suitable set of initial conditions $x_0$, an $N$-body simulation allows us to calculate (or at least approximate) the solution to Eq \[eq:de\]. We denote this solution $\phi(x_0,t)$, a $6N$ dimensional vector that depends on the initial conditions $x_0\equiv x(0)$ and time $t$.
Our goal is to estimate the solution vector $\phi$ for different initial conditions, i.e. we want to approximate $\phi(y_0,t)$. One way to do that is to simply solve the differential equation in Eq. \[eq:de\] with the new initial conditions $y_0$. However, depending on the problem, finding the new solution with an $N$-body integration can be either very inefficient or inaccurate[^1].
Thus, we are looking for a better way to estimate solutions for the initial conditions $y_0$ in a neighbourhood of $x_0$. The approach we consider here uses the fact that one can expand $\phi(y_0,t)$ around a reference solution $\phi(x_0,t)$ in a power series. For simplicity, we first consider the case of varying a single scalar parameter $\alpha$ on which the initial conditions depend, $y_0(\alpha)$. If $\alpha=0$, then $y_0(\alpha) = x_0$. This could correspond to varying a Cartesian component of $x_0$, or a parameter that mixes Cartesian components, such as a planet’s orbital eccentricity. Then each component of $\phi(y_0,t)$ can be expanded around the reference solution as a power series in $\alpha$, $$\begin{aligned}
\phi(y_0, t) = \sum_{m\geq 0} \frac{1}{m!}\; \phi^{(m)}\; \alpha^m, \label{eq:powerseries}\end{aligned}$$ In the above equation, $\phi^{(0)} \equiv \phi(x_0,t)$, i.e. the reference solution, and $$\begin{aligned}
\phi^{(m)} \equiv \left.\frac{\partial^m \phi(y_0(\alpha), t)}{\partial \alpha^m}\right|_{\alpha=0}, \label{eq:phi-m}\end{aligned}$$ i.e. a vector of the $m$-th derivative of each of the reference solution’s components with respect to the parameter $\alpha$. For sufficiently small $\alpha$ and $t$, this approximation is accurate even if we terminate the series at a finite $m=m_{\rm max}$ . The precise domain on which the solution can be trusted depends on the system and the initial conditions. For example, in chaotic dynamical systems, $\phi^{(1)}$ might grow exponentially fast, limiting the domain to relatively short times or small $\alpha$.
In conclusion, if one can obtain the $\phi^{(m)}$, one can approximate all nearby solutions of $\phi(x_0, t)$. Each $\phi^{(m)}$ is a function of time and must be numerically integrated. We therefore seek their governing differential equations.
We henceforth denote the reference solution $\phi(x_0, t)$ simply as $\phi$. The solution $\phi$, by definition, satisfies the original Eq. \[eq:de\], in other words $ \dot{\phi} = X\left(\phi\right)$. We now take the derivative of this equation with respect $\alpha$, the parameter we are varying, $$\begin{aligned}
\frac{\partial}{\partial t}\; \frac{\partial \phi}{\partial \alpha} =
\frac{\partial X(\phi)}{\partial \alpha} =
\sum_b \frac{\partial X(\phi)}{\partial \phi_b}
\frac{\partial \phi_b}{\partial \alpha},\end{aligned}$$ where we changed the order of the derivatives and made use of the chain rule. The summation index $b$ runs over all $6N$ elements of the vector $\phi$. The derivative of $\phi$ with respect to $\alpha$ is the $\phi^{(1)}$ we seek for use in Eq. \[eq:powerseries\]. Let us define the $6N$ by $6N$ matrix $X^{(1)}$ with components $$\begin{aligned}
X_{ab}^{(1)}(\phi) \equiv \frac{\partial X_a(\phi)}{\partial \phi_b}. \label{eq:x1}\end{aligned}$$ Using the matrix $X^{(1)}$ we then arrive at a compact set of differential equations for the vector $\phi^{(1)}$: $$\begin{aligned}
\dot{\phi}^{(1)} = X^{(1)}(\phi)\; \phi^{(1)}. \label{eq:ve1}\end{aligned}$$ This equation is the first-order variational equation. We will later calculate the components of the matrix $X^{(1)}$ explicitly. We then solve for the vector $\phi^{(1)}$ by integrating the differential equation numerically. Note that $X^{(1)}$ depends on the time-dependent reference solution $\phi$ but not on $\phi^{(1)}$. It is a linear operator acting on $\phi^{(1)}$.
Repeating the steps above but differentiating Eq. \[eq:de\] twice instead of once, we can write down the differential equation for $\phi^{(2)}$. Because we apply the chain rule in the process, one finds that the time derivative of the second-order variations $\phi^{(2)}$ depends not only on $\phi^{(2)}$ but also on $\phi^{(1)}$. Explicitly, the differential equation after two derivatives becomes $$\begin{aligned}
&\frac{\partial}{\partial t}\; \frac{\partial^2 \phi}{\partial \alpha^2} =
\frac{\partial^2 X(\phi)}{\partial \alpha^2}
=
\sum_{b,c}
\frac{\partial^2 X(\phi)}{\partial \phi_b\partial \phi_c}
\frac{\partial \phi_b}{\partial \alpha}
\frac{\partial \phi_c}{\partial \alpha}
+ \sum_{b}
\frac{\partial X(\phi)}{\partial \phi_b}
\frac{\partial^2 \phi_b}{\partial \alpha^2}.\end{aligned}$$ Defining the $6N$ by $6N$ by $6N$ tensor $X^{(2)}$ with components $$\begin{aligned}
X_{abc}^{(2)}(\phi) \equiv \frac{\partial X_a(\phi)}{\partial \phi_b \partial \phi_c},\end{aligned}$$ and using a short hand notation that suppresses the summation indices as well as the function arguments (we give explicit component forms for the general case in Sec. \[sec:multi\]), we have $$\begin{aligned}
\dot{\phi}^{(1)} &= X^{(1)}\; \phi^{(1)}, \nonumber \\
\dot{\phi}^{(2)} &= X^{(1)}\; \phi^{(2)} + X^{(2)}\; \left[\phi^{(1)}\right]^2. \label{eq:ve2}\end{aligned}$$ This set of equations is not linear anymore. But note that the linear term of the second line is the same as in the first line.
Higher order equations can be constructed in a straightforward way. Using the shorthand notation makes this particularly easy. One can reintroduce the indices at the end of the calculation. In this paper, we will only use variational equations up to second order [see e.g. @MoralesRuiz2007 for equations up to order 3].
Initial conditions
------------------
To integrate a differential equation forward in time, one needs appropriate initial conditions. To obtain the initial conditions for $\phi^{(1)}$ and $\phi^{(2)}$, one simply applies the chain rule to Eq. \[eq:phi-m\] and evaluates it at $t=0$, $$\begin{aligned}
\phi^{(1)}(x_0,0) &= \left. \frac{\partial y_0(\alpha)}{\partial \alpha}, \label{eq:phi1ic} \right|_{\alpha=0}\\
\phi^{(2)}(x_0,0) &= \left. \frac{\partial^2 y_0(\alpha)}{\partial \alpha^2}\right|_{\alpha=0}. \label{eq:phi2ic}\end{aligned}$$
In the case where the varied parameter $\alpha$ corresponds to a Cartesian component, choosing the initial conditions for $\phi^{(1)}$ and $\phi^{(2)}$ is straightforward. If we assume the varied parameter has the coordinate index $b$, then the initial conditions for $\phi(y_0(\alpha),t)$ in component form are $$\begin{aligned}
\phi_a(y_0(\alpha),0) = \phi_a(x_0,0) + \alpha \delta_{ab}, \end{aligned}$$ where $\delta_{ab}$ is the Kronecker delta. Thus $$\begin{aligned}
\phi_a^{(1)}(x_0,0) = \delta_{ab}
\quad\quad\text{and} \quad\quad
\phi_a^{(2)}(x_0,0) = 0. \label{eq:initcart}\end{aligned}$$
In practice, the function $y_0(\alpha)$ can be very complicated. As an example, let us consider a planetary system with one planet of mass $m$ on an initially circular and coplanar orbit around a star with mass $M$. The initial conditions of the planet might then be defined through the semi-major axis $a$ as $$\begin{aligned}
\begin{pmatrix}
{r}_x\\
{r}_y\\
{r}_z
\end{pmatrix}
=
\begin{pmatrix}
a\\
0\\
0
\end{pmatrix},
\quad
\quad
\begin{pmatrix}
v_x\\
v_y\\
v_z
\end{pmatrix}
=
\begin{pmatrix}
0\\
\sqrt{G(m+M)/a}\\
0
\end{pmatrix}.\end{aligned}$$ If we vary the initial semi-major axis by some length $\alpha$, then the initial conditions for the first-order variation are given by Eq. \[eq:phi1ic\], in our case $$\begin{aligned}
\label{eq:init1st}
\begin{pmatrix}
\mathbbm{r}_x\\
\mathbbm{r}_y\\
\mathbbm{r}_z
\end{pmatrix}^{(1)}
=
\begin{pmatrix}
1\\
0\\
0
\end{pmatrix},
\quad
\quad
\begin{pmatrix}
\mathbbm{v}_x\\
\mathbbm{v}_y\\
\mathbbm{v}_z
\end{pmatrix}^{(1)}
&=
\begin{pmatrix}
0\\
-\frac12 \sqrt{G(m+M)/a^3}\\
0
\end{pmatrix}.\end{aligned}$$ The initial conditions for the second-order variation are given by Eq. \[eq:phi2ic\], which for the present case are $$\begin{aligned}
\label{eq:init2nd}
\begin{pmatrix}
\mathbbm{r}_x\\
\mathbbm{r}_y\\
\mathbbm{r}_z
\end{pmatrix}^{(2)}
=
\begin{pmatrix}
0\\
0\\
0
\end{pmatrix},
\quad
\quad
\begin{pmatrix}
\mathbbm{v}_x\\
\mathbbm{v}_y\\
\mathbbm{v}_z
\end{pmatrix}^{(2)}
&=
\begin{pmatrix}
0\\
\frac34 \sqrt{G(m+M)/a^5}\\
0
\end{pmatrix}.\end{aligned}$$ The components of $\phi^{(1)}$ and $\phi^{(2)}$ that we calculated above correspond to the planet. All components corresponding to the star are 0.
The initialization can quickly get complicated. Suppose we work in the centre-of-mass frame. Then the star’s initial conditions will also depend on the semi-major axis of the planet. Similarly, if we add an additional outer planet and work in Jacobi coordinates, the outer planet’s initial conditions depend on the inner planet’s orbital parameters. For that reason we’ve implemented convenience functions for the initialization of orbits which we present later in Sec. \[sec:syntax\].
Multiple sets of variational equations {#sec:multi}
--------------------------------------
The above derivation of variational equations can be straightforwardly generalized when varying multiple parameters. Consider varying the initial value of $N_{\rm par}$ separate parameters $\alpha_\xi$. Here and in the rest of this paper Greek variables indicate to variations with respect to one parameter and will run over the interval $[0,N_{\rm par}-1]$. We write all equations in this section in component form for direct comparison with our later results.
When varying several parameters, the coupled set of differential equations, Eq. \[eq:ve2\], becomes $$\begin{aligned}
\dot{\phi}_{a,\xi}^{(1)} &= \sum_{b} X_{ab}^{(1)}\; \phi_{b,\xi}^{(1)}, \nonumber \\
\dot{\phi}_{a,\xi\eta}^{(2)} &= \sum_{b} X_{ab}^{(1)}\; \phi_{b,\xi\eta}^{(2)} + \sum_{b,c} X_{abc}^{(2)}\; \phi_{b,\xi}^{(1)} \phi_{c,\eta}^{(1)}, \label{eq:ve2multi}\end{aligned}$$ where $\phi_{a,\xi}^{(1)} \equiv \frac{\partial \phi_a(x_0, t)}{\partial \alpha_\xi}$, and $\phi_{a,\xi\eta}^{(2)} \equiv \frac{\partial^2 \phi_a(x_0, t)}{\partial \alpha_\xi \partial \alpha_\eta}$. Therefore, when varying $N_{\rm par}$ parameters, there are $N_{\rm par}$ sets of first-order variational equations, one for each of the vectors $\phi_{\xi}^{(1)}$. There are $N_{\rm par}^2$ sets of second-order variational equations (one for each of the vectors $\phi_{\xi\eta}^{(2)}$). Each set of variational equations has $6N$ components.
Once numerically integrated, these variations can then be plugged into a multi-variate power series expansion analog to Eq. \[eq:powerseries\] to obtain trajectories for arbitrary nearby initial conditions. Explicitly, to second order, $$\begin{aligned}
\phi_a(y_0, t) \approx \phi_a(x_0, t) + \sum_{\xi} \phi_{a,\xi}^{(1)} \alpha_\xi+ \frac12 \sum_{\xi,\eta}\phi_{a,\xi\eta}^{(2)} \alpha_\xi \alpha_\eta. \label{eq:powerseriesmulti}\end{aligned}$$ Note that because derivatives commute we find that $\phi^{(2)}_{\xi \eta} = \phi^{(2)}_{\eta \xi}$. Thus the total number of differential equations we need to integrate for the second order variations can be reduced from $6\,N\,N^2_{\rm par}$ to $3\,N\,N_{\rm par} (N_{\rm par}+1)$. This is in addition to the $6N$ differential equations for the reference simulation and $6\,N\,N_{\rm par}$ equations for the first order variations.
Index convention
----------------
As we saw above, the number of indices in second-order expressions is high. We therefore adopt a consistent index notation for the remainder of this paper. Specifically, we will consider a dynamical system consisting of $N$ particles and use the indices $i,j, k $ and $l$ to label different particles. These indices thus run from $0$ to $N-1$. The indices $a,b,c$ and $d$ label coordinate axes. Above, these indices ran over the $6N$ coordinates of the $N$-body system. We will find below that for the $N$-body system, it is simpler to consider positions and velocities separately. Therefore, in what follows $a,b,c$ and $d$ will run over the Cartesian $x$, $y$, and $z$ components only. As before, we will also make use of Greek characters $\xi$ and $\eta$ to indicate different *sets* of variational equations corresponding to different varied parameters (different *variations*). In the following sections we explicitly write summation symbols, i.e., we do not use a summation convention over repeating indices.
Variational Equations for the $N$-body System {#sec:varnbody}
=============================================
Let us now derive the differential equations from above for a specific problem: the dynamical system of $N$ gravitationally interacting particles. The differential equation for the $N$-body problem in vector notation is $$\begin{aligned}
\ddot{\mathbf{r}}_i &=& -\sum_{j,\;j\neq i} \frac{G m_j{\pmb r}_{ij}}{r_{ij}^3}\label{eq:nbody}\end{aligned}$$ where ${\pmb r}_{ij} = {\pmb r}_i - {\pmb r}_j$ and $r_{ij}$ is the norm of ${\pmb r}_{ij}$. We can also write the equation in component form $$\begin{aligned}
\dot{v}_{i,a} &=& -\sum_{j,\;j\neq i}\frac{ G m_j\,r_{ij,a}}{r_{ij}^3}. \label{eq:nbodycomp}\end{aligned}$$ where ${r}_{ij,a} = {r}_{i,a} - {r}_{j,a}$ is the relative position between particles $i$ and $j$. This is a second-order differential equation. However, note that the first time-derivative of the position is just the velocity, $\dot{\pmb r} = {\pmb v}$. The differential equation can thus easily be brought into the form of Eq. \[eq:de\] by introducing $$\begin{aligned}
\arraycolsep=2.4pt\def\arraystretch{1.}
x \equiv
\begin{pmatrix}
r_{0,x}&
r_{0,y}&
\hdots&
r_{N-1,y}&
r_{N-1,z}&
v_{0,x}&
v_{0,y}&
\hdots&
v_{N-1,y}&
v_{N-1,z}
\end{pmatrix}^T
\label{eq:x}\end{aligned}$$ such that $$\begin{aligned}
\arraycolsep=2.4pt\def\arraystretch{1.}
\dot x =
\begin{pmatrix}
v_{0,x}&
v_{0,y}&
\hdots&
v_{N-1,y}&
v_{N-1,z}&
\ddot r_{0,x}&
\ddot r_{0,y}&
\hdots&
\ddot r_{N-1,y}&
\ddot r_{N-1,z}
\end{pmatrix}^T.\end{aligned}$$ We end up with a first-order differential equation with twice as many variables as Eq. \[eq:nbody\] ($6N$ compared to $3N$). This set of differential equations together with suitable initial conditions completely describes the $N$-body problem.
First-order variational equations {#sec:derivation1}
---------------------------------
To derive the first-order variational equation for the $N$-body problem, we start by differentiating Eq. \[eq:nbodycomp\] with respect to $r_{j,b}$. To be as explicit as possible we do this in component form. We end up with an equation with four indices. Two of the indices run over coordinates, and two over particles, $$\begin{aligned}
\frac{\partial \dot{v}_{i,a}}{\partial r_{j,b}} &= \frac{\partial}{\partial r_{j,b}}\left(
-\sum_{k,\;k\neq i}\frac{ G m_k\,r_{ik,a}}{r_{ik}^3}. \nonumber
\right)
\\
&= -\sum_{k,\;k\neq i}\frac{Gm_k\,(\delta_{ij}-\delta_{kj})\delta_{ab} }{r_{ik}^3}
+ \sum_{k,\;k\neq i}3 \frac{G m_k r_{ik,a}\,r_{ik,b}}{r_{ik}^5}(\delta_{ij}-\delta_{kj})\nonumber
\\
&=
\begin{cases}
\sum_{k,\;k\neq i}\left( -\frac{Gm_k\,\delta_{ab} }{r_{ik}^3}
+ 3 \frac{G m_k r_{ik,a}\,r_{ik,b}}{r_{ik}^5}\right) &\mbox{ if } i=j\\
+\frac{Gm_j\,\delta_{ab} }{r_{ij}^3}
- 3 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^5} & \mbox{ if } i\neq j
\end{cases}.\end{aligned}$$ The above expression gives us the matrix elements of $X^{(1)}$ (Eq. \[eq:x1\]). Note that two indices $j$ and $b$ combined correspond to the row of that matrix, the other two ($i$ and $a$) correspond to the column of the matrix.
We also want to consider the influence of varying masses. Note that one can think of the masses $m_i$ as part of the initial conditions. However, we assume that the masses do not vary with time after the system has been initialized, thus $\dot m_i=0$. We need the derivative of the force with respect to the mass: $$\begin{aligned}
\frac{\partial \dot{v}_{i,a}}{\partial m_j} =
\frac{\partial}{\partial m_j}\left(-\sum_{k,\;k\neq i}\frac{G\,m_k\,{r}_{ik,a}}{r_{ik}^3}\right)
=
\begin{cases}
-\frac{G\,{r}_{ij,a} }{r_{ij}^3} & \mbox{if } j\neq i \\
0 & \mbox{if } j = i.
\end{cases} \end{aligned}$$
We can now write down the differential equation for the first-order variational equation using Eq. \[eq:ve1\]. We could do this in terms of $\phi^{(1)}$ and its components, but choose to use two separate vectors for the variational position and velocity components (to be consistent with Eq. \[eq:nbody\]). Variational quantities are denoted by double-striped symbols ${\ensuremath{\mathbbm{r}}}$, ${\ensuremath{\mathbbm{v}}}$ and ${\ensuremath{\mathbbm{m}}}$. First-order variational quantities receive the superscript ${}^{(1)}$. Thus, one might write $\phi^{(1)}$ in terms of ${\ensuremath{\mathbbm{r}}}^{(1)}$ and ${\ensuremath{\mathbbm{v}}}^{(1)}$: $$\begin{aligned}
\phi^{(1)}= \begin{pmatrix}
{\ensuremath{\mathbbm{r}}}_0^{(1)}&
\hdots&
{\ensuremath{\mathbbm{r}}}^{(1)}_{N-1}&
{\ensuremath{\mathbbm{v}}}^{(1)}_0&
\hdots&
{\ensuremath{\mathbbm{v}}}^{(1)}_{N-1}
\end{pmatrix}^T\end{aligned}$$ which should be compared to Eq. \[eq:x\]. The units of ${\ensuremath{\mathbbm{r}}}^{(1)}$ and ${\ensuremath{\mathbbm{v}}}^{(1)}$ depend on the variation we are considering. In general the units are not the same as those of $r$ and $v$ (see Eq. \[eq:phi1ic\]). We end up with the following set of equations for the components of $\dot {\ensuremath{\mathbbm{v}}}^{(1)}$ (corresponding to the second half of the components of $\dot \phi^{(1)}$): $$\begin{aligned}
\dot{{\ensuremath{\mathbbm{v}}}} ^{(1)}_{i,a} &=
\sum_j
\sum_b
\frac{\partial \dot{v}_{i,a}}{\partial r_{j,b}}
{\ensuremath{\mathbbm{r}}}^{(1)}_{j,b}
+
\sum_j
\frac{\partial \dot{v}_{i,a}}{\partial m_j}
{\ensuremath{\mathbbm{m}}}_j^{(1)},\nonumber
\\
&=
\sum_{j,\,j\neq i}
\sum_b
\frac{\partial \dot{v}_{i,a}}{\partial r_{j,b}}
{\ensuremath{\mathbbm{r}}}^{(1)}_{j,b},
+
\sum_b
\frac{\partial \dot{v}_{i,a}}{\partial r_{i,b}}
{\ensuremath{\mathbbm{r}}}^{(1)}_{i,b}
+
\sum_j
\frac{\partial \dot{v}_{i,a}}{\partial m_j}
{\ensuremath{\mathbbm{m}}}_j^{(1)},
\nonumber
\\
&=
\sum_{j,\,j\neq i}
\left[
\sum_b
\left(
-\frac{Gm_j\,\delta_{ab} }{r_{ij}^3}
+ 3 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^5}
\right)
{\ensuremath{\mathbbm{r}}}^{(1)}_{ij,b}
-
\frac{G\,{r}_{ij,a} }{r_{ij}^3}
{\ensuremath{\mathbbm{m}}}_j^{(1)}
\right]
,\end{aligned}$$ where ${\ensuremath{\mathbbm{r}}}^{(1)}_{ij,b} = {\ensuremath{\mathbbm{r}}}^{(1)}_{i,b} - {\ensuremath{\mathbbm{r}}}^{(1)}_{j,b}$. We can rewrite this in vector notation, which allows us to drop the indices $a$ and $b$: $$\begin{aligned}
\dot{\pmb {\ensuremath{\mathbbm{v}}}} ^{(1)}_{i} &=&
\sum_{j,\,j\neq i}
\left(
-\frac{Gm_j}{r_{ij}^3}
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij}
+ 3 \frac{G m_j {\pmb r}_{ij}}{r_{ij}^5}
\left( {\pmb r}_{ij} \cdot {\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij} \right)
-\frac{G\,{\pmb r}_{ij} }{r_{ij}^3} {\ensuremath{\mathbbm{m}}}_j^{(1)}
\right)\label{eq:var}\end{aligned}$$ The $\cdot$ represent the usual vector product. The equations for the variational positions, ${\ensuremath{\mathbbm{r}}}^{(1)}$ (the first half of the components of $\phi^{(1)}$), are significantly easier to write down: $$\begin{aligned}
\dot{\pmb {\ensuremath{\mathbbm{r}}}} ^{(1)}_{i} &=& {\pmb {\ensuremath{\mathbbm{v}}}}^{(1)}_{i}\end{aligned}$$ As mentioned before, the masses are assumed to be constant throughout a simulation; thus, the variational equation for the mass coordinates are not dynamic: $$\begin{aligned}
\dot{{\ensuremath{\mathbbm{m}}}} ^{(1)}_{i} &=& 0.\end{aligned}$$ The solutions are trivial, ${\ensuremath{\mathbbm{m}}}_i(t) = {\ensuremath{\mathbbm{m}}}_i(0)$, and we therefore do not evolve the quantities ${\ensuremath{\mathbbm{m}}}_i$.
Second-order variational equations {#sec:derivation2}
----------------------------------
We now derive the second-order variational equations. As a warning to the reader: this will get messy. We nevertheless present the calculation in full detail as it is easy to get confused with up to 7 indices in a single term. This should ease derivations for alternate dynamical systems, for example if one want to include additional non-gravitational effects. Conceptually, this is the same procedure as in the previous section, just to second order. Because of the chain rule, we end up with significantly more terms.
We begin by calculating various second-order derivatives that we will need later. The second-order derivative of the force with respect to the positions is $$\begin{aligned}
\frac{\partial^2 \dot{v}_{i,a}}{\partial r_{j,b} \partial r_{k,c}} &=&
\frac{\partial}{\partial r_{k,c}}
\begin{cases}
\sum_{l,\;l\neq i}\left( -\frac{Gm_l\,\delta_{ab} }{r_{il}^3}
+ 3 \frac{G m_l r_{il,a}\,r_{il,b}}{r_{il}^5}\right) &\mbox{ if } i=j\\
+\frac{Gm_j\,\delta_{ab} }{r_{ij}^3}
- 3 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^5} & \mbox{ if } i\neq j
\end{cases}\end{aligned}$$ We look at both cases individually. The first case, $i=j$, gives
$$\begin{aligned}
\frac{\partial^2 \dot{v}_{i,a}}{\partial r_{i,b} \partial r_{k,c}}
&=
\frac{\partial}{\partial r_{k,c}}
\sum_{l,\;l\neq i}\left( -\frac{Gm_l\,\delta_{ab} }{r_{il}^3}
+ 3 \frac{G m_l r_{il,a}\,r_{il,b}}{r_{il}^5}\right)
\nonumber
\\
&=
\sum_{l,\;l\neq i} 3\frac{Gm_l\,\delta_{ab} }{r_{il}^5} r_{il,c} (\delta_{ik} - \delta_{lk})
+
\sum_{l,\;l\neq i}
3 \frac{G m_l \delta_{ac}}{r_{il}^5} r_{il,b}(\delta_{ik}-\delta_{lk})
+
\sum_{l,\;l\neq i}
3 \frac{G m_l \delta_{bc}}{r_{il}^5} r_{il,a}(\delta_{ik}-\delta_{lk})
-
\sum_{l,\;l\neq i}
15 \frac{G m_l r_{il,a}\,r_{il,b}}{r_{il}^7} r_{il,c} (\delta_{ik}-\delta_{lk})
\nonumber
\\
&=
\begin{cases}
\sum_{l,\;l\neq i}\left(
3\frac{Gm_l\,\delta_{ab} r_{il,c} }{r_{il}^5}
+
3 \frac{G m_l \delta_{ac}r_{il,b}}{r_{il}^5}
+
3 \frac{G m_l \delta_{bc}r_{il,a}}{r_{il}^5}
-
15 \frac{G m_l r_{il,a}\,r_{il,b}r_{il,c}}{r_{il}^7}
\right)
& \mbox{ if } i=k\\
-\left(
3\frac{Gm_k\,\delta_{ab} r_{ik,c}}{r_{ik}^5}
+
3 \frac{G m_k \delta_{ac} r_{ik,b}}{r_{ik}^5}
+
3 \frac{G m_k \delta_{bc} r_{ik,a}}{r_{ik}^5}
-
15 \frac{G m_k r_{ik,a}\,r_{ik,b} r_{ik,c} }{r_{ik}^7}
\right)
& \mbox{ if } i\neq k
\end{cases}\end{aligned}$$
The second case, $i\neq j$, is similar but with the sign reversed and without the summation:
$$\begin{aligned}
\frac{\partial^2 \dot{v}_{i,a}}{\partial r_{j,b} \partial r_{k,c}}
&= -
3\frac{Gm_j\,\delta_{ab} }{r_{ij}^5} r_{ij,c} (\delta_{ik} - \delta_{jk})
-
3 \frac{G m_j \delta_{ac}}{r_{ij}^5} r_{ij,b}(\delta_{ik}-\delta_{jk})
-
3 \frac{G m_j \delta_{bc}}{r_{ij}^5} r_{ij,a}(\delta_{ik}-\delta_{jk})
+
15 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^7} r_{ij,c} (\delta_{ik}-\delta_{jk})
\nonumber
\\
&=
\begin{cases}
-
3\frac{Gm_j\,\delta_{ab} }{r_{ij}^5} r_{ij,c}
-
3 \frac{G m_j \delta_{ac}}{r_{ij}^5} r_{ij,b}
-
3 \frac{G m_j \delta_{bc}}{r_{ij}^5} r_{ij,a}
+
15 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^7} r_{ij,c}
& \mbox{ if } i= k\\
+
3\frac{Gm_j\,\delta_{ab} }{r_{ij}^5} r_{ij,c}
+
3 \frac{G m_j \delta_{ac}}{r_{ij}^5} r_{ij,b}
+
3 \frac{G m_j \delta_{bc}}{r_{ij}^5} r_{ij,a}
-
15 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^7} r_{ij,c}
& \mbox{ if } i\neq k \mbox{ and } j=k\\
0 & \mbox{ otherwise}
\end{cases}\end{aligned}$$
We also need the derivatives with respect to the particles’ masses. Luckily, if we differentiate the force twice with respect to mass, we get zero: $$\begin{aligned}
\frac{\partial^2 \dot{v}_{i,a}}{\partial m_j \partial m_k} &=& 0.\end{aligned}$$ However, other second derivatives involving the mass are not zero: $$\begin{aligned}
\frac{\partial^2 \dot{ v}_{i,a}}{\partial { r}_{k,b} \partial m_j}
= \frac{\partial}{\partial r_{k,b} }\left(
\frac{\partial \dot{v}_{i,a}}{\partial m_k} \right)
= \frac{\partial}{\partial r_{k,b} }
\begin{cases}
-\frac{G\,{r}_{ij,a} }{r_{ij}^3} & \mbox{if } j\neq i \\
0 & \mbox{if } j = i.
\end{cases} \end{aligned}$$ Restricting ourselves to the $j\neq i$ case, $$\begin{aligned}
\frac{\partial^2 \dot{ v}_{i,a}}{\partial { r}_{k,b} \partial m_j}
= \frac{\partial}{\partial r_{k,b} }
\left(-\frac{G\,{r}_{ij,a} }{r_{ij}^3} \right)
&=
-\frac{G}{r_{ij}^3} \delta_{ab} (\delta_{ik}-\delta_{jk})
+3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b} (\delta_{ik}-\delta_{jk})
=
\begin{cases}
+\frac{G}{r_{ij}^3} \delta_{ab} \delta_{jk}
-3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b} \delta_{jk}
& \mbox{if } k\neq i \\
-\frac{G}{r_{ij}^3} \delta_{ab}
+3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b}
& \mbox{if } k = i,
\end{cases} \end{aligned}$$ such that after putting all cases together we arrive at $$\begin{aligned}
\frac{\partial^2 \dot{ v}_{i,a}}{\partial { r}_{k,b} \partial m_j}
&=\begin{cases}
+\frac{G}{r_{ij}^3} \delta_{ab}
-3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b}
& \mbox{if } j\neq i \mbox{ and } k = j \\
-\frac{G}{r_{ij}^3} \delta_{ab}
+3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b}
& \mbox{if } j\neq i \mbox{ and } k = i\\
0
& \mbox{otherwise} .
\end{cases} \end{aligned}$$
With the above expressions of the second order force derivatives, we can now construct the second-order variational equations. At this point we introduce two more indices that describe the variation under consideration, $\xi$ and $\eta$. They run over all the variations that we want to consider. In vector notation Eq. \[eq:ve2\] can be expressed as
$$\begin{aligned}
\dot{{\ensuremath{\mathbbm{v}}}} ^{(2)}_{i,a,\eta\xi} &=
\sum_j
\sum_b
\frac{\partial \dot{v}_{i,a}}{\partial r_{j,b}}
{\ensuremath{\mathbbm{r}}}^{(2)}_{j,b,\eta\xi}
+
\sum_j
\sum_k
\sum_b
\sum_c
\frac{\partial^2 \dot{v}_{i,a}}{\partial r_{j,b} \partial r_{k,c}}
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{j,b,\xi}{{\ensuremath{\mathbbm{r}}}}^{(1)}_{k,c,\eta}
\nonumber
\\
&
+
\sum_j
\sum_k
\sum_b
\frac{\partial^2 \dot{ v}_{i,a}}{\partial { r}_{k,b} \partial m_j}
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{k,b,\xi}{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\eta}
+
\sum_j
\sum_k
\sum_b
\frac{\partial^2 \dot{ v}_{i,a}}{\partial m_k\partial { r}_{j,b} }
{{\ensuremath{\mathbbm{m}}}}^{(1)}_{k,\xi}
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{j,b,\eta}
+
\sum_j
\frac{\partial \dot{v}_{i,a}}{\partial m_j}
{\ensuremath{\mathbbm{m}}}_{j,\eta\xi}^{(2)}
+
\underbrace{
\sum_j
\sum_k
\frac{\partial^2 \dot{ v}_{i,a}}{\partial { m}_{k} \partial m_j}
{{\ensuremath{\mathbbm{m}}}}^{(1)}_{k,\xi}{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\eta}
}_{=0}.\end{aligned}$$
We replace the derivatives with what we calculated above. The result is a rather long expression with 7 different (summation) indices:
$$\begin{aligned}
\dot{{\ensuremath{\mathbbm{v}}}} ^{(2)}_{i,a,\xi\eta} = &
\sum_{j,\,j\neq i}
\left(
-\frac{Gm_j\, {\ensuremath{\mathbbm{r}}}^{(2)}_{ij,a,\xi\eta} }{r_{ij}^3}
+ 3 \frac{G m_j r_{ij,a}}{r_{ij}^5}
\sum_b
\left(
r_{ij,b}
\cdot
{\ensuremath{\mathbbm{r}}}^{(2)}_{ij,b,\xi\eta}
\right)
-
\frac{G\,{r}_{ij,a} }{r_{ij}^3} {\ensuremath{\mathbbm{m}}}_{j,\xi\eta}^{(2)}
\right)
\nonumber
\\
&
+
\sum_b
\sum_c
\sum_{l,\;l\neq i}\left(
3\frac{Gm_l\,\delta_{ab} r_{il,c} }{r_{il}^5}
+
3 \frac{G m_l \delta_{ac}r_{il,b}}{r_{il}^5}
+
3 \frac{G m_l \delta_{bc}r_{il,a}}{r_{il}^5}
-
15 \frac{G m_l r_{il,a}\,r_{il,b}r_{il,c}}{r_{il}^7}
\right)
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{i,b,\xi}{{\ensuremath{\mathbbm{r}}}}^{(1)}_{i,c,\eta}
\nonumber
\\
&
+
\sum_{k,\;k\neq i}
\sum_b
\sum_c
\left(
-3\frac{Gm_k\,\delta_{ab} r_{ik,c}}{r_{ik}^5}
-
3 \frac{G m_k \delta_{ac} r_{ik,b}}{r_{ik}^5}
-
3 \frac{G m_k \delta_{bc} r_{ik,a}}{r_{ik}^5}
+
15 \frac{G m_k r_{ik,a}\,r_{ik,b} r_{ik,c} }{r_{ik}^7}
\right)
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{i,b,\xi}{{\ensuremath{\mathbbm{r}}}}^{(1)}_{k,c,\eta}
\nonumber
\\
&
+
\sum_{j,\;j\neq i}
\sum_b
\sum_c
\left(
-
3\frac{Gm_j\,\delta_{ab} }{r_{ij}^5} r_{ij,c}
-
3 \frac{G m_j \delta_{ac}}{r_{ij}^5} r_{ij,b}
-
3 \frac{G m_j \delta_{bc}}{r_{ij}^5} r_{ij,a}
+
15 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^7} r_{ij,c}
\right)
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{j,b,\xi}{{\ensuremath{\mathbbm{r}}}}^{(1)}_{i,c,\eta}
\nonumber
\\
&
+
\sum_{j,\;j\neq i}
\sum_b
\sum_c
\left(
+
3\frac{Gm_j\,\delta_{ab} }{r_{ij}^5} r_{ij,c}
+
3 \frac{G m_j \delta_{ac}}{r_{ij}^5} r_{ij,b}
+
3 \frac{G m_j \delta_{bc}}{r_{ij}^5} r_{ij,a}
-
15 \frac{G m_j r_{ij,a}\,r_{ij,b}}{r_{ij}^7} r_{ij,c}
\right)
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{j,b,\xi}{{\ensuremath{\mathbbm{r}}}}^{(1)}_{j,c,\eta}
\nonumber
\\
&
+
\sum_b
\sum_{j,\;j\neq i}
\left(
+\frac{G}{r_{ij}^3} \delta_{ab}
-3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b}
\right)
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{ji,b,\xi}{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\eta}
+
\sum_b
\sum_{j,\;j\neq i}
\left(
+\frac{G}{r_{ij}^3} \delta_{ab}
-3\frac{G\,{r}_{ij,a} }{r_{ij}^5} r_{ij,b}
\right)
{{\ensuremath{\mathbbm{r}}}}^{(1)}_{ji,b,\eta}{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\xi}.\end{aligned}$$
Note that the first line has the same form as for the first-order variational equations. This is the linear part of the second-order variational equation. The other lines correspond to the non-linear part that couples to the first-order differential equations of the variations $\xi$ and $\eta$. We can simplify the above expression slightly and convert it to a somewhat more readable vector notation, arriving at $$\begin{aligned}
\dot{\pmb {\ensuremath{\mathbbm{v}}}} ^{(2)}_{i,\xi\eta} =&
\sum_{j,\,j\neq i}
\left(
-\frac{Gm_j\, {\pmb {\ensuremath{\mathbbm{r}}}}^{(2)}_{ij,\xi\eta} }{r_{ij}^3}
+ 3 \frac{G m_j {\pmb r}_{ij}}{r_{ij}^5}
\left(
{\pmb r}_{ij}
\cdot
{\pmb {\ensuremath{\mathbbm{r}}}}^{(2)}_{ij,\xi\eta}
\right)
-
\frac{G\,{\pmb r}_{ij} }{r_{ij}^3} {\ensuremath{\mathbbm{m}}}_{j,\xi\eta}^{(2)}
\right)
\nonumber
\\
&
+
\sum_{j,\;j\neq i}\left(
3\frac{Gm_j {\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\xi} }{r_{ij}^5}
\left(
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\eta} \cdot {\pmb r}_{ij}
\right)
+
3 \frac{G m_j{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\eta}}{r_{ij}^5}
\left(
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\xi} \cdot {\pmb r}_{ij}
\right)
+
3 \frac{G m_j {\pmb r}_{ij}}{r_{ij}^5}
\left(
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\xi}\cdot {\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\eta}
\right)
-
15 \frac{G m_j {\pmb r}_{ij}}{r_{ij}^7}
\left(
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\xi}
\cdot
{\pmb r}_{ij}
\right)
\left(
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\eta}
\cdot
{\pmb r}_{ij}
\right)
\right)
\nonumber
\\
&
+
\sum_{j,\;j\neq i}
\left(
-
\frac{G}{r_{ij}^3}
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\xi}{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\eta}
+3\frac{G\,{\pmb r}_{ij} }{r_{ij}^5}
\left(
{\pmb r}_{ij} \cdot
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\xi}
\right)
{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\eta}
\right)
+
\sum_{j,\;j\neq i}
\left(
-
\frac{G}{r_{ij}^3}
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\eta}{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\xi}
+3\frac{G\,{\pmb r}_{ij} }{r_{ij}^5}
\left(
{\pmb r}_{ij} \cdot
{\pmb {\ensuremath{\mathbbm{r}}}}^{(1)}_{ij,\eta}
\right)
{{\ensuremath{\mathbbm{m}}}}^{(1)}_{j,\xi}
\right). \label{eq:vdot2ndorder}\end{aligned}$$ We can use this equation to read off the matrix elements of $X^{(1)}$ and $X^{(2)}$ by comparing the above with Eqs. \[eq:ve1\] and \[eq:ve2\].
Implementation {#sec:implementation}
==============
We have implemented first and second-order variational equations into the $N$-body code [[REBOUND]{}]{}[@ReinLiu2012]. [[REBOUND]{}]{}is very modular and allows the user to choose from different numerical integrators. What we describe here has been tested for the high-accuracy integrator [[IAS15]{}]{}, which is based on a 15th-order Gauß-Radau quadrature [@ReinSpiegel2015]. First-order variational equations have also been implemented for the symplectic WHFast integrator [@ReinTamayo2015] as a symplectic tangent map [@MikkolaInnanen1999]. In principle, higher-order variational equations could also be implemented as a symplectic tangent map. However, the complexity of such a higher-order tangent map goes beyond what we expect to be useful in practice. We therefore exclusively focus on the general-purpose [[IAS15]{}]{}integrator for the remainder of this paper.
We implement the variational equations in terms of *variational particles*. This provides an elegant implementation where variational particles follow a structurally similar (though conceptually different) set of differential equations to the real particles (cf. Eqs. \[eq:var\] and \[eq:vdot2ndorder\]). For each first and second order variation that we consider we add $N$ such variational particles to the simulation. The Cartesian components of a variational particle are then the derivatives of the corresponding real particle’s components with respect to the parameter we are varying. This implies that the units for different variations will vary (compare with Eq. \[eq:phi-m\]).
Initialization routines
-----------------------
In addition to the variational equations themselves, we have implemented convenience methods for initializing the variational particles. If one is interested in varying one of the cartesian coordinates of a particle, initializing variational particles is as easy as setting all of the coordinates to 0 except one which is set to 1, see Eq. \[eq:initcart\]. However, as shown above, varying parameters that are non-linear functions of the cartesian coordinates involves calculating first and second derivatives and can quickly become cumbersome. We are particularly interested in applications involving planetary systems. We therefore provide routines that allow the initialization of variational particles with respect to changing a particle’s mass $m$, as well as its orbit’s semi-major axis $a$, eccentricity $e$, inclination $i$, longitude of the ascending node $\Omega$, argument of pericenter $\omega$ and true anomaly $f$.
Since we are doing this to second order for 7 orbital elements[^2], we thus have $7+ 7\cdot (7+1)/2=35$ different functions. In principle one could also initialize variational particles by calculating finite differences, i.e. creating a second particle with one orbital parameter shifted by a small amount $\alpha$, subtracting each component from the un-shifted particle and then dividing by $\alpha$. The problem is that this procedure easily leads to numerical issues as the shift $\alpha$ needs to be small enough to be in the linear (quadratic) regime, but large enough to avoid any rounding error due to limited floating point precision. Our functions that calculate the derivatives analytically avoid this issue. Our current implementation does not support Jacobi coordinates and assumes that all parameters are given with respect to a fixed central object (heliocentric frame).
We have also implemented a routine that moves the entire system to the centre of mass frame and corrects the variational particles’ positions and velocity coordinates consistently.
It is worth pointing out that automatic differentiation (AD) would be well suited to automate this task for even more complicated initialization routines [@Neidinger2010].
Test particles
--------------
We call a particle in an $N$-body simulation a test particle if it does not affect other particles in the simulation because it has no mass, $m_i=0$. If one is interested in the effect of varying the initial conditions of test particles, then the variational equations simplify significantly. Because variations of a test particle do not affect the variations of other particles, one can reduce the dimensionality of the first-order variational differential equation, Eq. \[eq:ve1\], from $6N$ to $6$. This speeds up the calculation significantly and we have implemented this as an optional flag that can be set when a set of variational equations is initialized. This might become particularly useful if an approximation of the derivatives is sufficient for a given application.
Syntax {#sec:syntax}
------
Here, we briefly demonstrate how to initialize and run a simulation using the python interface to [[REBOUND]{}]{}. We do this because the layer of abstraction that we came up with to hide the complicated expressions for second order variational equations is essential in making this tool useable in a real world scenario. We provide the full documentation for how to use variational equations within [[REBOUND]{}]{}online at <http://rebound.readthedocs.org>.
A simulation of one planet orbiting a central star can be setup with the following code in `python`:
import rebound
sim = rebound.Simulation()
sim.add(m=1.)
sim.add(m=0.001, a=1.)
By default [[REBOUND]{}]{}uses units in which $G=1$. One set of first-order variational particles can be set up with a single command:
var_i = sim.add_variation()
The `var_i` object contains all the information of this set of variational particles, e.g. the order, the location of variational particles, etc. By default, the variational particles’ position, velocity and mass coordinates are initialized to zero. For this example, let us assume that we want to vary the planet’s semi-major axis. We initialise the planet’s variational particle using the following command:
var_i.vary(1,"a")
In the background, this command first calculates the orbital parameters of the particle with index 1 (the planet) in heliocentric coordinates. Then, the variational particle is initialized using the analytic derivative with respect to the semi-major axis, see Eq.\[eq:init1st\]. We can now integrate the system forward in time for, say, 100 time units:
sim.integrate(100.)
The planet’s x-position after the integration can be accessed via `sim.particles[1].x`. Let us use the result of our integration to estimate the planet’s x-position assuming its initial semi-major axis was shifted by $\Delta a = 0.01$. This can be achieved with the following code
Delta_a = 0.01
print sim.particles[1].x
+ Delta_a * var_i.particles[+1].x
which should be compared with Eq. \[eq:powerseries\]. To go beyond first order and include second-order variational equations, we setup the second order variational equations (before the integration) with
var_ii = sim.add_variation(order=2,
first_order=var_i)
Note that we need to specify the the corresponding first-order variational particles. This is because second-order variational particles depend on the first-order variational particles and in principle there can be many different first oder variational particles for different parameters. The initialization of the particle is identical to before
var_ii.vary(1,"a")
The final position can then be estimated by applying Eq. \[eq:powerseries\] as before, but now accurate to second order,
print sim.particles[1].x
+ Delta_a * var_i.particles[1].x
+ 0.5*Delta_a*Delta_a * var_ii.particles[1].x
More complicated and realistic examples are available in the documentation at <http://rebound.readthedocs.org>.
Tests {#sec:tests}
=====
In this section we present various tests of our implementation. These show not only that the implementation is working correctly, but also what second-order variational equations can be used for. We plan to follow up on several of these ideas in much more detail in future work.
Varying one orbital parameter
-----------------------------
As a first test, we study a two-planet system and vary the initial semi-major axis of the outer planet. We use the first and second-order variational equations to approximate the $x$-position of the inner planet after 10 orbits using Eq. \[eq:powerseriesmulti\]. The inner planet’s position changes with time because of the planet-star as well as the planet-planet interactions.
We plot the results in Fig. \[fig:test1\]. The bold black line corresponds to the final $x$-position of the inner planet using a direct $N$-body integration. The results for both the first and second-order variational equations are shown as a green dashed and blue dotted line, respectively. To arrive at these approximations, only one $N$-body simulation with variational equations was run. Note that this is in contrast to 400 individual $N$-body simulations which were carried out to generate the black curve. The red dot indicates the initial semi-major axis used for the single run with variational equations.
As the plot clearly shows, we can use the results from second-order variational equations to accurately predict the final position of the inner planet to within a few percent for initial conditions that are not too far from the original simulation. Also note that as expected, the second-order variational equations give a significantly better estimate than the first-order equations alone.
Optimization problem with one one orbital parameter {#sec:onepar}
---------------------------------------------------
We continue to work with the above two-planet system. We now attempt to find the initial semi-major axis $a_0$ of the outer planet that minimizes the $x$ coordinate of the inner planet at the end of the simulation. This test therefore represents a simple case of a wide range of optimization problems. Instead of minimizing the $x$ coordinate of the planet, we could also minimize the distance to another planet, or maximize the velocity. Furthermore, one could replace one of the planets with a spacecraft and then search for an optimal spacecraft trajectory that uses a minimal amount of fuel to reach a final point, and so on.
We use the standard Newton’s method to find the optimal value, $x_{\rm min}$. For that we need the first and second derivatives of the planet’s $x$ position (we are looking for the root of the first derivative). We calculate these using the variational equations. As a starting point in Newton’s method, we use the red dot in Fig. \[fig:test1\].
We plot the results in Fig. \[fig:test2\]. The vertical axis shows the relative position offset $\bar x = \left| (x-x_{\rm min})/{x_{\rm min}}\right|$ as a function of the iteration. After four iterations, the method has converged to machine precision. With any derivative free method such as the bisection method we would need more iterations to achieve machine precision.
Fitting a radial velocity curve {#sec:rvtest}
-------------------------------
We now present a more complicated example in which we attempt to fit the reflex motion of a star in a two-planet system to a synthetic radial velocity data set. In Fig. \[fig:test3a\] we show the synthetic radial velocity curve of the star as a function of time (the units are irrelevant for this discussion). The red dots show where an observation is taken. For simplicity we only vary two orbital parameters, the semi-major axis $a$ and the eccentricity $e$ of the inner planet. All other parameters are the same as in the reference simulation. Our goal is to match these synthetic observations and thus to find the true parameters $(a, e)$, starting from an arbitrary initial guess of semi-major axis and eccentricity, $(a_0, e_0)$.
We label the synthetic observation at time $t_i$ with $o_i$ and the radial velocity in our simulation with $v_i$. The problem can be expressed again as an optimization problem by defining a goodness of fit, e.g., $$\begin{aligned}
\chi^2 = \sum_i \left(v_i-o_i\right)^2,\end{aligned}$$ which we try to minimize. Note that $v_i$ are functions of the initial $a$ and $e$. More complicated and realistic $\chi^2$ functions that take into account observational uncertainties can be easily constructed, but we here work with the simplest case. The chain rule yields $$\begin{aligned}
\frac{\partial}{\partial a}\chi^2 &= 2 \sum_i \left(v_i-o_i\right) \frac{\partial v_i}{\partial a}\\
\frac{\partial^2}{\partial a^2}\chi^2 &= 2 \sum_i\left( \left(v_i-o_i\right) \frac{\partial^2 v_i}{\partial a^2}+ \left(\frac{\partial v_i}{\partial a}\right)^2\right)\end{aligned}$$ and similar expressions for the derivatives with respect to $e$ and the cross-term $\nicefrac{\partial^2 \chi^2}{\partial a \partial e}$. The second-order variational equations are used to calculate the derivatives involving $v_i$. We can then use the standard Newton’s method to iterate and find the extremum in $\chi^2$: $$\begin{aligned}
\begin{pmatrix}
a_{n+1}\\
e_{n+1}
\end{pmatrix}
=
\begin{pmatrix}
a_n\\
e_n
\end{pmatrix}
-
\begin{pmatrix}
\frac{\partial^2 \chi^2}{\partial a^2} &
\frac{\partial^2 \chi^2}{\partial a\partial e} \\
\frac{\partial^2 \chi^2}{\partial a\partial e}&
\frac{\partial^2 \chi^2}{\partial e^2}
\end{pmatrix}^{-1}
\cdot
\begin{pmatrix}
\frac{\partial \chi^2}{\partial a} \\
\frac{\partial \chi^2}{\partial e}
\end{pmatrix}.\end{aligned}$$ The matrix in the above equation is the inverse of the Hessian of $\chi^2$, $H^{-1}$. Newton’s method will only converge where $\chi^2$ is convex, or in other words where the matrix $H$ is positive definite. To increase the convergence region we use a trick to ensure that $H$ is positive definite everywhere by using the softabs metric of the Hessian ${\ensuremath { \raisebox{0.3pt}{$\wr$} \mbox{{$H$}} \raisebox{0.3pt}{$\wr$} } }$, instead of $H$ itself [@Betancourt2013]. The modified Newton’s method becomes $$\begin{aligned}
\begin{pmatrix}
a_{n+1}\\
e_{n+1}
\end{pmatrix}
=
\begin{pmatrix}
a_n\\
e_n
\end{pmatrix}
-
{\ensuremath { \raisebox{0.3pt}{$\wr$} \mbox{{$H$}} \raisebox{0.3pt}{$\wr$} } }^{-1}
\cdot
\begin{pmatrix}
\frac{\partial \chi^2}{\partial a} \\
\frac{\partial \chi^2}{\partial e}
\end{pmatrix}.\end{aligned}$$
In Fig. \[fig:test3b\] we plot the relevant part of the parameter space. The colours and contours correspond to the logarithm of $\chi^2$. We start the iteration at $(a_0,e_0) = (0.96,0.2)$ and converge to the true minimum within machine precision in less than ten iterations. Newton’s method converges to the global minimum for most nearby starting values (those near the centre in Fig. \[fig:test3b\]). If the initial conditions are far from the global optimum, then, as expected, the method might not converge to the global minimum. We note that this problem of non-convergence is a feature of the adopted optimization algorithm (Newton’s method), and not of the variational equations. In particular, even in cases where Newton’s method does not converge, the derivatives are calculated exactly (to machine precision).
For the above reasons, the method presented in this example is not well suited for finding the global minimum within a complex parameter space. Other methods such as simulated annealing or parallel tempering are most likely faster and more reliable. However, as we discuss below, a combination of methods is a promising future area of research if one can make use the second-order variational equations to converge to a local optimum within almost constant time (or $\mathcal{O}(1)$ iterations)[^3].
Comparison to a finite difference approach {#sec:finitedifference}
------------------------------------------
In the above optimization problem, we use variational equations to calculate the first and second derivatives of $\chi^2$. One can also use a finite difference approach to estimate the derivatives. As we show in this section, this is not viable in most scenarios as two separate competing constraints require fine tuning of the finite difference parameters.
Let us try to calculate all the first and second order derivatives that we need in the radial velocity fit problem from Sec. \[sec:rvtest\]: $\nicefrac{\partial \chi^2}{\partial a}$, $\nicefrac{\partial \chi^2}{\partial e}$, $\nicefrac{\partial^2 \chi^2}{\partial a^2}$, $\nicefrac{\partial^2 \chi^2}{\partial e^2}$ and the cross term $\nicefrac{\partial^2 \chi^2}{\partial a \partial e}$. To use the finite difference method, we need to choose a *finite* initial differences $\delta a$ and $\delta e$. These are then used to initialize the orbits of shadow particles which are integrated using the normal equations of motion.
The actual value of $\delta a$ and $\delta e$ is crucial. It has to be small enough to ensure the simulation remains in a linear regime (quadratic for second order). However, making the finite differences too small results in loss of accuracy due to finite floating point precision. Thus there is an optimum between these two competing effects. The precise value is problem specific. For the radial velocity test case we find an optimum around $\delta a \sim 10^{-8}$ for first-order derivatives and $\delta a \sim 10^{-6}$ for second-order derivatives.
This problem is illustrated in Fig. \[fig:test4\]. We plot the relative error of the first and second-order derivative as a function of the initial finite differences $\delta a$ and $\delta e$. One can see that the best possible estimate of the first derivative is only accurate to within $10^{-7}$. Worse yet, the best estimate of the second derivative is only accurate to within $10^{-4}$. Using the finite difference approach we cannot obtain a better estimate.
The problem gets even worse if one is interested in the cross-term in the Jacobian, e.g. $\nicefrac{\partial^2 \chi^2}{\partial a \partial e}$. The relative error of this quantity is plotted in the right panel of Fig. \[fig:test4\]. The cross term depends on both finite differences $\delta a$ and $\delta e$. There is only a small area in the $\delta a$/$\delta e$ space that gives reasonably accurate results. Finding the best combination of the initial finite differences is difficult, requires problem-specific fine tuning and becomes quickly infeasible, especially for applications where a wide range of the physical parameter space is explored.
None of these problems exist using the variational equation approach that we present in this paper. Note that we also do not use finite difference for the initialization of variational particles. The entire framework does not contain any small parameters that could lead to numerical problems (cf. $\delta a, \delta e$).
It is worth pointing out that the [[IAS15]{}]{}integrator that we use for the normal $N$-body integration as well as for the variational equations is accurate to machine precision [@ReinSpiegel2015]. Furthermore, the energy error in long-term simulations grows sub-linearly and follows Brouwer’s Law [@Newcomb1899; @Brouwer1937]. [[IAS15]{}]{}is therefore as exact as any integrator can possibly be[^4], using only double precision arithmatic. This statement also applies to the variational equations and therefor to the derivatives we calculate with their help. All derivatives are exact up to machine precision. They can not be calculated more accurately without going to extended precision.
Runtime {#sec:rvtest}
-------
One thing to keep in mind for optimization problems is the computational complexity of a simulation with first and second-order variational equations. If there are $N$ particles and $N_{\rm par}$ free parameters, then the computation time for a simulation with second-order variational equations scales as $N^2(1+N_{\rm par} + \frac12 N_{\rm par} (N_{\rm par}+1)$.
We tested this scaling in a simulation of two planets in which we vary all 14 planet parameters (all orbital parameters and the masses). The results are plotted in Fig. \[fig:test5\] and agree with our estimate.
If every parameter of every particle is varied, one ends up with a runtime that scales approximately as $\frac12 N^4$. This indicates that using variational equations might only be competitive when combined with other methods. However, if another method brings us close to a local minimum (using, e.g., simulated annealing, parallel tempering), then an approach based on variational equations can converge to the local optimum within just a few iterations, in (almost) constant time.
Conclusions {#sec:conclusions}
===========
In this paper we presented the theoretical framework for using second-order variational equations in $N$-body simulations to estimate how particle trajectories vary with respect to their initial conditions. We described a flexible implementation of these equations within the [[REBOUND]{}]{}integrator package.
A major motivation for developing first-order variational equations was to overcome the numerical inaccuracies associated with finite-difference methods that use shadow particles [e.g., @Tancredi2001]. We showed in Sec. \[sec:finitedifference\] that this problem is exacerbated at second order, requiring careful problem-dependent fine tuning. Additionally, the number of shadow particles required by the finite-difference approach is always the same as the corresponding number of variational equations to follow. The variational approach is therefore much more robust and effectively equal in speed.
An important application for second-order variational equations is in solving optimization problems. First derivatives furnish only the right [*direction*]{} to move in a parameter landscape toward a minimum; second derivatives provide a [*scale*]{} for how far one must jump to reach that minimum. If near a minimum, the first and second derivatives furnished by the variational equations can converge to within machine precision of the minimum in just a few iterations. We illustrated this behaviour in both a simple two-planet case (Sec. \[sec:onepar\]) and in the fitting of a radial velocity curve (Sec. \[sec:rvtest\]). Variational equations might also be applied to spacecraft trajectory optimization or asteroid deflection.
The optimization problem presented in this paper uses second order variational equations in connection with the classical optimization algorithm of Newton. One can also use the second-order variational equations in connection with a Markov Chain Monte Carlo (MCMC) method. Specifically, for both Riemann Manifold Langevin and Hamiltonian Monte Carlo methods, higher order derivatives, and therefore higher order variational equations, are essential [@GirolamiCalderhead2011]. A full discussion of these MCMC methods and their application goes beyond the scope of this paper but we note that our initial tests of these methods show great promise. In particular, we observe very short auto-correlation times when using a Riemann Manifold Langevin MCMC to sample the posterior of radial velocity curves.
For a long time, first-order variational equations have been widely used to calculate Lyapunov exponents and the Mean Exponential Growth of Nearby Orbits (MEGNO, @Cincotta2003) in the astrophysics community [@Tancredi2001; @Hinse2010]. We speculate that higher-order variational equations may be able to improve such chaos indicators. Since only one set of variational equations is needed for the calculation of the Lyapunov exponent, including second-order variational equations will keep the numerical scaling $\mathcal{O}(N^2)$ and only increase the computational cost by 50%.
The latest version of [[REBOUND]{}]{}includes the second-order variational equations and can be downloaded at <https://github/com/hannorein/rebound>. The package is free to use under an open source license. We provide also a `git` repository with `jupiter` notebooks to reproduce the figures in this paper at <https://github.com/hannorein/variations>. The notebooks can be run interactively in the web browser without the need to install any software locally.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Eric Ford, Benjamin Nelson and Scott Tremaine for many helpful discussions and Gottfried Wilhelm Leibniz for the chain rule. This research has been supported by the NSERC Discovery Grant RGPIN-2014-04553. D.T. is grateful for support from the Jeffrey L. Bishop Fellowship. This research made use of `iPython` [@ipython], `SciPy` [@scipy] and `matplotlib` [@Hunter2007].
[^1]: In particular, it might be inaccurate if we are interested in the difference of the two solutions. See the discussion in Sec. \[sec:intro\] about shadow particles.
[^2]: This includes the mass of the particle.
[^3]: Newton’s method converges quadratically, i.e. the number of significant digits roughly doubles after every iteration. Thus, if we are close to a local minimum and we work in double floating point precision with 16 significant digits, we need $\sim 4$ iterations to converge to machine precision.
[^4]: To within a constant factor of a few.
|
{
"pile_set_name": "ArXiv"
}
|
---
address:
- |
RIKEN BNL Research Center,\
Brookhaven National Laboratory,\
Upton, New York 11973-5000, USA\
- |
Department of Physics,\
Brookhaven National Laboratory,\
Upton, New York 11973-5000, USA\
- |
Service de Physique Théorique, CP 225,\
Université Libre de Bruxelles,\
Blvd. du Triomphe, 1050 Bruxelles, Belgium
author:
- 'DMITRI KHARZEEV,'
- 'ROBERT D. PISARSKI,'
- 'MICHEL H. G. TYTGAT'
title: PARITY ODD BUBBLES IN HOT QCD
---
=cmr8
1.5pt \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
Introduction
============
In this paper we give a pedagogical introduction to recent work of ours.[@pap] We consider an $SU(N)$ gauge theory in the limit of a large number of colors, $N\rightarrow \infty$. This is, of course, a familiar limit.[@largeN] We use the large $N$ expansion to investigate the behavior of the theory at nonzero temperature, especially for the topological susceptibility. The results depend crucially upon the order of the deconfining phase transition; if it is first order, nothing very interesting happens. If the deconfining transition is of second order, however, the topological susceptibility vanishes in a calculable fashion.
This implies that metastable states, which act like regions with nonzero $\theta$, can appear. Parity is spontaneously broken in such parity odd bubbles, and produces novel physics. The $\eta$ meson becomes very light, at most a few hundred $MeV$, and so is easily produced. As parity is broken, the $\eta$ can decay into two pions, instead of the usual three. We also propose a global variable which can be used to measure an asymmetry in parity.
Large $N$
=========
Holding the number of fermion flavors fixed as $N \rightarrow \infty$, the large $N$ limit is very much a gluonic limit, as the $\sim N^2$ gluons totally dominate the $\sim N$ quarks. This is the basis for most of the conclusions which we can draw at large $N$: given what happens to the gluons, what happens to the quarks follows almost immediately.
The standard assumptions at large $N$ is that the physics is like that for $N=3$: confinement occurs, so at low temperature we can speak entirely of mesons and glueballs. Their masses are assumed to be of order one as $N \rightarrow \infty$; as usually occurs in any large $N$ limit, interactions between either mesons and/or glueballs are suppressed by powers of $1/N$.
It is also natural to assume that the degeneracy of mesons and glueballs is of order one. This is, after all, what we mean by confinement: all trace of the color indices disappear, leaving bound states which are characterized only by spin, parity, etc. Thus in the low temperature phase at a temperature $T$, since each meson or glueball has a free energy which is $\sim N^0 T^4$, the total free energy is also of order one. In contrast, in the deconfined phase at high temperature, the free energy is of order $\sim N^2$. As pointed out by Thorn,[@thorn] this allows us to use the free energy itself as an order parameter for the phase transition: the transition, at a temperature $T_d$, occurs when the term in the free energy $\sim N^2$ turns on.[@rplargeN; @otherlargeN]
Rigorously, the true order parameter for the deconfining phase transition is associated with the spontaneous breaking of the global $Z(N)$ symmetry above $T_d$; this symmetry becomes $O(2)$ as $N\rightarrow \infty$.
We also make a further assumption, namely that [*any*]{} other phase transitions occur at the same time as deconfinement, at $T_d$. We have no proof of this statement, although we suspect that a proof can be constructed in the limit of large $N$. Nevertheless, it strains credulity to image that at the point where this huge increase in the free energy occurs, that that alone doesn’t force any other phase transitions in the theory.
In particular, assume that we couple massless quarks to the gluons. At zero temperature, it is known that the quarks’ chiral symmetry must be broken in the familiar pattern, to a diagonal subgroup of flavor.[@coleman] Then we assume that the chiral symmetry is restored at a temperature $T_\chi$, with $T_\chi = T_d$.
We take the scale of the deconfining transition to be the same as for the glueball masses; thus $T_d$ is of order one as $N\rightarrow \infty$. This turns out to be a remarkably powerful assumption. Consider the large $N$ limit of a theory without confinement, such as a $N$-component vector with coupling $g^2$, holding $g^2 N$ fixed as $N\rightarrow \infty$. We assume that the masses of the fields are of order one. Then the only way for a transition to occur in what is, after all, free field theory, is to go to temperatures which grow with $N$; a simple one loop estimate gives $T_\chi \sim 1/\sqrt{g^2} \sim \sqrt{N}$ (this is also the scale of $f_\pi$, which is natural). What happens in a confining theory is far more dramatic: the transition occurs at temperatures of order one, not $\sim \sqrt{N}$. This implies that the hadronic phase is “cold” at large $N$: interactions are small, so that effects from the thermal bath, such as the loss of manifest Lorentz invariance, can be neglected.
The crucial thing which we do not know about the large $N$ limit is the order of the deconfining phase transition. (The effect of quarks can be neglected, since the gluons dominate the free energy above $T_d$; this will be elaborated later.) For this we must look to the lattice, which as always provides the true intellectual basis for our understanding.
In the early days of Monte Carlo simulations on the lattice,[@Nfour] it was generally agreed that the deconfining phase transition is of first order when $N=4$. It is not clear, however, if these simulations are definitive. In particular, they were done at $n_t=4$, where $n_t$ is the number of steps in the imaginary time direction. For the standard Wilson action, at this value of $n_t$ there is a bulk transition close to the finite temperature transition. The bulk transition can be avoided by going to larger values of $n_t$, but at the time this was computationally difficult to do. Recently, however, Ohta and Wingate[@ohta] have computed for $N=4$ and $n_t=6$; they find that the strong first order transition at $n_t = 4$ is gone for $n_t=6$. Of course, to really establish that there is a true second order phase transition is a difficult matter, requiring lengthy study. But these results do suggest that it may be hasty to conclude from $n_t = 4$ that the deconfining phase transition is of first order.
There are also results on the large $N$ limit of gauge theories on the lattice.[@reduced] Such reduced models appear to reliably predict the ratio of the critical temperature to the square root of the string tension. They predict a first order transition, but only under the technical assumption that the coupling between spacelike plaquettes can be neglected. It is not apparent to us how strong this assumption is.
Previously, Pisarski and Tytgat[@rdptyt] suggested that the large $N$ deconfining phase transition is of second order. Their argument was rudimentary: the easiest way to understand why the deconfining transition is [*weakly*]{} first order for $N=3$ is if the large $N$ expansion is a good approximation, and if the transition is of second order for $N=\infty$. Then the cubic invariant, which drives the transition first order at $N=3$, is suppressed by $\sim 1/N$. Of course the first assumption is rather strong: perhaps the large $N$ expansion is not a good guide to thermodynamic properties.
In the following we assume that the deconfining transition is of second order for all $N \geq 4$, but comment upon how our results change if the transition is of first order.
The principal object we are interested in is the topological suceptibility.[@witten1; @meggiolaro] From the topological charge density, $$Q(x) = (g^2/32\pi^2) tr
(G_{\alpha \beta} \widetilde G^{\alpha \beta}) = \partial_\alpha K^\alpha \; .
\label{eq:ea}$$ The current $K^\alpha$ is gauge dependent. The topological susceptibility is the two point function of $Q$, $$\lambda_{YM}(T) \equiv \partial^2 F(\theta, T)/\partial\theta^2
= \int d^4 x \, Q(x) Q(0) \; ;
\label{eq:eb}$$ $F(\theta,T)$ is the free energy, and the $\theta$ parameter is conjugate to $Q$. At zero temperature, the free energy reduces to the energy, $F(\theta,0)=E(\theta)$.
Since $Q$ is a total derivative, $\lambda_{YM}(T)$ vanishes order by order in perturbation theory. It receives contributions entirely from nonperturbative effects, such as instantons. The action of a single instanton with fixed scale size is $8\pi^2/g^2$. In the large $N$ limit, $g^2 N$ is held fixed as $N\rightarrow \infty$, so the contribution of an instanton to the topological susceptibility is $\lambda_{YM}(T) \sim exp(- a N)$, with $a = 8 \pi^2/(g^2 N)$. Thus the contribution of instantons vanishes exponentially in the large $N$ limit. This naive argument assumes that the integral over instanton scale size is well behaved. This is certainly true in the limit of high temperature; then the theory is weakly coupled, and instantons are suppressed by the Debye screening of electric fluctuations. This naive picture was verified, at all temperatures, by Affleck in a soluble asymptotically free theory, the $CP^N$ model in $1+1$ dimensions[@affleck].
Thus at large $N$, in the deconfined phase the topological susceptibility is exponentially small in $1/N$, and so essentially vanishes. At zero temperature, Witten[@witten1] suggested that instead of semiclassical fluctuations, that quantum fluctuations generate a nonzero value for $\lambda_{YM}(0) \sim N^0$.[@witten1] It is natural to assume that the topological susceptibility is $\sim N^0$ throughout the deconfined phase, and changes to $\sim exp(-a N)$ at the deconfining phase transition. This was previously argued by Affleck[@affleck] and by Davis and Matheson.[@davis]
How it changes depends upon the order of the phase transition. If the deconfining transition is of first order, then as the hadronic phase is “cold”, the most natural possibility is that the topological susceptibility is essentially constant in the hadronic phase, and changes discontinuously to zero at $T_d$. Recent lattice results[@DiGiacomo] in full QCD with four flavors indicate a sharp drop for the topologocal susceptibility across the phase transition, and thus seem to support this conjecture.
If the deconfining transition is of second order, a more extended analysis is necessary.[@pap] Generalizing the results of Witten[@witten1] and Veneziano[@veneziano] to nonzero temperature, and using results on the anomalous couplings of mesons,[@anom] we find that the free energy depends upon $\theta$ as $$F(\theta,T) \;
\sim_{\!\!\!\!\!\!\!\!\!_{ T\rightarrow T_d^-}}
(1 + c \, \theta^2) (T_d - T)^{2 - \alpha} \;\; .
\label{eq:ec}$$ Here $\alpha$ is the critical exponent for the deconfining phase transition, $\alpha \approx -.013$. We then use Witten’s[@witten1] formula for the $\eta'$ mass to conclude that the $\eta'$ mass vanishes at $T_d$, $$m^2_{\eta '}(T) = \frac{4 N_f}{f^2_\pi(T)} \; \lambda_{YM}(T) \;
\sim_{\!\!\!\!\!\!\!\!\!_{ T\rightarrow T_d^-}}
\;(T_d - T)^{1 - \alpha} \; .
\label{eq:ed}$$ Implicitly, we have used the fact that the hadronic phase is cold, so that zero temperature formulas, such as (\[eq:ed\]), generalize trivially.
Parity odd bubbles
==================
We now use this result on the $\eta'$ mass to investigate the nature of the theory in the hadronic phase, just below $T_d$. At zero temperature, a successful phenomenology of the $\eta'$ was developed with a chiral lagrangian formalism.[@witten1; @witten2] For $N_f$ flavors, a $U(N_f)$ matrix $U$ is introduced, satisfying $U^\dagger U = 1$. $U$ describes the $N_f^2 - 1$ pions and the $\eta'$. The effects of the anomaly are represented solely by a mass term for the $\eta'$, $(tr \; ln \; U)^2$.
This is stark contrast to how effects of the anomaly due to instantons are included. Consider a linear sigma model with a field $\Phi$. Then the effects of the anomaly enter exclusively through a term $\sim det(\Phi)$. As in the nonlinear sigma model, when the chiral symmetry is spontaneously broken, this term generates a mass for the $\eta'$. However, in the nonlinear sigma model, at large $N$, there is [*only*]{} a mass term for the $\eta'$; four point interactions between $\eta'$’s are induced by the anomaly, but are suppressed by higher powers of $1/N^2$. As emphasized by Witten,[@witten2] a term $\sim det(\Phi)$ violates this large $N$ counting. This is subject to the trivial qualification that $N_f \geq 4$, so that there are quartic interactions between the $\eta'$’s.
For $U$ fields which are constant in spacetime, the potential for $U$ is $$V(U) = \frac{f_\pi^2}{2}
\left( tr\left( M(U + U^\dagger) \right)
- a (tr \; ln \; U - \theta)^2 \right) \; ;
\label{eq:ee}$$ The pion decay constant $f_\pi = 93\ MeV$, while $M$ is the quark mass matrix. When $M=0$, $m^2_{\eta '} \sim a$, so $a \sim \lambda_{\eta'}^2/N$.
Taking $M_{i j} = \mu_i^2 \delta^{i j}$, any vacuum expectation value (v.e.v) of $U$ can be assumed to be diagonal, $U_{i j} = e^{i \phi_i} \delta^{i j}$. The potential reduces to $$V(\phi_i) = f_\pi^2 \left(
- \sum_{i} \mu_i^2 \; cos(\phi_i)
+ \frac{a}{2} (\sum_i \phi_i - \theta )^2 \right) \; .
\label{eq:ef}$$ This is minimized for $$\mu_i^2 \; sin(\phi_i) + a (\sum \phi_i - \theta) = 0 \; .
\label{eq:eg}$$ Note that as $\sum \phi_i$ arises from $tr \, ln \, U$, it is defined modulo $2 \pi$.
Previously, several authors studied how the v.e.v.’s of the $\phi$’s change as a function of $\theta$.[@largeNphen1; @largeNphen2] In the present work, we consider $\theta=0$, but consider how the $\mu_i$ and $a$ change with temperature. Witten[@witten2] pointed out that when the anomaly term $a$ becomes small, metastable states in the $\phi$’s can arise. From our arguments in the previous section, this happens naturally if the phase transition at large $N$ is of second order.
The presence of these metastable states can be easily understood for a single flavor, as discussed by Witten[@witten2] (for a recent discussion see[@shifman; @hz]). From (\[eq:eg\]), the v.e.v. arises from a balance between a term $\sim sin(\phi)$ and a term $\sim \phi$. For large $a$, the term linear in $\phi$ wins, and there is no possibility for a metastable point. Now consider the opposite limit, of vanishing $a$: then there automatically other solutions besides $\phi=0$, $\phi = 2 \pi, 4 \pi$, etc. These solutions are equivalent to the trivial vacuum, and so there is nothing new. But for small values of $a$, the term linear in $a$ will only move the stationary point a little bit from $2 \pi$, $4\pi$, etc. Because $a$ is nonzero, they will become metastable, distinct from the usual vacuum.
From (\[eq:eg\]), these states will act like regions of nonzero $\theta$. Parity and CP are both violated spontaneously in such a region.
The condition for metastable states to arise with more than one flavor is not apparent, and a new result of our analysis.[@pap] It is easiest understood by analogy. At zero temperature, and nonzero $\theta$, if any quark mass vanishes, the $\theta$ parameter can be eliminated by a chiral rotation through that quark flavor. Thus it is the lightest quark mass which controls $\theta$ dependence. We found a similar phenomenon for metastable states: they only occur when the anomaly term is small relative to the lightest quark masses.
This means that metastable states only arise when the anomaly term becomes very small. At zero temperature, the anomaly term is on the order of the strange quark mass. The previous argument indicates that it must become on the order of the up and down quark masses. Putting in the numbers, we find that metastable states only arise when the anomaly term becomes on the order of $1\%$ of its value at zero temperature. Clearly this is a strong variation of the topological susceptibility with temperature; nevertheless, it is interesting to investigate the possible implications for phenomenology.
Most notably, when the anomaly term $a$ becomes small, there is maximal violation of isospin.[@largeNphen1; @ren1; @small1; @small2] At zero temperature, the nonet of pseudo-Goldstone bosons — the $\pi$’s, $K$’s, $\eta$, and $\eta'$, are, to a good approximation, eigenstates of $SU(3)$ flavor. It is not often appreciated, but this is really due to the fact that the anomaly term is large, splitting off the $\eta'$ to be entirely an $SU(3)$ singlet. When the anomaly term becomes small, however, while the charged pseudo-Goldstone bosons remain approximate eigenstates of flavor, the neutral ones do not. Without the anomaly, the $\pi^0$ becomes pure $\overline{u} u$, the $\eta$ pure $\overline{d} d$, and the $\eta'$ pure $\overline{s} s$. Consequently, these three mesons become light. This is especially pronounced for the $\eta$, as it sheds all of its strangeness, to become purely $\overline{d} d$. Thus the $\eta$ and $\eta'$ would be produced copiously, and would manifest itself in at least two ways. First, light $\eta$’s and $\eta'$’s decay into two photons, and so produce an excess at low momentum. Secondly, these mesons decay into pions, which would be seen in Bose-Einstein correlations[@bose]. Further, through Dalitz decays, the enhanced production of $\eta$’s and $\eta'$’s will enhance the yield of low mass dileptons [@small2].
This maximal violation of isospin is true whenever the anomaly term becomes small. There are other signals which only appear when parity odd bubbles are produced. Since parity is spontaneously violated in such a bubble, various decays, not allowed in the parity symmetric vacuum, are possible. Most notably, the $\eta$ can decay not just to three pions, as at zero temperature, but to two pions. Because of the kinematics, in a parity odd bubble, $\eta \rightarrow \pi^0 \pi^0$ is allowed, but $\eta \rightarrow \pi^+ \pi^-$ is not.
There is another measure of how parity may be violated. We first argue by analogy. Consider propagation in a background magnetic field. As charged particles propagate in the magnetic field, those with positive charge are bent one way, and those with negative charge, the other. This could be observed by measuring the following variable globally, on an event–by–event basis: $${\cal P} = \sum_{\pi^+\pi^-}
\frac{(\vec{p}_{\pi^+} \times \vec{p}_{\pi^-})\cdot \vec{z}}
{ |\vec{p}_{\pi^+}| |\vec{p}_{\pi^-}|} \; ;
\label{eq:eh}$$ here $\vec{z}$ is the beam axis, and $\vec{p}$ are the three momenta of the pions.
If the quarks were propagating through a background chromo-magnetic field, then ${\cal P}$, which is like handedness in jet physics,[@hand] is precisely the right quantity. However, a parity odd bubble is not directly analogous to a background chromo–magnetic field: $\pi^+$’s and $\pi^-$’s propagate in a region with constant but nonzero $\phi$ in the same fashion. Consider, however, the edge of the parity odd bubble: in such a region, $U^\dagger \partial_\mu U$ is nonzero, and does rotate $\pi^+$ and $\pi^-$ in opposite directions. Thus it is the edges of parity odd bubbles which contribute to the parity odd asymmetry of (\[eq:eh\]). Purely on geometric grounds, this suggests that a reasonable estimate for the maximal value of ${\cal P}$ is on the order of a few percent.
We conclude by noting that what appears to be a rather technical subject — the $\theta$ dependence of the free energy — is related to interesting and novel experimental signatures in heavy ion collisions. Within our assumptions, we find that parity odd bubbles only arise very near the point of the phase transition. This is very much tied to the fact that we limit ourselves to an analysis at large $N$. For finite $N$, it is a long standing question of how to reconcile the known limit at large $N$ with periodicity in $\theta$, with period $2 \pi$. A probable solution involves “glued” potentials, which are a sum of cosines; see[@shifman; @hz] and recently.[@theta] The precise form of the potential at finite $N$ could dramatically alter our results, and, as discussed by Halperin and Zhitnitsky,[@theta] make the emergence of parity odd bubbles far more likely.
References {#references .unnumbered}
==========
[99]{} D. Kharzeev, R. D. Pisarski, and M. H. G. Tytgat, . G. ’t Hooft, ; E. Witten, ; A. V. Manohar, [*hep-ph/9802419*]{}; M. Teper, ; [*hep-lat/9804008*]{}. C. B. Thorn, . R. D. Pisarski, . J. J. Atick and E. Witten, ; J. Polchinski, ; T. H. Hansson and I. Zahed, . S. Coleman and E. Witten, . G. G. Batrouni and B. Svetitsky, ; A. Gocksch and M. Okawa, ; F. Green, ; J. F. Wheater and M. Gross, ;. M. Billo, M. Caselle, A. D’Adda, and S. Panzeri, [*Int. J. Mod. Phys. A*]{} [**12**]{}, 1783 (1997). S. Ohta and M. Wingate, [*hep-lat/9808022*]{}. R. D. Pisarski and M. H. G. Tytgat, [*hep-ph/9702340*]{}. E. Witten, [*Nucl. Phys. B*]{} [**156**]{}, 269 (1979). G. Veneziano, [*Nucl. Phys. B*]{} [**159**]{}, 213 (1980). P. Di Vecchia and G. Veneziano, [*Nucl. Phys. B*]{} [**171**]{}, 253 (1980). P. Di Vecchia, [*Phys. Lett. B*]{} [**85**]{}, 357 (1979); P. Di Vecchia, F. Nicodemi, R. Pettorino, and G. Veneziano, [*Nucl. Phys. B*]{} [**181**]{}, 318 (1981); P. Nath and R. Arnowitt, [*ibid.*]{} [**209**]{}, 234, 251 (1982); C. Rosenzweig, J. Schechter, and C. G. Trahern, [*Phys. Rev. D*]{} [**21**]{}, 3388 (1980). E. Witten, [*Annals Phys.*]{} [**128**]{}, 363 (1980). N. Evans, S. D. H. Hsu, and M. Schwetz, [*Nucl. Phys. B* ]{} [**484**]{}, 124 (1997); [*ibid.*]{}, [**494**]{}, 200 (1997); M. Shifman, [*Prog. Part. Nucl. Phys.*]{} [**39**]{}, 1 (1997); I. I. Kogan, A. Kovner, and M. Shifman, [*Phys. Rev. D*]{} [**57**]{}, 5195 (1998). I. Halperin and A. Zhitnitsky, [*hep-ph/9707286; hep-ph/9803301*]{}. I. Affleck, ; [*ibid.*]{}, [**171**]{}, 420 (1980). A. C. Davis and A. M. Matheson, [*Phys. Lett. B*]{} [**179**]{}, 135 (1986); [*Nucl. Phys. B*]{} [**258**]{}, 373 (1985). P. de Forcrand, M. G. Perez, J. E. Hetrick, and I.-O. Stamatescu, [*hep-lat/9802017*]{}; B. Alles, M. D’Elia, A. Di Giacomo, and P. W. Stephenson, [*hep-lat/9808004*]{}. E. Meggiolaro, [*Zeit. fur Phys.*]{} [**62**]{}, 679, 669 (1994); [*ibid.*]{} [**64**]{}, 323 (1994); [*hep-th/9802114*]{}. R.D. Pisarski, in [*From thermal field theory to neural networks: a day to remember Tanguy Altherr*]{}, edited by P. Aurenche, P. Sorba, and G. Veneziano. (World Scientific Publishing, Singapore, 1996); [*Phys. Rev. Lett.*]{} [**76**]{}, 3084 (1996); R. Baier, M. Dirks, and O. Kober, [*Phys. Rev. D*]{} [**54**]{}, 2222 (1996); R. D. Pisarski, T. L. Trueman, and M. H. G. Tytgat, [*ibid.*]{}, [**56**]{}, 7077 (1997). R. D. Pisarski and F. Wilczek, [*Phys. Rev. D*]{} [**29**]{} (1984) 338. Z. Huang, [*Phys. Rev. D*]{} [**49**]{}, 16 (1994). J. Kapusta, D. Kharzeev and L. McLerran, [*Phys. Rev. D*]{} [**53**]{}, 5028 (1996); Z. Huang and X.-N. Wang, [*ibid.,*]{} [**53**]{}, 5034 (1996). S. E. Vance, T. Csörgő and D. Kharzeev, [*nucl-th/9802074; Phys. Rev. Lett.*]{}, to appear. O. Nachtmann, [*Nucl. Phys. B*]{} [**127**]{}, 314 (1977); A.V. Efremov, [*Sov. J. Nucl. Phys.*]{} [**28**]{}, 83 (1978); A. Efremov and D. Kharzeev, [*Phys. Lett. B*]{} [**366**]{}, 311 (1996). A. V. Smilga, [*hep-ph/9805214*]{}; E. Witten, [*hep-th/9807109*]{}; I. Halperin and A. Zhitnitsky, [*hep-ph/9807335*]{}.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The complex physical properties of highly deformable materials like clothes pose significant challenges for autonomous robotic manipulation systems. We present a novel visual feedback dictionary-based method for manipulating deformable objects towards a desired configuration. Our approach is based on visual servoing and we use an efficient technique to extract key features from the RGB sensor stream in the form of histogram of deformable model features. These histogram features serve as high-level representations of the state of the deformable material. Next, we collect manipulation data and use a visual feedback dictionary that maps the velocity in the high-dimensional feature space to the velocity of the robotic end-effectors for manipulation. We have evaluated our approach on a set of complex manipulation tasks as well as human-robot manipulation tasks on different cloth pieces with varying material characteristics.'
author:
- 'Biao Jia, Zhe Hu, Jia Pan, Dinesh Manocha'
bibliography:
- 'ref.bib'
title: Manipulating Highly Deformable Materials Using a Visual Feedback Dictionary
---
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we present a new calculation of composition-dependent radiative cooling and heating curves of low-density gas, intended primarily for use in numerical simulations of galaxy formation and evolution. These curves depend on only five parameters: temperature, density, redshift, \[Fe/H\], and \[Mg/Fe\]. They are easily tabulated and can be efficiently interpolated during a simulation.
The ionization equilibrium of 14 key elements is determined for temperatures between $10$ K and $10^9$ K and densities up to 100 amu cm$^{-3}$ taking into account collisional and radiative ionization, by the cosmic UV background and an interstellar radiation field, and by charge-transfer reactions. These elements, ranging from H to Ni, are the ones most abundantly produced and/or released by SN[i]{}a, SN[ii]{}, and intermediate-mass stars. Self-shielding of the gas at high densities by neutral Hydrogen is taken into account in an approximate way by exponentially suppressing the H-ionizing part of the cosmic UV background for H[i]{} densities above a threshold density of $n_{\rm HI, crit}=0.007$ cm$^{-3}$. We discuss how the ionization equilibrium, and the cooling and heating curves depend on the physical properties of the gas.
The main advantage of the work presented here is that, within the confines of a well-defined chemical evolution model and adopting the ionization equilibrium approximation, it provides accurate cooling and heating curves for a wide range of physical and chemical gas properties, including the effects of self-shielding. The latter is key to resolving the formation of cold, neutral, high-density clouds suitable for star formation in galaxy simulations.
author:
- |
S. De Rijcke$^{1}$[^1], J. Schroyen$^{1}$, B. Vandenbroucke$^{1}$, N. Jachowicz$^{2}$, J. Decroos$^{1}$, A. Cloet-Osselaer$^{1}$, M. Koleva$^{1}$\
$^{1}$Ghent University, Dept. Physics & Astronomy, Krijgslaan 281, S9, B-9000, Ghent, Belgium\
$^{2}$ Ghent University, Dept. Physics & Astronomy, Proeftuinstraat 86, B-9000, Ghent, Belgium
title: New composition dependent cooling and heating curves for galaxy evolution simulations
---
\[firstpage\]
Physical Data and Processes: atomic processes, hydrodynamics, plasmas, ISM: general
Introduction {#intro}
============
Numerical simulations of galaxy evolution require basic physical input regarding the (thermo-)dynamical behavior of the interstellar gas. A crucial ingredient of the energy, or entropy, equation is the cooling rate of the gas. This quantity is, in principle, a complex function of the temperature, composition, and irradiation of the gas. An often used assumption is that the gas is in collisional ionization equilibrium (CIE). In that case, collisions with free electrons are deemed solely responsible for keeping atoms ionized. Since both the recombination rate and the ionization rate are in that case directly proportional to the electron density, the latter cancels from the equations and the ionization equilibrium becomes a function of temperature only (for a given elemental abundance mix). For low gas densities, each collisional ionization/excitation is followed by a radiative de-excitation, creating an escaping photon, and the cooling rate becomes proportional to the density squared (or to Hydrogen density times electron density, $n_{\rm H} n_{\rm e}$) times a temperature-dependent function.
{width="\textwidth"}
Many state-of-the-art simulation codes [@re09; @sa11; @sc11; @sch11; @co12; @gd12; @ki12] rely on the cooling curves compiled by @sd93. The latter authors calculated cooling rates, excluding a forefactor $n_{\rm H} n_{\rm e}$, as a function of temperature for a number of metallicities. During a simulation, the cooling rate of a gas parcel can be rapidly determined by simple two-dimensional interpolation on these curves. However, while this work was monumental and has spawned a large volume of literature based on simulations making use of these curves, one needs to be aware of the assumptions on which these cooling curves are based and simulators need to assess whether they can be used for the application at hand. To be clear: this is no criticism of the @sd93 cooling curves.
- Metallicity is quantified using \[Fe/H\], the Iron abundance. For \[Fe/H\]$\le -1$ the abundance ratios are taken to reflect those of a SN[ii]{}, with \[$\alpha$/Fe\]$>0$; for \[Fe/H\]$=0$ the solar abundance ratios are adopted, with \[$\alpha$/Fe\]$=0$; for other metallicities, the abundance ratios are interpolated linearly between these two sets of abundance ratios. Hence, [*adopting the @sd93 cooling curves immediately implies adopting the solar neighborhood’s chemical enrichment history.*]{} In some cases, this may not be a good approximation of reality. E.g., dwarf galaxies have low metallicities but also low \[$\alpha$/Fe\] (see e.g. @tht09 and references therein). Using the low-metallicity @sd93 cooling curves will then quite strongly overestimate the cooling contributed by $\alpha$-elements such as O, Si, and Mg. The centers of giant elliptical galaxies, on the other hand, have high metallicities and high \[$\alpha$/Fe\] [@wfg92]. With the @sd93 cooling curves, the contribution of the $\alpha$ elements will be strongly underestimated.
- Another issue is whether in the presence of a cosmological UV background (UVB) CIE is still an acceptable approximation. The UVB tends to keep (part of) the Hydrogen and metals ionized, thus lowering the fraction of H[i]{} and raising $n_{\rm e}$. This dramatically influences the shape of the cooling curve, as is well known [@wi09] and as we will also show below. @tp11 compared the net cooling rates computed assuming only CIE with those calculated including photo-ionization. These authors showed that the equilibrium temperature of the gas could be off by an order of magnitude at low densities and high metallicities when using CIE.
- Simulations nowadays achieve sufficiently high spatial resolution to be able to follow the formation of cold, dense star-forming clouds. This requires extending the @sd93 cooling curves to temperatures below $10^4$ K. An often used extension is the set of cooling curves of @ma07. These authors calculate the level populations of Fe[ii]{}, O[i]{}, Si[ii]{}, and C[ii]{} and the cooling rates due to the low-lying finestructure emission lines of these ions. These level populations are set by collisions with free electrons and thus by the ionization fraction of the gas. Again, this is a quantity which is very sensitive to the presence of a cosmic UVB. Clearly, the often used approach of “gluing” the @ma07 cooling curves ($10$ K$<T<10^4$ K) to the @sd93 cooling curves ($10^4$ K$<T<10^9$ K) will produce unreliable results in simulations with a cosmic UVB.
Recent calculations of cooling curves with CLOUDY, such as @wi09 (used in the OverWhelmingly Large Simulations project [@sch10]) and @sh10 (used in the simulation code GASOLINE [@sh10; @br12]), include the cosmic UVB in determining the ionization balance. Here, the total cooling rate is written as the sum of approximately independent terms: the cooling due to H and He, the cooling due to metals, and inverse Compton cooling. @wi09 advocate an element-by-element approach, necessitating the tracing of the abundances of a set of 11 elements during a simulation in order to calculate the independent contribution of each to the total cooling rate. In reality, these terms are linked by the free electron density, by charge-exchange reactions, and by other reactions between different elements (such as molecule and dust particle formation). Moreover, many authors still adopt the solar abundance ratios and scale the metal cooling rate proportional to metallicity which, as we have argued, can be expected to be a bad representation of reality for certain types of stellar systems. Likewise, @ssa08, and @sm11 only provide cooling curves for solar abundance ratios of the heavy elements (H and He ions are followed explicitly during simulations). In the former paper, the influence of the UVB is not taken into account while the latter focuses precisely on this issue. @gh12 incorporate the radiation field through four well-chosen normalized photoionization rates. This, together with a Taylor expansion of the curves up to quadratic terms in metallicity, yields an approximation to the cooling and heating curves with a median error of 10 % but with (although very rare) errors of up to a factor of six.
In this paper, we try to improve on several aspects of the existing cooling curve calculations.
Numerical details {#sect:num}
=================
Below, we give a list of the most prominent ingredients of our calculations:
- We adopt a chemical enrichment model that is self-consistent in the sense that, in an $N$-body/SPH simulation, gas particles can be enriched by stellar particles in only two ways: fast (by SN[ii]{} and massive intermediate-mass stars (IMS), with $M \sim
8$ M$_\odot$) and slow (by SN[i]{}a and less massive IMS). Thus, the chemical abundance mix of a gas particle depends solely on the ratio of the “slow” and “fast” contributions. The cooling and heating rates can then be tabulated for a small number of different ratios of “slow” to “fast” contributions, covering all possibilities that can occur in a simulation.
SN[ii]{} yields are taken from @no97 and modified according to the prescriptions detailed in @fr04; IMS yields come from @ga05; SN[i]{}a yields have been adopted from @ts95. For SN[ii]{}, the fraction of the initial mass of a stellar population that is returned in the form of element $X$ to the interstellar medium (ISM) is given by $$y_{X, \rm SNII} = \frac{ \int_{m_{\rm lII}}^{m_{\rm uII}} M_X(m) \phi(m) dm }{\int_{m_{\rm l}}^{m_{\rm u}} m \phi(m) dm }.$$ Here, $m_{\rm lII}=8$ M$_\odot$ and $m_{\rm uII}=70$ M$_\odot$ are the lower and upper bounds of the masses of stars that turn into SN[ii]{}; $m_{\rm l}=0.1$ M$_\odot$ and $m_{\rm u}=70$ M$_\odot$ are the adopted lower and upper bounds of the masses of stars. The mass returned in the form of element $X$ by a star with initial mass $m$ is denoted by $M_X(m)$. For the initial-mass function, or IMF, denoted by $\phi(m)$, we take the parameterization by @cha.
For SN[i]{}a, the fraction of the initial mass of a stellar population that is returned in the form of element $X$ to the ISM is given by $$y_{X, \rm SNIa} = A_{\rm Ia} M_X \frac{ \int_{m_{\rm lIa}}^{m_{\rm
lII}} \phi(m) dm }{\int_{m_{\rm l}}^{m_{\rm u}} m
\phi(m) dm }.$$ Here, $m_{\rm lIa}=3$ M$_\odot$ is the lower bound of the masses of stars that can turn into SN[i]{}a and $M_X$ is the SN[i]{}a yield of element $X$. The forefactor $A_{\rm Ia}$ was determined by demanding that the calculated ratio of the occurrence of SN[ i]{}a to that of SN[ii]{} reproduces that derived for the solar neighborhood by @ts95: $$\frac{N_{\rm Ia}}{N_{\rm II}} = 0.15 = A_{\rm Ia} \frac{ \int_{m_{\rm lIa}}^{m_{\rm
lII}} \phi(m) dm }{\int_{m_{\rm lII}}^{m_{\rm uII}} \phi(m) dm }.$$
For IMS, the yield is given analogously by $$y_{X, \rm IMS} = \frac{ \int_{m_{\rm lIMS}}^{m_{\rm uIMS}} M_X(m)
\phi(m) dm }{\int_{m_{\rm l}}^{m_{\rm u}} m \phi(m) dm },$$ with $m_{\rm lIMS}=0.8$ M$_\odot$ and $m_{\rm
uIMS}=8$ M$_\odot$. The yields of elements contributed by the most massive IMS, such as ${}^{13}$C and N, are added to the corresponding SN[ii]{} yields. Those of elements produced by longer-lived stars are added to the corresponding SN[i]{}a yields. This way, there are two contributions to the yield of a given element: a “fast” one (encompassing the contributions from SN[ii]{} and massive IMS) and a “slow” one (encompassing the contributions from SN[i]{}a and less massive IMS). The abundance of each chemical element in a gas parcel is then simply the weighted sum of these two contributions.
[|c|c|c|]{} element & slow & fast\
O & 0.000136 & 0.000937\
C & 0.000146 & 0.000143\
Ne & 2.41e-05 & 0.00013\
Mg & 1.06e-05 & 6.23e-05\
Si & 1.66e-05 & 4.89e-05\
Fe & 2.96e-05 & 1.67e-05\
S & 9.15e-06 & 1.41e-05\
N & 5.87e-05 & 9.65e-06\
Al & 6.46e-08 & 6.51e-06\
Na & 4.86e-09 & 3.36e-06\
Ni & 2.99e-07 & 1.67e-06\
Ca & 9.99e-07 & 1.64e-06\
Since this simple chemical evolution model contains two sets of yields, the elemental abundance ratios in a given gas parcel can also be quantified by just two numbers. Here, we choose \[Fe/H\] as a tracer of the overall metallicity and \[Mg/Fe\] as a second parameter. Mg is an $\alpha$-element released abundantly by SN[ ii]{} explosions but produced in only very small amounts by SN[ i]{}a explosions so it is a good tracer of the relative weights of the “fast” and “slow” contributions. Moreover, there are now quite advanced techniques available to determine the abundances of both Fe and Mg in a stellar population from absorption lines in optical spectra. Obviously, Oxygen would also make a good discriminator between the “fast” and “slow” contributions. We present the element yields used in this paper, expressed as a fraction of the initial mass of a single stellar population, in Table \[tab:yields\].
In Fig. \[fig:compsun\], we compare the observed abundances of a number of well-studied stars, taken from @wo09 [@gr10; @fi10; @la11], with those predicted by our chemical evolution model for the appropriate \[Fe/H\] and \[Mg/Fe\] values. For the sun and Arcturus, the largest discrepancies are about $0.2$ dex. For Sirius and Vega, there are larger deviations between data and model, although the author-to-author scatter on the measured abundances of these stars is, admittedly, quite substantive (e.g. the Na abundance in the atmosphere of Sirius varies by more than one dex between different authors, see e.g. @la11). Taking into account observational uncertainties and genuine cosmic scatter, this is reassuring evidence that the simple two-yield chemical evolution model employed here works adequately.
- The ionization balance, electron density, level populations, and cooling rates are calculated self-consistently in the presence of a cosmic UVB. Here, we adopt the UVB calculated by @fg09. We use the UV spectra available from this author’s webpage[^2] to calculate the ionization and heating rates of all elements.
- Stars also generate an interstellar radiation field (ISRF), capable of ionizing atoms with small ionizing potentials, such as C[i]{}, Si[i]{}, Mg[i]{}, Ca[i]{}, Ca[ii]{}, Fe[i]{}, etc. even at very low gas temperatures. Since the light of newly formed stars had to make its way through the H[ii]{} regions surrounding these stars, it no longer has an H-ionizing component. We include the parameterized ISRF of @ma83, appropriate for the solar neighborhood. The ISRF’s main task is to keep the elements mentioned above ionized while its precise form has shown to be of little consequence. At low gas densities, the UVB is the dominant photo-ionizing radiation field while at high densities, where star formation becomes important and neutral Hydrogen can shield the gas from the UVB, the ISRF gains importance.
- Through Hydrogen ionizations and cascade recombinations, the H-ionizing portion of diffuse UV radiation impacting on a gas cloud is converted into lower-frequency radiation. Thus, for sufficiently high densities, gas may become self-shielding against H-ionizing UV radiation once Hydrogen recombines. Therefore, self-shielding will generally be insignificant for temperatures $T \gtrsim 10^4$ K since then Hydrogen is collisionally ionized anyway.
This is not a straightforward problem since it in principle requires solving the radiative transfer equation. However, one can estimate the critical H density above which self-shielding can be expected to block most of the ionizing UV radiation. @tu98 put forward $n_{\rm H}
\sim 0.01$ cm$^{-3}$, @at10 estimate that $n_{\rm
H}=0.007$ cm$^{-3}$, and @ya11 quote $n_{\rm H} =
0.00634$ cm$^{-3}$. With this cut-off density, it is possible to reproduce the observed mass- and volume-averaged neutral fraction of the universe at a redshift $z\sim 6$ [@at10], the H[i]{} column density distribution of damped Ly$\alpha$ systems $z = 3$ [@na10], and the Ly$\alpha$ luminosity of forming galaxies [@f10].
We have implemented an approximate scheme for self-shielding by exponentially suppressing the UV radiation field with frequencies above $h\nu=\chi_{\rm HI}$, with $\chi_{\rm HI}$ the Hydrogen ionization potential, as $$\begin{aligned}
J_\nu(\nu, n_{\rm HI}) &=& J_\nu(\nu)\exp( -n_{\rm HI}/n_{\rm HI, crit} ) \hspace*{1em} h\nu>1\textrm{Ry} \nonumber \\
&=& J_\nu(\nu) \hspace*{9.2em} h\nu\le 1\textrm{Ry},\end{aligned}$$ with $n_{\rm HI, crit}=0.007$ cm$^{-3}$ and $J_\nu(\nu)$ the original UV spectrum, as in @at10. Note that we use the [*neutral*]{} H density here, not the [*total*]{} H density, since it is only the neutral fraction which is responsible for absorbing H-ionizing UV radiation.
One could worry that, when the gas at high densities and low temperatures becomes self-shielding against the external UVB, the cooling radiation itself may become trapped and be re-absorbed, affecting the cooling rate and the ionization equilibrium [@gs07]. However, in a self-shielding H[i]{} cloud below $T
\sim 10^4$ K, only low-energy UV photons unable to photo-heat the Hydrogen gas are emitted. Moreover, with most of the star-formation prescriptions currently popular in galaxy evolution and cosmological simulations [@go10; @sch11; @co12], such clouds will begin to form stars before reaching densities exceeding $100$ amu cm$^{-3}$ and stellar and supernova feedback will rapidly overwhelm any internal diffuse radiation field. Therefore, we expect this to be a minor issue.
- Charge-exchange reactions can efficiently transfer electrons between ions with similar ionization potentials. Given their high abundances, H[i]{} and H[ii]{} are the ions’ most likely reaction partners. Some of the ions that play an important role in gas cooling below $10^4$ K via fine-structure line emission, such as O[i]{}, are particularly affected by these reactions. We adopt the charge-transfer reaction rates for C[i]{}, C[ii]{}, O[i-v]{}, Si[i-v]{}, and Fe[i-v]{} from @kf96 [@st98; @st99] and from the online ORNL/UGA Charge Transfer Database for Astrophysics (http://www-cfadc.phy.ornl.gov/astro/ps/data/).
- Charge-exchange reactions are but one example of reactions that involve ions of different elements. Other examples are molecule and dust particle formation. We opted not to include molecular processes in the present work for the following reasons. Judging from e.g. @ma07 and @va13, cooling below $T \sim
10^3$ K by molecules is dominated by cooling by metals once the latter are present at levels $Z \gtrsim 10^{-3} Z_\odot$. However, at the low metallicities where H$_2$ cooling might be important, the low dust content strongly inhibits the formation of H$_2$ and the timescale for the conversion of H[i]{} to H$_2$ becomes much larger than the local free-fall time [@kr12]. Hence, in this metallicity range, star formation will precede the formation of H$_2$. Also, the rates at which dust forms and at which H$_2$ molecules form via grain catalysis contain many extra free parameters (such as the timescales for dust formation, growth, and destruction, the dust grain size, the probability for an H atom to stick to a grain, the probability that two H atoms on a grain join and detach themselves from the grain as a single H$_2$ molecule, etc.) that need to be constrained by experiments and observations [@b13]. Still, by neglecting molecular cooling we might be under-estimating the cooling rate at temperatures below $T \sim
10^3$ K for metallicities below $Z \sim 10^{-3} Z_\odot$. However, by not using the “independent element” approximation, as in @wi09 and @sh10, to calculate the ionization balance and the cooling curves, our approach can straightforwardly be extended in future work.
- Using all these ingredients, we calculate cooling and heating curves for the temperature range $10~{\rm K} < T < 10^9$ K. This way, the cooling and heating rates are calculated in perfect consistence with the ionization equilibrium over a very wide range of temperatures. Thus, there is no need to stitch together cooling curves from different authors with potentially very different (and inconsistent) input physics.
For this work, we have extended the capabilities of ChiantiPy, a Python interface to the CHIANTI atomic database [@de09], available from [http://chiantipy.sourceforge.net/]{}. For all ions, we use the recombination rates, collisional ionization rates, and energy level populations provided by standard ChiantiPy. Photo-ionization cross-sections $\sigma_i(\nu)$ are adopted from @ve96 and integrated over the stellar and cosmic UV backgrounds in order to obtain the photo-ionization rate $$\Gamma_i = 4\pi \int_{\nu_i}^\infty \sigma_i(\nu) J_\nu(\nu) \frac{d\nu}{h\nu},$$ and the photo-heating rate $$\dot{q}_i = 4 \pi \int_{\nu_i}^\infty \sigma_i(\nu) J_\nu(\nu)
(h\nu-h\nu_i) \frac{d\nu}{h\nu},$$ with $\nu_i$ the ion’s ionization threshold. The integral over the radiation backgrounds is split in two parts: the part for photon energies above 1 Ry, which can be suppressed by H[i]{} self-shielding, and the part for photon energies below 1 Ry, which is assumed to be unaffected by self-shielding. For instance, the photo-ionization rate of an element can then be written as $$\begin{aligned}
\Gamma_i &=& {\rm e}^{-n_{\rm HI}/n_{\rm HI, crit}}4\pi \int_{1~\rm Ry}^\infty
\sigma_i(\nu) J_\nu(\nu) \frac{d\nu}{h\nu} \nonumber \\
&& \hspace*{2.5em}+ 4\pi \int_{\nu_i}^{1~\rm Ry}
\sigma_i(\nu) J_\nu(\nu) \frac{d\nu}{h\nu}.\end{aligned}$$
![Carbon ions ionization timescale, $\tau_{\sf ion}$ (colored curves), and cooling timescale, $\tau_{\sf cool}$ (red curve), as a function of temperature for different redshifts and densities (as indicated in the panels). $\tau_{\sf ion}$ is only plotted if the fraction of the corresponding C ion is above $0.001$. The dashed curves trace the ionization timescale for pure collisional ionzation (this only makes a difference for C[i]{} at high densities). \[fig:timescales\]](fig2a.eps "fig:"){width="48.50000%"} ![Carbon ions ionization timescale, $\tau_{\sf ion}$ (colored curves), and cooling timescale, $\tau_{\sf cool}$ (red curve), as a function of temperature for different redshifts and densities (as indicated in the panels). $\tau_{\sf ion}$ is only plotted if the fraction of the corresponding C ion is above $0.001$. The dashed curves trace the ionization timescale for pure collisional ionzation (this only makes a difference for C[i]{} at high densities). \[fig:timescales\]](fig2b.eps "fig:"){width="48.50000%"} ![Carbon ions ionization timescale, $\tau_{\sf ion}$ (colored curves), and cooling timescale, $\tau_{\sf cool}$ (red curve), as a function of temperature for different redshifts and densities (as indicated in the panels). $\tau_{\sf ion}$ is only plotted if the fraction of the corresponding C ion is above $0.001$. The dashed curves trace the ionization timescale for pure collisional ionzation (this only makes a difference for C[i]{} at high densities). \[fig:timescales\]](fig2c.eps "fig:"){width="48.50000%"}
{width="102.00000%"}
{width="102.00000%"}
For a given temperature, the ionization equilibrium, i.e. the density of the $r$-times ionized ion of any element with atomic number $Z$, denoted by $n_{Z,r}$, is found employing a multi-dimensional Newton-Raphson technique using back-tracking. At each iteration, the electron density is given by $$n_e = \sum_{Z\ge 1} \sum_{r= 0}^Z r n_{Z,r}.$$ The general equilibrium condition then becomes $$\begin{aligned}
\lefteqn{ n_{Z,r} \biggl[ n_e {\sf Rec}_{Z,r}(T) + \sum_c n_c {\sf
Ion}^{c}_{Z,r}(T) + \Gamma_{Z,r}} \nonumber \\&& \hspace*{4em} +
n_{\rm HI}{\sf CT}^{\rm HI}_{Z,r}(T)+ n_{\rm HII}{\sf CT}^{\rm
HII}_{Z,r}(T) \biggr] \nonumber \\&&= n_{Z,r+1} \left[ n_e {\sf
Rec}_{Z,r+1}(T) + n_{\rm HI} {\sf CT}^{\rm HI}_{Z,r+1}(T) \right]
\nonumber \\&&+ n_{Z,r-1} \biggl[ \sum_c n_c {\sf Ion}^{c}_{Z,r-1}(T)
+ n_{\rm HII} {\sf CT}^{\rm
HII}_{Z,r-1}(T) \nonumber \\&& \hspace*{4em} + \Gamma_{Z,r-1} \biggr].\end{aligned}$$
Here, ${\sf Rec}$ indicates the ionic recombination rate, ${\sf
Ion}^c$ represents the collisional ionization rate with collisional partner $c$ (which could be electrons, protons, Hydrogen atoms, …), and ${\sf CT}$ stand for charge-transfer reaction rates (some of which are obviously zero, such as the reaction rate between H[ i]{} and a neutral atom). Given the strong, in this case exponential, dependence of the self-shielding on the neutral Hydrogen fraction this is clearly a very non-linear set of equations.
Moreover, the UVB, at least at sufficiently low gas densities, can be expected to keep a large fraction of the Hydrogen gas ionized, thereby suppressing the bound-bound and free-bound cooling contributed by Hydrogen. As a consequence, the pronounced peak in the CIE cooling rate around $T \sim 10^4$ K which is caused by Hydrogen, will be absent. Since non-equilibrium cooling generally leads to over-ionization compared with CIE, it also tends to suppress the H cooling peak. Hence, the ionization equilibrium assumption will have a smaller impact on cooling rates calculated in the presence of an UVB than on cooling rates calculated assuming CIE [@wi09]. As check on the assumption of ionization equilibrium, we calculated the ionization timescales, $\tau_{\sf ion}$, of e.g. the Carbon ions to be compared with the cooling timescale, $\tau_{\sf cool}$. The former is given by $$\begin{aligned}
\frac{1}{\tau_{\sf ion}} &\approx& n_e {\sf Rec}_{Z,r}(T) + \sum_c n_c {\sf
Ion}^{c}_{Z,r}(T) + \Gamma_{Z,r} \nonumber \\ && \hspace*{4em} +
n_{\rm HI}{\sf CT}^{\rm HI}_{Z,r}(T)+ n_{\rm HII}{\sf CT}^{\rm
HII}_{Z,r}(T)\end{aligned}$$ while for the latter we use $$\tau_{\sf cool} = \frac{3}{2} \frac{nkT}{\left| \Lambda(T)-{\cal H}(T) \right| }
= \frac{3}{2} \frac{nkT}{ \Lambda_{\sf net}(T) }$$ with $\Lambda(T)$ the cooling rate and ${\cal H}(T)$ the heating rate. The ionization equilibrium approximation is valid if $\tau_{\sf cool}
\gg \tau_{\sf ion}$ for all ions. Judging from Fig. \[fig:timescales\], this constraint is more easily fulfilled at low densities, when the UVB irradiates the gas unimpeded and keeps most of the Hydrogen ionized. Hence, the net cooling rate $\Lambda_{\sf net}(T)$ is small. At higher densities the UVB is attenuated, Hydrogen recombines and $\Lambda_{\sf net}(T)$ is large. This makes it much harder for this constraint to be fulfilled, especially at lower temperatures. This is a caveat that should, of course, be kept in mind when using any set of cooling and heating tables calculated assuming ionization equilibrium.
Ionization balance
==================
Density dependence
------------------
In Fig. \[fig:ioneq\], we show the ionization equilibrium (i.e. the fraction of the atoms of a given element that come in the form of a given ion) of Hydrogen (H), Oxygen (O), Carbon (C), and Iron (Fe) as a function of temperature, calculated for the UVB at redshift $z=0$ and a density of $n_{\sf H}=100$ cm$^{-3}$ (full lines) and $n_{\sf
H}=10^{-4}$ cm$^{-3}$ (dashed lines). At densities above the self-shielding density threshold, Hydrogen can recombine at temperatures below about 20,000 K, thus shielding the gas from the photo-ionizing UVB. Likewise, Oxygen, with its first ionization potential very close to that of Hydrogen, recombines to its neutral form. Only elements with first ionization potentials smaller than that of Hydrogen, such as Carbon and Iron in this example, remain once ionized below $\sim 10,000$ K.
At low densities, below the self-shielding density threshold of Hydrogen, the ionizing UVB can flood the gas unimpeded, keeping over 99 % of the Hydrogen ionized. This completely erases the contribution of Hydrogen to the cooling via its free-bound and bound-bound transitions. In this example the UVB affects essentially all ions of Carbon and Oxygen, with C[iii]{}, O[iii]{}, and Fe[iv]{} the most common ionization stages of these elements at low temperatures. The high abundance of free photo-electrons provides extra cooling through radiative free-bound transitions, more than making up for the lack of efficient coolants such as the C[ii]{} and Fe[ii]{} infrared fine-structure lines at low temperatures (see below, in section \[sect:cooling\]).
![Logarithm of the cooling rate as a function of temperature for \[Fe/H\]$=0.0$, and for the $z=0$ (top panel), $z=2$ (middle panel), and $z=15$ (bottom panel) cosmic UVB, plotted for different densities (from the top curve downwards: $n_{\sf H}=100$, $10$, $1$, $10^{-1}$, $10^{-2}$, $10^{-4}$, $10^{-6}$, $10^{-9}$ cm$^{-3}$) and \[Mg/Fe\]-values (color code). \[fig:cooling\_dens\_z2\_feh0.eps\]](fig5.eps){width="49.00000%"}
Redshift dependence
-------------------
For densities below the Hydrogen self-shielding density threshold, the ionizing strength of the UVB, which varies significantly with redshift, has a profound impact on the ionization balance. In Fig. \[fig:ioneqz\], we show the ionization equilibrium of Hydrogen (H), Oxygen (O), Carbon (C), and Iron (Fe) as a function of temperature, calculated for a gas with a density of $n_{\sf
H}=10^{-4}$ cm$^{-3}$ and subjected to a $z=2$ UVB (full lines) and a $z=0$ UVB (dashed lines). Clearly, the stronger $z=2$ UVB leads to more ionization of the various elements with C[iii]{} and C[iv]{} the dominant Carbon ions at low temperatures while Oxygen is found predominatly in its O[iv]{} and O[v]{} forms. Below $10^5$ K Iron exists mostly as Fe[vii]{} and Fe[viii]{}.
Cooling {#sect:cooling}
=======
Density dependence
------------------
In the top panel of Fig. \[fig:cooling\_dens\_z2\_feh0.eps\], we compare cooling curves, denoted by $\Lambda(T)$, calculated using the $z=0$ UVB, \[Fe/H\]=0, and for different densities between $n_{\rm
H}=10^{-9}$ cm$^{-3}$ and $n_{\rm H}=100$ cm$^{-3}$, as indicated in the figure.
At the lowest densities, inverse Compton scattering dominates the cooling rate at all temperatures and the cooling curve is virtually featureless. At higher densities, in the range $10^{-9} < n_{\rm H}
\ll 10^{-2}$ cm$^{-3}$, cooling via radiative free-bound and bound-bound transitions becomes important at temperatures below $\sim
10^7$ K. At these densities, Hydrogen is essentially completely photo-ionized by the UVB and does not contribute to the cooling rate via these transitions. The many free electrons keep the rate of free-bound transitions high even at the lowest temperatures considered here. At the highest temperatures, upwards of $\sim 10^7$ K, the cooling curve’s temperature slowly changes from the $\Lambda \propto
T$ behaviour of inverse Compton scattering to the less steep $\Lambda
\propto \sqrt{T}$ behaviour that is characteristic of free-free transitions. At the highest densities, for $n_{\rm H} \gtrsim
10^{-2}$ cm$^{-3}$, Hydrogen is able to recombine and shield the other elements from the UVB. Consequently, the $10,000-20,000$ K peak in the cooling rate contributed by free-bound and bound-bound transitions of Hydrogen appears. As a result of the strong reduction of the free electron density, the cooling rate is likewise strongly reduced below $T\sim 100$ K. Below $10^4$ K, cooling via infrared finestructure emission lines, predominantly the 157.7 $\mu$m line of C[ii]{} dominates.
Redshift dependence
-------------------
{width="49.00000%"} {width="49.00000%"}
The cosmological redshift enters the picture in two distinct ways: on the one hand, it controls the strength of the UVB, and on the other, it determines the contribution to the cooling rate from inverse Compton scattering through its $(1+z)^4$-dependence. For the zero-strength $z=15$ UVB, H always fully recombines below $T \sim
10^4$ K and the cooling rate plummets by 4 orders of magnitude, as is evident in the bottom panel of Fig. \[fig:cooling\_dens\_z2\_feh0.eps\]. The only contribution to the inverse Compton scattering cooling rate now comes from the handful of ionization electrons from ions with small ionizing potentials, such as C[i]{}, Fe[i]{}, etc.
The stronger the UVB, the more highly ionized the different elements are. This lack of lowly ionized species leads to a dramatic decrease of the cooling rate due to free-bound and bound-bound emission (compare e.g. the cooling curves for densities $n_{\sf H}=10^{-6\rightarrow
-4}$ cm$^{-3}$ between $z=0$ and $z=2$).
\[Fe/H\] dependence
-------------------
In Fig. \[fig:cooling\_dens\_z0.0\_feh0.eps\], we compare the cooling rates at metallicities between \[Fe/H\]$=-\infty$ and \[Fe/H\]$=0.5$ (as indicated in the figure), calculated for Mg-abundance \[Mg/Fe\]$=0.0$, density $n_{\sf H}=0.01$ cm$^{-3}$, and for redshifts $z=0$, $z=8$, and $z=15$.
At high temperatures, inverse Compton scattering and free-free interactions dominate the cooling rate at high redshift. In this regime, the metallicity only affects the density of free electrons and, as a result, the cooling curves are not very sensitive to metallicity. At temperatures $T \lesssim 10^7$ K, partially ionized atoms can exist and free-bound and bound-bound transitions contribute greatly to the cooling rate. Likewise, for $T \lesssim 10^4$ K, the cooling rate via infrared fine-structure lines is a strong function of metallicity since in this regime the cooling depends crucially on the presence of a few key ions. This is especially true when the UVB is very weak or even absent (as at $z=15$). In that case, Hydrogen fully recombines below $T \sim 10^4$ K, consuming all free electrons, and the cooling rate drops sharply. The cooling contributed by infrared fine-structure lines of metal ions can then make a huge difference.
\[Mg/Fe\] dependence
--------------------
In Fig. \[fig:cooling\_dens\_z2\_feh0.eps\], the cooling curves are color coded according to their \[Mg/Fe\]-value. The amount of $\alpha$-enhancement clearly has a great impact on the cooling rate in those temperature ranges where key ions of the $\alpha$-elements contribute free-bound and bound-bound cooling. Around $T \sim
200,000$ K, depending on the abundance of O and Ne, the cooling rate can vary by up to an order or magnitude. Around $T \sim 10^6$ K, Si lets its presence be felt. In the range $10^2 \lesssim T \lesssim
10^4$ K, infrared fine-structure emission lines from $\alpha$-element ions such as O[i]{} and Si[ii]{} contribute to the cooling rate [@ma07] and can make an order of magnitude difference depending on whether a gas parcel has been enriched only the by SN[i]{}a and low-mass stars (the “slow” contribution to the yield) or only by SN[ii]{} and massive stars (the “fast” contribution to the yield). Since the “fast” and “slow” yields of Carbon happen to be very similar, the cooling rate is relatively unsensitive to \[Mg/Fe\] below $100$ K. For a very weak or absent UVB, ions of $\alpha$-elements with small first ionization potentials, ionized by the ISRF, contribute free electrons and slightly raise the cooling rate below $T \sim 10^4$ K (this is most noticeable in the bottom panel of Fig. \[fig:cooling\_dens\_z2\_feh0.eps\]).
Heating
=======
The heating rate is a very strong function of gas density and metallicity. Both parameters determine the density of partially ionized atoms capable of absorbing energy from the UVB through further ionization while the former, moreover, sets the amount of self-shielding. At low densities, the heating rate is essentially a monotonically declining function of temperature: the higher the temperature the lower the densities of the lowly-ionized species that absorb heat most efficiently. The heating rate also increases with increasing metallicity since this obviously raises the number of heat absorbing ions.
At high densities, the heating rate shows a much more complex behavior. While the densities of near-neutral species increase towards lower temperatures, potentially raising the heat-absorbing capabilities of the gas, the self-shielding by neutral Hydrogen suppresses the UVB. This can lead to a plateau in the heating rate for temperatures below $10^4$ K, see e.g. at $n_{\sf H}=0.1$ cm$^{-3}$ in Fig. \[fig:heating\]. This plateau is simply the heating rate of a fully neutral Hydrogen gas irradiated by a strongly reduced UVB. For higher metallicities, lowly ionized metals provide a small amount of extra heating above this plateau.
{width="\textwidth"}
At the highest densities, the Hydrogen-ionizing part of the UVB is almost completely suppressed and other sources of heating, such as the ISRF, become significant. Since Hydrogen cannot extract heat from the ISRF employed in our calculations, the heating rate drops to zero for \[Fe/H\]$=-\infty$ when Hydrogen recombines. For non-zero metallicities, the ISRF can ionize and thus heat those elements that have ionization potentials smaller than that of Hydrogen, explaining the complex behavior of the heating rate between $10^3$ and $10^4$ K.
The \[Mg/Fe\]-dependence of the heating rates is much weaker than their metallicity dependence. As an example, we show the heating rate as a function of temperature for Fe-abundance \[Fe/H\]$=0.0$, redshift $z=0$, gas density $n_{\sf H}=100$ cm$^{-3}$, and different \[Mg/Fe\] ratios in Fig. \[fig:heatingmgfe\]. While \[Mg/Fe\] varies with one dex, the heating rate changes by a factor of 5 at most.
In Fig. \[fig:shielding\], we plot the net cooling rates, defined as $\Lambda_{\sf net} = \left| \Lambda-{\cal H} \right|$, of solar metallicity gas exposed to the $z=0$ (top panel), $z=2$ (middle panel), and $z=8$ (bottom panel) UVB for different densities, with self-shielding (full lines) and without self-shielding (dotted lines). Clearly, for densities below the self-shielding threshold density of $n_{\rm HI, crit}=0.007$ cm$^{-3}$ switching shielding on or off makes no difference: the UVB can fully penetrate, ionize, and heat the gas. Above the self-shielding threshold (roughly the top four curves in each panel), the gas is dense enough to recombine and, with self-shielding switched on, to strongly suppress the UVB. As a consequence, the heating rate drops steeply below $T \sim 10^4$ K. With self-shielding switched off, Hydrogen also largely recombines but now does not suppress the UVB. This has a profound influence on the heating rate which keeps on increasing towards lower temperatures until reaching a plateau below $T \sim 10^4$ K, as discussed earlier. Without self-shielding against the UV radiation, ionization levels tend to be higher, leading to higher abundances of important cooling ions such as C[ii]{} and, consequently, to higher cooling rates in the $T < 10^4$ K temperature regime. Self-shielding also has a strong effect on the equilibrium temperature of the gas. If, without self-shielding, the UVB can flood the gas unimpeded, the equilibrium temperature can be over two orders of magnitude higher than in the self-shielded case (e.g. $T_{\sf eq}=6500$ K without self-shielding versus $T_{\sf
eq}=20$ K with self-shielding for the $z=2$ UVB at gas density $n_{\rm HI}=0.1$ cm$^{-3}$.
The most striking effect of the self-shielding prescription is the equilibrium temperature to which the gas would evolve, given sufficient time. The sudden downward break in each curve $\Lambda_{\sf net}$ indicates where $\Lambda={\cal H}$ and hence marks the equilibrium temperature. In the middle panel of Fig. \[fig:shielding\], the strong $z=2$ UVB is employed while in the top and bottom panels, the results for the much weaker $z=0$ and $z=8$ UVBs are shown. Without self-shielding, gas is incapable of radiatively cooling significantly below $T \sim 10^4$ K, except for sufficiently high densities and at large enough redshifts where the UVB is still weak. Except at late and at very early cosmic times, the UVB severely inhibits the formation of the cold, high-density clouds in which stars are thought to form while it facilitates the removal of low-density gas by ram-pressure stripping and galaxy interactions @may07 [@go10]. Together with supernova feedback, which most strongly affects high-density regions, this almost completely extinguishes star formation in simulated dwarf galaxies after $z \sim
2$ [@si12]. However, while the specific star-formation rate of Local Volume dwarf galaxies was generally larger before $z \sim 2$ compared with later times, they do show a wide variety of more or less continuous star-formation histories over their full lifetimes [@wi11]. It remains to be seen whether a proper, self-consistent treatment of the effects of photo-heating by the UVB, including the effects of self-shielding, on the ionization equilibrium and the resultant cooling and heating rates, can alleviate this problem.
![Logarithm of the heating rate as a function of temperature for Fe-abundance \[Fe/H\]$=0.0$, redshift $z=0$, gas density $n_{\sf
H}=100$ cm$^{-3}$, and different \[Mg/Fe\] ratios (color code). \[fig:heatingmgfe\]](fig8.eps){width="50.00000%"}
Moreover, the non-inclusion of self-shielding makes it virtually impossible to clearly identify different phases in the ISM of a simulated galaxy and to compare them with observed galaxies. For instance, in their analysis of a fully cosmological hydrodynamical dwarf galaxy which was simulated including an UVB but neglecting self-shielding, @pi11 find that gas particles typically have temperatures of the order of $7,000-9,000$ K. In order to investigate the simulated dwarf’s H[i]{} properties, these authors are forced to select as “cold” gas particles those with temperatures below $15,000$ K. Since the calculation of the cooling and heating curves presented here involves the determination of the ionization equilibrium as a function of the gas properties, we immediately have the neutral Hydrogen fraction at our disposal. As can be seen Fig. \[fig:neutral\], in the presence of a UVB, the neutral fraction of a gas parcel is not only temperature dependent but also strongly density and redshift dependent. The black dots in this figure indicate the neutral fraction at the equilibrium temperature for each redshift and gas density. Clearly, the Hydrogen gas has to be denser than several $0.01$ cm$^{-3}$ before it can become neutral (i.e. $n_{\rm
HI}/n_{\rm H} \gtrsim 0.95$). This broadly agrees with the estimated lower bound on the density of the Warm Neutral Medium in the Milky Way [@wo03].
Discussion {#sect:conc}
==========
In this paper, we have presented a new calculation of composition dependent cooling and heating curves intended for use in numerical simulations of galaxy formation and evolution. For each elemental mix, density, temperature, and cosmological time the ionization equilibrium was determined using a modified version of ChiantiPy, equipped with collisional and radiative ionization, by the cosmic UV background and an interstellar radiation field, and charge-transfer reactions. We believe these curves address several drawbacks of currently available tabulations of cooling rates.
We have shown that the full range of abundance variations likely to be encountered in stars and neutral and ionized gas in a galaxy can be described adequately by a simplified chemical-evolution model in which there are only two contributions to the elemental yields: a “fast” one (encompassing the contributions from SN[ii]{} and massive IMS) and a “slow” one (encompassing the contributions from SN[i]{}a and less massive IMS). The ratio of both contributions can be linked directly to the Fe and Mg abundances which provide us with two strong handles on [*all*]{} other element abundances.
Thus, the cooling and heating curves depend on only five parameters (temperature, density, redshift, \[Fe/H\], and \[Mg/Fe\]). They are easily tabulated, and can be efficiently interpolated during a simulation. We have implemented a five-dimensional interpolator in our own simulation code enabling us to employ the cooling and heating curves presented here in galaxy evolution simulations. A detailed analysis of the effects of using these new curves on such simulations, especially regarding the evolution of the star-formation rate in dwarf galaxies, will be presented in a forthcoming paper.
![Logarithm of the net cooling rate of solar metallicity gas exposed to the $z=0$ UVB (top panel), the $z=2$ UVB (middle panel), and the weak $z=8$ UVB (bottom panel) with self-shielding (full lines) and without self-shielding (dotted lines). The line colors reflect gas density (from the top curve downwards: $n_{\sf H}=100$, $10$, $1$, $10^{-1}$, $10^{-2}$, $10^{-4}$, $10^{-6}$, and $10^{-9}$ cm$^{-3}$). \[fig:shielding\]](fig9.eps){width="50.00000%"}
![Neutral fraction of solar metallicity gas exposed to the $z=0$ UVB (top panel), the $z=2$ UVB (middle panel), and the weak $z=8$ UVB (bottom panel), as a function of temperature and density (indicated in the figure, from the top curve downwards: $n_{\sf
H}=100$, $10$, $1$, $10^{-1}$, $10^{-2}$, $10^{-4}$, $10^{-6}$, and $10^{-9}$ cm$^{-3}$). The black dots indicate the neutral fraction at the equilibrium temperature for each density and redshift. \[fig:neutral\]](fig10.eps){width="50.00000%"}
We believe that the main advantage of this work is that, within the confines of a well-defined chemical evolution model and adopting the ionization equilibrium approximation, it provides accurate cooling and heating curves for a wide range of physical and chemical gas properties. These should be valid as long as the gas is neutral or, partially, ionized (molecule and dust formation have not been taken into account). Moreover, during a numerical simulation, one need only follow the evolution of the Fe and Mg abundances, leading to a reduction in memory requirements and computational cost compared to the element-by-element approach. It takes into account self-shielding in an approximate but motivated and well-tested way. This is key to resolving the formation of cold, neutral, high-density clouds suitable for star formation and to studying the structure of the multi-phase ISM in galaxy simulations. Moreover, since we have stored the ionization equilibrium for each combination of temperature, density, redshift, \[Fe/H\], and \[Mg/Fe\], we can in principle calculate any desired physical property of the gas using ChiantiPy.
Pre-compiled tables of the cooling and heating curves are available to the community as online-only supporting information. These tables, future updates, and new material can also be found on and downloaded from http://users.ugent.be/$^{_\sim}$sdrijcke.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank the anonymous referee for his/her supportive comments and suggestions. They improved the content and readability of the paper.
Aubert D. & Teyssier T., 2010, ApJ, 724, 244-266 Bekki K., 2013, arXiv1304.1633B Brook C. B., Stinson G., Gibson B. K., Roškar R., Wadsley J., Quinn T., 2012, MNRAS, 419, 771-779 Chabrier G., 2003, ApJ, 586, L133-L136 Cloet-Osselaer A., De Rijcke S., Schroyen J., Dury V., 2012arXiv1203.1863C Dere K. P., Landi E., Young P. R., Del Zanna G., Landini M., Mason H. E., 2009, A&A, 498, 915-929 Faucher-Giguère C.-A., Lidz A., Zaldarriaga M., Hernquist L., 2009, ApJ, 703, 1416-1443 Faucher-Giguère C.-A., Kereš D., Dijkstra M., Hernquist L., Zaldarriaga M, 2010, ApJ, 725, 633-657 Fitzpatrick E. L., 2010, ApJ, 725, 2401-2416 François P., Matteucci F., Cayrel R., Spite M., Spite F., Chiappini C., 2004, A&A, 421, 613-621 Gabor J. M. & Davé R., 2012arXiv1202.5315G Gavilán M., Buell J. F., Mollá M., 2005, A&A, 432, 861-877 Gnat O. & Sternberg A., 2007, ApJSS, 168, 213-230 Gneding N. Y. & Hollon N., 2012, ApJSS, 202, 13-20 Governato F. Brook C., Mayer L., Brooks A., Rhee G., Wadsley J., Jonsson P., Willman B., Stinson G., Quinn T., Madau P., 2010, Nature, 463, 203-206 Grevesse N., Asplund M., Sauval A. J., Scott P., 2010, ApSS, 328, 179-183 Kim J.-H., Krumholz M. R., Wise J. H., Turk M. J., Goldbaum N. J., Abel T., 2012, arXiv1210.3361 Kingdon J. B. & Ferland G. J., 1996, ApJSS, 106, 205-211 Krumholz M. R., 2012, ApJ, 759, 9-17 Landstreed J. D., 2011, A&A, 528, A132 Maio U., Dolag K., Ciardi B., Tornatore L., 2007, MNRAS, 379, 963-973 Mathis J. S., Mezger P. G., Panagia N., 1983, A&A, 128, 212-229 Mayer L., Kazantzidis S., Mastropietro C., Wadsley J., 2007, Nature, 445, 738-740 Nagamine K., Choi J., Yajima H., 2010, ApJ, 725, L219-L222 Nomoto K., Hashimoto M., Tsujimoto T., Thielemann F.-K., Kishimoto N., Kubo Y., Nakasato N., 1997, NuPha, 616, 79-90 Pilkington K., Gibson B. K., Calura F., Brooks A. M., Mayer L., Brook C. B., Stinson G. S., Thacker R. J., Few C. G., Cunnama D., Wadsley J., 2011, MNRAS, 417, 2891-2898 Revaz Y., Jablonka P., Sawala T., Hill V., Letarte B., Irwin M., Battaglia G., Helmi A., Shetrone M. D., Tolstoy E., Venn K. A., 2009, A&A, 501, 189-206 Sawala T., Guo Q., Scannapieco C., Jenkins A., White S., 2011, MNRAS, 413, 659-668 Scannapieco C., White S. D. M., Springel V., Tissera P. B., 2011, MNRAS, 417, 154-171 Schaye J., Dalla Vecchia C., Booth C. M., Wiersma R. P. C., Theuns T., Haas M. R., Bertone S., Duffy A. R., McCarthy I. G., van de Voort F., 2010, MNRAS, 402, 1536-1560 Schroyen J., De Rijcke S., Valcke S., Cloet-Osselaer A., Dejonghe H., 2011, MNRAS, 416, 601-617 Shen S., Wadsley J., Stinson G., 2010, MNRAS, 407, 1581-1596 Simpson C. M., Bryan G. L., Johnston K. V., Smith B. D., Mac Low M.-M., Sharma S., Tumlinson J., 2012, arXiv:1211.1071v1 Smith B., Sigurdsson S., Abel T., 2008, MNRAS, 385, 1443-1454 Smith B. D., Hallman E. J., Shull J. M., O’Shea B. W., 2011, ApJ, 731, 6-26 Stancil P. C., Havener P. S., Krstić P. S., Schultz D. R., Kimura M., Gu J.-P., Hirsch G., Buenker R. J., Zygelman B., 1998, ApJ, 502, 1006-1009 Stancil P. C., Schultz D. R., Kimura M., Gu J.-P., Hirsch G., Buenker R. J., 1999, A&ASS, 140, 225-234 Sutherland R. S. & Dopita M. A., 1993, ApJS, 88, 253-327 Tajiri Y. & Umemura M., 1998, ApJ, 502, 59-62 Tepper-García T., Richter P., Schaye J., Booth C. M., Dalla Vecchia C., Theuns T., Wiersma R. P. C., 2011, MNRAS, 413, 190-212 Tolstoy E., Hill V., Tosi M., 2009, ARA&A, 47, 371-425 Tsujimoto T., Nomoto K., Yoshii Y., Hashimoto M., Yanagida S., Thielemann F.-K., 1995, MNRAS, 277, 945-958 Vasiliev E. O., 2013, MNRAS, 431, 638-647 Verner D. A., Ferland G. J., Korista K. T., Yakovlev D. G., 1996, ApJ, 465, 487-498 Weisz D. R., Dalcanton J. J., Williams B. F., Gilbert K. M., Skillman E. D., Seth A. C., Dolphin A. E., McQuinn K. B. W., Gogarten S. M., Holtzman J., Rosema K., Cole A., Karachentsev I. D., Zaritsky D., 2011, ApJ, 739, article id. 5, 27 pp. Wiersma R. P. C., Schaye J., Smith B. D., 2009, MNRAS, 393, 99-107 Wolfire M. G., McKee C. F., Hollenbach D., Tielens A. G. G. M., 2003, ApJ, 587, 278-311 Worley C. C., Cottrell P. L., Freeman K. C., Wylie-de Boer E. C., MNRAS, 400, 1039-1048 Worthey g., Faber S. M., Gonzalez J. J., 1992, ApJ, 398, 69-73 Yajima H., Jun-Hwan C., Nagamine K., 2011, MNRAS, 412, 411-422
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: https://www.cfa.harvard.edu/$\sim$cgiguere/UVB.html
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Proton and deuteron correlation functions have been investigated with both impact parameter and emission source selections. The correlations of the system $ ^{129}Xe+^{Nat}Sn $ at 50 AMeV have been measured with the 4$ \pi $ INDRA which provides a complete kinematical description of each event. The emission time scale analyzed with a quantum model reveals the time sequence of the light particles emitted by the projectile-like fragment. The short and constant emission time of the proton, independent of the impact parameter, can be attributed to a preequilibrium process.'
title: 'Emission time scale of light particles in the system Xe+Sn at 50 AMeV. A probe for dynamical emission ?'
---
[$ ^{1} $GSI mbH, D-64291 Darmstadt, Germany.]{}
[$ ^{2} $SUBATECH, IN2P3-CNRS et Université, F-44070 Nantes Cedex, France.]{}
[$ ^{3} $GANIL, CEA et IN2P3-CNRS, B.P. 5027, F-14076 Caen Cedex, France.]{}
[$ ^{4} $IPN, IN2P3-CNRS, F-91406 Orsay Cedex, France.]{}
[$ ^{5} $LPC, IN2P3-CNRS, ISMRA et Université, F-14050 Caen Cedex, France.]{}
[$ ^{6} $DAPNIA/SPhN, CEA/Saclay, F-91191 Gif sur Yvette Cedex, France.]{}
[$ ^{7} $IPN, IN2P3-CNRS et Université, F-69622 Villeurbanne Cedex, France.]{}
[$ ^{9} $DSFSezione INFN, Università di Napoli “Federico II”, I-80126 Napoli, Italy.]{}
Introduction
=============
The microscopic description of hot nuclear matter is fundamental to the understanding of energetic heavy ion collisions. Many of the theoretical models developed so far suffer from a shortage of observables to directly test the properties of nuclear matter early after the reaction. An example is the time scale of the reaction. By the very nature, experimenters observe only the signals reaching the detectors at infinite time. Subsequently only asymptotic properties of particles leaving the reaction zone are monitored. This shortcoming is partly bypassed in studies using interferometry of light particles [@Koonin]-[@Review_Corinne]. It has been shown that the two-body correlation function is not only sensitive to emission time and source size, but also to charge and slope parameter of the source [@Martin_3Body] which affect the velocity distribution of the particles and thus the relative distance between them.
It is quite obvious that a 4$ \pi $ detector array[@Miniball][@Fopi] can be superior to a finite angle hodoscope, as used in many previous investigations, provided the granularity (angular resolution), energy threshold, energy resolution and the particle identification are of sufficient quality. A good 4$ \pi $ setup allows minimally biased event selection avoiding many possible distortions.
****elator, which fully exploits the event topology. This correlator continuously adapts to the kinematical configuration of every single collision. So each emission process can be characterized ****individually**.**
This study focuses on two-proton (p-p), two-deuteron (d-d) and proton-deuteron (p-d) correlation functions for the system $ ^{129}Xe+^{Nat}Sn $ at 50 AMeV **** occurring at the early stage of the reaction (up to 12 AMeV) is covered and the light particle multiplicity which can be measured is adequate for our selection criteria.
With a conventional reaction picture in mind, and supported by recent studies showing that a sizeable fraction of fragments are emitted in the mid-rapidity region [@Lukasik], only thermalized particles are expected in the forward hemisphere of the momentum space of the projectile-like source. Such an assumption has for example important consequences for the estimation of the excitation energy and the slope parameter of the projectile source [@Caloric; @curve], [@Dissipative; @Collision]. Conversely semiclassical calculations of heavy ion reactions in this energy domain have shown that the projectile-like and the prompt emissions from present a large overlap in their rapidity distributions [@LV_Eudes]. We have constructed our correlation functions with particles selected in this forward region to find out if the thermalized component is really the single contribution.
The extraction of the emission time was performed with a quantum model whose interesting feature is to take into account the Coulomb effect of the source charge by analytically solving the three-body problem [@FSI1], [@FSI2].
INDRA setup and light particle resolutions
===========================================
The experiment was performed at the GANIL facility where the INDRA detector has been installed with a target of 350 $ \mu g/cm^{2} $. The beam intensity was limited to 0.4 nA
INDRA [@Indra] has been designed to maximize the detection efficiency of charged particles at intermediate energy. It reaches a total detection efficiency 90%. The fine granularity chosen is such that double counts stay below 5%. INDRA consists of an array of 336 modules reparted on 17 rings centered along the beam axis. Each module is a telescope composed of an ionization chamber (ChIo) filled with $ C_{3}F_{8} $ gas followed by a Cesium Iodide (CsI) scintillator. For forward angles below 45 degrees, the resolution is further improved by insertion of a 300 $ \mu m $ silicon (Si) wafer between the ChIo and the CsI. With $ \Delta E-E $ methods in the telescope, the charge identification goes up to Z=54. Isotopic resolution (PID) is obtained for Z=1,2 (and up to Z=5 for ring 2 to 8) by pulse shape analysis of the CsI light output. Lowest energy threshold for the identification of protons and deuterons is 6 using the matrix [\[]{}CsI(fast) + Si[\]]{} versus [\[]{}CsI(slow)[\]]{}.
In particular p-p correlation functions require the resolution of very small relative momenta, less than 20 MeV/c. A minimum relative momentum of 10 MeV/c can be reached for forward angles under $ 20^{o} $. The energy resolution of light particles is between 100 keV and 200 keV depending on the module. With exception of INDRA’s first ring ($ \Delta \Theta =2^{o}-3^{o} $) which consists of plastic phoswich detectors (NE102 and NE115) for standing higher particle rates in this region, the light isotope separation could be performed on the overall domain.
Fig. 1 shows the isotope resolution summed up for rings 2 to 9 ($ 3^{o}<\theta <45^{o} $). The insert shows it separately for rings 2, 5 and 9. The projectile-like fragments at small angles come out at larger energy. To avoid saturation there, the photomultiplier gains steadily increase from ring 2 to ring 17 Subsequently ring 2 has less PID resolution, seen in the insert to Fig. 1. However the angular resolution of the first rings is superior and therefore they contribute important information to the correlation function at small relative momentum.
The p-p correlation function of Fig. 2 shows the data from the forward hemisphere of the projectile source (FHPS, see section 3) without impact parameter selection. **It can be continuously constructed from 10 MeV/c up to 250 MeV/c relative momenta due to the 4$ \pi $ coverage and the good angular resolution (yet the forward source selection slightly increases the minimal relative momentum). The structure in the correlation function at 20 MeV/c is due to the attractive s-wave p-p interaction [@Koonin]. There is a Coulomb suppression at very small relative momentum and possibly in the range between 50-75 MeV/c. The normalization has been applied to the data points between 100 and 120 MeV/c which is above any remaining two-body effects and below any kinematical effects at higher momentum. A normalization at lower relative momentum would have introduced a misinterpretation of the correlation effect.
Event sorting and source selection
==================================
We take in our analysis events in which the total longitudinal momentum of detected ejectiles is above 80% of the initial momentum. We refrain from further cuts, for instance the totally detected charge $ Z_{tot} $, to conserve a representative impact parameter distribution. For most of the events the target like ejectile is lost due to This ives an unbiased starting point for the analysis of light particle correlations.
The calculation of the impact parameter is based on the total transverse energy of the light charged particles (Z$ \leq $2) is quite independent of the reaction mechanism.
In order to observe impact parameter dependent properties while maintaining sufficient statistics, we have defined three $ E_{T} $ bins for which the correlation function is constructed. The “peripheral”, “intermediate‘" and “central” events have a $ E_{T} $ range between 0-280 MeV, 280-420 MeV and above 420 MeV corresponding to reduced impact parameters in [\[]{}1-0.65[\]]{}, [\[]{}0.65-0.35[\]]{} and smaller than 0.35 respectively.
INDRA allows to build event by event a momentum tensor [@Ellipsoide] by :
$$Q_{ij}=\sum ^{M}_{k=1}\frac{1}{p}p_{i}(k)p_{j}(k)$$ where M is the multiplicity of fragments with a charge Z greater than 2, p is the momentum of the k’th particle in M and $ p_{i} $, $ p_{j} $ two of the Cartesian momentum components. The of this tensor establish a reference frame. The main axis (eigenvector of the largest eigenvalue) gives the average direction of nuclear matter emission. The eigenvectors associated to the two largest eigenvalues define a reaction plane. The FHPS selections and the calculations have been performed with regards to this new reference system. Fig. 3 shows the transverse versus the parallel velocity of the protons when $ E_{T} $ is smaller than 40 MeV. Since this selection implies only very peripheral events, a clearer separation of the sources is exhibited.
The next task consists of recognizing the fragments emitted either from the target-like or from the projectile-like source. The thrust variable defined by $$T=\max _{c_{1},c_{2}}\frac{\mid \sum _{i\in c_{1}}\overrightarrow{P_{i}}\mid +\mid \sum _{j\in c_{2}}\overrightarrow{P_{j}}\mid }{\sum ^{M}_{k=1}|\overrightarrow{P_{k}}\mid }$$
divides up the fragments in two groups $ c_{1} $ and $ c_{2} $ corresponding to the two emitting sources. The velocity of each of them is determined by a kinematic reconstruction within these two ensembles [@Metivier_Thrust]. As an example, the average velocity of the projectile-like source is drawn with a vertical bold line in Fig. 3.
To disentangle the projectile-like source emission from mid-rapidity contributions, we have taken particles with a parallel velocity larger than the projectile source velocity. This region \[FHPS\](FHPS) is on the right of the bold line in Fig. 3.
Source parameters : disentangle size and time
=============================================
*Size*
-------
The shape and the height of the correlation function is given by the strength of the interactions which themselves depend on the average distance between the two particles detected in coincidence. This distance depends on the average emission time and on the spatial source extension. The double parametrization can be circumvented only for two extreme cases: at high energy the emission time is set to zero while at low energy the emission time is very long and consequently the source size is negligible [@Fixed_source_size]. In our energy domain, both parameters are relevant. Using the complete detection by INDRA of all charged products we alternatively can determine the source size directly.
*Slope parameter*
-----------------
** ****
Description of the quantum model
================================
The extraction of the emission time has been performed by using the three-body quantum model developed by R. Lednicky [@FSI1], [@FSI2]. This code calculates the quantum statistics for identical particles and the final state interaction by taking into account the nuclear and the Coulomb potentials. The Coulomb repulsion on the particle pair due to the emitting projectile-like source is also included [@Martin_3Body]. The three-body problem is analytically solved by making an adiabatic assumption : the relative motion between the two particles has to be much slower than their velocity in the source reference frame. Thus the adiabatic assumption
The introduction of the emitter Coulomb effect in the quantum calculation is a new feature brought by this model. We feel that this type of description is required to correctly reproduce the experimental data because the presence of the remnant source charge is intrinsic to the model. In so far it may surpass models where the Coulomb influence of the emitter is only treated as a correction [@Gamov_factor].
The time distribution for particle emission follows an exponential decay law with average emission times to be determined from comparison with the data. The energy distribution is taken from the experimental slope parameter. The distributions have been filtered according to the geometrical granularity, including also double counts. In addition, energy and particle identification thresholds as well as the source selection criteria (FHPS) have been folded in.
Analysis of the light particle correlation functions
====================================================
The experimental correlation function is defined by the yield of true coincidences as a function of the two particle relative momentum divided by the so-called false coincidences obtained by the technique of event mixing [@Mixing]. It consists of taking two particles from two different events which assures a full decorrelation and has the advantage to use the same sample of events for true and false coincidences. Possible distortions coming from our FHPS selection are then largely eliminated. In general, event mixing introduces an additional term in the relative momentum due to different source velocities. This influence which otherwise would disturb the correlation function is reduced by the impact parameter selection.
The Proton-Proton correlation function
--------------------------------------
The p-p correlation functions for the three intervals in impact parameter are represented in Fig. ****6. ****The statistics of 2.3 millions of reconstructed events give a reasonable correlation function above 10 MeV/c relative momentum. In all three event classes, the resonance at 20 MeV/c is clearly visible and well described by a time value of $ \tau $=80 fm/c A time variation of 25% changes the height of the resonance by a factor of two. This demonstrates the high sensitivity of the correlation function on the emission time parameter.
****
One must ask here, how the emission time alters for a given error in source size. To test this we have reduced the charge from Z=46 to Z=36, assuming all the Z=1 particles do not belong to the projectile-like emission. The result for the emission time is then 100 fm/c instead of 80 fm/c which still stays short enough to be compatible with our conclusion saying, these light particles characterize a prompt process of pre-equilibrium emission which covers a large domain of rapidity.
The Deuteron-Deuteron correlation function
------------------------------------------
The d-d correlation function has been constructed on the same base of events as p-p. the results are shown on Fig. ****7. We immediately observe the anticorrelation effect in the d-d correlation function for small relative momenta. Despite the fact that data do not go below 30 MeV/c the fit of the quantum model gives the following results : for peripheral reactions the emission time is as least 200 fm/c, for semi-central reactions it is 100 fm/c and for central it is 25 fm/c. This behavior can be interpreted as an increasing contribution of out-of-equilibrium emission.
Yet the creation of barely bound particles is not the prefered mechanism of hot nuclear matter to dissipate energy. out-of-equilibrium emission. In the picture of the coalescence model [@Coalescnence_Deuteron], the deuteron formation is directly connected to the proton creation. Consequently it is not surprising to find also two components in the deuteron energy spectra.
In conclusion the process of deuteron production remains an open question. Still we tend to favor the preequilibrium emission over the other explanations. Unfortunatelyfficient to disentangle more.
The Proton-Deuteron correlation function and the emission chronology
--------------------------------------------------------------------
The correlation function of non-identical particles can give model independent information about their mean order of emission simply making velocity selections [@Order_lednicky], [@Order_gelderlos], [@Order_Soff]. We have applied this method to the p-d correlation function for particles emitted in the FHPS region. The principle is to compare two functions. The first, ( $ 1+R^{+} $), is constructed with pairs where the proton is faster than the deuteron in the projectile-like source reference frame. The second function, ( $ 1+R^{-} $), corresponds to the reverse situation . When the first emitted particle is slower than the second, the average distance will be reduced and the Coulomb suppression effect enhanced, and vice versa. The comparison of the two functions gives the mean order of emission as it is shown in Fig. 8 for the peripheral collisions. The Coulomb suppression is more pronounced in $ 1+R^{+} $, which the ratio clearly demonstrates. This means that the deuteron is on average emitted earlier than the proton, namely $ \tau _{deuteron}<\tau _{proton} $. The same time sequence is observed for the two other impact parameter selections. this result is only validated between 0-120 MeV/c relative momentum where the anticorrelation effect leaves a measurable signal.
The chronology of emission in p-d spectra of peripheral collision as surprising since we just learned from previous paragraphs that the mean emission time is 80 fm/c for p-p and 200 fm/c for d-d. This apparent contradiction can be resolved by postulating that the protons which contribute to p-p are not identical with those contributing to p-d. Indeed the proton energy spectra show a fast and a slow component (Fig. 5) from the
o satisfying agreement can be found between the quantum model and the p-d data for peripheral and int
Although both components of protons simultaneously play a role in the correlation function, we can deduce a chronology pattern of the light particle emission. The fast protons from the interacting zone come first, then the deuterons and still later the protons thermally emitted by the projectile-like source. It is possible to better separate the two proton components by making a cut on the parallel velocity as Fig. 11 shows. The clear enhancement of the p-p resonance at 20 MeV/c and the disappearance of the Coulomb suppression at 50 MeV/c reveal the enlarged part of fast protons not feeling **** ****
Conclusion
==========
We have taken the large data set for collisions of Xe on Sn at 50 AMeV which INDRA has accumulated at GANIL to examine correlations of protons and deuterons. Such studies have many attractive aspects in view of the complete detection of all collision residues by a 4$ \pi $ detector. Foremost to name is the unique possibility to well determine the emitting source in particular for symmetric systems at intermediate energy. In addition each event can be individually characterized in its own frame of reference, due to full charge, angular and energy coverage of INDRA.
The time scale as well as the chronology of emission of light projectile-like particles could be determined from two particles correlation functions interpreted by a full three body quantum code. Total charge measurements as a function of impact parameter indicate possible out-of-equilibrium emission of protons from the forward hemisphere of the projectile source. The observation that the slope parameter of the energy spectra exhibits two components points to a similar conclusion. Strong confirmation of these findings stems from the very short emission time extracted from p-p correlation functions. While this process is expected to dominate in central collisions our study unveils that it also contributes to the forward zone of binary peripheral collisions. We explained in the text how both proton components are not equally shared in p-p and p-d correlation functions. So the whole emission chronology pattern remains self-consistent. It is also in good agreement with the measurement of slope parameters. The short time scale in p-p reveals the presence of fast hot protons from an out-of-equilibrium process. Protons emitted later than the deuterons correspond to the really equilibrated production from the projectile-like source. The light particle emission chronology including the deuteron formation via the $ NNN\rightarrow dN $ process has been calculated in the Boltzmann-Uehling-Uhlenbeck (BUU) approach [@Order_BUU]. The theoretical results although for lighter systems are in excellent agreement with the present experimental study.
We would have liked to look into the hydrogenic correlation function in more detail by finer selecting impact parameter intervals. Furthermore, the inclusion of tritons could have given valuable additional information. This task cannot yet be performed on the same footing as with protons and deuterons due to a serious shortage of data statistics.
We therefore suggest to perform a high statistics experiment especially dedicated to light particle correlations. INDRA parameters, optimized toward this goal, could contribute important and still better information on the dynamics of light particle emission.
Acknowledgments
===============
The authors whish to thank H.Orth for his careful reading of the manuscript, R.Lednicky for discussions and providing his code. D.G acknowledges the support of the ALADIN group at GSI.
Table **1** : Parameters of the projectile-like source. The radius has been calculated from the total reconstructed charge of the source, by assuming a A/Z ratio in the valley of nuclear stability and a normal nuclear density. The slope parameter has been extracted from the energy spectra in the source reference frame.
[\*]{}) Only one slope could be extracted.
Fig. 1 : Light isotopes resolution (Z=1) of the 8 first rings and in the insert for the rings 2, 5 and 9 separately.
Fig. 2 : Two-proton correlation function for the FHPS selection.
Fig. 3 : Invariant velocity plot of the protons in the center of mass for a total light particles transverse energy smaller than 40 MeV which corresponds to a normalized impact parameter larger than 0.9. The average value of the reconstructed forward source is represented by a vertical line at $ V_{//}=0.155c $. This line position is in accordance to the middle of the Coulomb circle. The FHPS is defined by the particles in each event which are faster than the reconstructed forward source velocity.
Fig. 5 : The energy spectra of the protons (top) and deuterons (bottom) for the ring 2, 4 and 6 in the case of the peripheral collisions. The shapes clearly exhibit two components, better separated in the case of the protons. For comparison reasons the relative scale is the same for all panels.
Fig. 6 : The experimental p-p correlation functions (black circles) of the projectile-like source for the three impact parameter intervals. Each case has been calculated with different emission times using the source parameters of Table 1. A time of 80 fm/c was found to be the best for all impact parameter bins. The resonance of the calculated function decreases with the centrality because the slope parameter of the emitter increases.
Fig. 7 : The experimental d-d correlation functions of the projectile-like source for the three impact parameter bins (black circles). Each case has been calculated with different emission times using the source parameters of Table 1. There is only a weak dependence on the long emission time parameter for the peripheral reactions.
Fig. ****8 ****: The experimental proton-deuteron correlation functions of the projectile-like source for the peripheral collisions grouped into two velocity bins (left and middle panel). $ 1+R^{+} $ contains all the pairs of particles where the proton was faster than the deuteron. $ 1+R^{-} $ is the reverse situation. The ratio of both functions (right panel) which is smaller than unity indicates that the deuteron is on the average emitted earlier than the proton.
Fig.9 : The kinetic energy of the protons in the projectile-like source frame which contribute to p-p (open circle) or to p-d (black circle) with a relative momentum selection smaller than 50 MeV/c for peripheral collisions. For p-p the average energy and the slope parameter are higher than for p-d.
Fig.10 : The experimental proton-deuteron correlation functions of the projectile-like source for the three impact parameter selections (black circles). The calculation (lines) fails to reproduce the data for the peripheral and the intermediate collisions which may reveal the double contribution of fast and slow protons in the interference pattern.
Fig. 11 : The experimental proton-proton correlation functions for all impact parameters with increasing selection on their longitudinal velocity in the projectile-like reference frame. The higher resonance indicates a faster emission time.
S. Koonin, Phys. Lett. 70B (1977) 43. D. Ardouin, Int. Journ. Mod. Phys. E, Vol. 6, N.2 (1997) 3. G.J. Kunde et al., Phys. Rev. Lett. 70 (1993) 2545. M.A. Lisa et al., Phys. Rev. Lett. 71 (1993) 2863. J. Quebert, Ann. de Physique 17 (1992) 99. W. Bauer et al., Annu. Rev. Nucl. Part. Sci. 42 (1992) 77. W.G. Gong et al., Phys. Rev. C43 (1991) 781. D. Boal et al., Rev. Mod. Phys. 62 (1990) 553. J. Pochodzalla et al., Phys. Rev. C35 (1987) 1695. CORINNE I, Int. Workshop. on Particle Correlations and Interferometry. in Nucl Collisions, Nantes, France (1990). Ed. D. Ardouin (World Scientific) and references therein ; CORINNE II, Int. Work.shop on Multiparticle Correlation and Nucl. React. Collisions, Nantes, France (1994). Ed. J. Aichelin, D. Ardouin (World Scientific) and references therein. L. Martin et al., Nucl. Phys. A604 (1996) 69. R.T. De Souza et al., Nucl. Instr. Meth. A295 (1990) 109. A. Gobbi et al., Nucl. Instr. Meth. A324 (1993) 156. N. Marie et al., Phys. Rev. C58 (1998) 256-269 N. Marie et al., Phys. Lett. B391 (1997) 15. J. Lukasik et al., Phys. Rev. C55 (1997) 1906. D. Gourio, Thesis of University, Nantes France (1996) unpublished. Y.G. Ma et al., Phys. Lett. B390 (1997) 41. J. Peter et al., Nucl. Phys. A593 (1995) 95. P. Eudes et al., Phys. Rev. C56 (1997) 2003. R. Lednicky and V.L. Lyuboshitz, Sov. Jour. Nucl. Phys. 35 (1982) 770. R. Lednicky et al., submitted to Nucl. Phys. A. J. Pouthas et al., Nucl. Instr. and Meth. A357 (1995) 95. J.D Bjorken and S. J. Brodsky, Phys Rev D1 (1970) 1416 ; S. Brandt et al., Phys. Lett. 35 (1977) 1609. V. Metivier et al., XXXIII Int. Wint. Meet. on Nucl. Phys., Bormio, Italy (1995).; V. Metivier et al. subm. to Nucl. Phys. A. J. Alexander et al., Phys. Rev. C48 (1993) 2874. R. Bougault et al., XXXV Int. Wint. Meet. on Nucl. Phys., Bormio, Italy (1997). H. Fuchs and K. Mohing, Rep. Prog. Phys. 57 (1994) 231 and references therein. R. Bougault et al., XXVII Int. Work. on Gross Properties of Nuclei and Nuclear Excitations, Hirschegg, Austria (1999). B. Tamain et al., Heavy-Ion Dynamics and Hot Nuclei. Ed. G. Nebbia and M.N. Namboodiri (World Scientific). Y.M. Sinyukov et al., Physics Letters B432 (1998) 248. D. Brown and P. Danielewicz, Phys. Lett. B 398 (1997) 252 ; ****D. Brown and P. Danielewicz, Phys. Rev. C57 (1998) 2474 D. Drijard et al. Nucl. Instr. and Meth. 225 (1984) 367. R. Nebauer and J. Aichelin, Nucl. Phys. A650 (1999) 65 ; R. Nebauer et al., to be published in Nucl. Phys. A. M.A. Lisa et al., Phys. Rev. Lett. 70 (1993) 3709. M. Baldo et al., Phys. Rev. C52 (1995) 975. S.T. Bulter et al., Phys. Rev. Lett. 7 (1961) 69; A. Schwarzschild et al., Phys. Rev. 129 (1963) 836. R. Lednicky et al., Phys. Lett. B 373 (1996) 30. C.J. Gelderlos et al., Phys. Rev. C52 (1995) 2834. S. Soff et al., Phys. Lett. B449 (1999) 191.; P. Sapienza, Nucl. Phys. A630 (1998) 215 W.G. Gong et al., Phys. Rev. C47 (1993) 429.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present detections of absorption from the ground state and excited states of C I in the z = 1.9731 damped Ly[$\alpha$]{} system of the QSO 0013$-$004. The excitation temperature between the J = 0 and J = 1 fine-structure levels of C I is 11.6 $\pm$ 1.0 K. We estimate other contributions to the excitation of the C I fine-structure levels, and use the population ratio of the excited state to the ground state to derive an estimate for the cosmic microwave background radiation (CMBR) temperature of T = 7.9 $\pm$ 1.0 K at 0.61 mm and z = 1.9731, which is consistent with the predicted value of T = 8.105 $\pm$ 0.030 K from the standard cosmology.'
author:
- 'Jian Ge, Jill Bechtold, John H. Black'
title: 'A New Measurement of the Cosmic Microwave Background Radiation Temperature at z = 1.97[^1]'
---
Introduction
============
The standard Friedman cosmology predicts a simple relationship between the temperature of the cosmic microwave background radiation (CMBR) and redshift, z: $$T_{CMBR}(z) = T_{CMBR}(0)(1+z),$$ where T$_{CMBR}$(0) is the CMBR temperature today (e.g. Peebles 1993). The present-day CMBR temperature has been measured precisely with the FIRAS instrument on the Cosmic Background Explorer (COBE), with T$_{CMBR}$(0) = 2.726$\pm 0.010$ K (at the 95% confidence level, Mather et al. 1994).
The CMBR temperature at higher redshifts can be measured indirectly by using atomic fine-structure transitions in absorbers toward high redshift quasars (Bahcall & Wolf 1968). The first attempt to measure the CMBR temperature in this way gave an upper limit for the CMBR temperature, T$_{CMBR} <$ 45 K, at z = 2.309 from limits on the fine-structure excitation of C II toward PHL 957 (Bahcall et al. 1973). Compared with other abundant species (such as O I, C II, Si II, N II), C I is a better species to use because it has the smallest energy separations in its fine-structure levels. The ground term of C I is split into three levels (J = 0, 1, 2) with J = 0 - 1 and J = 1 - 2 separations of 23.6 K and 38.9 K (or 0.61 mm and 0.37 mm). Meyer et al. (1986) used the C I fine structure lines of a damped Ly$\alpha$ system in the spectrum of the QSO 1331+170 to obtain an upper limit (2$\sigma$) of T$_{CMBR} < $ 16 K at z = 1.776. More recently, Songaila et al. (1994b) have observed QSO 1331+170 again and obtained T$_{CMBR}$ = 7.4$\pm 0.8$ K, which agrees with the predicted value of 7.58 K. C II is another good species to use for the CMBR measurements at high redshift because it has reasonably small energy separation between its fine-structure levels, 91.3 K. Songaila et al. (1994) obtained a 2 $\sigma$ upper limit of T$_{CMBR} < 13.5$ K at z = 2.909 toward QSO 0636+680 based on upper limits to C II fine-structure. Lu et al. (1995) achieved a 3 $\sigma$ upper limit of T$_{CMBR} < 19.6$ K at z = 4.3829 toward QSO 1202$-$07 by measuring upper limits for the excited states of C II.
There are several difficulties in carrying out measurements of T$_{CMBR}$(z) with quasar absorbers. First, the ground state C I absorption lines are often weak and difficult to detect in quasar absorbers at high redshift. Second, other non-cosmological sources such as collisions and pumping by UV radiation can also populate the excited fine-structure levels of C I. Thus, the excitation temperature derived is an upper limit to the CMBR temperature, unless the local excitation can be estimated. Third, most absorption lines from abundant species such as O I, C II, Si II, N II show strong saturation in their ground state transitions and hence the population ratio of their excited state to the ground state cannot be accurately determined.
In this paper we present spectra obtained at the Multiple Mirror Telescope (MMT) of C I and C I$^*$ absorption in the z = 1.9731 damped Ly$\alpha$ system toward the QSO 0013$-$004 and estimate the contributions of the various sources of excitation. The neutral hydrogen column density of the z = 1.9731 damped system is N(H I) = $5(\pm 1)\times 10^{20}$ cm$^{-2}$ (Pettini et al. 1994). The metal abundance is about 1/4 of the solar value and the heavy element depletion by dust is more than 20% of the Milky Way value (Pettini et al. 1994). These properties suggested to us that this system was a good candidate for a search for C I absorption.
Observations
============
The observations of QSO 0013$-$004 were obtained on October 9 and December 8, 1994 with the Blue Channel Spectrograph and the Loral 3072$\times$1024 CCD on the MMT. The 832 l/mm grating was used in second order. A CuSO$_4$ filter was used to block the first order light. In October, we took three 50-minute and one 60-minute exposures with wavelength coverage from 3860 Å to 4960 Å. Because of poor seeing conditions, a 1.5$''\times 180''$ slit was used to get a spectral resolution of 1.3 Å (FWHM). In December, we took four 50-minute exposures with wavelength coverage from 4380 Å to 5459 Å. A 1$''\times 180''$ slit was used to obtain a spectral resolution of 1 Å (FWHM). In all our observations, the quasar was moved a few arcseconds along the slit between each exposure to smooth out any residual irregularities in the detector response which remained after flat-fielding. An exposure of a He-Ne-Ar lamp and a quartz lamp were taken before and after each exposure of the object to provide an accurate wavelength reference, a measure of the instrumental resolution, and a flat-field correction. The spectra were reduced using standard routines in IRAF, and were summed with individual exposures weighted by the signal-to-noise ratio (S/N). We then summed the spectra with the wavelength coverage from 4550 to 5940 Å from our two runs to reach S/N of about 40.
Figure 1 shows the total spectrum obtained. All reported wavelengths are vacuum and have been corrected to the heliocentric frame. The continuum was fitted and significant absorption features were identified and measured in the way described by Bechtold (1994). The spectra shown were normalized by their fitted continuum. All absorption lines with more than 5 $\sigma$ significance are marked. Table 1 shows the equivalent widths of the absorption lines and their identifications. The equivalent widths were measured by specifying start and stop wavelengths for each absorption feature by hand (cf. Bechtold 1994). The central wavelength of each line is the centroid, weighted by the depth of each pixel in the line profile below the continuum. The error for the central wavelength shown in this table is from the uncertainty in the measurement of the equivalent width. There are at least four velocity components associated with the z = 1.9731 damped system. The redshifts are z = 1.9673, 1.9700, 1.9714, 1.9731. Two components (z = 1.9673, 1.9731) clearly show absorption lines from the C I ground state levels. Since some important lines such as C II [$\lambda$]{}1334 and C I [$\lambda$]{} 1560 lines are blended lines, we have also tried to fit the absorption lines with Gaussians. This method is similar to the method described by Schneider et al. (1993). The equivalent widths measured in this way are consistent with the ones listed in Table 1 within the 1 $\sigma$ errors even for heavily blended absorption lines. Figure 2 shows the Gaussian fits for the C II [$\lambda$]{} 1334 line and C I [$\lambda$]{} 1560 line at z = 1.97.
Table 2 lists the rest wavelengths, predicted wavelengths, and f-values for the two strongest C I multiplets and the strongest C II multiplet in the z = 1.9731 component. The f-values are from the compilation of Morton (1991). We also list the observed equivalent widths of these lines. Figure 3 shows our spectrum of QSO 0013$-$004 in the vicinity of the two C I multiplets and one C II multiplet listed in Table 2. The fit of the continuum and the 1 $\sigma$ deviation of each pixel are also displayed. C I J = 0 absorption lines are clearly present in UV multiplet 2 at 1656.93 Å and multiplet 3 at 1560.31 Å. C II J = 1/2 (1334.53 Å), J = 3/2 (1335.70 Å) absorption lines of multiplet 1 are also present. C I J = 1 absorption is present in the multiplet 2 at 1656.27 Å and 1657.91 Å and multiplet 3 at 1560.68 Å. The C I and C II lines are observed at the wavelengths expected from the redshift of other low-ionization ions, such as Zn II (Pettini et al. 1994), Fe II, and Si II (Table 1), within the wavelength uncertainty of about 0.1 Å. The C I J = 0 absorption line at $\lambda =1656.928$ Å is blended with one of the C I J = 1 lines at $\lambda = 1657.379$ Å and also one of the C I J = 2 lines at $\lambda = 1657.008 $ Å. No absorption features for J = 2 at $\lambda = 1658.121$ Å and $\lambda\lambda\lambda = 1561.340, 1561.367,
1561.438$ Å are detected. The third strongest C I multiplet at [$\lambda$]{} = 1329 Å is blended with Si II [$\lambda$]{} 1304 Å from another absorber at z = 2.0290.
Results
=======
We can use the relative population ratios of the J = 1 and J = 0 levels in the multiplets 2 and 3 to obtain the excitation temperature of the C I fine-structure levels in its ground state and to derive limits on the CMBR temperature at z = 1.9731. Since our spectral resolution is insufficient to resolve the profiles of the lines, we used observed Si II lines, at [$\lambda$]{} 1260, [$\lambda$]{} 1304, [$\lambda$]{} 1526 and [$\lambda$]{} 1808, to construct an empirical curve-of-growth (Figure 4). The measurement of Si II [$\lambda$]{} 1206 is from another observation by Bechtold, who found a rest frame equivalent width for this line of 0.8691 $\pm$ 0.0188 Å. The Si II curve-of-growth provides a Doppler parameter b = 42 $\pm$ 2 [km s$^{-1}$]{} which we then used to infer the column densities of different absorption lines. The results of calculated column densities are shown in Table 3. We have also shown central optical depths for different C I and C II lines. The central optical depths for C I and C I$^*$ absorption lines indicate that all these lines are on the linear part of the curve-of-growth. Thus, the derived column densities for the C I fine structure levels are independent of the derived b value. However, the optical depths for the C II and C II$^*$ lines indicate that they are saturated, and so the derived column densities for the C II and C II$^*$ line depend on the b-value. The derived b-value indicates that there are probably several velocity components blended with each other. However, the uncertainties in the column densities from single b-value curve-of-growth analyses are usually on the order of a factor of 2 (Jenkins 1986). We therefore use this b-value to derive the column densities for the saturated C II and C II$^*$ lines. In our calculation, because of our limited resolution, we have combined the f-values of the two J = 1 lines of C I multiplet 3, [$\lambda$]{} [$\lambda$]{} 1560.682 Å, 1560.709 Å, and also the f-values of the two J = 3/2 lines of C II multiplet 1, [$\lambda$]{} [$\lambda$]{} 1335.663 Å, 1335.708 Å, and further derived the relative population ratios of their fine structure levels. We have assumed that the absorption at $\lambda = 4926.313$ Å is only from the J = 0, $\lambda = 1656.928$ Å of C I multiplet 2 since the strengths of other blended lines such as J = 1, $\lambda = 1657.379$ Å and J = 2, $\lambda = 1657.008$ Å are much weaker than that of the J = 0 line. The other two J = 1 lines of C I multiplet 2, [$\lambda$]{} [$\lambda$]{} 1656.267 Å, 1657.907 Åare detected at about the 2-3 $\sigma$ level, and in the weighted mean are present at the 4 $\sigma$ level. We therefore have used this weighted mean to derive the population ratio of the J = 1 and J = 0 levels of C I multiplet 2, as shown in Table 3.
Next, we can use the relative population ratios to derive the excitation temperature of the C I and C II fine-structure levels. According to the Boltzmann equation, an excitation temperature T$_{ex}$ can be expressed in terms of the column densities N$_e$ and N$_g$ in the excited and the ground state levels, $$N_e/N_g=g_e/g_g~exp(-\Delta~E_{eg}/kT_{ex}),$$ where $\Delta$E$_{eg}$ is the energy difference between the excited and ground levels. $\Delta$E$_{eg}$ is 23.6 K for the difference between J = 1 and J = 0 in C I and 91.2 K for the difference between J = 3/2 and J= 1/2 in C II. The weights are $g_J = 2J + 1$. Thus, the population ratios N(J = 1)/N(J =0) in the C I multiplets 2 and 3 indicate excitation temperatures, T$_{ex}$ = 11.6 $\pm$ 1.6 K and 11.6 $\pm$ 1.4 K for multiplets 2 and 3, respectively. The weighted mean value is T$_{ex}$ = 11.6 $\pm$ 1.0 K for the C I fine structure. The population ratio N(J = 3/2)/N(J = 1/2) of the C II fine-structure levels indicates an excitation temperature, T$_{ex} = 16.1\pm1.4$ K. Because the C I and C II fine-structure levels can be excited by not only the CMBR field but also other excitation sources such as collision and UV pumping, the derived excitation temperatures are upper limits to the CMBR temperature at z = 1.9731. Thus, the upper limits of the CMBR temperature at 0.61 mm and 0.16 mm are 11.6 K and 16.1 K, respectively, consistent with the predicted value at this redshift, T$_{CMBR}$ = 8.105 K.
We can estimate the contribution from collisional and UV pumping to the excitation of C I by modeling the absorption region. The equilibrium between the excitation and de-excitation of the C I J = 0$\rightarrow$1 fine structure can be expressed as $$N_0[\sum_{j}<\sigma_{01} v>_{j} n_j + B_{01}~I_\nu + \Gamma_{01}] = N_1[A_{10} +B_{10}~I_\nu +\sum_{j}<\sigma_{10} v>_{j} n_j + \Gamma_{10}],$$ where j = H, e, p, He and H$_2$, $\Gamma_{10}$ is the UV pumping rate from J = 0 to J = 1, the $\Gamma_{01}$ is the UV pumping rate from J = 1 to J = 0. The spontaneous transition probability for the C I J = 1$\rightarrow$0 transition $A_{10}=7.93\times 10^{-8}$ s$^{-1}$ (Bahcall & Wolf, 1968). The collisional excitation rates due to different collision partners are given by Launay et al. (1977); Keenan et al. (1986); Johnson et al. (1987); Roueff et al. (1990); Flower (1990); Staemmler et al. (1991) and Schröder et al. (1991). The collisional de-excitation rate $$<\sigma_{10} v>_{j}=1/3<\sigma_{01} v>_{j}exp(23.6~K/T_k),$$ for j = H, e, p, He and H$_2$. $B_{10}=1/3~B_{01}$. The J = 0$\rightarrow$1 excitation rate due to the absorption of the CMBR, $B_{01}$, can be expressed as $$I_\nu B_{01} = 2.38\times 10^{-7}/[exp(23.6~K/T_{CMBR})-1]~s^{-1}.$$ The UV pumping rate depends on the strength of UV radiation field. $\Gamma_{01} = 7.55\times 10^{-10}$ s$^{-1}$ and $\Gamma_{10} = 2.52\times 10^{-10}$ s$^{-1}$ if the UV field intensity in the z = 1.9731 is the same as that in the Milky Way which is about $4.7\times 10^{-19}$ ergs s$^{-1}$ cm$^{-2}$ Hz$^{-1}$ at 912 Å(Jenkins & Shaya 1979; Mathis et al. 1983).
In order to solve Eq. (3), we have to know n$_H$, n$_e$, n$_{He}$, n$_{H_2}$ and the UV pumping rates. To estimate plausible values for the z = 1.9731 absorber we constructed a photoionization model with the CLOUDY program (Ferland 1993). For the input to CLOUDY, we adopted a metallicity of 25% of the solar value, i.e., \[Zn/H\] = $-$0.61, for all the elements in the z = 1.9731 damped system (Pettini et al. 1994). We have also considered depletion by dust grains. The dust-to-gas ratio is about 20% of the Milky Way, estimated from the relative depletion of Cr and Ni to Zn, i.e. \[Cr/Zn\] $\le$ $-$1.15 (Pettini et al. 1994) and \[Ni/Zn\] $\le$ $-$0.98 from our data. For the shape of the spectral energy distribution (SED), we adopted a parameterization of the Milky Way SED given by Black (1987). Because we are interested in the low ionization species (C I, C II and H$_2$), the results are sensitive to the UV flux adopted at wavelengths from $\sim 500-1100$ Å which is probably dominated by local sources within the galaxy. The adopted flux at the Lyman limit is about one order of magnitude higher than the metagalactic UV flux at z $\approx$ 2, estimated to be J(912 Å) $\approx 3.8 \times 10^{20}$ ergs s$^{-1}$ cm$^{-2}$ Hz$^{-1}$ (e.g. Bechtold 1994), so we have neglected the ionization contribution from the metagalactic radiation field. The results are shown in Figure 5. The ionization parameter, U = $\phi(H)/n_H c = 2.7\times 10^{-4}$, gives the best fit to the observational results, where $\phi(H)$ is the surface flux of hydrogen-ionizing photons (cm$^{-2}$ s$^{-1}$). The photoionization model with this U-value indicates that the column density of molecular hydrogen, N(H$_2) = 5\times 10^{19}$ cm$^{-2}$, or n$_{H_2} = 0.1 n_H$; the electron temperature, T$_e \sim 1\times 10^3$ - 80 K and n$_e$/n$_H$ $\sim 1.0\times 10^{-2}$ - $5\times 10^{-4}$ from the outer region to the inner region of the absorber. T$_e$ of $\sim$ 100 K and n$_e$/n$_H$ of $\sim$ $5.0\times 10^{-4}$ dominate most regions of the cloud. In the following discussions we adopted two sets of extreme limit values: n$_e$/n$_H = 1.0\times 10^{-2}$ and T$_e$ = 1000 K; n$_e$/n$_H$ = $5.0\times 10^{-4}$ and T$_e$ = 100 K.
In order to estimate n$_H$ we use derived from the relative population ratio of the C II fine-structure levels. In the H I dominant region with n$_H \la 3\times 10^3$ cm$^{-3}$ (Flower 1990; Bahcall & Wolf 1968), the ratio of excited C II$^*$ relative to the ground state C II populations can be expressed as $$\frac{N(C II^*)}{N(C II)} \approx \frac{n_H <\sigma_{01} v>_{H} +
n_e<\sigma_{01} v>_{e} + n_{H_2}<\sigma_{01} v>_{H_2}}{A_{10}},$$ where $A_{10} = 2.29\times 10^{-6}$ s$^{-1}$ is the spontaneous transition probability, and where we have neglected the excitation term due to proton collisions because this term is much less than the others at T$_e <
2\times 10^5$ K (Bahcall & Wolf 1968). As a result, the neutral hydrogen density, n$_H = 21.0\pm 9.6$ cm$^{-3}$ when T$_e = 100$ K, and n$_H = 4.5\pm2.0$ cm$^{-3}$ when T$_e = 1000$ K. Thus, the H-ionization photon flux $\phi(H)$, is $17.0\times 10^7$ cm$^{-2}$s$^{-1}$ when T$_e =100$ K, and is $ 3.6\times 10^7$ cm$^{-2}$s$^{-1}$ when T$_e = 1000$ K. For comparison, the Milky Way H-ionization flux is about $1\times 10^7$ cm$^{-2}$s$^{-1}$ calculated from the SED given by Black (1987). So, the UV pumping rates in our calculations are 17.0 and 3.6 times of the Milky Way rate for the 100 K and 1000 K cases, respectively. We obtain n$_e$ = 1.05$\times 10^{-2}$ cm$^{-3}$, n$_{He}$ = 1.68 cm$^{-3}$ and n$_{H_2}$ = 2.1 cm$^{-3}$ for the T$_e$ = 100 K case and n$_e$ = 4.5$\times
10^{-2}$ cm$^{-3}$, n$_{He}$ = 0.36 cm$^{-3}$ and n$_{H_2}$ = 0.45 cm$^{-3}$ for the T$_e$ = 1000 K case.
Substituting into Eq. (4), we finally estimate the contribution to the excitation of the C I fine structure levels from collisions and UV pumping. After these contributions are removed, the CMBR temperature at z = 1.9731 T$_{CMBR} = 7.9 \pm 1.0$ K when T$_e = 100$ K is adopted, and T$_{CMBR} = 10.6 \pm 1.0$ K when T$_e = 1000$ K is used. Since the electron temperature in most regions of the z = 1.9731 absorber is around 100 K, our best guess for the CMBR temperature at z = 1.9731 is 7.9 $\pm$ 1.0 K.
The above results are based on the assumption of a single homogeneous zone model, which is probably different from the real case. Previous high resolution observations of the QSO 1331 + 170 have shown that the C I absorption lines split into two components with different excitation temperature (Songaila et al. 1994). There may be two or more different velocity components associated with C I absorption in the QSO 0013$-$004 system. Without knowledge of the individual cloud structure, there may be some uncertainties in the correction of local excitation from only considering the C II fine structure excitation. Ultimately, a higher resolution spectrum is needed to get an improved measurement of the CMBR temperature at z = 1.9731.
Discussion
==========
We have estimated the local contributions to the excitation of C I , which can contribute $\sim 1-3$ K to the excitation temperature of the C I ground state fine-structure levels at z $\sim$ 2 in reasonable physical and chemical conditions for C I to exist. After estimating these local contributions, our best guess for the CMBR temperature is 7.9 $\pm 1.0$ K, which is consistent with the predicted value of 8.105 K, at z = 1.9731.
Our study shows that the local contributions to the excitation of the C I fine-structure levels are dominated by collisions with neutral hydrogen and UV pumping. However, if the number density of molecular hydrogen is comparable to that of neutral hydrogen, H$_2$ can also be an important collisional partner for C I excitation. At high electron temperature (e.g. 1000 K or higher) electrons can also be important collisional partners.
Our study also shows that the UV radiation field in the z = 1.9731 absorber is about 10 times stronger than the average value in the Milky Way. This could be the result of a higher star formation rate in this system.
Figure 6 illustrates the measurements of the CMBR temperature at different redshifts. All high redshift measures are essentially upper limits, since local contributions to the C I and C II excitation may be significant. So far, all measurements are consistent with the Big Bang predictions.
We thank Dr. A. Songaila for pointing out an important point which improved the paper. We thank G. Ferland for providing his CLOUDY program. We also thank the staff of MMTO for their help. This research was supported by NSF AST-9058510 and NASA grant NAGW-2201.
Bahcall, J. N., & Wolf, R. A., 1968, ApJ, 152, 701 Bahcall, J. N., Joss, P. C., & Lynds, R., 1973, ApJ, 182, L95 Bechtold, J., 1994, ApJS, 91, 1 Black, J. H., Chaffee, F. H., & Foltz, C. B. 1987, ApJ, 317, 442 Black, J. H., 1987, in Interstellar Processes, eds. Hollenbach, D. J., & Thronson, Jr. H. A., (Reidel), 931 Ferland, G. J., 1993, Univ. of Kentucky, Dept. of Physics and Astronomy Internal Report. Flower, D., 1990, Molecular Collisions in the Interstellar Medium, (Cambridge: The University Press) Jenkins, E. B., & Shaya, E. J., 1979, ApJ, 231, 55 Jenkins, E. B., 1986, ApJ, 304, 739 Johnson, C. T., Burke, P. G., & Kingston, A. E., 1987, J. Phys. B, 20, 2553 Keenan, F. P., Lennon, D. J., Johnson, C. T., & Kingston, A. E., 1986, MNRAS, 220, 571 Launay, J. M., & Roueff, E., 1977, A&A, 56, 289 Lu, L., Sargent, W. L. W., Womble, D. S., & Barlow, T. A., 1995, Preprint Mather, J. C., et al., 1994, ApJ, 420, 439 Mathis, J. S., Mezger, P. G., & Panagia, N., 1983, A&A, 128, 212 Meyer, D. M., Black, J. H., Chaffee, F. H., Foltz, C. B., & York, D. G., 1986, ApJ, 308, L37 Morton, D. C., ApJS, 1991, 77, 119 Peebles, P, J. E., 1993, Principles of Physical Cosmology (Princeton: Princeton University Press) Pettini, M., Smith, L. J., Hunstead, R. W., & King, D. L., 1994, ApJ, 426, 79 Roueff, E., & Le Bourlet, J., 1990, A&A, 236, 515 Schneider, D. P., et al., 1993, ApJS, 87, 45 Schröder, K., Staemmler, V., Smith, M. D., Flower, D. R., & Jaquet, R., 1991, J. Phys. B, 24, 2487 Songaila, A., et al., 1994a, Nature, 368, 599 Songaila, A., et al., 1994b, Nature, 371, 43 Staemmler, V., & Flower, D. R., 1991, J. Phys. B, 24, 2343
**Figure Captions**
Figure 1.—Spectrum of QSO 0013$-$004 with significant absorption lines marked. The solid line is the fit to the continuum. The dotted line is the 1 $\sigma$ error. The features marked with an asterisk are bad columns or traps in the CCD.
Figure 2.—(a). Profile fit to the C II lines at z = 1.9672, 1.9711 and 1.9733, and C II$^*$ line at z = 1.9732 . The solid line is the continuum. The dot-dashed line is the 1 $\sigma$ errors. The dashed line is the Gaussian fit to the absorption lines with four velocity components (see Table 1 and text). (b). Profile fit to the C I [$\lambda$]{} 1560.31 Å and C I$^*$ [$\lambda$]{} [$\lambda$]{} 1560.68, 1560.70 Å lines at z = 1.9731. The solid line is the continuum. The dashed line is the Gaussian fit to the absorption lines.
Figure 3. —Spectrum of QSO 0013$-$004 showing the C I multiplets 2 and 3 and C II multiplet 1 listed in Table 1. The dotted line shows the 1 $\sigma$ error. The expected positions of the ground state (J = 0 or 1/2) and excited states (J =1, 2 or 3/2) are marked.
Figure 4.—Curve of growth for singly ionized and neutral species in the z = 1.9731 system. The solid line which best fits the data from Si II absorption lines is the theoretical curve of growth for b = 42 $\pm$ 2 [km s$^{-1}$]{}.
Figure 5.—Results of ionization model for a H I dominant region with N(H I) = $5\times 10^{20}$ cm$^{-2}$. The ordinate is the column density of various ions, the abscissa is the log of ionization parameter U. The solid lines are predicted values from CLOUDY. The dotted lines correspond to measured column densities of C II and C I. The error bars are 1 $\sigma$ errors on the column densities of C I and C II which include errors from photon statistics and uncertainty in the b-value (42 $\pm$ 2 [km s$^{-1}$]{}).
Figure 6.—Measurements of the CMBR temperature as a function of redshifts. The solid line is the predicted relation. The filled circle is from the COBE measurement (Mather et al. 1994). The open squares are upper limits obtained by Songaila et al. (1994a,b). The filled square is obtained here. The filled hexagon is obtained by Lu et al. (1995).
-10 mm
Table 1. The identifications of absorption lines of QSO 0013-004
No. $\lambda_{obs}$(Å) $\sigma(\lambda)$ W$_{obs}$ (Å) $\sigma(W)$ W$_{obs}$ (Å)$^a$ $\sigma(W)^a$ Significance Level ID z$_{abs}$
----- -------------------- ------------------- --------------- ------------- ------------------- --------------- -------------------- --------------- -----------
1 3869.22 0.03 2.145 0.081 26.62 O I 1302 1.9714
2 3871.58 0.03 2.719 0.093 29.28 O I 1302 1.9732
Si II 1304 1.9673
3 3874.02 0.03 0.537 0.057 9.41 Si II 1304 1.9700
O I 1302 1.9756
4 3875.63 0.04 1.336 0.074 1.22 0.23 18.16 Si II 1304 1.9713
O I 1302 1.9776
5 3878.03 0.05 1.974 0.086 2.12 0.22 22.31 Si II 1304 1.9731
6 3936.93 0.09 0.564 0.055 10.32 C IV 1548 1.5429
7 3943.25 0.05 0.286 0.037 7.79 C IV 1550 1.5428
8 3944.27 0.09 0.257 0.041 6.23 O I 1304 2.0290
9 3951.09 0.08 0.252 0.040 6.28 Si II 1304 2.0291
C I 1329 1.9731
10 3959.78 0.02 2.853 0.051 2.91 0.19 55.44 C II 1334 1.9672
11 3964.48 0.02 4.581 0.052 4.80 0.70 88.57 C II 1334 1.9711
12 3967.95 0.02 2.664 0.042 2.31 0.58 63.68 C II 1334 1.9733
13 3971.34 0.05 0.797 0.045 0.80 0.09 17.81 C II$^*$ 1335 1.9732
14 4042.49 0.06 0.855 0.046 18.54 C II 1334 2.0292
15 4135.63 0.07 0.730 0.046 15.80 Si IV 1393 1.9674
16 4139.41 0.03 1.137 0.039 29.10 Si IV 1393 1.9700
17 4141.78 0.02 1.577 0.041 38.54 Si IV 1393 1.9717
18 4162.39 0.10 0.406 0.043 9.34 Si IV 1402 1.9673
19 4166.10 0.04 0.654 0.039 16.99 Si IV 1402 1.9699
20 4168.29 0.04 1.031 0.043 23.97 Si IV 1402 1.9715
21 4201.21 0.02 1.961 0.040 48.43 C IV 1548 1.7136
22 4208.07 0.03 1.610 0.043 37.07 C IV 1402 1.7135
23 4221.58 0.03 1.045 0.036 29.02 Si IV 1393 2.0289
24 4248.90 0.08 0.578 0.044 13.20 Si IV 1402 2.0290
25 4530.40 0.02 1.980 0.043 46.48 Si II 1526 1.9674
26 4534.19 0.06 0.478 0.038 12.53 Si II 1526 1.9699
27 4536.38 0.02 1.754 0.038 1.89 0.14 46.13 Si II 1526 1.9714
28 4538.98 0.02 2.486 0.042 2.55 0.14 66.14 Si II 1526 1.9731
29 4572.32 0.11 0.146 0.026 5.59
30 4587.70 0.16 0.158 0.031 5.11
31 4594.00 0.03 1.880 0.037 51.47 C IV 1548 1.9673
32 4597.70 0.02 1.746 0.030 58.65 C IV 1548 1.9697
33 4601.13 0.02 3.724 0.039 95.74 C IV 1548 1.9719
34 4605.54 0.01 1.804 0.027 66.51 C IV 1550 1.9698
35 4608.24 0.01 2.396 0.031 77.52 C IV 1550 1.9716
C IV 1548 1.9756
36 4610.95 0.04 0.515 0.026 19.60 C IV 1550 1.9733
C IV 1548 1.9777
37 4614.64 0.05 0.592 0.032 18.45 C IV 1550 1.9757
38 4617.95 0.08 0.250 0.029 8.72 C IV 1550 1.9778
39 4630.00 0.015 0.162 0.030 5.31 C I 1560 1.9674
40 4638.81 0.08 0.198 0.026 0.17 0.02 7.57 C I 1560 1.9730
41 4689.12 0.03 1.760 0.036 49.04 C IV 1548 2.0288
42 4696.79 0.03 1.292 0.032 40.91 C IV 1550 2.0287
43 4772.77 0.06 0.372 0.018 20.73 Fe II 1608 1.9673
44 4777.48 0.06 0.074 0.011 7.04 Fe II 1608 1.9702
45 4779.34 0.02 0.548 0.013 41.00 Fe II 1608 1.9714
46 4782.21 0.02 0.992 0.016 60.25 Fe II 1608 1.9731
47 4916.83 0.21 0.321 0.047 6.77 C I 1656 1.9674
48 4926.08 0.10 0.432 0.043 0.44 0.05 10.13 C I 1656 1.9730
49 4957.65 0.05 1.531 0.066 23.29 Al III 1670 1.9673
50 4962.44 0.08 0.923 0.062 14.83 Al III 1670 1.9701
51 4964.81 0.02 1.753 0.048 36.25 Al III 1670 1.9715
52 4967.65 0.05 2.232 0.067 33.11 Al III 1670 1.9732
53 5061.16 0.15 0.425 0.061 6.94 Al III 1670 2.0292
54 5208.79 0.16 0.735 0.091 8.12
55 5211.96 0.17 0.523 0.085 6.16
$^a$The eqivalent widths are measured through Gaussian profile fitting.
-10 mm -5mm
Table 2. The expected wavelengths for C I and C II in the z = 1.9731 absorber
Multiplet J $\lambda_{rest}$(Å) $\lambda_{rest}(1+z)$(Å)$^a$ f W$_{obs}$(Å) $\sigma(W)$ SL
------------------ ----- --------------------- ------------------------------ ------- ----------------- ----------------- -------
C I multiplet 2 0 1656.928 4926.213 0.141 0.432 0.043 10.1
1 1656.267 4924.247 0.059 0.088 0.027 3.3
1 1657.379$^b$ 4927.554 0.035
1 1657.907 4929.123 0.047 0.054 0.028 1.9
2 1657.008$^b$ 4926.451 0.105
2 1658.121$^c$ 4929.594 0.035 $<$0.111
C I multiplet 3 0 1560.309 4638.955 0.080 0.198(0.17)$^f$ 0.026(0.02)$^f$ 7.6
1 1560.682 4639.908 0.060 0.082(0.09)$^f$ 0.019(0.02)$^f$ 4.4
1 1560.709$^d$ 4640.144 0.020
2 1561.340$^c$ 4642.020 0.012 $< 0.087$
2 1561.367$^c$ 4642.100 0.001 $<0.087$
2 1561.438$^c$ 4642.311 0.068 $<0.087$
C II multiplet 1 1/2 1334.532 3967.698 0.128 2.664 0.039 68.23
3/2 1335.663$^e$ 3971.059 0.013
3/2 1335.708 3971.193 0.115 0.797 0.042 18.92
$^a$These wavelengths are vacuum, heliocentric values.
$^b$The line is blended with J = 0 line of $\lambda = 1656.928$ Å.
$^c$The upper limits are 3 $\sigma$.
$^d$The line is blended with J = 1 line of $\lambda = 1560.709$ Å.
$^e$The line is blended with J = 3/2 line of $\lambda = 1335.708$ Å.
$^f$The measurement in the bracket is from the Gaussian profile fitting.
-10 mm -5mm
Table 3. Excitation Temperature of C I and C II at z = 1.9731.
------------------------------------------ -------------------------------- ----------
C I multiplet 2 $\tau_0$
N($\times 10^{13}$ cm$^{-2}$):
J = 0, $\lambda = 1656.928$ Å 4.9 $\pm$ 0.5 0.43
J = 1, $\lambda = 1656.267$ Å 2.1 $\pm$ 0.6 0.078
J = 1, $\lambda = 1657.907$ Å 1.6 $\pm$ 0.8 0.047
$<$J = 1$>_{Weighted}$ 1.9 $\pm$ 0.5
N(J = 1)/N(J = 0) 0.39 $\pm 0.11$
T$_{ex}$(K) 11.6 $\pm$ 1.6
C I multiplet 3
N($\times 10^{13}$ cm$^{-2}$):
J = 0, $\lambda = 1560.309$ Å 4.1 $\pm$ 0.5 0.19
J = 1, $\lambda = 1560.695$ Å 1.6 $\pm$ 0.4 0.076
N(J = 1)/N(J = 0) 0.39 $\pm$ 0.10
T$_{ex}$(K) 11.6 $\pm$ 1.4
Weighted Mean of C I Multiplets
N(J = 0)($\times 10^{13}$ cm$^{-2}$) 4.5 $\pm$ 0.4
N(J = 1)($\times 10^{13}$ cm$^{-2}$) 1.8 $\pm$ 0.3
N(J = 1)/N(J = 0) 0.39 $\pm$ 0.07
T$_{ex}$(K) 11.6 $\pm$ 1.0
C II multiplet 1
N(J = 1/2)($\times 10^{16}$ cm$^{-2})$ 2.7 $\pm$ 1.2 164
N(J = 3/2)($\times 10^{14}$ cm$^{-2})$ 1.9 $\pm$ 0.2 1.1
N(J = 3/2)/N(J = 1/2) 7.0($\pm$ 3.2)$\times 10^{-3}$
T$_{ex}$(K) 16.1 $\pm$ 1.4
------------------------------------------ -------------------------------- ----------
[^1]: Observations here were obtained with the Multiple Mirror Telescope, a joint facility of the University of Arizona and the Smithsonian Institution.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
A [*private learner*]{} is an algorithm that given a sample of labeled individual examples outputs a generalizing hypothesis while preserving the privacy of each individual. In 2008, Kasiviswanathan et al. (FOCS 2008) gave a generic construction of private learners, in which the sample complexity is (generally) higher than what is needed for non-private learners. This gap in the sample complexity was then further studied in several followup papers, showing that (at least in some cases) this gap is unavoidable. Moreover, those papers considered ways to overcome the gap, by relaxing either the privacy or the learning guarantees of the learner.\
We suggest an alternative approach, inspired by the (non-private) models of [*semi-supervised learning*]{} and [*active-learning*]{}, where the focus is on the sample complexity of [*labeled*]{} examples whereas [*unlabeled*]{} examples are of a significantly lower cost. We consider private semi-supervised learners that operate on a random sample, where only a (hopefully small) portion of this sample is labeled. The learners have no control over which of the sample elements are labeled. Our main result is that the labeled sample complexity of private learners is characterized by the VC dimension.\
We present two generic constructions of private semi-supervised learners. The first construction is of learners where the labeled sample complexity is proportional to the VC dimension of the concept class, however, the unlabeled sample complexity of the algorithm is as big as the representation length of domain elements. Our second construction presents a new technique for decreasing the labeled sample complexity of a given private learner, while roughly maintaining its unlabeled sample complexity. In addition, we show that in some settings the labeled sample complexity does not depend on the privacy parameters of the learner.
author:
- |
Amos Beimel[^1] Kobbi Nissim[^2] Uri Stemmer[^3]\
\
Dept. of Computer Science\
Ben-Gurion University of the Negev\
[{beimel|kobbi|stemmer}@cs.bgu.ac.il]{}
title: Learning Privately with Labeled and Unlabeled Examples
---
Introduction
============
A [*private learner*]{} is an algorithm that given a sample of labeled examples, where each example represents an individual, outputs a generalizing hypothesis while preserving the privacy of each individual. This formal notion, combining the requirements of PAC learning [@Valiant84] and Differential Privacy [@DMNS06], was presented in 2008 by Kasiviswanathan et al. [@KLNRS08], who also gave a generic construction of private learners. However, the sample complexity of the learner of [@KLNRS08] is (generally) higher than what is needed for non-private learners. Namely, their construction requires $O(\log|C|)$ samples for learning a concept class $C$, as opposed to the non-private sample complexity of $\Theta({{\rm VC}}(C))$.
This gap in the sample complexity was studied in several followup papers. For [*pure*]{} differential privacy, it was shown that in some cases this gap can be closed with the price of giving up proper learning – where the output hypothesis should be from the learned concept class – for [*improper*]{} learning. Indeed, it was shown that for the class of point functions over domain of size $2^d$, the sample complexity of every proper private learner is $\Omega(d)$ (matching the upper bound of [@KLNRS08]), whereas there exist improper private learners with sample complexity $O(1)$ that use pseudorandom or pairwise independent functions as their output hypotheses [@BBKN12; @BNS13].[^4] A complete characterization for the sample complexity of pure-private improper-learners was given in [@BNS13] in terms of a new dimension – the Representation Dimension. They showed that $\Theta({{\rm RepDim}}(C))$ examples are both necessary and sufficient for a pure-private improper-learner for a class $C$. Following that, Feldman and Xiao [@FX14] separated the sample complexity of pure-private learners from that of non-private ones, and showed that the representation dimension can sometimes be significantly bigger then the VC dimension. For example, they showed that every pure-private learner (proper or improper) for the class of thresholds over $\{0,1\}^d$ requires $\Omega(d)$ samples [@FX14] (while there exists a non-private proper-learner with sample complexity $O(1)$).
Another approach for reducing the sample complexity of private learners is to relax the privacy requirement to [*approximate*]{} differential privacy. This relaxation was shown to be significant as it allows privately and [*properly*]{} learning point functions with $O(1)$ sample complexity, and threshold functions with sample complexity $2^{O(\log^* d)}$ [@BNS13b]. Recently, Bun et al. [@BNSV14] showed that the dependency in $\log^* d$ in necessary. Namely, they showed that every approximate-private proper-learner for the class of thresholds over $\{0,1\}^d$ requires $\Omega(\log^* d)$ samples. This separates the sample complexity of approximate-private proper-learners from that of non-private learners.
Tables \[table:upper\] and \[table:lower\] summarize the currently known bounds on the sample complexity of private learners. Table \[table:upper\] specifies [*general*]{} upper bounds, and table \[table:lower\] specifies known upper and lower bounds on the sample complexity of privately learning thresholds over $\{0,1\}^d$.
[c|c|c]{} & Pure-privacy & Approximate-privacy\
----------
Proper
learning
----------
: General upper bounds on the sample complexity of private learners for a class $C$.[]{data-label="table:upper"}
& $O(\log|C|)$ & $O(\log|C|)$\
----------
Improper
learning
----------
: General upper bounds on the sample complexity of private learners for a class $C$.[]{data-label="table:upper"}
& $\Theta({{\rm RepDim}}(C))$ & $O({{\rm RepDim}}(C))$\
[c|c|c]{} & Pure-privacy & Approximate-privacy\
----------
Proper
learning
----------
: Bounds on the sample complexity of private learners for a thresholds over $\{0,1\}^d$. While the VC dimension of this class is constant, its representation dimension is $\Theta(d)$.[]{data-label="table:lower"}
& $\Theta(d)$ &
--------------------------------
Upper bound: $2^{O(\log^*d)}$
Lower bound: $\Omega(\log^*d)$
--------------------------------
: Bounds on the sample complexity of private learners for a thresholds over $\{0,1\}^d$. While the VC dimension of this class is constant, its representation dimension is $\Theta(d)$.[]{data-label="table:lower"}
\
----------
Improper
learning
----------
: Bounds on the sample complexity of private learners for a thresholds over $\{0,1\}^d$. While the VC dimension of this class is constant, its representation dimension is $\Theta(d)$.[]{data-label="table:lower"}
& $\Theta(d)$ &
-------------------------------
Upper bound: $2^{O(\log^*d)}$
Lower bound: $\Omega(1)$
-------------------------------
: Bounds on the sample complexity of private learners for a thresholds over $\{0,1\}^d$. While the VC dimension of this class is constant, its representation dimension is $\Theta(d)$.[]{data-label="table:lower"}
\
This Work
---------
In this work we examine an alternative approach for reducing the costs of private learning, inspired by the (non-private) models of [*semi-supervised learning*]{} [@SemiSupervised] and [*active learning*]{} [@ActiveLearning].[^5] In both models, the focus is on reducing the sample complexity of [*labeled*]{} examples whereas it is assumed that [*unlabeled*]{} examples can be obtained with a significantly lower cost. In this vein, a recent work by Balcan and Feldman [@BF13] suggested a generic conversion of active learners in the model of statistical queries [@Kearns98] into learners that also provide differential privacy. For example, Balcan and Feldman showed an active pure-private proper-learner for the class of thresholds over $\{0,1\}^d$ that uses $O(1)$ labeled examples and $O(d)$ unlabeled examples.
We show that while the unlabeled sample complexity of private learners is subject to the lower bounds mentioned in tables \[table:upper\] and \[table:lower\], the [*labeled*]{} sample complexity is characterized by the VC dimension of the target concept class. We present two generic constructions of private semi-supervised learners via an approach that deviates from most of the research in semi-supervised and active learning: (1) Semi-supervised learning algorithms and heuristics often rely on strong assumptions about the data, e.g., that close points are likely to be labeled similarly, that the data is clustered, or that the data lies on a low dimensional subspace of the input space. In contrast, we work in the standard PAC learning model, and need not make any further assumptions. (2) Active learners examine their pool of unlabeled data and then choose (maybe adaptively) which data examples to label. Our learners have no control over which of the sample elements are labeled. Our main result is that the labeled sample complexity of such learners is characterized by the VC dimension. Our first generic construction is of learners where the labeled sample complexity is proportional to the VC dimension of the concept class. However, the unlabeled sample complexity of the algorithm is as big as the representation length of domain elements. The learner for a class $C$ starts with an unlabeled database and uses private sanitization to create a synthetic database, with roughly ${{\rm VC}}(C)$ points, that can answer queries in a class related to $C$. It then uses this database to choose a subset of the hypotheses of size $2^{O({{\rm VC}}(C))}$ and then uses the exponential mechanism [@MT07] to choose from these hypotheses using $O({{\rm VC}}(C))$ labeled examples.
As an example, applying this technique with the private sanitizer for threshold functions from [@BNS13b] we get a (semi-supervised) approximate-private proper-learner for thresholds over $\{0,1\}^d$ with optimal $O(1)$ labeled sample complexity and near optimal $2^{O(\log^*d)}$ unlabeled sample complexity. This matches the labeled sample complexity of Balcan and Feldman [@BF13] (ignoring the dependency in all parameters except for $d$), and improves on the unlabeled sample complexity.[^6]
Our second construction presents a new technique for decreasing the labeled sample complexity of a given private learner ${\mathcal A}$. At the heart of this construction is a technique for choosing (non-privately) a hypothesis using a small labeled database; this hypothesis is used to label a bigger database, which is given to the private learner ${\mathcal A}$.
Consider, for example, the concept class ${{\tt RECTANGLE}}_d^{\ell}$ containing all axis-aligned rectangles over $\ell$ dimensions, where each dimension consists of $2^d$ points. Applying our techniques on the learner from [@BNS13b] results in a non-active semi-supervised private learner with optimal $O(\ell)$ labeled sample complexity and with $\widetilde{O}(\ell^3 \cdot 8^{\log^*d})$ unlabeled sample complexity. This matches the labeled sample complexity of Balcan and Feldman [@BF13], and improves the unlabeled sample complexity whenever the dimension $\ell$ is not too big (roughly, $\ell \leq \sqrt{d}$).
#### Private Active Learners.
We study the labeled sample complexity of private [*active*]{} learners, i.e., learners that operate on a pool of unlabeled examples (individuals’ data) and adaptively query the labels of specific examples. As those queries depend on individuals’ data, they may breach privacy if exposed. We, therefore, introduce a stronger definition for private active learners that remedies this potential risk, and show that (most of) our learners satisfy this stronger definition, while the learners of [@BF13] do not. This strong definition has its downside, as we show that (at least in some cases) it introduces a $\frac{1}{\alpha}$ blowup to the labeled sample complexity (where $\alpha$ is the approximation parameter). On the other hand, when considering private active learners that only satisfy the definition of [@BF13] (which is still a reasonable definition), we show that the labeled sample complexity has no dependency on the privacy parameters.
Related Work
------------
Differential privacy was defined in [@DMNS06] and the relaxation to approximate differential privacy is from [@DKMMN06]. Most related to our work is the work on private learning and its sample complexity [@BDMN05; @KLNRS08; @CH11; @DRV10; @BBKN12; @BNS13; @BNS13b; @FX14] and the early work on sanitization [@BLR08full]. Blum et al. [@BDMN05] showed that computationally efficient private-learners exist for all concept classes that can be efficiently learned in the [*statistical queries*]{} model of [@Kearns98]. Kasiviswanathan et al. [@KLNRS08] showed an example of a concept class – the class of parity functions – that is not learnable in the statistical queries model but can be learned privately and efficiently. These positive results show that many “natural” learning tasks that are efficiently learned non-privately can be learned privately and efficiently.
Chaudhuri and Hsu [@CH11] presented upper and lower bounds on the sample complexity of [*label-private*]{} learners, a relaxation of private learning where the learner is required to only protect the privacy of the labels in the sample. Following that, Beimel et al. [@BNS13b] showed that the VC dimension completely characterizes the sample complexity of such learners.
Dwork et al. [@DRV10] showed how to boost the accuracy of private learning algorithms. That is, given a [*private*]{} learning algorithm that has a big classification error, they produced a [*private*]{} learning algorithm with small error. Other tools for private learning include, e.g., private SVM [@RBHT09], private logistic regression [@CM08], and private empirical risk minimization [@CMS11].
Preliminaries
=============
In this section we define differential privacy and semi-supervised (private) learning. Additional preliminaries on the VC dimension and on data sanitization are deferred to the appendix.
#### Notation.
We use $O_{\gamma}(g(n))$ as a shorthand for $O(h(\gamma) \cdot g(n))$ for some non-negative function $h$. In informal discussions, we sometimes write $\widetilde{O}(g(n))$ to indicate that $g(n)$ is missing lower order terms. We use $X$ to denote an arbitrary domain, and $X_d$ for the domain $\{0,1\}^d$.
#### Differential Privacy.
Consider a database where each entry contains information pertaining to an individual. An algorithm operating on such databases is said to preserve [*differential privacy*]{} if its outcome is insensitive to any modification in a single entry. Formally:
\[def:dp\] Databases $S_1\in X^n$ and $S_2\in X^n$ over a domain $X$ are called [*neighboring*]{} if they differ in exactly one entry. A randomized algorithm ${\mathcal A}$ is $(\epsilon,\delta)$-differentially private if for all neighboring databases ${S}_1,{S}_2\in X^n$, and for all sets $F$ of outputs, $$\begin{aligned}
\label{eqn:diffPrivDef}
& \Pr[{\mathcal A}({S}_1) \in F] \leq \exp(\epsilon) \cdot \Pr[{\mathcal A}({S}_2) \in F] + \delta. &\end{aligned}$$ The probability is taken over the random coins of ${\mathcal A}$. When $\delta{=}0$ we omit it and say that ${\mathcal A}$ preserves [*pure*]{} differential privacy, otherwise (when $\delta>0$) we say that ${\mathcal A}$ preserves [*approximate*]{} differential privacy.
See Appendix \[sec:dp\_mech\] for basic differentially private mechanisms.
#### Semi-Supervised PAC Learning.
The standard PAC model (and similarly private PAC) focuses on learning a class of concepts from a sample of labeled examples. In a situation where labeled examples are significantly more costly than unlabeled ones, it is natural to attempt to use a combination of labeled and unlabeled data to reduce the number of labeled examples needed. Such learners may have no control over which of the examples are labeled, as in [*semi-supervised learning*]{}, or may specifically choose which examples to label, as in [*active learning*]{}. In this section we focus on semi-supervised learning. Active learning will be discussed in Section \[sec:privActive\].
A concept $c:X\rightarrow \{0,1\}$ is a predicate that labels [*examples*]{} taken from the domain $X$ by either 0 or 1. A *concept class* $C$ over $X$ is a set of concepts (predicates) mapping $X$ to $\{0,1\}$. A semi-supervised learner is given $n$ examples sampled according to an unknown probability distribution $\mu$ over $X$, where $m\leq n$ of these examples are labeled according to an unknown [*target*]{} concept $c\in C$. The learner succeeds if it outputs a hypothesis $h$ that is a good approximation of the target concept according to the distribution $\mu$. Formally:
Let $c$ and $\mu$ be a concept and a distribution over a domain $X$. The [*generalization error*]{} of a hypothesis $h:X\rightarrow\{0,1\}$ w.r.t. $c$ and $\mu$ is defined as ${{\rm error}}_{\mu}(c,h)=\Pr_{x \sim \mu}[h(x)\neq c(x)].$ When ${{\rm error}}_{\mu}(c,h)\leq\alpha$ we say that $h$ is [*$\alpha$-good*]{} for $c$ and $\mu$.
\[def:PAC\] Let $C$ be a concept class over a domain $X$, and let ${\mathcal A}$ be an algorithm operating on (partially) labeled databases. Algorithm ${\mathcal A}$ is an [*$(\alpha,\beta,n,m)$-SSL (semi-supervised learner)*]{} for $C$ if for all concepts $c \in C$ and all distributions $\mu$ on $X$ the following holds.
Let $D=(x_i,y_i)_{i=1}^n\in(X\times\{0,1,\bot\})^n$ be a database s.t. (1) each $x_i$ is drawn i.i.d. from $\mu$; (2) in the first $m$ entries $y_i=c(x_i)$; (3) in the last $(n-m)$ entries $y_i=\bot$. Then, $$\Pr[{\mathcal A}(D){=}h \text{ s.t.\ } {{\rm error}}_{\mu}(c,h) > \alpha] \leq \beta.$$ The probability is taken over the choice of the samples from $\mu$ and the coin tosses of ${\mathcal A}$.
If a semi-supervised learner is restricted to only output hypotheses from the target concept class $C$, then it is called a [*proper*]{} learner. Otherwise, it is called an [*improper*]{} learner. We sometimes refer to the input of a semi-supervised learner as two databases $D\in (X \times \{\bot\})^{n-m}$ and $S\in(X \times \{0,1\})^m$, where $m$ and $n$ are the [*labeled*]{} and [*unlabeled*]{} sample complexities of the learner.
Given a [*labeled*]{} sample $S=(x_i,y_i)_{i=1}^m$, the [*empirical error*]{} of a hypothesis $h$ on $S$ is ${{\rm error}}_S(h) = \frac{1}{m} |\{i : h(x_i) \neq y_i\}|$. Given an [*unlabeled*]{} sample $D=(x_i)_{i=1}^n$ and a target concept $c$, the [*empirical error*]{} of $h$ w.r.t. $D$ and $c$ is ${{\rm error}}_D(h,c) = \frac{1}{n} |\{i : h(x_i) \neq c(x_i)\}|$.
Semi-supervised learning algorithms operate on a (partially) labeled sample with the goal of choosing a hypothesis with a small [*generalization*]{} error. Standard arguments in learning theory (see Appendix \[sec:VC\]) state that the generalization of a hypothesis $h$ and its [*empirical*]{} error (observed on a large enough sample) are similar. Hence, in order to output a hypothesis with small generalization error it suffices to output a hypothesis with small empirical error.
#### Agnostic Learner.
Consider an SSL for an [*unknown*]{} class $C$ that uses a (known) hypotheses class $H$. If $H\neq C$, then a hypothesis with small empirical error might not exist in $H$. Such learners are referred to in the literature as [*agnostic*]{}-learners, and are only required to produce a hypothesis $f\in H$ (approximately) minimizing ${{\rm error}}_{\mu}(c,f)$, where $c$ is the (unknown) target concept.
\[def:PACagnostic\] Let $H$ be a concept class over a domain $X$, and let ${\mathcal A}$ be an algorithm operating on (partially) labeled databases. Algorithm ${\mathcal A}$ is an [*$(\alpha,\beta,n,m)$-agnostic-SSL*]{} using $H$ if for all concepts $c$ (not necessarily in $H$) and all distributions $\mu$ on $X$ the following holds.
Let $D=(x_i,y_i)_{i=1}^n\in(X\times\{0,1,\bot\})^n$ be a database s.t. (1) each $x_i$ is drawn i.i.d. from $\mu$; (2) in the first $m$ entries $y_i=c(x_i)$; (3) in the last $(n-m)$ entries $y_i=\bot$. Then, ${\mathcal A}(D)$ outputs a hypothesis $h\in H$ satisfying $\Pr[{{\rm error}}_{\mu}(c,h) \leq \min_{f\in H}\{{{\rm error}}_{\mu}(c,f)\}+\alpha] \geq 1-\beta.$ The probability is taken over the choice of the samples from $\mu$ and the coin tosses of ${\mathcal A}$.
#### Private Semi-Supervised PAC learning. {#sec:PPAC}
Similarly to [@KLNRS08] we define private semi-supervised learning as the combination of Definitions \[def:dp\] and \[def:PAC\].
\[def:private-SSL\] Let ${\mathcal A}$ be an algorithm that gets an input $S\in(X\times\{0,1,\bot\})^n$. Algorithm ${\mathcal A}$ is an [*$(\alpha,\beta,\epsilon,\delta,n,m)$-PSSL (private SSL)*]{} for a concept class $C$ over $X$ if ${\mathcal A}$ is an $(\alpha,\beta,n,m)$-SSL for $C$ [*and*]{} ${\mathcal A}$ is $(\epsilon,\delta)$-differentially private.
#### Active Learning.
Semi-supervised learners are a subset of the larger family of [*active learners*]{}. Such learners can adaptively request to reveal the labels of specific examples. See formal definition and discussion in Section \[sec:privActive\].
A Generic Construction Achieving Low Labeled Sample Complexity {#sec:semiSuper}
==============================================================
We next study the labeled sample complexity of private semi-supervised learners. We begin with a generic algorithm showing that for every concept class $C$ there exist a pure-private proper-learner with labeled sample complexity (roughly) ${{\rm VC}}(C)$. This algorithm, called $GenericLearner$, is described in Algorithm \[alg:genericPrivate\]. The algorithm operates on a labeled database $S$ and on an unlabeled database $D$. First, the algorithm produces a sanitization $\widetilde{D}$ of the unlabeled database $D$ w.r.t. $C^{\oplus}$ (to be defined). Afterwards, the algorithm uses $\widetilde{D}$ to construct a small set of hypotheses $H$ (we will show that $H$ contains at least one good hypothesis). Finally, the algorithm uses the exponential mechanism to choose a hypothesis out of $H$.
Similar ideas have appeared in [@CH11; @BNS13b] in the context of [*label-private*]{} learners, i.e., learners that are only required to protect the privacy of the [*labels*]{} in the sample (and not the privacy of the elements themselves). Like $GenericLearner$, the learners of [@CH11; @BNS13b] construct a small set of hypotheses $H$ that “covers” the hypothesis space and then use the exponential mechanism in order to choose a hypothesis $h\in H$. However, $GenericLearner$ differs in that it protects the privacy of the entire sample (both the labels and the elements themselves).
Given two concepts $h,f\in C$, we denote $(h {\oplus} f): X_d \rightarrow \{0,1\} $, where $(h {\oplus} f)(x)=1$ if and only if $h(x)\neq f(x)$. Let $C^{\oplus}=\{ (h {\oplus} f) \; : \; h,f\in C \}.$
To preserve the privacy of the examples in $D$, we first create a sanitized version of it – $\widetilde{D}$. If the entries of $D$ are drawn i.i.d. according to the underlying distribution (and if $D$ is big enough), then a hypothesis with small empirical error on $D$ also has small generalization error (see Theorem \[thm:generalization\]). Our learner classifies the sanitized database $\widetilde{D}$ with small error, thus we require that a small error on $\widetilde{D}$ implies a small error on $D$. Specifically, if $c$ is the target concept, then we require that for every $f \in C$, ${{\rm error}}_D(f,c) = \frac{1}{|D|} \left| \{ x\in D \; : \; f(x)\neq c(x) \} \right|$ is approximately the same as ${{\rm error}}_{\widetilde{D}}(f,c) = \frac{1}{|\widetilde{D}|} \left| \{ x\in \widetilde{D} \; : \; f(x)\neq c(x) \} \right|$. Observe that this is exactly what we would get from a sanitization of $D$ w.r.t. the concept class $C^{\oplus c}=\{ (f {\oplus} c) \; : \; f\in C \}$. As the target concept $c$ is unknown, we let $\widetilde{D}$ be a sanitization of $D$ w.r.t. $C^{\oplus}$, which contains $C^{\oplus c}$.
To apply the sanitization of Blum et al. [@BLR08full] to $D$ w.r.t. the class $C^{\oplus}$, we analyze the VC dimension of $C^{\oplus}$ in the next observation.
\[obs:vcdim\] For any concept class $C$ over $X_d$ it holds that ${{\rm VC}}(C^{\oplus})=O({{\rm VC}}(C))$.
Recall that the projection of $C$ on a set of domain points $B=\{b_1,\ldots,b_\ell\}\subseteq X_d$ is $\Pi_C(B)=\{\left\langle c(b_1),\ldots,c(b_\ell) \right\rangle :c\in C\}.$ Now note that for every $B=\{b_1,\ldots,b_\ell\}\subseteq X_d$ $$\begin{aligned}
\Pi_{C^{\oplus}}(B)&=&\{ \left\langle (h \oplus f)(b_1),\ldots,(h \oplus f)(b_\ell) \right\rangle :h,f\in C\}\\
&=&\{ \left\langle h(b_1),\ldots,h(b_\ell) \right\rangle \oplus \left\langle f(b_1),\ldots,f(b_\ell) \right\rangle :h,f\in C\}\\
&=&\{ \left\langle h(b_1),...,h(b_\ell) \right\rangle :h{\in} C\} \oplus \{ \left\langle f(b_1),...,f(b_\ell) \right\rangle :f{\in} C\}\\
&=&\Pi_{C}(B) \oplus \Pi_{C}(B).\end{aligned}$$
Therefore, by Sauer’s lemma \[thm:sauer\], $|\Pi_{C^{\oplus}}(B)| \leq |\Pi_{C}(B)|^2 \leq \left(\frac{e \ell}{{{\rm VC}}(C)}\right)^{2{{\rm VC}}(C)}$. Hence, for $C^{\oplus}$ to shatter a subset $B\subseteq X_d$ of size $\ell$ it must be that $\left(\frac{e \ell}{{{\rm VC}}(C)}\right)^{2{{\rm VC}}(C)} \geq 2^\ell$. For $\ell\geq 10{{\rm VC}}(C)$ this inequality does not hold, and we can conclude that ${{\rm VC}}(C^{\oplus})\leq10{{\rm VC}}(C)$.
[**Input:**]{} parameter $\epsilon$, an unlabeled database $D=(x_i)_{i=1}^{n-m}$, and a labeled database $S=(x_i,y_i)_{i=1}^m$.
1. Initialize $H=\emptyset$.
2. Construct an $\epsilon$-private sanitization $\widetilde{D}$ of $D$ w.r.t. $C^{\oplus}$, where $|\widetilde{D}|=O\left( \frac{{{\rm VC}}(C^{\oplus})}{\alpha^2}\log(\frac{1}{\alpha}) \right) = O\left( \frac{{{\rm VC}}(C)}{\alpha^2}\log(\frac{1}{\alpha}) \right)$ (e.g., using Theorem \[thm:BlumUp\]).
3. Let $B=\{b_1,\ldots,b_\ell\}$ be the set of all (unlabeled) points appearing at least once in $\widetilde{D}$.
4. For every $(z_1,\ldots,z_\ell)\in \Pi_C(B)=\{\left( c(j_1),\ldots,c(j_\ell) \right) :c\in C\}$, add to $H$ an arbitrary concept $c\in C$ s.t. $c(b_i)=z_i$ for every $1\leq i\leq\ell$.
5. Choose and return $h\in H$ using the exponential mechanism with inputs $\epsilon,H,S$.
\[thm:SampleComplexity\] Let $C$ be a concept class over $X_d$. For every $\alpha,\beta,\epsilon$, there exists an $(\alpha,\beta,\epsilon,\delta{=}0,n,m)$-private semi-supervised proper-learner for $C$, where $m=O\left(\frac{{{\rm VC}}(C)}{\alpha^3\epsilon}\log(\frac{1}{\alpha}) +\frac{1}{\alpha\epsilon}\log(\frac{1}{\beta})\right)$, and $n=O\left(\frac{d\cdot{{\rm VC}}(C)}{\alpha^3\epsilon}\log(\frac{1}{\alpha}) +\frac{1}{\alpha\epsilon}\log(\frac{1}{\beta})\right)$. The learner might not be efficient.
Note that $GenericLearner$ only accesses $D$ via a sanitizer, and only accesses $S$ using the exponential mechanism (on Step 5). As each of those two mechanisms is $\epsilon$-differentially private, and as $D$ and $S$ are two disjoint samples, $GenericLearner$ is $\epsilon$-differentially private. We, thus, only need to prove that with high probability the learner returns a good hypothesis.
Fix a target concept $c\in C$ and a distribution $\mu$ over $X$, and define the following three “good” events:
1. For every $h\in C$ it holds that $|{{\rm error}}_S(h)-{{\rm error}}_{\widetilde{D}}(h,c)|\leq\frac{3\alpha}{5}$.
2. The exponential mechanism chooses an $h\in H$ such that ${{\rm error}}_S(h) \leq \frac{\alpha}{5} + \min_{f\in H}\left\{{{\rm error}}_S(f)\right\}$.
3. For every $h\in H$ s.t. ${{\rm error}}_S(h)\leq\frac{4\alpha}{5}$, it holds that ${{\rm error}}_{\mu}(c,h)\leq\alpha$.
We first observe that when these three events happen algorithm $GenericLearner$ returns an $\alpha$-good hypothesis: For every $(y_1,\ldots,y_\ell)\in \Pi_C(B)$, algorithm $GenericLearner$ adds to $H$ a hypothesis $f$ s.t. $\forall 1\leq i \leq \ell,\;f(b_i)=y_i$. In particular, $H$ contains a hypothesis $h^*$ s.t. $h^*(x)=c(x)$ for every $x\in B$, that is, a hypothesis $h^*$ s.t. ${{\rm error}}_{\widetilde{D}}(h^*,c)=0$. As event $E_1$ has occur we have that this $h^*$ satisfies ${{\rm error}}_S(h^*)\leq \frac{3\alpha}{5}$. Thus, event $E_1 \cap E_2$ ensures that algorithm $GenericLearner$ chooses (using the exponential mechanism) a hypothesis $h\in H$ s.t. ${{\rm error}}_S(h)\leq\frac{4\alpha}{5}$. Event $E_3$ ensures, therefore, that this $h$ satisfies ${{\rm error}}_{\mu}(c,h)\leq\alpha$. We will now show $E_1 \cap E_2 \cap E_3$ happens with high probability.
Standard arguments in learning theory state that (w.h.p.) the empirical error on a (large enough) random sample is close to the generalization error (see Theorem \[thm:generalization\]). Specifically, by setting $n$ and $m$ to be at least $\frac{1250}{\alpha^2}{{\rm VC}}(C)\ln(\frac{25}{\alpha\beta})$, Theorem \[thm:generalization\] ensures that with probability at least $(1-\frac{2}{5}\beta)$, for every $h\in C$ the following two inequalities hold. $$\begin{aligned}
&&|{{\rm error}}_S(h)-{{\rm error}}_{\mu}(h,c)|\leq\frac{\alpha}{5} \label{eq:errorSD}\\
&&|{{\rm error}}_D(h,c)-{{\rm error}}_{\mu}(h,c)|\leq\frac{\alpha}{5} \label{eq:errorDD}\end{aligned}$$ Note that Event $E_3$ occurs whenever Inequality (\[eq:errorSD\]) holds (since $H\subseteq C$). Moreover, by setting the size of the unlabeled database $(n-m)$ to be at least $$\begin{aligned}
(n-m)&\geq& O\left(\frac{d \cdot {{\rm VC}}(C^{\oplus})\log(\frac{1}{\alpha})}{\alpha^3\epsilon}+\frac{\log(\frac{1}{\beta})}{\epsilon\alpha}\right)\\
&=&O\left(\frac{d \cdot {{\rm VC}}(C)\log(\frac{1}{\alpha})}{\alpha^3\epsilon}+\frac{\log(\frac{1}{\beta})}{\epsilon\alpha}\right).\end{aligned}$$ Theorem \[thm:BlumUp\] ensures that with probability at least $(1-\frac{\beta}{5})$ for every $(h\oplus f)\in C^{\oplus}$ (i.e., for every $h,f\in C$) it holds that $$\begin{aligned}
\frac{\alpha}{5}&\geq& | Q_{(h{\oplus} f)}(D) - Q_{(h{\oplus} f)}(\widetilde{D}) | \\
&=& \left| \frac{|\{x\in D : (h{\oplus} f)(x){=}1 \}|}{|D|} - \frac{|\{x\in \widetilde{D} : (h{\oplus} f)(x){=}1 \}|}{|\widetilde{D}|} \right| \\
&=& \left| \frac{|\{x\in D : h(x){\neq} f(x) \}|}{|D|} - \frac{|\{x\in \widetilde{D} : h(x){\neq} f(x) \}|}{|\widetilde{D}|} \right| \\
&=& \left| {{\rm error}}_D(h,f) - {{\rm error}}_{\widetilde{D}}(h,f) \right|. \end{aligned}$$ In particular, for every $h\in C$ it holds that $$\begin{aligned}
\left| {{\rm error}}_D(h,c) - {{\rm error}}_{\widetilde{D}}(h,c) \right|\leq\frac{\alpha}{5}.
\label{eq:errorDD'}\end{aligned}$$ Therefore (using Inequalities (\[eq:errorSD\]),(\[eq:errorDD\]),(\[eq:errorDD’\]) and the triangle inequality), Event $E_1\cap E_3$ occurs with probability at least $(1-\frac{3\beta}{5})$.
The exponential mechanism ensures that the probability of event $E_2$ is at least $1-|H| \cdot \exp(-\epsilon \alpha m /10)$ (see Proposition \[prop:expMech\]). Note that $\log|H|\leq|B|\leq|\widetilde{D}| = O\left( \frac{{{\rm VC}}(C)}{\alpha^2}\log(\frac{1}{\alpha}) \right)$. Therefore, for $m \geq O\left(\frac{{{\rm VC}}(C)}{\alpha^3\epsilon}\log(\frac{1}{\alpha}) +\frac{1}{\alpha\epsilon}\log(\frac{1}{\beta})\right)$, Event $E_2$ occurs with probability at least $(1-\frac{\beta}{5})$.
All in all, setting $n\geq O\left(\frac{d \cdot {{\rm VC}}(C)\log(\frac{1}{\alpha})}{\alpha^3\epsilon}+\frac{\log(\frac{1}{\beta})}{\epsilon\alpha}\right)$, and $m \geq O\left(\frac{{{\rm VC}}(C)}{\alpha^3\epsilon}\log(\frac{1}{\alpha}) +\frac{1}{\alpha\epsilon}\log(\frac{1}{\beta})\right)$, ensures that the probability of $GenericLearner$ failing to output an $\alpha$-good hypothesis is at most $\beta$.
Note that the labeled sample complexity in Theorem \[thm:SampleComplexity\] is optimal (ignoring the dependency in $\alpha,\beta,\epsilon$), as even without the privacy requirement every PAC learner for a class $C$ must have [*labeled*]{} sample complexity $\Omega({{\rm VC}}(C))$. However, the unlabeled sample complexity is as big as the representation length of domain elements, that is, $O(d\cdot{{\rm VC}}(C))$. Such a blowup in the unlabeled sample complexity is unavoidable in any generic construction of pure-private learners.[^7]
To show the usefulness of Theorem \[thm:SampleComplexity\], we consider the concept class ${{\tt THRESH}}_d$ defined as follows. For $0\leq j\leq 2^d$ let $c_j:X_d \rightarrow\{0,1\}$ be defined as $c_j(x)=1$ if $x<j$ and $c_j(x)=0$ otherwise. Define the concept class ${{\tt THRESH}}_d = \{c_j \, : \, 0\leq j\leq 2^d\}$. Balcan and Feldman [@BF13] showed an efficient pure-private proper-learner for ${{\tt THRESH}}_d$ with labeled sample complexity $O_{\alpha,\beta,\epsilon}(1)$ and unlabeled sample complexity $O_{\alpha,\beta,\epsilon}(d)$. At the cost of preserving approximate-privacy, and using the efficient approximate-private sanitizer for thresholds from [@BNS13b] (in Step 2 of Algorithm $GenericLearner$ instead on the sanitizer of [@BLR08full]), we get the following lemma (as $GenericLearner$ requires unlabeled examples only in Step 2, and the sanitizer of [@BNS13b] requires a database of size $\widetilde{O}_{\alpha,\beta,\epsilon,\delta}(8^{\log^*d})$).
\[intervalLeraner\] There exists an efficient approximate-private proper-learner for ${{\tt THRESH}}_d$ with labeled sample complexity $O_{\alpha,\beta,\epsilon}(1)$ and unlabeled sample complexity $\widetilde{O}_{\alpha,\beta,\epsilon,\delta}(8^{\log^*d})$.
Beimel et al. [@BNS13b] showed an efficient approximate-private proper-learner for ${{\tt THRESH}}_d$ with (both labeled and unlabeled) sample complexity $\widetilde{O}_{\alpha,\beta,\epsilon,\delta}(16^{\log^*d})$. The learner from Corollary \[intervalLeraner\] has similar unlabeled sample complexity, but improves on the labeled complexity.
Boosting the Labeled Sample Complexity of Private Learners {#sec:boost}
==========================================================
We now show a generic transformation of a private learning algorithm ${\mathcal A}$ for a class $C$ into a private learner with reduced labeled sample complexity (roughly ${{\rm VC}}(C)$), while maintaining its unlabeled sample complexity. This transformation could be applied to a proper or an improper learner, and to a learner that preserves pure or approximated privacy.
The main ingredient of the transformation is algorithm $LabelBoostProcedure$ (Algorithm \[alg:LabelBoostProcedure\]), where the labeled sample complexity is reduced logarithmically. We will later use this procedure iteratively to get our learner with labeled sample complexity $O_{\alpha,\beta,\epsilon}({{\rm VC}}(C))$.
Given a partially labeled sample $B$ of size $n$, algorithm $LabelBoostProcedure$ chooses a small subset $H$ of $C$ that strongly depends on the points in $B$ so outputting a hypothesis $h\in H$ may breach privacy. Nevertheless, $LabelBoostProcedure$ does choose a good hypothesis $h\in H$ (using the exponential mechanism) and use it to relabel part of the sample $B$. In Lemma \[lemma:TransformationPrivacy\], we analyze the privacy guarantees of Algorithm $LabelBoostProcedure$.
[**Input:**]{} A partially labeled database $B=S{\circ}T{\circ}D\in(X\times\{0,1,\bot\})^*$.
1.
<!-- -->
1. Initialize $H=\emptyset$.
2. Let $P=\{p_1,\ldots,p_\ell\}$ be the set of all points $p\in X$ appearing at least once in $S{\circ}T$.
3. For every $(z_1,\ldots,z_\ell)\in \Pi_C(P)=\{\left( c(p_1),\ldots,c(p_\ell) \right) :c\in C\}$, add to $H$ an arbitrary concept $c\in C$ s.t. $c(p_i)=z_i$ for every $1\leq i\leq\ell$.
4. \[step:Oneexpmech\] Choose $h\in H$ using the exponential mechanism with privacy parameter $\epsilon{=}1$, solution set $H$, and the database $S$.
5. \[step:Onerelabel\] Relabel $S{\circ}T$ using $h$, and denote this relabeled database as $(S{\circ}T)^h$, that is, if $S{\circ}T=(x_i,y_i)_{i=1}^t$ then $(S{\circ}T)^h=(x_i,y'_i)_{i=1}^t$ where $y'_i=h(x_i)$.
6. \[step:OneAAA\] Output $(S{\circ}T)^h {\circ} D$.
\[lemma:TransformationPrivacy\] Let ${\mathcal A}$ be an $(\epsilon,\delta)$-differentially private algorithm operating on partially labeled databases. Construct an algorithm ${\mathcal B}$ that on input a database $S{\circ}T{\circ}D\in(X\times\{0,1,\bot\})^*$ applies ${\mathcal A}$ on the outcome of $LabelBoostProcedure(S{\circ}T{\circ}D)$. Then, ${\mathcal B}$ is $(\epsilon+3,4e\delta)$-differentially private.
Consider the executions of ${\mathcal B}$ on two neighboring inputs $S_1{\circ}T_1{\circ}D_1$ and $S_2{\circ}T_2{\circ}D_2$. If these two neighboring inputs differ (only) on the last portion $D$ then the executions of $LabelBoostProcedure$ on these neighboring inputs are identical, and hence Inequality (\[eqn:diffPrivDef\]) (approximate differential privacy) follows from the privacy of ${\mathcal A}$. We, therefore, assume that $D_1=D_2=D$ (and that $S_1{\circ}T_1,S_2{\circ}T_2$ differ in at most one entry).
Denote by $H_1,P_1$ and by $H_2,P_2$ the elements $H,P$ as they are in the executions of algorithm $LabelBoostProcedure$ on $S_1{\circ}T_1{\circ}D$ and on $S_2{\circ}T_2{\circ}D$. The main difficulty in proving differential privacy is that $H_1$ and $H_2$ can significantly differ. We show, however, that the distribution on relabeled databases $(S{\circ}T)^h$ generated in Step \[step:Onerelabel\] of the two executions are similar in the sense that for each relabeled database in one of the distributions there exist one or two databases in the other s.t. (1) all these databases have, roughly, the same probability, and (2) they differ on at most one entry. Thus, executing the differentially private algorithm ${\mathcal A}$ on $(S{\circ}T)^h {\circ} D$ preserves differential privacy. We now make this argument formal.
Note that $|P_1\setminus P_2|\in\{0,1\}$, and let ${\mathsf p}_1$ be the element in $P_1\setminus P_2$ if such an element exists. If this is the case, then ${\mathsf p}_1$ appears exactly once in $S_1{\circ}T_1$. Similarly, let ${\mathsf p}_2$ be the element in $P_2 \setminus P_1$ if such an element exists. Let $K=P_1\cap P_2$, hence $P_i = K$ or $P_i = K \cup \{{\mathsf p}_i\}$. Therefore, $|\Pi_C(K)|\leq|\Pi_C(P_i)|\leq2|\Pi_C(K)|$. Thus, $|H_1|\leq2|H_2|$ and similarly $|H_2|\leq2|H_1|$.
More specifically, for every $\vec{z}\in\Pi_C(K)$ there are either one or two (but not more) hypotheses in $H_1$ that agree with $\vec{z}$ on $K$. We denote these one or two hypotheses by $h_{1,\vec{z}}$ and $h'_{1,\vec{z}}$, which may be identical if only one unique hypothesis exists. Similarly, we denote $h_{2,\vec{z}}$ and $h'_{2,\vec{z}}$ as the hypotheses corresponding to $H_2$. For every $\vec{z}\in\Pi_C(K)$ we have that $|q(S_i,h_{i,\vec{z}})-q(S_i,h'_{i,\vec{z}})|\leq1$ because if $h_{i,\vec{z}}=h'_{i,\vec{z}}$ then the difference is clearly zero and otherwise they differ only on ${\mathsf p}_i$, which appears at most once in $S_i$. Moreover, for every $\vec{z}\in\Pi_C(K)$ we have that $|q(S_1,h_{1,\vec{z}})-q(S_2,h_{2,\vec{z}})|\leq1$ because $h_{1,\vec{z}}$ and $h_{2,\vec{z}}$ disagree on at most two points ${\mathsf p}_1, {\mathsf p}_2$ such that at most one of them appears in $S_1$ and at most one of them appears in $S_2$. The same is true for every pair in $\{ h_{1,\vec{z}}, h'_{1,\vec{z}} \} \times \{ h_{2,\vec{z}} , h'_{2,\vec{z}} \}$.
Let $w_{i,\vec{z}}$ be the probability that the exponential mechanism chooses $h_{i,\vec{z}}$ or $h'_{i,\vec{z}}$ in Step \[step:Oneexpmech\] of the execution on $S_i{\circ}T_i{\circ}D$. We get that for every $\vec{z}\in\Pi_C(K)$, $$\begin{aligned}
w_{1,\vec{z}}
&\leq& \frac{\exp(\frac{1}{2}\cdot q(S_1,h_{1,\vec{z}}))+\exp(\frac{1}{2}\cdot q(S_1,h'_{1,\vec{z}}))}{\sum_{f\in H_1}{\exp(\frac{1}{2}\cdot q(S_1,f))}}\\
&\leq& \frac{\exp(\frac{1}{2}\cdot q(S_1,h_{1,\vec{z}}))+\exp(\frac{1}{2}\cdot q(S_1,h'_{1,\vec{z}}))}{\sum_{\vec{r}\in\Pi_C(K)}{\exp(\frac{1}{2}\cdot q(S_1,h_{1,\vec{r}}))}}\\
&\leq& \frac{\exp(\frac{1}{2}\cdot [q(S_2,h_{2,\vec{z}})+1])+\exp(\frac{1}{2}\cdot [q(S_2,h'_{2,\vec{z}})+1])}{\frac{1}{2}\sum\limits_{\vec{r}\in\Pi_C(K)}\left({\exp(\frac{q(S_2,h_{2,\vec{r}})-1}{2})+\exp(\frac{q(S_2,h'_{2,\vec{r}})-1}{2})}\right)}\\
&\leq& 2 e\cdot \frac{\exp(\frac{1}{2}\cdot [q(S_2,h_{2,\vec{z}})])+\exp(\frac{1}{2}\cdot [q(S_2,h'_{2,\vec{z}})])}{\sum_{f\in H_2}{\exp(\frac{1}{2}\cdot q(S_2,f))}}\\
&\leq& 4e \cdot
w_{2,\vec{z}}.
$$
We can now conclude the proof by noting that for every $\vec{z}\in\Pi_C(K)$ the databases $(S_1{\circ}T_1)^{h_{1,\vec{z}}}$ and $(S_2{\circ}T_2)^{h_{2,\vec{z}}}$ are neighboring, and, therefore, $(S_1{\circ}T_1)^{h_{1,\vec{z}}} {\circ}D$ and $(S_2{\circ}T_2)^{h_{2,\vec{z}}} {\circ}D$ are neighboring. For every $\vec{z}\in\Pi_C(K)$, let ${\mathsf h}_{i,\vec{z}}$ denote the event that the exponential mechanism chooses $h_{i,\vec{z}}$ or $h'_{i,\vec{z}}$ in Step \[step:Oneexpmech\] of the execution on $S_i{\circ}T_i{\circ}D$. By the privacy properties of algorithm ${\mathcal A}$ we have that for any set $F$ of possible outputs of algorithm ${\mathcal B}$ $$\begin{aligned}
\Pr[{\mathcal B}\left( S_1{\circ}T_1{\circ}D \right)\in F]&=& \sum_{\vec{z}\in\Pi_C(K)} w_{1,\vec{z}}
\cdot\Pr\left[{\mathcal A}\left( (S_1{\circ}T_1)^h {\circ}D \right) \in F \Big| {\mathsf h}_{1,\vec{z}} \right]\\
&\leq& \sum_{\vec{z}\in\Pi_C(K)}4e \; w_{2,\vec{z}}
\left(e^{\epsilon}\Pr\left[{\mathcal A}\left( (S_2{\circ}T_2)^h {\circ}D \right) \in F \Big| {\mathsf h}_{2,\vec{z}} \right]+\delta\right)\\
&\leq& e^{\epsilon+3}\cdot\Pr[{\mathcal B}\left( S_2{\circ}T_2{\circ}D \right)\in F]+4e\delta. \end{aligned}$$
Consider an execution of $LabelBoostProcedure$ on a database $S{\circ}T{\circ}D$, and assume that the examples in $S$ are labeled by some target concept $c\in C$. Recall that for every possible labeling $\vec{z}$ of the elements in $S$ and in $T$, algorithm $LabelBoostProcedure$ adds to $H$ a hypothesis from $C$ that agrees with $\vec{z}$. In particular, $H$ contains a hypothesis that agrees with the target concept $c$ on $S$ (and on $T$). That is, $\exists f\in H$ s.t. ${{\rm error}}_S(f)=0$. Hence, the exponential mechanism (on Step \[step:Oneexpmech\]) chooses (w.h.p.) a hypothesis $h\in H$ s.t. ${{\rm error}}_S(h)$ is small, provided that $|S|$ is roughly $\log|H|$, which is roughly ${{\rm VC}}(C)\cdot\log(|S|+|T|)$ by Sauer’s lemma. So, algorithm $LabelBoostProcedure$ takes an input database where only a small portion of it is labeled, and returns a similar database in which the labeled portion grows exponentially.
\[claim:LabelBoostProcedureUtility\] Fix $\alpha$ and $\beta$, and let $S{\circ}T{\circ}D$ be s.t. $S$ is labeled by some target concept $c\in C$, and s.t. $$|T|\leq\frac{\beta}{e} {{\rm VC}}(C)\exp(\frac{\alpha |S|}{2{{\rm VC}}(C)})-|S|.$$ Consider the execution of $LabelBoostProcedure$ on $S{\circ}T{\circ}D$, and let $h$ denote the hypothesis chosen on Step \[step:Oneexpmech\]. With probability at least $(1-\beta)$ we have that ${{\rm error}}_{S}(h)\leq\alpha$.
Note that by Sauer’s lemma, $$\begin{aligned}
|H|&=&|\Pi_C(P)| \leq \left(\frac{e|P|}{{{\rm VC}}(C)}\right)^{{{\rm VC}}(C)}\\
&\leq&\left(\frac{e(|T|+|S|)}{{{\rm VC}}(C)}\right)^{{{\rm VC}}(C)}\\
&\leq&\left(\beta\exp(\frac{\alpha |S|}{2{{\rm VC}}(C)})\right)^{{{\rm VC}}(C)}\\
&\leq&\beta\exp(\frac{\alpha |S|}{2}).\end{aligned}$$
For every $(z_1,\ldots,z_\ell)\in \Pi_C(P)$, algorithm $LabelBoostProcedure$ adds to $H$ a hypothesis $f$ s.t. $\forall 1\leq j \leq \ell,\;f(p_j)=z_j$. In particular, $H$ contains a hypothesis $f^*$ s.t. ${{\rm error}}_{S}(f^*)=0$. Hence, Proposition \[prop:expMech\] (properties of the exponential mechanism) ensures that the probability of the exponential mechanism choosing an $h$ s.t. ${{\rm error}}_{S}(h)>\alpha$ is at most $$|H|\cdot\exp(-\frac{\alpha |S|}{2})\leq\beta.$$
We next embed algorithm $LabelBoostProcedure$ in a wrapper algorithm, called $LabelBoost$, that iteratively applies $LabelBoostProcedure$ in order to enlarge the labeled portion of the database. Every such application deteriorates the privacy parameters, and hence, every iteration includes a sub-sampling step, which compensates for those privacy losses. In a nutshell, the learner $LabelBoost$ could be described as follows. It starts by training on the given labeled data. In each step, a part of the unlabeled points is labeled using the current hypothesis (previously labeled points are also relabeled); then the learner retrains using its own predictions as a (larger) labeled sample. Variants of this idea (known as self-training) have appeared in the literature for non-private learners (e.g., [@Scudder65; @Fralick67; @Agrawala70]). As we will see, in the context of [*private*]{} learners, this technique provably reduces the labeled sample complexity (while maintaining utility).
[**Setting:**]{} Algorithm ${\mathcal A}$ with (labeled and unlabeled) sample complexity $n$.\
[**Input:**]{} An unlabeled database $D\in X^{90000n}$ and a labeled database $S\in (X\times\{0,1\})^m$.
1. Set $i=1$.
2. \[step:while\] While $|S|<300n$:
1.
<!-- -->
1. Denote $\alpha_i=\frac{\alpha}{10\cdot2^i}$, and $\beta_i=\frac{\beta}{4\cdot2^i}$.
2. \[step:addElements\] Set $v{=}\min\hspace{-2pt}\left\{\hspace{-1pt}30000n \, , \, \beta_i {{\rm VC}}(C) e^{\frac{\alpha_i |S|}{200{{\rm VC}}(C)}}-|S|\hspace{-1pt}\right\}$. Let $T$ be the first $v$ elements of $D$, and remove $T$ from $D$. Fail if there are not enough elements in $D$.
1.
3. \[step:subsampling\] Delete (permanently) $\frac{99}{100}|T|$ random entries from $T$, and $\frac{99}{100}|S|$ random entries from $S$.
1.
4. \[step:procedure\] $S{\circ}T{\circ}D \leftarrow LabelBoostProcedure(S{\circ}T{\circ}D)$.
1.
5. \[step:uniteTS\] Add every element of $T$ to $S$.
6. Set $i=i+1$.
3. \[step:subsamplingFinal\] Delete $\frac{299}{300}|S|$ random entries from $S$.
1.
4. \[step:iidSampling\] Let $S'$ denote the outcome of $|S|$ i.i.d. samples from $S$.
1.
5. \[step:AAA\] Execute ${\mathcal A}$ on $S'$.
Before analyzing algorithm $LabelBoost$ we recall the sub-sampling technique from [@KLNRS08; @BBKN12].
\[claim:boostPrivacy\] Let ${\mathcal A}$ be an $(\epsilon^*,\delta)$-differentially private algorithm operating on databases of size $n$. Fix $\epsilon\leq1$, and denote $t=\frac{n}{\epsilon}(3+\exp(\epsilon^*))$. Construct an algorithm ${\mathcal B}$ that on input a database $D=(z_i)_{i=1}^t$ uniformly at random selects a subset $J\subseteq\{1,2,...,t\}$ of size $n$, and runs ${\mathcal A}$ on the multiset $D_J=(z_i)_{i\in J}$. Then, ${\mathcal B}$ is $\left(\epsilon,\frac{4\epsilon}{3+\exp(\epsilon^*)}\delta\right)$-differentially private.
In Claim \[claim:boostPrivacy\] we assume that ${\mathcal A}$ treats its input as a multiset. If this is not the case, then algorithm ${\mathcal B}$ should be modified to randomly shuffle the elements in $D_J$ before applying ${\mathcal A}$ on $D_j$.
Claim \[claim:boostPrivacy\] boosts privacy by selecting random elements from the database and ignoring the rest of the database. The intuition is simple: Fix two neighboring databases $D,D'$ differing (only) on their $i^\text{th}$ entry. If the $i^\text{th}$ entry is ignored (which happens with high probability), then the executions on $D$ and on $D'$ are the same (i.e., perfect privacy). Otherwise, $(\epsilon^*,\delta)$-privacy is preserved.
In algorithm $LabelBoost$ we apply the learner ${\mathcal A}$ on a database containing $n$ i.i.d. samples from the database $S$ (Step \[step:iidSampling\]). Consider two neighboring databases $D,D'$ differing on their $i^\text{th}$ entry. Unlike in Claim \[claim:boostPrivacy\], the risk is that this entry will appear several times in the database on which ${\mathcal A}$ is executed. As the next claim states, the affects on the privacy guarantees are small. The intuition is that the probability of the $i^\text{th}$ entry appearing “too many” times is negligible.
\[claim:iidSampling\] Let $\epsilon\leq1$ and ${\mathcal A}$ be an $(\epsilon,\delta)$-differentially private algorithm operating on databases of size $n$. Construct an algorithm ${\mathcal B}$ that on input a database $D=(z_i)_{i=1}^n$ applies ${\mathcal A}$ on a database $D'$ containing $n$ i.i.d. samples from $D$. Then, ${\mathcal B}$ is $\left(\ln(244),2467\delta\right)$-differentially private.
We next prove the privacy properties of algorithm $LabelBoost$.
\[lemma:LabelBoostPrivacy\] If ${\mathcal A}$ is $(1,\delta)$-differentially private, then $LabelBoost$ is $(1,41\delta)$-differentially private.
We think of the input of $LabelBoost$ as one database $B\in(X\times\{0,1,\bot\})^{90000n+m}$. Note that the number of iterations performed on neighboring databases is identical (determined by the parameters $\alpha,\beta,n,m$), and denote this number as $N$. Throughout the execution, random elements from the input database are deleted (on Step \[step:subsampling\]). Note however, that the size of the database at any moment throughout the execution does not depend on the database content (determined by the parameters $\alpha,\beta,n,m$). We denote the size of the database at the beginning of the $i^\text{th}$ iteration as $n(i)$, e.g., $n(1)=90000n+m$.
Let ${\mathcal L}_t$ denote an algorithm similar to $LabelBoost$, except that only the last $t$ iterations are performed. The input of ${\mathcal L}_t$ is a database in $(X\times\{0,1,\bot\})^{n(N-t+1)}$. We next show (by induction on $t$) that ${\mathcal L}_t$ is $(1,41\delta)$-differentially private. To this end, note that an execution of ${\mathcal L}_0$ consists of sub-sampling (as in Claim \[claim:boostPrivacy\]), i.i.d. sampling (as in Claim \[claim:iidSampling\]), and applying the $(1,\delta)$-private algorithm ${\mathcal A}$. By Claim \[claim:iidSampling\], steps \[step:iidSampling\]–\[step:AAA\] preserve $(\ln(244),2476)$-differential privacy, and, hence, by Claim \[claim:boostPrivacy\], we have that ${\mathcal L}_0$ is $(1,41\delta)$-differentially private.
Assume that ${\mathcal L}_{t-1}$ is $(1,41\delta)$-differentially private, and observe that ${\mathcal L}_t$ could be restated as an algorithm that first performs one iteration of algorithm $LabelBoost$ and then applies ${\mathcal L}_{t-1}$ on the databases $D,S$ as they are at the end of that iteration. Now fix two neighboring databases $B_1,B_2$ and consider the execution of ${\mathcal L}_t$ on $B_1$ and on $B_2$.
Let $S_1^b,T_1^b,D_1^b$ and $S_2^b,T_2^b,D_2^b$ be the databases $S,T,D$ after Step \[step:addElements\] of the first iteration of ${\mathcal L}_t$ on $B_1$ and on $B_2$ (note that $B_1=S_1^b{\circ} T_1^b{\circ} D_1^b$ and $B_2=S_2^b{\circ} T_2^b{\circ} D_2^b$). If $B_1$ and $B_2$ differ (only) on their last portion, denoted as $D_1^b,D_2^b$, then the execution of ${\mathcal L}_t$ on these neighboring inputs differs only in the execution of ${\mathcal L}_{t-1}$, and hence Inequality (\[eqn:diffPrivDef\]) (approximate differential privacy) follows from the privacy of ${\mathcal L}_{t-1}$. We, therefore, assume that $D_1^b=D_2^b$ (and that $S_1^b{\circ} T_1^b$ and $S_2^b{\circ} T_2^b$ differ in at most one entry). Now, note that an execution of ${\mathcal L}_t$ consists of sub-sampling (as in Claim \[claim:boostPrivacy\]), applying algorithm $LabelBoostProcedure$ on the inputs, and executing the $(1,41\delta)$-private algorithm ${\mathcal L}_{t-1}$. By Lemma \[lemma:TransformationPrivacy\] (privacy properties of $LabelBoostProcedure$), the application of ${\mathcal L}_{t-1}$ on top of $LabelBoostProcedure$ preserves $(4,446\delta)$-differential privacy, and, hence, by Claim \[claim:boostPrivacy\] (sub-sampling), we have that ${\mathcal L}_t$ is $(1,41\delta)$-differentially private.
Before proceeding with the utility analysis, we introduce to following notations.
#### Notation.
Consider the $i^{\text{th}}$ iteration of $LabelBoost$. We let $S_i^b,T_i^b$ and $S_i^c,T_i^c$ denote the elements $S,T$ as they are after Steps \[step:addElements\] and \[step:subsampling\], and let $h_i$ denote the the hypothesis $h$ chosen in the execution of $LabelBoostProcedure$ in the $i^{\text{th}}$ iteration.
\[obs:LabelBoostEmpiricalError\] In every iteration $i$, with probability at least $(1-\beta_i)$ we have that ${{\rm error}}_{S_i^c}(h_i)\leq\alpha_i$.
Follows from Claim \[claim:LabelBoostProcedureUtility\].
Let $LabelBoost$ be executed with a base learner with sample complexity $n$, and on databases $D,S$. If $|D|\geq90000n$, then $LabelBoost$ never fails on Step \[step:addElements\].
Denote the number of iterations throughout the execution as $N$. We need to show that $\sum_{i=1}^N T_i^b\leq90000n$. Clearly, $|T_N^b|,|T_{N-1}^b|\leq30000n$. Moreover, for every $1<i<N$ we have that $|T_i^b|\geq2|T_{i-1}^b|$. Hence, $$\sum_{i=1}^N T_i^b \leq 30000n+30000n\sum_{i=0}^{\infty}\frac{1}{2^i}=90000n.\qedhere$$
\[claim:LabelBoostDatabaseSize\] Fix $\alpha,\beta$. Let $LabelBoost$ be executed on a base learner with sample complexity $n$, and on databases $D,S$, where $|D|\geq90000n$ and $|S|\geq\frac{96000}{\alpha}{{\rm VC}}(C)\log(\frac{2240}{\alpha\beta})$. In every iteration $i$ $$|S_i^b|\geq\frac{4800}{\alpha_i}{{\rm VC}}(C)\log(\frac{14}{\alpha_i \beta_i}).$$
The proof is by induction on $i$. Note that the base case (for $i=1$) trivially holds, and assume that the claim holds for $i-1$. We have that $$\begin{aligned}
|S_i^b| &=& |S_{i-1}^c|+|T_{i-1}^c| = \frac{1}{100}(|S_{i-1}^b|+|T_{i-1}^b|)\\
&=& \frac{1}{100}\beta_{i-1}{{\rm VC}}(C)\exp\left(\frac{\alpha_{i-1}|S_{i-1}^b|}{200{{\rm VC}}(C)}\right)\\
&\geq& \frac{1}{100}\beta_{i-1}{{\rm VC}}(C)\exp\left(24 \log(\frac{14}{\alpha_{i-1} \beta_{i-1}})\right)\\
&\geq& \frac{1}{100}\beta_{i-1}{{\rm VC}}(C)\cdot\left(\frac{14}{\alpha_{i-1} \beta_{i-1}}\right)^{24}\\
&\geq& \frac{4800}{\alpha_i}{{\rm VC}}(C)\log(\frac{14}{\alpha_i \beta_i}).\end{aligned}$$
The above analysis could easily be strengthen to show that $|S_i^b|$ grows as an exponentiation tower in $i$. This implies that there are at most $O(\log^*n)$ iterations throughout the execution of $LabelBoost$ on a base learner ${\mathcal A}$ with sample complexity $n$.
\[claim:LabelBoostGeneralization\] Let $LabelBoost$ be executed on databases $D,S$ containing i.i.d. samples from a fixed distribution $\mu$, where the examples in $S$ are labeled by some fixed target concept $c\in C$, and $|S|\geq\frac{96000}{\alpha}{{\rm VC}}(C)\log(\frac{2240}{\alpha\beta})$. For every $i$, the probability that ${{\rm error}}_{\mu}(c,h_i)>10\sum_{j=1}^i\alpha_j$ is at most $2\sum_{j=1}^i\beta_j$.
The proof is by induction on $i$. Note that for $i=1$ we have that $S_1^c$ contains $\frac{48}{\alpha_1}{{\rm VC}}(C)\log(\frac{14}{\alpha_1 \beta_1})$ i.i.d. samples from $\mu$ that are labeled by the target concept $c$. By Observation \[obs:LabelBoostEmpiricalError\], with probability at least $(1-\beta_1)$, we have that ${{\rm error}}_{S_1^c}(h_1)\leq\alpha_1$. In that case, Theorem \[thm:VCconsistant\] (the VC dimension bound) states that with probability at least $(1-\beta_1)$ it holds that ${{\rm error}}_{\mu}(c,h_1)\leq10\alpha_1$.
Now assume that the claim holds for $(i-1)$, and consider the $i^{\text{th}}$ iteration. Note that $S_i^c$ contains i.i.d. samples from $\mu$ that are labeled by $h_{i-1}$. Moreover, by Claim \[claim:LabelBoostDatabaseSize\], we have that $|S_i^c|=\frac{1}{100}|S_i^b|\geq\frac{48}{\alpha_i}{{\rm VC}}(C)\log(\frac{14}{\alpha_i \beta_i})$. By Observation \[obs:LabelBoostEmpiricalError\], with probability at least $(1-\beta_i)$, we have that ${{\rm error}}_{S_i^c}(h_i)\leq\alpha_i$. If that is the case, Theorem \[thm:VCconsistant\] states that with probability at least $(1-\beta_i)$ it holds that ${{\rm error}}_{\mu}(h_{i-1},h_i)\leq10\alpha_i$. So, with probability at least $(1-2\beta_i)$ we have that ${{\rm error}}_{\mu}(h_{i-1},h_i)\leq10\alpha_i$. Using the inductive assumption, the probability that ${{\rm error}}_{\mu}(c,h_i)\leq{{\rm error}}_{\mu}(c,h_{i-1})+{{\rm error}}_{\mu}(h_{i-1},h_i)\leq10\sum_{j=1}^i\alpha_j$ is at least $(1-2\sum_{j=1}^i\beta_j)$.
\[lemma:LabelBoostUtility\] Fix $\alpha,\beta$. Applying $LabelBoost$ on an $(\alpha,\beta,n,n)$-SSL for a class $C$ results in an $(11\alpha,2\beta,O(n),m)$-SSL for $C$, where $m=O(\frac{1}{\alpha}{{\rm VC}}(C)\log(\frac{1}{\alpha\beta}))$.
Let $LabelBoost$ be executed on databases $D,S$ containing i.i.d. samples from a fixed distribution $\mu$, where $|D|\geq90000n$ and $|S|\geq\frac{96000}{\alpha}{{\rm VC}}(C)\log(\frac{2240}{\alpha\beta})$. Moreover, assume that the examples in $S$ are labeled by some fixed target concept $c\in C$.
Consider the last iteration of Algorithm $LabelBoost$ (say $i=N$) on these inputs. The intuition is that after the last iteration, when reaching Step \[step:iidSampling\], the database $S$ is big enough s.t. ${\mathcal A}$ returns (w.h.p.) a hypothesis with small error on $S$. This hypothesis also has small generalization error as $S$ is labeled by $h_N$ which is close to the target concept (by Claim \[claim:LabelBoostGeneralization\]).
Formally, let $S^3$ denote the database $S$ as it after Step \[step:subsamplingFinal\] of the execution, and let $h_{\text{fin}}$ denote the hypothesis returned by the base learner ${\mathcal A}$ on Step \[step:AAA\]. By the while condition on Step \[step:while\], we have that $|S^3|\geq n$. Hence, by the utility guarantees of the base learner ${\mathcal A}$, with probability at least $(1-\beta)$ we have that ${{\rm error}}_{S^3}(h_{\text{fin}})\leq\alpha$. As $|S^3|\geq\frac{1}{300}|S|\geq\frac{640}{\alpha}{{\rm VC}}(C)\log(\frac{4480}{\alpha\beta})$, and as $S^3$ contains i.i.d. samples from $\mu$ labeled by $h_N$, Theorem \[thm:VCconsistant\] states that with probability at least $(1-\frac{\beta}{2})$ it holds that ${{\rm error}}_{\mu}(h_{\text{fin}},h_N)\leq10\alpha$. By Claim \[claim:LabelBoostGeneralization\], with probability at least $(1-2\sum_{i=1}^N\beta_i)\geq(1-\frac{\beta}{2})$ it holds that ${{\rm error}}_{\mu}(c,h_N)\leq10\sum_{j=1}^N\alpha_i\leq\alpha$. All in all (using the triangle inequality), with probability at least $(1-2\beta)$ we get that ${{\rm error}}_{\mu}(c,h_{\text{fin}})\leq11\alpha$.
Combining Lemma \[lemma:LabelBoostPrivacy\] and Lemma \[lemma:LabelBoostUtility\] we get the following theorem.
\[thm:LabelBoost1\] Fix $\alpha,\beta,\delta$. Applying $LabelBoost$ on an $(\alpha,\beta,\epsilon{=}1,\delta,n,n)$-PSSL for a class $C$ results in an $(11\alpha,2\beta,\epsilon{=}1,41\delta,O(n),m)$-PSSL for $C$, where $m=O(\frac{1}{\alpha}{{\rm VC}}(C)\log(\frac{1}{\alpha\beta}))$.
Using Claim \[claim:boostPrivacy\] to boost the privacy guarantees of the learner resulting from Theorem \[thm:LabelBoost1\], proves Theorem \[thm:LabelBoost2\]:
\[thm:LabelBoost2\] There exists a constant $\lambda$ such that: For every $\alpha,\beta,\epsilon,\delta,n$, if there exists an $(\alpha,\beta,1,\delta,n,n)$-PSSL for a concept class $C$, then there exists an $(\lambda\alpha,\lambda\beta,\epsilon,\delta,O(\frac{n}{\epsilon}),m)$-PSSL for $C$, where $m=O(\frac{1}{\alpha\epsilon}{{\rm VC}}(C)\log(\frac{1}{\alpha\beta}))$.
Let ${\mathcal B}$ be the learner resulting from applying $LabelBoost$ on a learner ${\mathcal A}$. Then (1) If ${\mathcal A}$ preserves pure-privacy, then so does ${\mathcal B}$; and (2) If ${\mathcal A}$ is a proper-learner, then so is ${\mathcal B}$.
Algorithm $LabelBoost$ can also be used as an [*agnostic*]{} learner, where the target class $C$ is unknown, and the learner outputs a hypothesis out of a set $F\neq C$. Note that given a labeled sample, a consistent hypothesis might not exist in $F$. Minor changes in the proof of Theorem \[thm:LabelBoost2\] show the following theorem.
There exists a constant $\lambda$ such that: For every $\alpha,\beta,\epsilon,\delta,n$, if there exists an $(\alpha,\beta,1,\delta,n,n)$-PSSL for a concept class $F$, then there exists an $(\lambda\alpha,\lambda\beta,\epsilon,\delta,O(\frac{n}{\epsilon}),m)$-agnostic-PSSL using $F$, where $m=O(\frac{1}{\alpha^2\epsilon}{{\rm VC}}(F)\log(\frac{1}{\alpha\beta}))$.
To show the usefulness of Theorem \[thm:LabelBoost2\], we consider (a discrete version of) the class of all axis-aligned rectangles (or hyperrectangles) in $\ell$ dimensions. Formally, let $X_d^{\ell}=(\{0,1\}^d)^{\ell}$ denote a discrete ${\ell}$-dimensional domain, in which every axis consists of $2^d$ points. For every $\vec{a}=(a_1,\ldots,a_{\ell}),\vec{b}=(b_1,\ldots,b_{\ell})\in X_d^{\ell}$ define the concept $c_{[\vec{a},\vec{b}]}:X_d^{\ell}\rightarrow\{0,1\}$ where $c_{[\vec{a},\vec{b}]}(\vec{x})=1$ if and only if for every $1\leq i\leq {\ell}$ it holds that $a_i\leq x_i\leq b_i$. Define the concept class of all axis-aligned rectangles over $X^{\ell}_d$ as ${{\tt RECTANGLE}}_d^{\ell}=\{ c_{[\vec{a},\vec{b}]} \}_{\vec{a},\vec{b}\in X_d^{\ell}}$. The VC dimension of this class is $2{\ell}$, and, thus, it can be learned non-privately with (labeled and unlabeled) sample complexity $O_{\alpha,\beta}({\ell})$. The best currently known private PAC learner for this class [@BNS13b] has (labeled and unlabeled) sample complexity $\widetilde{O}_{\alpha,\beta,\epsilon,\delta}({\ell}^3 \cdot 8^{\log^*d})$. Using $LabelBoost$ with the construction of [@BNS13b] reduces the labeled sample complexity while maintaining the unlabeled sample complexity.
There exists a private semi-supervised learner for ${{\tt RECTANGLE}}_d^{\ell}$ with unlabeled sample complexity $\widetilde{O}_{\alpha,\beta,\epsilon,\delta}({\ell}^3 \cdot 8^{\log^*d})$ and labeled sample complexity $O_{\alpha,\beta,\epsilon}(\ell)$. The learner is efficient (runs in polynomial time) whenever the dimension $\ell$ is small enough (roughly, $\ell \leq \log^{\frac{1}{3}} d$).
The [*labeled*]{} sample complexity in Theorem \[thm:LabelBoost2\] has no dependency in $\delta$.[^8] It would be helpful if we could also reduce the dependency on $\epsilon$. As we will later see, this can be achieved in the active learning model.
#### $\mathbf{LabelBoost}$ vs. $\mathbf{GenericLearner}$.
While both constructions result in learners with labeled sample complexity proportional to the VC dimension, they differ on their unlabeled sample complexity.
Recall the generic construction of Kasiviswanathan et al. [@KLNRS08] for private PAC learners, in which the (labeled and unlabeled) sample complexity is logarithmic in the size of the target concept class $C$ (better constructions are known for many specific cases). Using Algorithm $LabelBoost$ with their generic construction results in a private semi-supervised learner with unlabeled sample complexity (roughly) $\log|C|$, which is better than the bound achieved by $GenericLearner$ (whose unlabeled sample complexity is $O(\log|X|\cdot{{\rm VC}}(C))$). In cases where a sample-efficient private-PAC learner is known, applying $LabelBoost$ would give even better bounds.
Another difference is that (a direct use of) $GenericLearner$ only yields pure-private proper-learners, whereas $LabelBoost$ could be applied to every private learner (proper or improper, preserving pure or approximated privacy). To emphasize this difference, recall that the sample complexity of pure-private improper-PAC-learners is characterized by the Representation Dimension [@BNS13].
For every concept class $C$ there is a pure-private semi-supervised improper-learner with labeled sample complexity $O_{\alpha,\beta,\epsilon}({{\rm VC}}(C))$ and unlabeled sample complexity $O_{\alpha,\beta,\epsilon}({{\rm RepDim}}(C)) $.
Private Active Learners {#sec:privActive}
=======================
Semi-supervised learners are a subset of the larger family of active learners. Such learners can adaptively request to reveal the labels of specific examples. An active learner is given access to a pool of $n$ unlabeled examples, and adaptively chooses to label $m$ examples.
\[def:AL\] Let $C$ be a concept class over a domain $X$. Let ${\mathcal A}$ be an interactive (stateful) algorithm that holds an initial input database $D=(x_i)_{i=1}^n\in(X)^n$. For at most $m$ rounds, algorithm ${\mathcal A}$ outputs an index $i\in\{1,2,\ldots,n\}$ and receives an answer $y_i\in\{0,1\}$. Afterwards, algorithm ${\mathcal A}$ outputs a hypothesis $h$, and terminates.
Algorithm ${\mathcal A}$ is an [*$(\alpha,\beta,n,m)$-AL (Active learner)*]{} for $C$ if for all concepts $c\in C$ and all distributions $\mu$ on $X$: If ${\mathcal A}$ is initiated on an input $D=(x_i)_{i=1}^n$, where each $x_i$ is drawn i.i.d. from $\mu$, and if every index $i$ queried by ${\mathcal A}$ is answered by $y_i=c(x_i)$, then algorithm ${\mathcal A}$ outputs a hypothesis $h$ satisfying $\Pr[{{\rm error}}_{\mu}(c,h) \leq \alpha] \geq 1-\beta.$ The probability is taken over the random choice of the samples from $\mu$ and the coin tosses of the learner ${\mathcal A}$.
In the standard definition of active learners, the learners specify examples by their value (whereas in Definition \[def:AL\] the learner queries the labels of examples by their index). E.g., if $x_5=x_9=p$ then instead of asking for the label of $p$, algorithm ${\mathcal A}$ asks for the label example 5 (or 9). This deviation from the standard definition is because when privacy is introduced, every entry in $D$ corresponds to a single individual, and can be changed arbitrarily (and regardless of the other entries).
\[def:PAL\] An algorithm ${\mathcal A}$ is an [*$(\alpha,\beta,\epsilon,\delta,n,m)$-PAL*]{} (Private Active Learner) for a concept class $C$ if Algorithm ${\mathcal A}$ is an $(\alpha,\beta,n,m)$-active learner for $C$ and ${\mathcal A}$ is $(\epsilon,\delta)$-differentially private, where in the definition of privacy we consider the input of ${\mathcal A}$ to be a fully labeled sample $S=(x_i,y_i)_{i=1}^n\in(X\times\{0,1\})^n$ (and limit the number of labels $y_i$ it can access to $m$).
Note that the queries that an active learner makes depend on individuals’ data. Hence, if the indices that are queried are exposed, they may breach privacy. An example of how such an exposure may occur is a medical research of a new disease – a hospital may posses background information about individuals and hence can access a large pool of unlabeled examples, but to label an example an actual medical test is needed. Partial information about the labeling queries would hence be leaked to the tested individuals. More information about the queries may be leaked to an observer of the testing site. The following definition remedies this potential breach of privacy.
\[def:transcriptPrivate\] We define the transcript in an execution of an active learner ${\mathcal A}$ as the ordered sequence $L=(\ell_i)_{i=1}^m\in\{1,2,\ldots,n\}^m$ of indices that ${\mathcal A}$ outputs throughout the execution. We say that a learner ${\mathcal A}$ is $(\epsilon,\delta)$-transcript-differentially private if the algorithm whose input is the labeled sample and whose output is the output of ${\mathcal A}$ together with the transcript of the execution is $(\epsilon,\delta)$-differentially private. An algorithm ${\mathcal A}$ is an [*$(\alpha,\beta,\epsilon,\delta,n,m)$-TPAL (transcript-private active-learner)*]{} for a concept class $C$ if Algorithm ${\mathcal A}$ is an $(\alpha,\beta,n,m)$-Active learner for $C$ and ${\mathcal A}$ is $(\epsilon,\delta)$-transcript-differentially private.
Recall that a semi-supervised learner has no control over which of its examples are labeled, and the indices of the labeled examples are publicly known. Hence, a private semi-supervised learner is, in particular, a transcript-private active learner.
If ${\mathcal A}$ is an $(\alpha,\beta,\epsilon,\delta,n,m)$-PSSL, then ${\mathcal A}$ is an $(\alpha,\beta,\epsilon,\delta,n,m)$-TPAL.
In particular, our algorithms from Sections \[sec:semiSuper\] and \[sec:boost\] satisfy Definition \[def:transcriptPrivate\], suggesting that the strong privacy guarantees of Definition \[def:transcriptPrivate\] are achievable. However, as we will now see, this comes with a price. The work on (non-private) active learning has mainly focused on reducing the dependency of the labeled sample complexity in $\alpha$ (the approximation parameter). The classic result in this regime states that the labeled sample complexity of learning ${{\tt THRESH}}_d$ without privacy is $O(\log(\frac{1}{\alpha}))$, exhibiting an exponential improvement over the $\Omega(\frac{1}{\alpha})$ labeled sample complexity in the non-active model. As the next theorem states, the labeled sample complexity of every transcript-private active-learner for ${{\tt THRESH}}_d$ is lower bounded by $\Omega(\frac{1}{\alpha})$.
\[thm:TPALlowerBound\] Let $\alpha\leq\frac{1}{9}$ and $\beta\leq\frac{1}{4}$. In every $(\alpha,\beta,\epsilon,\delta,n,m)$-TPAL for ${{\tt THRESH}}_d$ the labeled sample complexity satisfies $m=\Omega\left(\frac{1}{\alpha}\right)$.
Let ${\mathcal A}$ be an $(\alpha,\beta,\epsilon,\delta,n,m)$-TPAL for ${{\tt THRESH}}_d$ with $\alpha\leq1/9$ and $\beta\leq1/4$. Without loss of generality, we can assume that $n\geq\frac{100}{\alpha^2}\ln(\frac{1}{\alpha\beta})$ (since ${\mathcal A}$ can ignore part of the sample). Denote $B=\{1,2,\ldots,8\alpha 2^d\}\subseteq X_d$, and consider the following thought experiment for randomly generating a labeled sample of size $n$.
The above process induces a distribution on labeled samples of size $n$, denoted as ${\mathcal P}$. Let $S\sim{\mathcal P}$, and consider the execution of ${\mathcal A}$ on $S$. Recall that ${\mathcal A}$ operates on the unlabeled portion of $S$ and actively queries for labels. Let $b$ denote the the number of elements from $B$ in the database $S$. Standard arguments in learning theory (see Theorem \[thm:generalization\]) state that with all but probability $\beta\leq\frac{1}{4}$ it holds that $7\alpha n\leq b\leq9\alpha n$. We continue with the proof assuming that this is the case. We first show that ${\mathcal A}$ must (w.h.p.) ask for the label of at least one example in $B$. To this end, note that even given the labels of all $x\notin B$, the target concept is distributed uniformly on $B$, and the probability that ${\mathcal A}$ fails to output an $\alpha$-good hypothesis is at least $\frac{3}{4}$. Hence,
$$\begin{aligned}
\beta&\geq&\Pr_{S,{\mathcal A}}[{\mathcal A}\text{ fails}]\\
&\geq&\Pr_{S,{\mathcal A}}\left[ \begin{array}{c}
{\mathcal A}\text{ does not ask for the label}\\
\text{of any point in } B \text{ and fails}
\end{array} \right]\\
&=&\Pr_{S,{\mathcal A}}\left[ \begin{array}{c}
{\mathcal A}\text{ does not ask for the}\\
\text{label of any point in } B
\end{array} \right]
\cdot
\Pr_{S,{\mathcal A}}\left[ {\mathcal A}\text{ fails} \middle\vert
\begin{array}{c}
{\mathcal A}\text{ does not ask for the}\\
\text{label of any point in } B
\end{array} \right]\\
&\geq&\Pr_{S,{\mathcal A}}\left[ \begin{array}{c}
{\mathcal A}\text{ does not ask for the}\\
\text{label of any point in } B
\end{array} \right]
\cdot
\frac{3}{4}\\
&\geq&\Pr_{S}[b\leq9\alpha n]\cdot\Pr_{S,{\mathcal A}}\left[ \begin{array}{c}
{\mathcal A}\text{ does not ask for the}\\
\text{label of any point in } B
\end{array} \middle\vert b\leq9\alpha n \right]
\cdot
\frac{3}{4}\\
&\geq&\frac{9}{16}\cdot\Pr_{S,{\mathcal A}}\left[ \begin{array}{c}
{\mathcal A}\text{ does not ask for the}\\
\text{label of any point in } B
\end{array} \middle\vert b\leq9\alpha n \right].\end{aligned}$$
Thus, assuming that $b\leq9\alpha n$, the probability that ${\mathcal A}$ asks for the label of a point in $B$ is at least $(1-\frac{16}{9}\beta)$. Now choose a random $x^*$ from $S$ s.t. $x^*\in B$. Note that $$\begin{aligned}
\Pr_{S,x^*,{\mathcal A}}\left[ {\mathcal A}(S) \text{ asks for the label of } x^* \right]
&\geq&\Pr_{S}[b\leq9\alpha n]\cdot\Pr_{S,x^*,{\mathcal A}}\left[
\begin{array}{c}
{\mathcal A}(S) \text{ asks for}\\
\text{the label of } x^*
\end{array}
\middle\vert
b\leq9\alpha n\right]\\
&\geq&(1-\beta)\cdot\frac{(1-\frac{16}{9}\beta)}{9\alpha n}\\
&\geq&\frac{1-\frac{25}{9}\beta}{9\alpha n}.\end{aligned}$$
Choose a random $\hat{x}$ from $S$ (uniformly), and construct a labeled sample $S'$ by swapping the entries $(x^*,c(x^*))$ and $(\hat{x},c(\hat{x}))$ in $S$. Note that $S'$ is also distributed according to ${\mathcal P}$, and that $\hat{x}$ is a uniformly random element of $S'$. Therefore, $$\Pr_{S,x^*,\hat{x},{\mathcal A}}\left[ {\mathcal A}(S') \text{ asks for the label of } \hat{x} \right]\leq\frac{m}{n}.$$ As $S$ and $S'$ differ in at most 2 entries, differential privacy states that $$\begin{aligned}
\frac{m}{n} &\geq& \Pr_{S,x^*,\hat{x},{\mathcal A}}\left[ {\mathcal A}(S') \text{ asks for the label of } \hat{x} \right]\\
&=& \sum_{S,x^*,\hat{x}}\Pr[S,x^*,\hat{x}]\cdot\Pr_{{\mathcal A}}\left[ {\mathcal A}(S') \text{ asks for the label of } \hat{x} \right]\\
&\geq& \sum_{S,x^*,\hat{x}}\Pr[S,x^*,\hat{x}] \; e^{-2\epsilon} \; \Pr_{{\mathcal A}}\left[ {\mathcal A}(S) \text{ asks for the label of } x^* \right]-\delta(1{+}e^{-\epsilon})\\
&=& e^{-2\epsilon}\cdot\Pr_{S,x^*,{\mathcal A}}\left[ {\mathcal A}(S) \text{ asks for the label of } x^* \right]-\delta(1+e^{-\epsilon})\\
&\geq& e^{-2\epsilon}\cdot\frac{1-\frac{25}{9}\beta}{9\alpha n}-\delta(1+e^{-\epsilon}).\end{aligned}$$ Solving for $m$, this yields $m=\Omega(\frac{1}{\alpha})$.
The private active learners presented in [@BF13] as well as the algorithm described in the next section only satisfy the weaker Definition \[def:PAL\].
Removing the Dependency on the Privacy Parameters
-------------------------------------------------
We next show how to transform a semi-supervised private learner ${\mathcal A}$ into an active learner ${\mathcal B}$ with better privacy guarantees without increasing the labeled sample complexity. Algorithm ${\mathcal B}$, on input an unlabeled database $D$, randomly chooses a subset of the inputs $D'\subseteq D$ and asks for the labels of the examples in $D'$ (denote the resulting labeled database as $S$). Algorithm ${\mathcal B}$ then applies ${\mathcal A}$ on $D,S$. As the next claim states, this eliminates the $\frac{1}{\epsilon}$ factor from the labeled sample complexity as the (perhaps adversarial) choice for the input database is independent of the queries chosen.
\[claim:activeBoostPrivacy\] If there exists an $(\alpha,\beta,\epsilon^*,\delta,n,m)$-PSSL for a concept class $C$, then for every $\epsilon$ there exists an $\left(\alpha,\beta,\epsilon,\frac{7+e^{\epsilon^*}}{3+e^{2\epsilon^*}}\epsilon\delta,t,m\right)$-PAL (private active learner) for $C$, where $t=\frac{n}{\epsilon}(3+\exp(2\epsilon^*))$.
[**Inputs:**]{} Base learner ${\mathcal A}$, privacy parameters $\epsilon^*,\epsilon$, and a database $D=(x_i)_{i=1}^t$ of $t$ unlabeled examples.
1. Uniformly at random select a subset $J\subseteq\{1,2,...,t\}$ of size $n$, and let $K\subseteq J$ denote the smallest $m$ indices in $J$.
2. Request the label of every index $i\in K$, and let $\{y_i\;:\;i\in K\}$ denote the received answers.
3. Run ${\mathcal A}$ an the multiset $D_J=\{ (x_i,\bot) : i\in J\setminus K \}\cup\{ (x_i,y_i) : i \in K \}$.
The proof is via the construction of Algorithm $SubSampling$ (Algorithm \[alg:algWrapper\]). The utility analysis is straight forward. Fix a target concept $c$ and a distribution $\mu$. Assume that $D$ contains $t$ i.i.d. samples from $\mu$ and that every query on an index $i$ is answered by $c(x_i)$. Therefore, algorithm ${\mathcal A}$ is executed on a multiset $D_J$ containing $n$ i.i.d. samples from $\mu$ where $m$ of those samples are labeled by $c$. By the utility properties of ${\mathcal A}$, an $\alpha$-good hypothesis is returned with probability at least $(1-\beta)$.
For the privacy analysis, fix two neighboring databases $S,S'\in(X\times\{0,1\})^t$ differing on their $i^{\text th}$ entry, and let $D,D'\in X^t$ denote the restriction of those two databases to $X$ (that is, $D$ contains an entry $x$ for every entry $(x,y)$ in $S$). Consider an execution of $SubSampling$ on $D$ (and on $D'$), and let $J\subseteq\{1,\ldots,t\}$ denote the random subset of size $n$ chosen on Step 1. Moreover, and let $D_J$ denote the multiset on which ${\mathcal A}$ in executed.
Since $S$ and $S'$ differ in just the $i^{\text th}$ entry, for any set of outcomes $F$ it holds that $\Pr[{\mathcal A}(D_J) \in F | i \not\in J] = \Pr[{\mathcal A}(D'_J) \in F | i \notin J]$. When $i\in J$ we have that $$\begin{aligned}
\Pr[SubSampling(D) \in F \wedge i\in J]
&=&\sum_{\begin{array}{c} {\scriptstyle R \subseteq [t]\setminus \{i\}}\\{\scriptstyle |R|=n-1}\end{array}}\hspace{-7pt}\Pr[J=R\cup\{i\}]\cdot\Pr[{\mathcal A}(D_J)\in F | J = R\cup \{i\}].\end{aligned}$$ Note that for every choice of $R \subseteq [t]\setminus \{i\}$ s.t. $|R|=(n-1)$, there are exactly $(t-n)$ choices for $Q \subseteq [t]\setminus \{i\}$ s.t. $|Q|=n$ and $R\subseteq Q$. Hence, $$\begin{aligned}
\Pr[SubSampling(D) \in F \wedge i\in J] &=&
\sum_{\begin{array}{c} {\scriptstyle R \subseteq [t]\setminus \{i\}}\\{\scriptstyle |R|=n-1}\end{array}} \hspace{-10pt} \frac{1}{t-n} \hspace{-10pt} \sum_{\begin{array}{c} {\scriptstyle Q \subseteq [t]\setminus \{i\}}\\{\scriptstyle |Q|=n}\\{\scriptstyle R\subseteq Q}\end{array}} \hspace{-13pt} \Pr[J{=}R{\cup}\{i\}]{\cdot}\Pr[{\mathcal A}(D_J){\in} F | J {=} R{\cup} \{i\}] \\
&\leq& \sum_{\begin{array}{c} \vspace{-6pt}\\{\scriptstyle R \subseteq [t]\setminus \{i\}}\\{\scriptstyle |R|=n-1}\end{array}} \hspace{-15pt} \frac{1}{t-n} \hspace{-15pt} \sum_{\begin{array}{c} \vspace{-6pt}\\{\scriptstyle Q \subseteq [t]\setminus \{i\}}\\{\scriptstyle |Q|=n}\\{\scriptstyle R\subseteq Q}\end{array}} \hspace{-15pt} \Pr[J{=}Q]\left({\mathrm{e}}^{2\epsilon^*}\Pr[{\mathcal A}(D_J){\in} F | J {=} Q]{+}\delta{+}\delta e^{\epsilon^*}\right).\end{aligned}$$ For the last inequality, note that $D_Q$ and $D_{R\cup\{i\}}$ differ in at most two entries, as they differ in one unlabeled example, and possibly one other example that is labeled in one multiset and unlabeled on the other. Now note that every choice of $Q$ will appear in the above sum exactly $n$ times (as the number of choices for appropriate $R$’s s.t. $R\subseteq Q$). Hence, $$\begin{aligned}
\Pr\left[\{SubSampling(D) \in F \} \wedge \{ i\in J \}\right]
&\leq&\frac{n}{t-n}\hspace{-10pt}\sum_{\begin{array}{c} {\scriptstyle Q \subseteq [t]\setminus \{i\}}\\{\scriptstyle |Q|=n}\end{array}}\hspace{-12pt}\Pr[J{=}Q]\left({\mathrm{e}}^{2\epsilon^*}\Pr[{\mathcal A}(D_J){\in} F | J {=} Q]{+}\delta{+}\delta e^{\epsilon^*}\right)\\
&=&\frac{n}{t-n}\cdot\Pr[i\notin J]\left(
e^{2\epsilon^*}\Pr[{\mathcal A}(D_J)\in F | i\notin J] {+} \delta {+} \delta e^{\epsilon^*}
\right)\\
&=&\frac{n}{t}e^{2\epsilon^*}\cdot\Pr[{\mathcal A}(D_J)\in F | i\notin J] + \frac{n}{t}(1+e^{\epsilon^*})\delta\\
&=&\frac{n}{t}e^{2\epsilon^*}\cdot\Pr[{\mathcal A}(D'_J)\in F | i\notin J] + \frac{n}{t}(1+e^{\epsilon^*})\delta.\end{aligned}$$ Therefore, $$\begin{aligned}
\Pr[SubSampling(D) \in F]
&=&\Pr\left[\{SubSampling {\in} F\} {\wedge} \{i{\in} J\}\right] {+} \Pr[i{\notin} J]{\cdot}\Pr[{\mathcal A}(D'_J){\in} F | i{\notin} J]\\
&\leq&\left(\frac{n}{t}e^{2\epsilon^*}+\frac{t-n}{t}\right)\cdot\Pr[{\mathcal A}(D'_J)\in F | i\notin J] + \frac{n}{t}(1+e^{\epsilon^*})\delta.\end{aligned}$$ Similar arguments show that $$\begin{aligned}
\Pr[SubSampling(D') \in F]
&\geq& \left(\frac{n}{t}e^{-2\epsilon^*}+\frac{t-n}{t}\right)\cdot\Pr[{\mathcal A}(D'_J)\in F | i\notin J] - \frac{n}{t}2\delta.\end{aligned}$$ For $t\geq\frac{n}{\epsilon}(3+\exp(2\epsilon^*))$, this yields $$\begin{aligned}
&&\hspace{-25pt}\Pr[SubSampling(D) \in F]\\
&&\hspace{-20pt} \leq e^{\epsilon}\cdot\Pr[SubSampling(D') \in F]+\frac{7+e^{\epsilon^*}}{3+e^{2\epsilon^*}}\epsilon\delta.\end{aligned}$$
The transformation of Claim \[claim:activeBoostPrivacy\] preserves the efficiency of the base (non-active) learner. Hence, a given (efficient) non-active private learner could always be transformed into an (efficient) active private learner whose labeled sample complexity does not depend on $\epsilon$. Applying Claim \[claim:activeBoostPrivacy\] to the learner from Theorem \[thm:LabelBoost2\] result in the following theorem, showing that the labeled sample complexity of private active learners has no dependency in the privacy parameters $\epsilon$ and $\delta$.
There exists a constant $\lambda$ such that: For every $\alpha,\beta,\epsilon,\delta,n$, if there exists an $(\alpha,\beta,1,\delta,n,n)$-PSSL for a concept class $C$, then there exists an $(\lambda\alpha,\lambda\beta,\epsilon,\delta,O(\frac{n}{\epsilon}),m)$-PAL for $C$, where $m=O(\frac{1}{\alpha}{{\rm VC}}(C)\log(\frac{1}{\alpha\beta}))$.
#### [**Acknowledgments.**]{}
We thank Aryeh Kontorovich, Adam Smith, and Salil Vadhan for helpful discussions of ideas in this work.
[10]{}
A. Agrawala. Learning with a probabilistic teacher. , 16(4):373–379, Jul 1970.
Martin Anthony and John Shawe-Taylor. A result of [V]{}apnik with applications. , 47(3):207–217, 1993.
Matin Anthony and Peter L. Bartlett. . Cambridge University Press, 2009.
Maria-Florina Balcan and Vitaly Feldman. Statistical active learning algorithms. In [*Advances in Neural Information Processing Systems 26*]{}, pages 1295–1303, 2013.
Amos Beimel, Hai Brenner, Shiva Prasad Kasiviswanathan, and Kobbi Nissim. Bounds on the sample complexity for private learning and private data release. , 94(3):401–437, 2014.
Amos Beimel, Kobbi Nissim, and Uri Stemmer. Characterizing the sample complexity of private learners. In Robert D. Kleinberg, editor, [*ITCS*]{}, pages 97–110. ACM, 2013.
Amos Beimel, Kobbi Nissim, and Uri Stemmer. Private learning and sanitization: Pure vs. approximate differential privacy. In Prasad Raghavendra, Sofya Raskhodnikova, Klaus Jansen, and Jos[é]{} D. P. Rolim, editors, [*APPROX-RANDOM*]{}, volume 8096 of [ *Lecture Notes in Computer Science*]{}, pages 363–378. Springer, 2013.
Avrim Blum, Cynthia Dwork, Frank McSherry, and Kobbi Nissim. Practical privacy: The [SuLQ]{} framework. In Chen Li, editor, [*PODS*]{}, pages 128–138. ACM, 2005.
Avrim Blum, Katrina Ligett, and Aaron Roth. A learning theory approach to noninteractive database privacy. , 60(2):12, 2013.
Anselm Blumer, Andrzej Ehrenfeucht, David Haussler, and Manfred K. Warmuth. Learnability and the vapnik-chervonenkis dimension. , 36(4):929–965, 1989.
Mark Bun, Kobbi Nissim, Uri Stemmer, and Salil P. Vadhan. Differentially private release and learning of threshold functions. , abs/1504.07553, 2015.
Kamalika Chaudhuri and Daniel Hsu. Sample complexity bounds for differentially private learning. In Sham M. Kakade and Ulrike von Luxburg, editors, [*COLT*]{}, volume 19 of [*JMLR Proceedings*]{}, pages 155–186. JMLR.org, 2011.
Kamalika Chaudhuri and Claire Monteleoni. Privacy-preserving logistic regression. In Daphne Koller, Dale Schuurmans, Yoshua Bengio, and L[é]{}on Bottou, editors, [*NIPS*]{}. MIT Press, 2008.
Kamalika Chaudhuri, Claire Monteleoni, and Anand D. Sarwate. Differentially private empirical risk minimization. , 12:1069–1109, July 2011.
Cynthia Dwork, Krishnaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In Serge Vaudenay, editor, [*EUROCRYPT*]{}, volume 4004 of [ *Lecture Notes in Computer Science*]{}, pages 486–503. Springer, 2006.
Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Shai Halevi and Tal Rabin, editors, [*TCC*]{}, volume 3876 of [ *Lecture Notes in Computer Science*]{}, pages 265–284. Springer, 2006.
Cynthia Dwork, Guy N. Rothblum, and Salil P. Vadhan. Boosting and differential privacy. In [*FOCS*]{}, pages 51–60. IEEE Computer Society, 2010.
Andrzej Ehrenfeucht, David Haussler, Michael J. Kearns, and Leslie G. Valiant. A general lower bound on the number of examples needed for learning. , 82(3):247–261, 1989.
Vitaly Feldman and David Xiao. Sample complexity bounds on differentially private learning via communication complexity. , abs/1402.6278, 2014.
S. Fralick. Learning to recognize patterns without a teacher. , 13(1):57–64, September 2006.
Shiva Prasad Kasiviswanathan, Homin K. Lee, Kobbi Nissim, Sofya Raskhodnikova, and Adam Smith. What can we learn privately? , 40(3):793–826, 2011.
Michael J. Kearns. Efficient noise-tolerant learning from statistical queries. , 45(6):983–1006, 1998.
Andrew McCallum and Kamal Nigam. Employing em and pool-based active learning for text classification. In [*Proceedings of the Fifteenth International Conference on Machine Learning*]{}, ICML ’98, pages 350–358, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc.
Frank McSherry and Kunal Talwar. Mechanism design via differential privacy. In [*FOCS*]{}, pages 94–103. IEEE Computer Society, 2007.
Benjamin I. P. Rubinstein, Peter L. Bartlett, Ling Huang, and Nina Taft. Learning in a large function space: Privacy-preserving mechanisms for svm learning. , abs/0911.5708, 2009.
N Sauer. On the density of families of sets. , 13(1):145 – 147, 1972.
III Scudder, H. Probability of error of some adaptive pattern-recognition machines. , 11(3):363–371, Jul 1965.
Leslie G. Valiant. A theory of the learnable. , 27(11):1134–1142, 1984.
Vladimir Vapnik and Alexey Chervonenkis. Theory of pattern recognition \[in russian\]. , 1974.
Vladimir N. Vapnik and Alexey Y. Chervonenkis. On the uniform convergence of relative frequencies of events to their probabilities. , 16(2):264–280, 1971.
Some Differentially Private Mechanisms {#sec:dp_mech}
======================================
The Exponential Mechanism [@MT07]
---------------------------------
We next describe the exponential mechanism of McSherry and Talwar [@MT07]. We present its private learning variant; however, it can be used in more general scenarios. The goal here is to chooses a hypothesis $h\in H$ approximately minimizing the empirical error. The choice is probabilistic, where the probability mass that is assigned to each hypothesis decreases exponentially with its empirical error.
[**Inputs:**]{} Privacy parameter $\epsilon$, finite hypothesis class $H$, and $m$ labeled examples $S=(x_i,y_i)_{i=1}^m$.
1. $\forall h\in H$ define $q(S,h)=|\{i:h(x_i)=y_i\}|$.
2. Randomly choose $h \in H$ with probability $\frac{\exp\left(\epsilon \cdot q(S,h) /2 \right)}{\sum_{f\in H}\exp\left(\epsilon \cdot q(S,f) /2 \right)}$.
3. Output $h$.
\[prop:expMech\] (i) The exponential mechanism is $\epsilon$-differentially private. (ii) Let $\hat e\triangleq\min_{f\in H}\{{{\rm error}}_S(f)\}$. For every $\Delta>0$, the probability that the exponential mechanism outputs a hypothesis $h$ such that ${{\rm error}}_S(h)>\hat e + \Delta$ is at most $|H| \cdot \exp(-\epsilon \Delta m /2)$.
Data Sanitization
-----------------
Given a database $S=(x_1,\ldots,x_m)$ containing elements from some domain $X$, the goal of data sanitization is to output (while preserving differential privacy) another database $\hat{S}$ that is in some sense similar to $S$. This returned database $\hat{S}$ is called a [*sanitized*]{} database, and the algorithm computing $\hat{S}$ is called a [*sanitizer*]{}.
For a concept $c:X\rightarrow\{0,1\}$ define $Q_c:X^*\rightarrow[0,1]$ as $Q_c({S}) = \frac{1}{|{S}|}\cdot \Big|\{i \,:\, c(x_i) =1\} \Big|.$ That is, $Q_c({S})$ is the fraction of the entries in ${S}$ that satisfy $c$. A sanitizer for a concept class $C$ is a differentially private algorithm that given a database ${S}$ outputs a database $\hat{{S}}$ s.t. $Q_c({S}) \approx Q_c(\hat{{S}})$ for every $c\in C$.
Let $C$ be a class of concepts mapping $X$ to $\{0,1\}$. Let ${\mathcal A}$ be an algorithm that on an input database $S\in X^*$ outputs another database $\hat{S}\in X^*$. Algorithm ${\mathcal A}$ is an $(\alpha,\beta,\epsilon,\delta,m)$-sanitizer for predicates in the class $C$, if
1. ${\mathcal A}$ is $(\epsilon,\delta)$-differentially private;
2. For every input $S\in X^m$, $$\Pr\limits_{{\mathcal A}}\left[ \exists c\in C \text{ s.t.\ } |Q_c(S)-Q_c(\hat{S})|>\alpha \right]\leq \beta.$$
The probability is over the coin tosses of algorithm ${\mathcal A}$. As before, when $\delta{=}0$ (pure privacy) we omit it from the set of parameters.
\[thm:BlumUp\] For any class of predicates $C$ over a domain $X$, and any parameters $\alpha,\beta,\epsilon$, there exists an $(\alpha,\beta,\epsilon,m)$-sanitizer for $C$, where the size of the database $m$ satisfies: $$m = O\left(\frac{\log|X|\cdot {{\rm VC}}(C)\cdot\log(1/\alpha)}{\alpha^3\epsilon}+\frac{\log(1/\beta)}{\epsilon\alpha}\right).$$ The returned sanitized database contains $O(\frac{{{\rm VC}}(C)}{\alpha^2}\log(\frac{1}{\alpha}))$ elements.
The Vapnik-Chervonenkis Dimension {#sec:VC}
=================================
The Vapnik-Chervonenkis (VC) Dimension is a combinatorial measure of concept classes that characterizes the sample size of PAC learners. Let $C$ be a concept class over a domain $X$, and let $B=\{b_1,\ldots,b_\ell\}\subseteq X$. The set of all dichotomies on $B$ that are realized by $C$ is $\Pi_C(B)=\Big\{(c(b_1),\ldots,c(b_\ell)):c\in C\Big\}$. A set $B\subseteq X$ is [*shattered*]{} by $C$ if $C$ realizes all possible dichotomies over $B$, i.e., $\Pi_C(B)=\{0,1\}^{|B|}$.
The ${{\rm VC}}(C)$ is the cardinality of the largest set $B\subseteq X$ shattered by $C$. If arbitrarily large finite sets can be shattered by $C$, then ${{\rm VC}}(C)=\infty$.
Sauer’s lemma bounds the cardinality of $\Pi_C(B)$ in terms of ${{\rm VC}}(C)$ and $|B|$.
\[thm:sauer\] Let $C$ be a concept class over a domain $X$, and let $B\subseteq X$ such that $|B|>{{\rm VC}}(C)$. It holds that $\Pi_C(B)\leq \left(\frac{e|B|}{{{\rm VC}}(C)}\right)^{{{\rm VC}}(C)}$.
VC Bounds
---------
Classical results in computational learning theory state that a sample of size $\Theta({{\rm VC}}(C))$ is both necessary and sufficient for the PAC learning of a concept class $C$. The following two theorems give upper and lower bounds on the sample complexity.
\[thm:VClower\] For any $(\alpha,\beta{<}\frac{1}{2},n,m)$-SSL for a class $C$ it holds that $m\geq\frac{{{\rm VC}}(C)-1}{16\alpha}$.
\[thm:VCconsistantOld\] Let $C$ and $\mu$ be a concept class and a distribution over a domain $X$. Let $\alpha,\beta>0$, and $m\geq\frac{8}{\alpha}({{\rm VC}}(C)\ln(\frac{16}{\alpha})+\ln(\frac{2}{\beta}))$. Fix a concept $c\in C$, and suppose that we draw a sample $S=(x_i,y_i)_{i=1}^m$, where $x_i$ are drawn i.i.d. from $\mu$ and $y_i=c(x_i)$. Then, $$\Pr\left[\exists h\in C \text{ s.t.\ } {{\rm error}}_{\mu}(h,c)>\alpha \; \wedge \; {{\rm error}}_S(h)=0 \right]\leq\beta.$$
Hence, an algorithm that takes a sample of $m=\Omega_{\alpha,\beta}({{\rm VC}}(C))$ labeled examples and outputs a concept $h\in C$ that agrees with the sample is a PAC learner for $C$. The following is a simple generalization of Theorem \[thm:VCconsistantOld\].
\[thm:VCconsistant\] Let $C$ and $\mu$ be a concept class and a distribution over a domain $X$. Let $\alpha,\beta>0$, and $m\geq\frac{48}{\alpha}\left( 10{{\rm VC}}(C)\log(\frac{48e}{\alpha})+\log(\frac{5}{\beta})) \right)$. Suppose that we draw a sample $S=(x_i)_{i=1}^m$, where each $x_i$ is drawn i.i.d. from $\mu$. Then, $$\Pr\left[ \begin{array}{c}
\exists c,h\in C \text{ s.t.\ } {{\rm error}}_{\mu}(c,h)\geq\alpha\\
\text{and } {{\rm error}}_S(c,h)\leq\alpha/10
\end{array} \right]\leq\beta.$$
The above theorem generalizes Theorem \[thm:VCconsistantOld\] in two aspects. First, it holds simultaneously for every pair $c,h\in C$, whereas in Theorem \[thm:VCconsistantOld\] the target concept $c$ is fixed before generating the sample. Second, Theorem \[thm:VCconsistantOld\] only ensures that a hypothesis $h$ has small generalization error if ${{\rm error}}_S(h)=0$. In Theorem \[thm:VCconsistant\] on the other hand, this is guaranteed even if ${{\rm error}}_S(h)$ is small (but non-zero).
The next theorem handles (in particular) the agnostic case, in which the concept class $C$ is unknown and the learner uses a hypotheses class $H$. In particular, given a labeled sample $S$ there may be no $h\in H$ for which ${{\rm error}}_S(h)$ is small.
\[thm:generalization\] Let $H$ and $\mu$ be a concept class and a distribution over a domain $X$, and let $f:X\rightarrow\{0,1\}$ be some concept, not necessarily in $H$. For a sample $S=(x_i,f(x_i))_{i=1}^m$ where $m\geq\frac{50 {{\rm VC}}(H)}{\alpha^2}\ln(\frac{1}{\alpha\beta})$ and each $x_i$ is drawn i.i.d. from $\mu$, it holds that $$\Pr\Big[\forall \; h\in H,\;\; \big|{{\rm error}}_\mu(h,f)-{{\rm error}}_S(h)\big|\leq\alpha\Big]\geq1-\beta.$$
Notice that the sample size in Theorem \[thm:VCconsistant\] is smaller than the sample size in Theorem \[thm:generalization\], where, basically, the former is proportional to $\frac{1}{\alpha}$ and the latter is proportional to $\frac{1}{\alpha^2}$.
[^1]: Supported by a grant from the Israeli Science and Technology ministry, by a Israel Science Foundation grant 544/13, and by the Frankel Center for Computer Science.
[^2]: Work done while the second author was a visiting scholar at the Harvard Center for Research on Computation and Society (supported by NSF grant CNS-1237235) and at the Boston University Hariri Institute for Computing and Computational Science & Engineering. Supported in part by Israel Science Foundation grant no. 276/12.
[^3]: Supported by the Ministry of Science and Technology (Israel), by the Check Point Institute for Information Security, by the IBM PhD Fellowship Awards Program, and by the Frankel Center for Computer Science.
[^4]: To simplify the exposition, we omit in this section dependency on all variables except for $d$, corresponding to the representation length of domain elements.
[^5]: A semi-supervised learner uses a small batch of labeled examples and a large batch of unlabeled examples, whereas an active-learner operates on a large batch of unlabeled example and chooses (maybe adaptively) which examples should be labeled.
[^6]: We remark that – unlike this work – the focus in [@BF13] is on the dependency of the labeled sample complexity in the approximation parameter. As our learners are non-active, their labeled sample complexity is lower bounded by $\Omega(\frac{1}{\alpha})$ (where $\alpha$ is the approximation parameter).
[^7]: Feldman and Xiao [@FX14] showed an example of a concept class $C$ over $X_d$ for which every pure-private learner must have unlabeled sample complexity $\Omega({{\rm VC}}(C)\cdot d)$. Hence, as a function of $d$ and ${{\rm VC}}(C)$, the unlabeled sample complexity in Theorem \[thm:SampleComplexity\] is the best possible for a generic construction of pure-private learners.
[^8]: The unlabeled sample complexity depends on $\delta$ as $n$ depends on $\delta$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we explore the 3D structure of light scattering from dark-field illuminated live 3T3 cells in the presence of 40 nm gold nanomarkers. For this purpose, we use a high resolution holographic microscope combining the off-axis heterodyne geometry and the phase-shifting acquisition of the digital holograms. A comparative study of the 3D reconstructions of the scattered fields allows us to locate the gold markers which yield, contrarily to the cell structures, well defined bright scattering patterns that are not angularly tilted and clearly located along the optical axis (z). This characterization is an unambiguous signature of the presence of gold biological nanomarkers, and validates the capability of digital holographic microscopy to discriminate them from background signals in live cells.'
author:
- Fadwa Joud
- Nilanthi Warnasooriya
- Philippe Bun
- Frédéric Verpillat
- 'Sarah Y. Suck'
- Gilles Tessier
- Michael Atlan
- Pierre Desbiolles
- 'Maïté Coppey-Moisan'
- Marie Abboud
- Michel Gross
-
date: 'Received: / Accepted: '
title: 3D exploration of light scattering from live cells in the presence of gold nanomarkers using holographic microscopy
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction {#intro}
============
Gold nanoparticles attract great scientific and technological interest because of their physical and chemical characteristics. In particular, the optical tracking of gold nanoparticles in biology have gained popularity for several reasons. Gold nanoparticles provide high scattering efficiencies ([@jain2006calculated]) and they can be detected directly using dark field or total internal reflection (TIR) illumination down to particle diameters of 40 nm as shown by [@sonnichsen2000spectroscopy]. Unlike fluorescent markers, they are immune to photo bleaching, and they are potentially non-cytotoxic (see [@west06]). Because of these properties, the use of gold nanoparticles as biomarkers for live cell imaging using photothermal tracking (see [@cognet2002], [@boyer2003] and [@lasne2006]) has a high potential.
As shown by [@atlan2008heterodyne], holography has proved its ability to image and localize gold nanoparticles in 3D, either for fixed particles spin coated on a glass substrate or in free motion within a water suspension. More recently, [@absil2010photothermal] have shown that heterodyne holography also allows the photothermal detection of 10 nm gold particles, and [@warnasooriya2010imaging] have imaged 40 nm gold particles in a cellular environment. In that last experiment the particle holographic signal is superimposed with the light scattered by the cell refractive index fluctuations, which yield a speckle field. For particles imaging, this induced speckle is a parasitic signal, but in many other situations, like in Dark Field microscopy , or in Differential Interference Contrast (DIC) microscopy (see [@goldberg1986stages] ) this speckle is the main source of contrast that is used to image the cell itself. It is thus important to discriminate the particle signal from the cell parasitic speckle.
In this paper, we have imaged biological samples (3T3 cells) labeled with 40 nm gold particles using the digital holographic setup described in [@warnasooriya2010imaging]. We have performed the 3D holographic reconstruction of the wave-field scattered by the samples, which are illuminated at 45$^\circ$ in a total internal reflection configuration, and we have shown that these wave-fields are noticeably different for the particle, and the speckle signal. We showed here that important information can be derived not only from the intensity of the bright spots caused by the gold particles, but also from the 3D shape of the light scattering pattern, which is easily accessed using holography. We showed, in particular, that the speckle signal keeps memory of the illumination direction, while the particle signal does not. This result has been confirmed by imaging samples of cells that have not been labeled with gold particles, and samples of free gold particles. The shape of the wave-field scattered by the sample can thus be used as a signature that helps to discriminate the particle signal from the speckle.
Materials and Methods {#sec:1}
=====================
Biological specimen preparation {#sec:2}
-------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Gold bioconjugates and fibroblasts preparation procedures. (a) Gold bioconjugates functionalisation. (b) Fibroblasts-gold nanoparticles coupling.[]{data-label="fig:1"}](spl_prep "fig:"){height="6cm"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The biological specimens that we imaged are monolayers of live NIH 3T3 mouse fibroblasts labeled with 40 nm gold particles via their integrin cellular surface receptors. Streptavidin-coated gold nanoparticles were attached to the surface cellular integrin receptors via biotin and fibronectin proteins: see Fig.\[fig:1\]. Streptavidin and biotin are very well known for their strong affinity towards each other, and fibronectin, an extracellular matrix protein, has the property of interacting specifically with cellular surface receptors of integrin family.
Fibronectin proteins (fibronectin from bovine plasma, Sigma, St Louis, MO) were biotinylated using EZ-Link$^{\textcircled {\scriptsize R}}$Sulfo-NHS-LC-Biotin according to the provider protocol (Pierce, Rockford, IL). The final concentration of biotinylated-fibronectin solution was 0.447 mg/mL. The streptavidin-coated gold conjugates of 40 nm average diameter (BioAssay, Gentaur, France) were rinsed twice with 1X PBS (Phosphate Buffered Saline) (pH = 7.25). We than diluted 10 $\mu$L of the gold solution in 990 $\mu$L of the same PBS buffer solution. Then the dilute gold solution was incubated with 50 $\mu$L of the biotinylated-fibronectin solution for four hours at room temperature to allow the specific streptavidin-biotin bonding. The final functionalized gold particles solution was kept at 4$^\circ$C and used within 24 hours after its preparation in order to ensure maximum functionality. Before every use, the functionalized gold particles solution was sonicated.
48 hours before the observation, monolayers of 3T3 cells were cultured in Duelbecco’s modified Eagle’s medium (DMEM Gibco, Invitrogen, Carlsbad, CA) supplemented with 10% fetal calf serum (PAA Laboratories GmbH) on 32 mm diameter fibronectin-coated glass cover slips (fibronectin from bovine plasma, Sigma, St Louis, MO) at 37$^\circ$C in a 5% CO2 atmosphere. After 24 hours of incubation, we added to each coverslip a solution composed of DMEM (2 mL) plus 500 $\mu$L of the functionalized gold particles solution. The binding of integrin and fibronectin occurs at this level allowing the cells to attach, on their surface, the functionalised gold nanoparticles.
The coverslip containing adherent 3T3 cells tagged with 40 nm gold nanoparticles was mounted on a specific observation chamber. In order to maintain the physiological pH condition during the experiments, cells were kept in DMEM-F12 medium (DMEM-F12 without Phenol red, B12 vitamin, Riboflavin, 0.5% fetal calf serum and supplemented with 20 mM of HEPES \[(4-(2-hydroxyethyl)-1-piperazineethanesulfonicacid)\] and L-Glutamine from PAA Laboratories). We measured the level of biotin incorporation on an HABA \[2-(4’ -Hydroxyazobenzene) Benzoic Acid)\] quantitation assay to verify the efficiency of the biotinylation protocol. Average number of biotin molecules obtained per fibronectin is 2.5.
Holographic Microscope Experimental Setup {#sec:3}
-----------------------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![(a) Experimental setup : AOM1, AOM2: acousto-optic modulators; M: mirror; MO: microscope objective (NA = 0.5); BS: beam splitter; PBS: polarizing beam splitter; CCD: CCD camera; $E_r$: reference field; $E_o$: scattered field; z = 0: CCD plane; z = $z_0$: CCD conjugate plane with respect to MO. (b) Details of the total internal reflection optical arrangement that is used for dark-field illumination of the biological sample.[]{data-label="fig_2setup"}](exp_mo_ "fig:"){height="7.2cm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Fig.\[fig\_2setup\] illustrates the optical setup. The illumination source is a single-mode near infrared laser diode emitting at $\lambda =785$nm (DL7140-201S 80 mW Laser Diode $@$90 mA current). A polarizing beam splitter cube (PBS) is used to split the original illumination laser light into two beams, a reference beam (complex field $E_R$, frequency $f_R$) and an object illumination beam (complex field $E_0$, frequency $f_0$) forming the two arms of a Mach-Zehnder interferometer. A combination of a half wave plate and two neutral density filters is used to prevent the saturation of the detector by controlling the optical power traveling in each arm. Two acousto-optic modulators (AOM1, AOM2) driven around 80 MHz and using the first order of diffraction, shift both frequencies at respectively $f_{AOM1}$ and $f_{AOM2}$ .
The object beam illuminates the sample, prepared as described in Section \[sec:2\], by provoking total internal reflection (TIR) at the medium-air interface in order to prevent direct illumination light from entering the system. The evanescent wave locally frustrated and the illumination wave directly scattered by the beads and cells give off a propagating scattered wave (complex field $E$), which is collected by a microscopic objective (MO, $50 \times$ magnification, NA=0.5, air). A beam splitter is then used to combine the scattered object wave and the reference wave which is slightly angularly tilted ($ \sim 1^\circ$) with respect to the propagation axis of the object wave in an off-axis configuration. A half wave plate on the object illumination arm aligns the polarization of the corresponding beam ensuring its optimal interference with the reference beam. A CCD camera (Roper Cascade 512F, $512 \times 512$ square pixels of 16 $\mu$m size, exposure time 100 ms, frame rate $f_{CCD}$ = 8 Hz) detects the interference pattern (hologram) and sends it to a computer. The hologram is then numerically treated and the complex field $E(x,y,z)$ is reconstructed numerically.
Holographic acquisition {#sec:acquistion}
-----------------------
In order to filter out unwanted parasitic signals, we use a heterodyne modulation. A four-phase demodulation method is used to record holograms. This method consists in acquiring a sequence of images with a relative phase shift $\Delta\varphi=\pi/2$ between two consecutive frames. The reference wave is frequency shifted by tuning the two acousto-optic modulators AOM1 and AOM2 as done by [@Leclerc2000], and we get, as shown by [@atlan2007aps], an accurate phase shift $\Delta\varphi$ that simplify the phase shifting digital holography demodulation. The heterodyne beat frequency is thus: $$\label{Eq_1a}
\Delta f={f}_{AOM1}-{f}_{AOM2}=\frac{f_{CCD}}{4}$$ where $f_{CCD}=$8Hz is the frame rate frequency of the CCD camera. The camera records a sequence of 32 frames $\emph{I}_{0}$,...,$\emph{I}_{31}$, and the object field $\emph{E}$ on the CCD plane ($z=0$) is given by: $$\label{eq2}
\emph{E}(x,y,z=0) = \sum_{n=1}^{M} j^{~n} \emph{I}_{n}$$ where $j^2=-1$, and $M=32$ is the number of frames used for the reconstruction. In Eq.\[eq2\], the coordinates $x,y$ (with $0<x,y<511$) are integers, which represent the pixel location within the CCD plane. The pixel size is then the physical CCD pixel size, i.e. 16 $\mu$m.
Holographic reconstruction {#sec:reconstruction}
--------------------------
The problem of the reconstruction in the context of holographic microsopy has been discussed in details by [@colomb2006automatic] and [@colomb2006]. Nevertheless, the Colomb et al. method refers implicitly to the phase-contrast imaging of [@marquet2005dhm], and is not well suited to the reconstruction of the 3D image of a wave-field as done by [@grilli2001whole]. Here, to get a 3D image of the wave-field scattered by the gold particle, we have used a slightly different reconstruction method, which is inspired from the reconstruction method used by [@mann2005].
First, we considered that the measured hologram represents the field $\emph{E}(x,y,z_0)$ within the CCD conjugate plane $z=z_0$ with respect to the Microscope Objective (MO), i.e., the plane, whose image is on focus on the CCD detector. Since we image the sample through a microscope objective MO, we must compensate the phase curvature, the phase tilt and the enlargement factor that are related to the presence of MO as disccused by [@colomb2006]. We have thus: $$\label{EQ_2b}
\emph{E}(x,y,z=z_0)\;=\; e^{j (K_x x + K_y y)} \; e^{j A (x^2 + y^2) } \sum_{n=0}^{M} j^{~n} \emph{I}_{n}$$ where $(K_x,K_y)$ and $A$ are the tilt and lens parameters respectively that must be determined.
We measured these parameters (or compensated their effect) by an original method that consists in reconstructing the image of the microscope objective output pupil by the one Fourier transform reconstruction method of [@schnars1994drh]. The lens parameters $A$ is then close to the lens parameter that is used in the pupil reconstruction by the [@schnars1994drh] method. On the other hand the tilt parameters $(K_x,K_y)$ are compensated by translating the pupil [@schnars1994drh] image in the center of the calculation grid.
The properly compensated measured hologram represents then the field $\emph{E}(x,y,z=z_0)$ in the conjugate plane $z_0$. Then, as done by [@mann2005] in holographic microscopy, the field $\emph{E}(x,y,z)$ in the vicinity of the conjugate plane (i.e. for $z \simeq z_0$) is calculated by the angular spectrum method, which involves two Fourier transforms (see for example [@Leclerc2000], [@leclerc2001] or [@yu2005]). This method is chosen here since it keeps the pixel size $\delta z$ constant whatever the reconstruction distance $z$ is.
The pixel size $\delta x=\delta y$, which must be calibrated to make a quantitative analysis of the holographic data, is measured by imaging with the same setup geometry a calibrated USAF target located in the CCD conjugate plane. We get $\delta x=\delta y=177$ nm. The reconstruction is then done for 512 different reconstruction distances $$\label{EQ_2c}
z=z_0+ (m_z-256) \delta z$$ where $\delta z=177$ nm and $m_z=0...511 $. By this way, we get 3D volume images with $512 \times 512\times 512$ voxels, with the same pixel size ($\delta x= \delta y=\delta
z=177$ nm) in the 3 directions $x,y$ and $z$.
Results and Discussion {#sec:}
======================
We have studied different samples of gold marked cell, unmarked cells and of free 40 nm gold particles.
Samples of marked cells
-----------------------
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Images of the first sample with two fibroblast cells marked with 40 nm gold particles. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image near the $z = z0$ conjugate plane (z = 255). The display is made in linear grey scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.9$, white to $\ln I = 16.3$ Digital Counts (DC). (c) 3D linear surface plot of the same data. []{data-label="fig_3"}](Fig3a.eps "fig:"){height="5cm"} ![Images of the first sample with two fibroblast cells marked with 40 nm gold particles. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image near the $z = z0$ conjugate plane (z = 255). The display is made in linear grey scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.9$, white to $\ln I = 16.3$ Digital Counts (DC). (c) 3D linear surface plot of the same data. []{data-label="fig_3"}](Fig3b.eps "fig:"){height="5cm"} ![Images of the first sample with two fibroblast cells marked with 40 nm gold particles. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image near the $z = z0$ conjugate plane (z = 255). The display is made in linear grey scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.9$, white to $\ln I = 16.3$ Digital Counts (DC). (c) 3D linear surface plot of the same data. []{data-label="fig_3"}](Fig3c.eps "fig:"){height="5cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Figures \[fig\_3\] show the images of a first sample, with two cells and several particles. Figure \[fig\_3\](a) shows a direct white light illumination image of the sample. The two cells can be seen, but the contrast is low. Figure \[fig\_3\] (b) shows the intensity holographic image of the same region of the sample reconstructed in the $z = z_0$ or $z=256$ conjugate plane (here and in the following we will express the $x$, $y$ and $z$ coordinates by the corresponding pixel index $m_x$, $m_y$ and $m_z$). The display is made in logarithmic color scale.
Because of the variation of the refractive index within the cells, the illumination light is diffracted yielding a speckle pattern that is superimposed with the particle signal. This speckle is visible on Fig. 3 (b), and, from the envelope of the speckled zone, one can guess the shape of the cells. We interpret the brightest points 1, 2 and 3 of Fig. \[fig\_3\] (b), which correspond to maximum intensities $I = 11.5 \times 10^6$, $4.6
\times 10^6$, and $8.4\times 10^6$ and Digital Count (DC) respectively, as being particles signal. Many other bright points are also visible, but it is not simple to determine, which points are particles, and which are speckle hot spots. This is especially true within circle 4, where many bright points, close together, are visible.
To better visualize the 40 nm gold particles, we have displayed, on Fig. \[fig\_3\] (c), by using the Interactive 3D Surface Plot plug-in of Image J (see [@abramoff2004image]), a 3D linear surface plot of the region of the sample that is displayed on Fig.\[fig\_3\] (b). As seen, the particles that correspond to sharp peaks can be easily visualized, but some ambiguity remains around the meaning of the lower peaks, which could be attributed either to particles, noises, or scattering by biological features of the cell.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Images of the second sample with one fibroblast cells marked with one 40 nm gold particle. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 325$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 4.48$, white to $\ln I = 15.56$ Digital Counts (DC). (c) 3D linear surface plot. []{data-label="fig_4"}](Fig4a.eps "fig:"){height="5cm"} ![Images of the second sample with one fibroblast cells marked with one 40 nm gold particle. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 325$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 4.48$, white to $\ln I = 15.56$ Digital Counts (DC). (c) 3D linear surface plot. []{data-label="fig_4"}](Fig4b.eps "fig:"){height="5cm"} ![Images of the second sample with one fibroblast cells marked with one 40 nm gold particle. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 325$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 4.48$, white to $\ln I = 15.56$ Digital Counts (DC). (c) 3D linear surface plot. []{data-label="fig_4"}](Fig4c.eps "fig:"){height="5cm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The images of Fig.\[fig\_4\] are obtained for a second sample with a single cell, and, as we will see, a single gold particle. Figure \[fig\_4\] (a) shows a white light image of the sample. One can see the cell, whose shape is triangular. Figures \[fig\_4\] (b) shows the intensity holographic image reconstructed for $z = 325$ (i.e. 6.93 $\mu$m above the $z = z_0$ conjugate plane), with logarithmic color scale display. Here again, the speckle related to the light diffracted by the cell is visible on Fig.\[fig\_4\] (b), and one can guess the triangular shape of the cell. The brightest point (arrow 1 on Fig. \[fig\_4\] (b)) is interpreted as a particle. Since the illumination intensity and focusing area is not well controlled, the particle maximum intensity $I = 2.9 \times 10^6$ DC obtained here is noticeably lower than for the first sample. Nevertheless, the signal obtained for the bright point marked by arrow 2 is more than 10 times lower (i.e. $I =
1.9\times 10^5$), so we can interpret it as a speckle hot spot. This result is confirmed by Fig. \[fig\_4\] (c) that shows a 3D linear surface plot of the sample.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 respectively. The display is made in the same logarithmic color scale than Fig.\[fig\_4\](b) by Volume Wiever. []{data-label="fig_5"}](Fig5a.eps "fig:"){height="6cm"}  respectively. The display is made in the same logarithmic color scale than Fig.\[fig\_4\](b) by Volume Wiever. []{data-label="fig_5"}](Fig5b.eps "fig:"){height="6cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
To go further, and to better characterize the particle’s signal with respect to hot spots, we have analyzed the 3D images of the wave-field obtained by performing the holographic reconstruction for the 512 different reconstruction distances of Eq.\[EQ\_2c\]. The 3D data were displayed using the Volume Viewer plug-in of Image J, which is able to display cuts of 3D data in arbitrary planes. By using this plug-in, we have performed cuts parallel to the $yz$ plane of incidence of the sample illumination beam (see Fig.\[fig\_2setup\] (b) ).
Figure \[fig\_5\] (a) shows the image of a cut made in the plane $x=128$, which intersects the particle signal 1 of Fig.\[fig\_4\] (b). The particle is seen as a bright spot on the $yz$ cut image. We can notice here that the particle signal is located at $z
\simeq 325$ coordinate, which corresponds to the reconstruction plane of Fig.\[fig\_4\] (b). This is expected, since we have chosen to display on Fig.\[fig\_4\] (b) the plane where the maximum intensity is reach in 3D, and since this maximum correspond to the particle we consider here.
The image of Fig.\[fig\_5\] (b) corresponds to $x=221$ i.e. to a cut that intersects the bright spot 2 of Fig.\[fig\_4\] (b) we interpret as a speckle hot spot. Contrarily to the particle, the hot spot signal extension along the $x$ axis (vertical axis on the Fig.\[fig\_5\] images) is quite large. Moreover, the hot spot image is angularly tilted in the $yz$ plane.
This angular tilt can be simply interpreted by describing how the light propagation is governed in the biological cells. This propagation is dependant on the phase function inside the illuminated cell. Because biological tissues are inhomogeneous, the form of their phase function is not well defined and is thus characterized by the anisotropy coefficient g, which is the average cosine of the phase function. This parameter g describes the asymmetry of the single scattering pattern ; it is thus null when the scattering is isotropic, equals 1 for forward scattering and equals -1 in the case of backward scattering. In our experiments, the illuminated cells are maintained in DMEM medium that consists mainly of water. Since the refractive index of cells is close to that of water, the cell anisotropy factor g is close to one ($g \simeq 0.9$ in biological tissues as mentioned by [@cheong1990review]). As a consequence, the light scattered by the cells mainly follows the forward scattering regime and the observed light scattering pattern appears to be tilted by approximately 45$^\circ$ since the incident illumination laser beam is initially tilted by 45$^\circ$ in conformity with the TIR illumination geometry (see Fig.\[fig\_2setup\] (b)).
The exact shape of the hot spot’s wave-field can be calculated, but it is quite complicated. It involves the calculation of the angular distribution of the scattered light, which depends on the cell anisotropy factor $g$. One must then calculate the refraction of the scattered light on the medium-air interface, and take into consideration the collection of light by the microscope objective. The quantitative analysis of the wave-field’s shape, which yields the angular tilt, is thus out of the scope in the present paper, and one can simply say that the hot spot signal keeps some memory of the illumination direction, and is thus angularly tilted in the $y z$ plane.
One can notice that a similar angular tilt effect has been observed recently on the photothermal signal of 50 nm and 10 nm gold particles by [@absil2010photothermal].
Control experiments performed on free particles or on unmarked-cells
--------------------------------------------------------------------
In order to confirm our interpretation of the angular tilt seen on Fig.\[fig\_5\] (b), we have performed some control experiments by imaging an unmarked cell sample and another sample of free gold particles.
### Unmarked cell sample
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Images of an unmarked cell sample. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 252$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.80$, white to $\ln I = 13.28$ Digital Counts (DC). (c) 3D linear surface plot. (d) Image obtained par performing cut of the 3D data in the plane $x=271$ that corresponds to the white dashed line seen in (b). The display is made by Volume Viewer with the same logarithmic color scale as in (b).[]{data-label="fig_6"}](Fig6a "fig:"){height="5cm"} ![ Images of an unmarked cell sample. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 252$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.80$, white to $\ln I = 13.28$ Digital Counts (DC). (c) 3D linear surface plot. (d) Image obtained par performing cut of the 3D data in the plane $x=271$ that corresponds to the white dashed line seen in (b). The display is made by Volume Viewer with the same logarithmic color scale as in (b).[]{data-label="fig_6"}](Fig6b "fig:"){height="5cm"} ![ Images of an unmarked cell sample. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 252$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.80$, white to $\ln I = 13.28$ Digital Counts (DC). (c) 3D linear surface plot. (d) Image obtained par performing cut of the 3D data in the plane $x=271$ that corresponds to the white dashed line seen in (b). The display is made by Volume Viewer with the same logarithmic color scale as in (b).[]{data-label="fig_6"}](Fig6c "fig:"){height="5cm"}
![ Images of an unmarked cell sample. (a) Direct white light illumination image. (b) Holographic intensity reconstructed image for the $z = 252$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.80$, white to $\ln I = 13.28$ Digital Counts (DC). (c) 3D linear surface plot. (d) Image obtained par performing cut of the 3D data in the plane $x=271$ that corresponds to the white dashed line seen in (b). The display is made by Volume Viewer with the same logarithmic color scale as in (b).[]{data-label="fig_6"}](Fig6d "fig:"){height="6cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Fig.\[fig\_6\] shows the results obtained with a cell sample without particle. The direct white light image is shown in Fig. \[fig\_6\](a). Figure \[fig\_6\](b) shows the holographic intensity reconstructed image that shows the light scattered by the cell refractive index inhomogeneities. This scattered light, which has a speckle structure, exhibits several hot spots that correspond to bright points on Fig.\[fig\_6\](b), and we have chosen to image on Fig.\[fig\_6\](b) the reconstruction plane $z=252$ that corresponds to the maximum hot spot intensity ($5.85\times 10^5$ DC).
The image we get here without particles is visually quite similar to Fig.\[fig\_3\](b) which is obtained with several particles. The 3D linear surface plot shown in Fig.\[fig\_3\](c) does not help since it is visually similar to Fig.\[fig\_3\](c). Nevertheless, the maximum of the bead signal ($11.6\times 10^6$ DC) in Fig.\[fig\_3\](b) and (c) is much higher than the maximum of the hot spot signal ($5.85\times 10^5$ DC) in Fig.\[fig\_6\](b) and (c). This makes the background noise visually higher in Fig.\[fig\_6\](c).
Figure \[fig\_6\](d) shows the $yz$ image obtained by performing a cut in the plane $x=271$ that intersects the two brightest speckle hot spots of the sample. Now the $yz$ image obtained in Fig. \[fig\_6\](d) without particle is qualitatively different than with a particle in Fig. \[fig\_5\](a). The signal is angularly tilted like in Fig. \[fig\_5\](b), and its extension in the $z$ direction (vertical axis) is larger.
### Free particles sample
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Images of free gold particles in a water and agarose suspension. (a) Holographic intensity reconstructed image for the $z = 251$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.85$, white to $\ln I = 17.61$ Digital Counts (DC). (b) 3D linear surface plot. (c) Image obtained par performing cut of the 3D data in the plane $x=258$ that corresponds to the white dashed line seen in (a). The display is made by Volume Viewer with the same logarithmic color scale than for (a).[]{data-label="fig_7"}](Fig7a "fig:"){height="5cm"} ![ Images of free gold particles in a water and agarose suspension. (a) Holographic intensity reconstructed image for the $z = 251$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.85$, white to $\ln I = 17.61$ Digital Counts (DC). (b) 3D linear surface plot. (c) Image obtained par performing cut of the 3D data in the plane $x=258$ that corresponds to the white dashed line seen in (a). The display is made by Volume Viewer with the same logarithmic color scale than for (a).[]{data-label="fig_7"}](Fig7b "fig:"){height="5cm"}
![ Images of free gold particles in a water and agarose suspension. (a) Holographic intensity reconstructed image for the $z = 251$. The display is made in logarithmic color scale for the intensity $I = |E|^2$. Black correspond to $\ln I = 5.85$, white to $\ln I = 17.61$ Digital Counts (DC). (b) 3D linear surface plot. (c) Image obtained par performing cut of the 3D data in the plane $x=258$ that corresponds to the white dashed line seen in (a). The display is made by Volume Viewer with the same logarithmic color scale than for (a).[]{data-label="fig_7"}](Fig7c "fig:"){height="6cm"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The images of Fig.\[fig\_7\] are obtained for free particles (uncoupled to cells) in brownian motion in a water and agarose suspension. Here, we used agarose to slow down the motion of the particles enough to make the displacement of the particles negligible during the time of acquisition of the sequence of $M=32$ frames.
Figure \[fig\_6\](a) is the holographic intensity reconstructed image. The brightest points correspond to particles that are located within the reconstruction plane, while the less bright points are interpreted as out of focus particles. Reconstruction is made here in the plane $z=251$ which corresponds to the maximum intensity ($4.4\times 10^7$ DC) for the brightest spot (marked by a white arrow on Fig.\[fig\_6\](a)). Many particles can be seen on Fig. \[fig\_6\](a) and on Fig. \[fig\_6\](b) that displays the holographic data with 3D linear surface plot.
Figure \[fig\_6\](c) shows the $yz$ image in the plane $x=258$ that intersects the brightest particle and corresponds to the white dashed line on Fig.\[fig\_6\](a). As expected, the image of the particle (white arrow) has a smaller extension along the $z$ direction (vertical direction) and does not exhibit the angular tilt we get with speckle hot spot in Fig.\[fig\_5\](b) and Fig.\[fig\_6\](d). This result confirms our interpretation of the angular tilt observed with the speckle hot spots.
Conclusion {#conc}
==========
To summarize, we have shown that heterodyne holographic microscopy, in the off-axis geometry, is well adapted to the detection of weakly scattering objects. The sensitivity, signal to noise ratio and selectivity of the technique allow the localization of light-scattering gold nanoparticles of a few tens of nanometers, which are good candidates of non-toxic and perennial markers. Biological environments, however, are difficult to address since cell features generate strong parasitic speckle. Here, we have reported the detection of 40 nm particles attached to the surface of live 3T3 mouse fibroblasts. A comparison of these signals with either non-labelled cells or simple gold particles in solution allowed us to unambiguously discriminate particles. We show that, in addition to a stronger scattering signal, gold particles induce a relatively isotropic scattering, whereas biological features are characterized by mostly forward scattering. This dissimilarity in the scattering patterns, explained by the inconsistency of the refractive indexes and anisotropy parameters g, is easily characterized by digital holography, making it an excellent tool for the 3D detection of gold labels in biological environments.
Authors wish to acknowledge the French Agence Naionale de la Recherche (ANR) and the “Centre de Compétence NanoSciences Ile de France” (C’nano IdF) for their financial support.
[22]{} \[1\][\#1]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{}
\[2\]\[\][[\#2](#2)]{}
Abramoff M, Magalhaes P, Ram S (2004) [Image processing with ImageJ]{}. Biophotonics International 11(7):36–43
Absil E, Tessier G, Gross M, Atlan M, Warnasooriya N, Suck S, Coppey-Moisan M, Fournier D (2010) [Photothermal heterodyne holography of gold nanoparticles]{}. Opt Express 18:780–786
Atlan M, Gross M, Absil E (2007) [Accurate phase-shifting digital interferometry]{}. Opt Lett 32:1456–1458
Atlan M, Gross M, Desbiolles P, Absil [É]{}, Tessier G, Coppey-Moisan M (2008) [Heterodyne holographic microscopy of gold particles]{}. Optics letters 33(5):500–502
Boyer D, Tamarat P, Maali A, Lounis B, Orrit M (2003) [Photothermal imaging of nanometer-sized metal particles among scatterers]{}. Science 297:1160–1163
Cheong W, Prahl S, Welch A (1990) [A review of the optical properties of biological tissues]{}. IEEE journal of quantum electronics 26(12):2166–2185
Cognet L, C Tardin C, Boyer D, Choquet D, Tamarat P, Lounis B (2002) [Single metallic nanoparticles imaging for protein detection in cells]{}. Proc Natl Acad Sci 100:11,350–11,355
Colomb T, Cuche E, Charri[è]{}re F, K[ü]{}hn J, Aspert N, Montfort F, Marquet P, Depeursinge C (2006) [Automatic procedure for aberration compensation in digital holographic microscopy and applications to specimen shape compensation]{}. Applied optics 45(5):851–863
Colomb T, Montfort F, Kuehn J, Aspert N, Cuche E, Marian A, Charri[é]{}re F, Bourquin S, Marquet P, Depeursinge C (2006) [Numerical parametric lens for shifting, magnification, and complete aberration compensation in digital holographic microscopy]{}. J Opt Soc Am A 23:3177–3190
Goldberg D, Burmeister D (1986) [Stages in axon formation: observations of growth of Aplysia axons in culture using video-enhanced contrast-differential interference contrast microscopy]{}. Journal of Cell Biology 103(5):1921
Grilli S, Ferraro P, De Nicola S, Finizio A, Pierattini G, Meucci R (2001) [Whole optical wavefields reconstruction by digital holography]{}. Optics Express 9(6):294–302
Jain P, Lee K, El-Sayed I, El-Sayed M (2006) [Calculated absorption and scattering properties of gold nanoparticles of different size, shape, and composition: applications in biological imaging and biomedicine]{}. J Phys Chem B 110(14):7238–7248
Lasne D, Blab GA, Berciaud S, Heine M, Groc L, Choquet D, Cognet L, Lounis B (2006) [Single nanoparticle photothermal tracking (SNaPT) of 5-nm gold beads in live cells]{}. Biophys J 91:4598–4604
LeClerc F, Collot L, Gross M (2000) Numerical heterodyne holography with two-dimensional photo-detector arrays. Opt Lett 25:716–718
LeClerc F, Gross M, Collot L (2001) [Synthetic-aperture experiment in the visible with on-axis digital heterodyne holography]{}. Opt Lett 26:1550–1552
Mann CJ, Yu L, Lo CM, Kim MK (2005) [High resolution quantitative phase-contrast microscopy by digital holography]{}. Opt Express 13:8693–8698
Marquet P, Rappaz B, Magistretti P, Cuche E, Emery Y, Colomb T, Depeursinge C (2005) [Digital holographic microscopy: a noninvasive contrast imaging technique allowing quantitative visualization of living cells with subwavelength axial accuracy]{}. Optics letters 30(5):468–470
Schnars U, J[ü]{}ptner W (1994) [Direct recording of holograms by a CCD target and numerical reconstruction]{}. Applied Optics 33(2):179–181
S[ö]{}nnichsen C, Geier S, Hecker N, Von Plessen G, Feldmann J, Ditlbacher H, Lamprecht B, Krenn J, Aussenegg F, Chan V, et al (2000) [Spectroscopy of single metallic nanoparticles using total internal reflection microscopy]{}. Applied Physics Letters 77:2949
Warnasooriya N, Joud F, Bun P, Tessier G, Coppey-Moisan M, Desbiolles P, Atlan M, Abboud M, Gross M (2010) [Imaging gold nanoparticles in living cell environments using heterodyne digital holographic microscopy]{}. Opt Express 18:3264–3273
West JL, Drezek RA, J HN (2006) Nanotechnology provides new tools for biomedical optics. In: Bronzino JD (ed) Tissue Engineering and Artifical Organs, 3rd Edition, CRC Press, pp 25–1–25–9
Yu L, Kim M (2005) [Wavelength-scanning digital interference holography for tomographic three-dimensional imaging by use of the angular spectrum method]{}. Opt Lett 30:2092–2094
|
{
"pile_set_name": "ArXiv"
}
|
[Spontaneous Symmetry Breaking through Mixing]{}
[Harald Fritzsch and Michael Spannowsky]{}
Ludwig Maximilian University
Sektion Physik, Theresienstr. 37, 80333 München Germany
**Abstract:**
We discuss a model, in which the negative mass square needed in the Higgs mechanism is generated by mixing with a heavy scalar. We have two scalar doublets in the standard model. Phenomenological properties of the heavy new scalar are discussed. The heavy scalar can be detected by the LHC.
In the Standard Model (SM) the masses of the W- and Z-bosons and of the fermions (leptons and quarks) are introduced by the mechanism of spontaneous symmetry breaking (“Higgs mechanism”). It is assumed that the squared mass term of the scalar particle is negative. This, together with the $\phi ^{4}$ self interaction of the scalar bosons leads to a non-zero vacuum expectation value.
The masses of the W- and Z-bosons are spontaneously generated via the gauge interaction, and the masses of the fermions appear due to their Yukawa-type interactions with the scalar. No prediction can be made about the fermion masses, since they depend on the Yukawa coupling constants, which are free parameters.
The negative mass square of the scalar boson is rather peculiar and looks to us like an ad-hoc assumption, not explaining why the electroweak gauge symmetry is broken. In this paper we would like to propose a possible underlying mechanism, able to generate the imaginary Higgs mass (see also [@Calmet]) at an energy scale testable at the LHC.
We assume that there is a massless scalar which mixes (seesaw like) with a heavy scalar. Through the mixing a negative mass for the scalar is introduced in a natural way. As a result of the mixing the spontaneous symmetry breaking arises as in the Standard Model. For a long time, the seesaw mechanism has been widely studied to generate light neutrino masses [@Mohap] as well as to explain the mass hierarchies in the whole fermion sector [@Kiyo]. Instead, the idea of applying this mechanism to the scalar sector of the SM beside incorporating new degrees of freedom is quite new and recieved increasing interest over the last years [@Atwood].
The scalar sector of this model is given by the following Lagrange density:$$L_{sc.}=D^{\mu }%
\begin{pmatrix}
\phi \\
\Phi%
\end{pmatrix}%
^{+}D_{\mu }%
\begin{pmatrix}
\phi \\
\Phi%
\end{pmatrix}%
-\left( \lambda \phi ^{+}\Phi +\lambda \Phi ^{+}\phi +\Lambda \Phi
^{2}-\kappa \left( \phi ^{+}\phi \right) ^{2}\right)$$
The mass matrix is given by
$$M=%
\begin{pmatrix}
0 & \lambda \\
\lambda & \Lambda%
\end{pmatrix}%
,$$
where we assume $\lambda \ll \Lambda $.
We diagonalize this mass matrix using$$S=%
\begin{pmatrix}
\cos \theta & -\sin \theta \\
\sin \theta & \cos \theta%
\end{pmatrix}%
.$$
The mixing angle $\theta $ is fixed by $\lambda $ and $\Lambda $.
$$\begin{aligned}
M^{\prime } &=&SMS^{-1} \\
M^{\prime } &=&%
\begin{pmatrix}
1/2(\Lambda -\sqrt{4\lambda ^{2}+\Lambda ^{2}}) & 0 \\
0 & 1/2(\Lambda +\sqrt{4\lambda ^{2}+\Lambda ^{2}})%
\end{pmatrix}%\end{aligned}$$
The assumption $\lambda \ll \Lambda $ ensures that $\left\vert \sin \theta
\right\vert \ll \left\vert \cos \theta \right\vert $.
Note that the negative mass squared for the light scalar is proportional to $%
\lambda ^{2}$ in first order.
$$\begin{aligned}
\mu ^{2}\left( \lambda ,\Lambda \right) &\equiv &-1/2(\Lambda -\sqrt{%
4\lambda ^{2}+\Lambda ^{2}}) \notag \\
\rho ^{2}\left( \lambda ,\Lambda \right) &\equiv &1/2(\Lambda +\sqrt{%
4\lambda ^{2}+\Lambda ^{2}})\end{aligned}$$
where $\mu ^{2},\rho ^{2}>0$.
Beside the light scalar there is a heavy scalar $\Phi $, which is an isospin doublet. As far as the light scalar is concerned, it acts like the Higgs particle in the Standard Model. Besides this particle there is, however, the new heavy scalar$$\Phi ^{\prime }=%
\begin{pmatrix}
H^{+} \\
H_{0}+iH_{1}%
\end{pmatrix}%
.$$
with $\phi =\phi ^{\prime }\cos \theta -\Phi ^{\prime }\sin \theta $.
$S$ is, of course, an orthogonal transformation. Thus the purely kinetic part of the lagrangian is not altered. There is no need to introduce for this particle a self-interaction proportional to the field in the fourth power or any Yukawa couplings. But because of the $\phi ^{4}$ term in the scalar potential and the Yukawa coupling of $\phi $ there are couplings between the $\Phi ^{\prime }$ and the $\phi ^{\prime }$ as well as Yukawa like couplings for the heavy scalar, induced by the mixing. These new interactions depend on the ratio $\lambda /\Lambda $. The larger the mass difference between those particles, the more are non-standard-model couplings suppressed in the Yukawa and scalar sector.
The minimum of the Higgs potential is given by $$\left\langle \phi \right\rangle =\sqrt{\frac{\mu ^{2}}{2\kappa \cos \theta }}$$
and we have the tree-level relations$$M_{h}=\sqrt{2}\mu$$
and$$M_{H}=\left( 3M_{h}^{2}\tan ^{2}\theta +2\rho ^{2}\right) ^{1/2}$$
If the masses of the scalars were much heavier than the Higgs boson mass, the mixing between those particles would be strongly suppressed and could be neglected. For example, if the Higgs mass is $M_{h}=116~GeV$ and the new scalars mass is $M_{\phi ^{\prime }}\approx 16~TeV$, these couplings are suppressed by powers of the factor $\left\vert \sin \theta \right\vert
\approx 0.008$.
Neglecting the mixing terms it follows from $\left\langle \Phi ^{\prime
}\right\rangle =0$ that there are just four particle couplings between gauge bosons and $\Phi ^{\prime }$, so the new scalars have to decay into fermions. It decays almost exclusively to $t\bar{t}$ which makes it easily possible to distinguish $\Phi ^{\prime }$ from $\phi ^{\prime }$ experimentally.
We would like to consider as an example the heavy scalar to have a mass light enough to be produced at the LHC and the Higgs particle to have a mass of about $140$ $GeV $. In this case the mixing cannot be neglected and the scalars decay into fermions and - although still suppressed - into gauge bosons.
The heavy neutral particle $H_{0}$ decays predominantly into fermionic channels: $$H_{0}\rightarrow \bar{f}f$$
For the partial widths we find$$\Gamma \left( H_{0}\rightarrow \bar{f}f\right) =\tan ^{2}\theta \frac{G_{\mu
}N_{c}}{2\sqrt{2}\pi }M_{H_{0}}m_{f}^{2}\beta _{f}^{3}$$
with $\beta =\left( 1-4m_{f}^{2}/M_{H_{0}}^{2}\right) ^{1/2}$ being the velocity of the fermions and $N_{c}$ the color factor $N_{c}=3\left(
1\right) $ for quarks (leptons).
The heavy charged scalar $H^{\pm }$ decays as follows:
$$H^{\pm }\rightarrow \bar{t}b.$$
This decay channel suffers from a large irreducible background$$gg\rightarrow t\bar{t}g,~gq\rightarrow t\bar{t}q.$$
The partial width for $H^{\pm }$ can be calculated numerically (see [Hahn]{}).
The decay widths of heavy scalars into gauge bosons are suppressed compard to the Standard Model Higgs particle by the factor $\sqrt{2}\left(
M_{h}/M_{H_{0}}\right) ^{4}\tan ^{2}\theta .$ The partial widths for $H^{0}$ are given by$$\Gamma \left( H_{0}\rightarrow VV\right) =\frac{G_{\mu }}{16\pi }\left(
\frac{M_{h}}{M_{H_{0}}}\right) ^{4}M_{H_{0}}^{3}\tan ^{2}\theta ~\delta _{V}%
\sqrt{1-4x}\left( 1-4x+12x^{2}\right) ,~x=\frac{M_{V}^{2}}{M_{H_{0}}^{2}}$$
with $\delta _{W}=2$ and $\delta _{Z}=1$ [@Lee].
The largest contributions for the production cross section of a very heavy scalar are the gluon and vector boson fusion processes. The associative production with $W$ or $Z$ bosons is at least two orders of magnitude smaller for a scalar particle of more than $1~TeV$.
Reduced by a factor $\arctan ^{2}\theta \,$compared to a Higgs particle the gluon fusion cross section for a single $H_{0}$ can be expressed in leading order by [@Georgi]$$\sigma \left( gg\rightarrow H_{0}\right) =\tan ^{2}\theta \frac{G_{\mu
}\alpha _{s}^{2}}{144\pi }\left\vert \frac{3}{4}\sum_{q}A_{1/2}\left( \tau
_{Q}\right) \right\vert ^{2}$$
with $\tau _{Q}=M_{H_{0}}^{2}/4m_{Q}^{2}$ and the form factor$$A_{1/2}\left( \tau _{Q}\right) =2\left[ \tau _{Q}-\frac{1}{4}\left( \tau
_{Q}-1\right) \left( \log \frac{1+\sqrt{1-\tau _{Q}^{-1}}}{1-\sqrt{1-\tau
_{Q}^{-1}}}-i\pi \right) ^{2}\right] .$$
For calculation of the hadronic cross section $pp\rightarrow H_{0}$ we can use ref. [@Spira].
The cross section for the vector boson fusion production channel $%
qq\rightarrow V^{\ast }V^{\ast }qq\rightarrow H_{0}qq$ is suppressed by $%
\sqrt{2}\left( M_{h}/M_{H}\right) ^{4}\tan ^{2}\theta .$ To calculate the hadronic cross section we use the program VV2H [@Spira2].
The heavy quark associated production is a third production channel [AssocProc]{}. But this channel is in the here considered mass region at least two orders of magnitude smaller than the gluon fusion channel.
The only heavy scalar production channel which is not suffering from any mixing induced suppression mechanism is the scalar pair production by vector boson fusion. But this cross-section is far below $1$ $fb$ and thus not large enough to enhance the total production cross section decisively [Djouadi]{}.
There is also a fusion production process for the $H^{+}$.$$u\bar{d}\rightarrow H^{+}$$
But this process should be even smaller because the coupling between $H^{+}$ and the quarks is proportional to the fermion masses. Numerical results are shown in (Table \[Ergebnis\]).
To distinguish the scalar $H_{0}$ from the pseudoscalar $H_{1}$ it is possible to use the process $$H_{0}/H_{1}\rightarrow t\bar{t}\rightarrow \left( W^{+}b\right) \left( W^{-}%
\bar{b}\right)$$analogously to the Higgs boson case [@parity].
----------------------------------------------------------------------------------------------
$M_{H_{0}}=504$ $M_{H_{0}}=800$
----------------------------------------------------- -------------------- -------------------
$\lambda $ $3.5\cdot 10^{4}$ $5.6\cdot 10^{4}$
$\Lambda $ $11.5\cdot 10^{4}$ $31\cdot 10^{4}$
$M_{h_{0}}$ $140.10$ $140.04$
$\theta $ $15.66$ $9.93$
$\Gamma \left( H_{0}\rightarrow \bar{t}t\right) $ $1.74$ $2.18$
$\Gamma \left( H^{-}\rightarrow \bar{t}b\right) $ $1.86$ $1.36$
$\Gamma \left( H_{0}\rightarrow W^{+}W^{-}\right) $ $3.3\cdot 10^{-2}$ $%
8.9\cdot 10^{-3}$
$\Gamma \left( H^{-}\rightarrow W^{-}Z\right) $ $8.0\cdot 10^{-4}$ $%
2.1\cdot 10^{-4}$
$\sigma \left( pp\rightarrow H_{0}\right) $ $310$ $13$
$\sigma \left( pp\rightarrow H_{0}qq\right) $ $5.3\cdot 10^{-1}$ $%
1.1\cdot 10^{-2}$
----------------------------------------------------------------------------------------------
: The hadronic cross sections for gluon fusion $\protect\sigma \left(
pp\rightarrow H_{0}\right) $ and for vector boson fusion $\protect\sigma %
\left( pp\rightarrow qqH_{0}\right) $ are in $fb$. The partial widths and masses are in $GeV$. $\protect\alpha _{s}$ was evaluated at the partonic center of mass energy.[]{data-label="Ergebnis"}
The scalars decay almost exclusively into fermionic channels and so their detectability suffers from a large irreducible background. Especially for heavy scalar particles above $2$ $TeV$ for which the production cross section is far below $1~fb$ it seems to be very peculiar to detect them.
But for scalars with a mass below $1~TeV$ there is a good chance to find them at the LHC after collecting enough integrated luminosity.
We would like to thank D.N. Gao for valuable discussions.
[99]{} X. Calmet Eur. Phys. J. C, **28** (2003) 451-454 ; X.Calmet and J.F. Oliver *hep-ph/0606209* (2006).
R.N. Mohapatra and G. Senjanovic Phys. Rev. Lett., **44** (1980) 165; D. Chang, R.N. Mohapatra Phys. Rev.D **32** (1985) 1248.
Y. Kiyo, T. Morozumi, P. Parada, M.N. Rebelo and M. Tanimoto Prog.Theor.Phys.* ***101**(1999) 671; Z.G. Berezhiani, R.Rattazzi Phys. Lett.* *B,* ***279** (1992)124*;* A. Davidson, K.C. Wali Phys.Rev.Lett., **60** (1988) 1813.
D. Atwood, S. Bar-Shalom and A. Soni Eur. Phys. J. C, **45** (2006) 219; S. Bar-Shalom, D.Atwood and A. Soni, PoS HEP2005 (2006) 358.
M. Kramer,.J.H. Khun, M.L. Stong, P.M. Zerwas Z.Phys.** **C **64** (1994) 21** **
T. Hahn, M. Perez-Victoria Comput. Phys. Commun. **118** (1999) 153.
B.W. Lee, C. Quigg and H.B. Thacker Phys. Rev. D **16** (1977) 1519.
H. Georgi, S. Glashow, M. Machacek and D. Nanopoulos Phys. Rev. Lett.* ***40** (1978) 692
M. Spira Nucl.Instrum.Meth. A **389** (1997) 357** **
K. Jakobs *et al* Eur.Phys.J. C **32**, S2 (2004) 19.
A. Djouadi, W. Kilian, M. Mühlleitner and P.M. Zerwas Eur. Phys. J. C.,* ***45** (1999)** **10
J.N. Ng and P. Zakarauskas, TRI-PP-82/85 (1982).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the effect of spin-orbit interaction on one-dimensional U(1)-invariant frustrated magnets with dominant critical nematic fluctuations. The spin-orbit coupling explicitly breaks the U(1) symmetry of arbitrary global spin rotations about the high-symmetry axis down to $Z_2$ (invariance under a $\pi$-rotation). Given that the nematic order parameter is invariant under a $\pi$-rotation, it is relevant to ask if other discrete symmetries can be spontaneously broken. Here we demonstrate that the spin-orbit coupling induces a bond density wave that spontaneously breaks the translational symmetry and opens a gap in the excitation spectrum.'
address:
- '$^1$Department of Physics and Astronomy, University of Tennessee, Knoxville, Tennessee 37996-1200, USA'
- '$^2$Materials Science and Technology Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA'
- '$^3$Quantum Condensed Matter Division and Shull-Wollan Center, Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831, USA'
author:
- 'Shang-Shun Zhang$^{1}$, Nitin Kaushal$^{1}$, Elbio Dagotto$^{1,2}$, and Cristian D. Batista$^{1,3}$'
title: 'Spin-orbit interaction driven dimerization in one dimensional frustrated magnets'
---
Introduction
============
Frustrated magnetism is a continuous source of exotic states of matter that challenge the existing characterization probes [@ref:review1; @ref:review2]. Once quantum fluctuations melt the traditional magnetic long-range order, it often happens that the remaining liquid or multipolar orderings do not couple directly to the usual experimental probes. A simple example is the spin nematic phase proposed to be the ground of the one-dimensional (1D) $J_1-J_2$ Heisenberg model near its saturation field [@ref:nematic1; @ref:nematic2; @ref:nematic3; @ref:nematic4; @ref:nematic5; @ref:nematic5b; @ref:nematic5c; @ref:nematic5d; @ref:nematic6; @ref:nematic7; @ref:nematic8; @ref:nematic9; @ref:nematic10; @ref:nematic11]. This phase arises from a Bose-Einstein condensation of magnon pairs right below the saturation field $h_{\rm sat}$ [@ref:nematic1; @ref:nematic2; @ref:nematic3; @ref:nematic4; @ref:nematic5; @ref:nematic5b; @ref:nematic5c; @ref:nematic5d]. The attractive magnon-magnon interaction is generated by a ferromagnetic (FM) nearest neighbor (NN) exchange ($J_1<0$), which competes against an antiferromagnetic (AFM) next-nearest-neighbor (NNN) exchange $J_2>0$. $J_2$ must be bigger than $|J_1|/4$ for the zero field ground state not to be ferromagnetic.
Several quasi-1D materials are approximately described by the $J_1-J_2$ model with FM and AFM exchange interactions $J_1$ and $J_2$, respectively. Known examples include Rb$_2$Cu$_2$Mo$_3$O$_{12}$, [@ref:1DmaterialA1; @ref:1DmaterialA2], LiCuVO$_4$, [@ref:1Dmaterial2; @ref:1Dmaterial3; @ref:1Dmaterial4; @ref:1Dmaterial6; @LiCuVO4; @LiCuVO4b; @ref:1Dmaterial7; @ref:1Dmaterial8; @ref:1Dmaterial9; @ref:1Dmaterial10; @ref:1Dmaterial11] LiCuSbO$_4$, [@LiCusbO4; @grafe2016] PbCuSO$_4$(OH)$_2$, [@PbCuSO4OH2] which span a wide range of $J_2/\rvert J_1\rvert$ values. However, a direct experimental observation of the predicted nematic ordering is rahter challenging. [@sato09; @sato11] Given the symmetry of the order parameter, the nematic spin ordering is expected to induce a local quadrupolar electric moment via the always present spin-orbit coupling combined with the lattice anisotropy. However, *the spin anisotropy induced by this combination has the additional effect of breaking the global U(1) symmetry of spin rotations along the magnetic field axis down to a finite group*. For most of the known compounds, this group is not bigger than $Z_2$ for any direction of the applied magnetic field (only C$_2$ rotation axes). This fact raises another concern because the nematic order parameter $\langle S_r^{+}S_{r+1}^+\rangle$ does not break the remaining $Z_2$ group. In other words, the nematic order parameter is invariant under $\pi$-rotations. This simple observation implies that if some form of ordering still exists right below the high field paramagnetic phase of these compounds, it should not be called “nematic ordering”. Nevertheless, the dominant nematic susceptibility of the U(1) invariant model may still induce additional symmetry breaking in the presence of spin anisotropy. If this is the case, it is important to identify those discrete symmetries.
![(Color online) (a) The two scenarios of nematic bond order parameter $\Psi_{r}=\langle S_{r}^{+}S_{r+1}^{+}\rangle = \langle {\cal O}^{a} \rangle + i \langle {\cal O}^{b} \rangle$: open and full circles represent the real, $\langle {\cal O}^{a} \rangle$, and imaginary, $\langle {\cal O}^{b} \rangle$ parts of the nematic order parameter, respectively. (b) Lattice distortions induced via spin-orbit coupling by the real (upper panel) and the imaginary (lower panel) parts of the bond nematic order parameter. Translational symmetry is broken in both cases, but the lattice distortion takes place along different directions. The bigger circles represent the magnetic transition-metal ions. The smaller circles represent the oxygen atoms that mediate the super-exchange interaction.[]{data-label="fig:dimerization"}](dimerization.pdf)
In this work we investigate the relevant effect of spin-orbit interaction on the 1D frustrated $J_1-J_2$ model. Based on our previous considerations, there are two possible scenarios: i) The field-induced transition from a quantum paramagnet to the nematic phase is replaced by a crossover (no discrete symmetry breaking); ii) The field-induced transition from a quantum paramagnet to the nematic phase is replaced by a discrete symmetry breaking. We can anticipate that the problem under consideration belongs to the second case because it is known that the magnon-pairs condense at a finite wave vector $\pm Q$. In other words, the nematic ordering breaks the translational symmetry, which is not affected by the inclusion of the spin-orbit interaction. In addition, the system has an additional $Z_2$ symmetry besides the $\pi$-rotation about the $z$-axis. This symmetry is the product of two operations, ${\cal T} {\cal R}(\pi)$, where ${\cal T}$ is the time reversal operator and ${\cal R}(\pi)$ is a $\pi$-rotation operator about an axis perpendicular to the field direction. As we explain below, the real part of the nematic order parameter, ${\cal O}^a$, preserves this $Z_2$ symmetry, while the imaginary part, ${\cal O}^b$, does not. Fig. \[fig:dimerization\] (a) shows the the different broken symmetry states associate with the stabilization of the real or imaginary part of the nematic order parameter.
In the following sections, we derive a phenomenological Ginzburg-Landau theory that is complemented by microscopic analytical and numerical calculations. Moreover, we demonstrate that the combination of a divergent nematic susceptibility and spin anisotropy stabilizes the real component of the nematic order parameter, ${\cal O}^a$, which in turn produces bond dimerization in most of the phase diagram. This bond ordering is accompanied by an orthorhombic distortion of the surrounding oxygen octahedron, as it is schematically shown in the upper panel of Fig. \[fig:dimerization\] (b). In contrast, the imaginary component of the “nematic” order parameter, ${\cal O}^b$, does not produce bond dimerization. If stabilized by other mechanisms, this phase would produce a local orthorhombic distortion of the surrounding oxygen octahedron along diagonal directions, as illustrated in the lower panel of Fig. \[fig:dimerization\] (b).
Given that bond dimerization couples to the lattice via magneto-elastic interaction and lowers the space group of the material under consideration, the combined effect of high spin nematic susceptibility and spin anisotropy can in principle be detected with X-rays. In addition, the incommensurate bond-density wave expected for smaller values of $J_2/|J_1|$ should lead to a double-horn shape of the nuclear magnetic resonance (NMR) line. These conclusions can shed light on the search for the spin “nematic ordering” predicted on the basis of a U(1) invariant $J_1-J_2$ Heisenberg model.
The structure of this paper is as follows. In Sec. II, we introduce a simple model Hamiltonian in which the U(1) symmetry is reduced to Z$_2$ via the inclusion of an Ising term (symmetric anisotropy). In Sec. III we consider a simple Ginzburg-Landau (GL) theory, which describes the possible scenarios. The phenomenological input parameters of the GL theory are calculated in Sec. IV by means of an analytical approach to the microscopic Hamiltonian. The results of this analytical approach are confirmed by numerical Density Matrix Renormalization Group (DMRG) calculations presented in Sec. V. The general implications of our results for experimental studies of unconventional magnetic ordering in quasi-1D frustrated compounds are discussed in Sec. VI.
Model Hamiltonian
=================
We consider a spin-$1/2$ anisotropic Heisenberg model on a 1D chain with ferromagnetic nearest-neighbor exchange, $J_{1}<0$, and antiferromagnetic next nearest neighbor exchange $J_{2}>0$: $$\begin{aligned}
\mathcal{H} & =J_{1}\sum_{j}{\bf S}_{j}\cdot{\bf S}_{j+1}+J_{2}\sum_{j}{\bf S}_{j}\cdot{\bf S}_{j+2}-h\sum_{j}S_{j}^{z} \nonumber \\
& \;\;+\alpha\sum_{j}S_{j}^{x}S_{j+1}^{x},\label{eq:model}\end{aligned}$$ The last term is an Ising interaction that in real materials arises from the combined effect of spin-orbit coupling and lattice anisotropy. This term reduces the U(1) symmetry of continuous spin rotations about the field axis to $Z_2$. We note that the U(1) symmetry is restored if the magnetic field is applied along the $x$-direction. However, in real quasi-1D materials, the U(1) symmetry is not present for any direction of the magnetic field because of the interaction with other chains [@ref:nematic1]. In spite of these considerations, the U(1) invariant model has been invoked to describe various quasi-1D transition metal compounds. [@sato13]
Based on the U(1) invariant model ($\alpha=0$), several authors proposed that nematic quasi-long range ordering should be observed right below the saturation field $h_{\rm sat}$ [@ref:nematic1; @ref:nematic2; @ref:nematic3; @ref:nematic4; @ref:nematic5; @ref:nematic5b; @ref:nematic5c; @ref:nematic5d]. $h_{\rm sat}$ is finite only for $J_2 > |J_1|/4$ because the zero field ground state is ferromagnetic for $J_2 \leq |J_1|/4$. The nematic ordering corresponds to a Bose-Einstein condensation of two-magnon bound states with a local order parameter $\langle S_{j}^{+}S_{j+1}^{+}\rangle$. The attractive magnon-magnon interaction is provided by the ferromagnetic nearest-neighbor exchange $J_1$. The ratio $ J_{2}/\lvert J_{1} \rvert$ controls the total momentum $\pm Q$ of the two-magnon bound state. $Q$ is incommensurate for small values of $ J_{2}/\lvert J_{1} \rvert$ and it is equal to $\pi$ for $ J_{2}/\lvert J_{1} \rvert \geq 0.375$.
In general, the continuous SU(2) symmetry of the Heisenberg interaction is broken down to a discrete symmetry group in real materials [@ref:1DmaterialA2; @ref:1Dmaterial2]. Even for an idealized single-chain system, the exchange interaction turns out to be anisotropic, instead of SU(2) invariant, once the spin-orbit interaction is included. This is so because an isolated chain has only one symmetry axis parallel to the chain direction ($x$-direction in our notation). In other words, the exchange interaction between spin components parallel to the chain direction is different from the exchange interaction between the spin components perpendicular to the chain direction, as it is clear from the $\alpha$-term in Eq. . Consequently, the pure 1D Hamiltonian has only discrete point group symmetries if the external magnetic field is not parallel to the chain direction. For the case under consideration (magnetic field perpendicular to the chain direction), the U(1) symmetry of the Heisenberg model is reduced to a discrete $Z_{2}$ symmetry corresponding to a $\pi$-rotation about the $z$-axis ${\cal R}_{z}(\pi):\;S_{r}^{z}\rightarrow S_{r}^{z},S_{r}^{x,y}\rightarrow-S_{r}^{x,y}$. Correspondingly, the spin components $S_{r}^{\pm}$ transform into $e^{\pm i\phi} S_{r}^{\pm}$ under a rotation by $\phi$ about the $z$-axis. This means that the nematic order parameter, $\langle S_{r}^{+} S_{r}^{+} \rangle$, transforms into $ e^{i 2 \phi} \langle S_{r}^{+} S_{r}^{+} \rangle$, implying that it is invariant under $\pi$-rotations, as expected for a director. The inclusion of spin anisotropy then forces us to reconsider the problem because other symmetries (different from rotations) have to be invoked to characterize the phase that replaces the nematic quasi-long range ordering.
Besides the above mentioned $\pi$ rotation about the $z$-axis, the Hamiltonian of Eq. is invariant under the product of the time reversal operation and a $\pi$-rotation about the $y$-axis, ${\cal T}{\cal R}_{y}(\pi)$, which changes the sign of the $y$ spin component: $S_{r}^{x,z}\rightarrow S_{r}^{x,z},S_{r}^{y}\rightarrow-S_{r}^{y}$. The real-part of the nematic order parameter, $$\Re\langle S_{r}^{+}S_{r+1}^{+}\rangle=\langle S_{r}^{x}S_{r+1}^{x}-S_{r}^{y}S_{r+1}^{y}\rangle \equiv \langle {\cal O}^a (r) \rangle,$$ remains invariant under this transformation. In contrast, the imaginary part, $$\Im\langle S_{r}^{+}S_{r+1}^{+}\rangle=\langle S_{r}^{x}S_{r+1}^{y}+S_{r}^{y}S_{r+1}^{x}\rangle,$$ changes sign. Finally, the nematic order parameter breaks the translational symmetry because the magnon-pairs condense at a finite momentum $\pm Q$. This symmetry is then expected to break spontaneously for $\alpha \neq 0$, as long as $Q$ is commensurate.
Based on this simple symmetry analysis, the spin anisotropy should stabilize a state that breaks the translational symmetry (in a strong or a weak sense) and select either the real or the imaginary part of the original nematic order parameter. Only one of these two components should be selected, as supposed to some linear combination, because they belong to different irreducible representations of the point group of $\mathcal{H}$.
Ginzburg-Landau theory {#GG}
======================
The attractive interaction between magnons arising from the ferromagnetic nearest-neighbor interaction, leads to two-magnon bound states for $J_{2}/|J_{1}|>1/4$. The minimum energy of the two-magnon bound state is achieved for a finite value, $\pm Q$, of the center of mass momentum. The two-magnon bound states condense for $ h < h_{c}$ (note that $h_c$ is higher than the field required to close the single-magnon gap). The two-magnon condensate is characterized by a two component complex order parameter $\Psi_{\pm Q}$ (macroscopic wave function of condensate) whenever $Q \neq -Q$. The spin-orbit interaction generates an effective coupling between these two components, as it can be inferred from the lowest order expansion of the Ginzburg-Landau free energy: $${\cal F}=r\left(|\Psi_{Q}|^{2}+|\Psi_{\bar{Q}}|^{2}\right)+u\left(\Psi_{Q}^{*}\Psi_{\bar{Q}}^{*}+\Psi_{\bar{Q}}\Psi_{Q}\right),$$ where $\bar{Q}=-Q$. Due to the $Z_{2}$ symmetry restriction, the complex field $\Psi_{\pm Q}$ is fixed up to a phase factor $\pm1$. We have also assumed that $u$ is real based on the underlying microscopic theory. We will first assume $Q=\pi$ which is the condensation wave vector for $J_{2}>J_{2c}$ ($J_{2c}\simeq 0.375\rvert J_1 \rvert$ for $\alpha=0$). Given that $Q=\pi$ is invariant under spatial inversion, we have $\Psi_{\pi}=\Psi_{-\pi}$. Then, upon minimization of the free energy, we obtain a real order parameter $\Psi_{\pi}$ for $u<0$, and a purely imaginary order parameter for $u>0$. In the original spin language, we have $\Psi_{\pi}=\frac{1}{N}\sum_{r}e^{i\pi r}\langle S_{r}^{-}S_{r+1}^{-}\rangle$, whose real and imaginary parts are $$\begin{aligned}
{\cal O}^{a} & =\sum(-1)^{r}\langle S_{r}^{x}S_{r+1}^{x}-S_{r}^{y}S_{r+1}^{y}\rangle,\\
{\cal O}^{b} & =-\sum(-1)^{r}\langle S_{r}^{x}S_{r+1}^{y}+S_{r}^{y}S_{r+1}^{x}\rangle.\end{aligned}$$ A real order parameter only breaks the translational symmetry, while an imaginary order parameter breaks additional symmetries, such as, ${\cal T}{\cal R}_{y}(\pi):\;S_{r}^{x,z}\rightarrow S_{r}^{x,z},S_{r}^{y}\rightarrow-S_{r}^{y}$. In both cases, the system should develop long-range ordering at $T=0$ because only discrete symmetries are broken. We note that the spin anisotropy corresponds to a uniform nematic field that couples linearly to the uniform component of the nematic order parameter $\Psi_{0}=\frac{1}{N}\sum_{r}\langle S_{r}^{-}S_{r+1}^{-}\rangle$, implying that $\Psi_{0}$ becomes finite for a finite $\alpha$. As we will see next, the interference between $\Psi_{0}$ and the $\pi$ component, $\Psi_{\pi}$, of the [*real*]{} part of the order parameter leads to a real space modulation (dimerization) of the expectation value of nearest-neighbor bond operators.
In general, the nematic order parameter is a complex number, $|\Psi_{\pi}|e^{i\theta}$, where $\theta=0$ and $\theta=\frac{\pi}{2}$ correspond to ${\cal O}^{a}$ and ${\cal O}^{b}$, respectively. The real space version of these order parameters is obtained via a Fourier transformation, $$\begin{aligned}
{\cal O}^{a}(r) & =\Psi_{0}+2(-1)^{r}\Psi_{\pi}\cos(\theta),\\
{\cal O}^{b}(r) & =2(-1)^{r}\Psi_{\pi}\sin(\theta),\end{aligned}$$ which gives the real and imaginary parts of $\Psi_{r}=\langle S_{r}^{-}S_{r+1}^{-}\rangle$. It is clear that the amplitude of the real component, ${\cal O}^{a}(r)$, is modulated in real space for finite values of the spin-orbit interaction ($\Psi_{0} \neq 0$). In contrast, only the phase is modulated for $\alpha=0$. In other words, the spin-orbit coupling induces a [*dimerized bond ordering*]{} if the real component of the original nematic order parameter is selected. This interference between the $0$ and $\pi$ components of ${\cal O}^{a}$ also leads to a [*magnon pair*]{} density wave: $$\begin{aligned}
n_{pair}(r) & =\langle\Psi_{r}^{\dagger}\Psi_{r}\rangle\simeq\langle\Psi_{r}\rangle^{*}\langle\Psi_{r}\rangle\nonumber \\
& =\Psi_{0}^{2}+4\Psi_{\pi}^{2}+4(-1)^{r}\Psi_{0}\Psi_{\pi}\cos(\theta).\end{aligned}$$ It follows that the long range ${\cal O}^{a}$ ordering is accompanied by another bond ordering associated with the longitudinal spin component $$\langle S_{r}^{z}S_{r+1}^{z}\rangle\simeq\langle\Psi_{r}\rangle^{*}\langle\Psi_{r}\rangle+ const..
\label{mod}$$ This is just the usual bond dimerization that appears in spin-Peierls systems. [@Miller82] Indeed, Eq. implies that the usual bond order parameter, $\langle{\bf S}_{r}\cdot{\bf S}_{r+1}\rangle$, must also exhibit dimerization.
The condensation wave vector becomes incommensurate ($Q<\pi$) for smaller values of $J_{2}/|J_{1}|$ (about $0.375$ for $\alpha=0$). In this case we need to consider a two component order parameter with phases $e^{i\theta_{\pm}}$ for $\pm Q$. Minimization of the free energy leads to $\theta_{+}+\theta_{-}=0\text{ mod }(2\pi)$ for $u<0$ and to $\phi_{+}+\phi_{-}=\pi\text{ mod }(2\pi)$ for $u>0$. The complex order parameter does not have a fixed phase because of the additional $U(1)$ phase factor $e^{\pm iQr}$, arising from translational symmetry. This $U(1)$ symmetry precludes long-range order for the single-chain problem. The free energy minimization also leads to the same amplitude for both components of the order parameter: $|\Psi_{Q}|=|\Psi_{\bar{Q}}|$, implying that the ground state must exhibit quasi-long range bond density wave order ${\cal O}^{a}(r)=\Psi_{0}+2|\Psi_{Q}|\cos(Qr)$ or ${\cal O}^{b}(r)=2|\Psi_{Q}|\sin(Qr)$.
Microscopic theory
==================
![(Color online) Single-magnon dispersion produced from the non-interacting part of the Hamiltonian \[see Eq \] with $\alpha=0$. The dashed line corresponds to the Lifshitz point where the single-magnon dispersion becomes quartic at low-enrgies because the single minimum splits into two minima.[]{data-label="fig:spectrum"}](magnon2.pdf)
Our discussion in the previous section indicates that two different kinds of bond order can be induced by spin anisotropy. To determine which order parameter is selected we need to consider the underlying microscopic theory. To this end, we use the Jordan-Wigner transformation to reformulate the spin Hamiltonian (\[eq:model\]) as a model for interacting spinless fermions: $$\begin{aligned}
S_{j}^{+} & =e^{-i\pi\sum_{k=1}^{j-1}n_{k}}c_{j}^{\dagger},\\
S_{j}^{-} & =e^{i\pi\sum_{k=1}^{j-1}n_{k}}c_{j},\\
S_{j}^{z} & =c_{j}^{\dagger}c_{j}-\frac{1}{2},\end{aligned}$$ where $c_{j}$ is the fermionic spinless operator which represents a spin flip on site $j$. The fermionic Hamiltonian is defined as $${\cal H}={\cal H}_{0}+{\cal H}_{int},$$ where $$\begin{aligned}
\label{eq:H0}
{\cal H}_{0} & =\sum_{k}c_{k}^{\dagger}c_{k}\left(J_{1}\cos k+J_{2}\cos2k-\left(h+J_{1}+J_{2}\right)\right)\nonumber \\
& +i\frac{\alpha}{4}\sum_{k}\sin k\left(c_{k}^{\dagger}c_{-k}^{\dagger}-c_{-k}c_{k}\right).\end{aligned}$$ For $\alpha=0$, the single-particle spectrum, corresponding to single-magnon excitations, has a minimum at $k_{0}=\arccos(-\frac{J_{1}}{4J_{2}})$ if $4J_{2}>|J_{1}|$ and at $k_{0}=0$ if $4J_{2}<|J_{1}|$. Fig. \[fig:spectrum\] shows the evolution of the single-particle spectrum with increasing $J_2/\rvert J_1\rvert$. The spin-orbit interaction ($\alpha \neq 0$) breaks the $U(1)$ symmetry and the fermion number is no longer conserved. The interacting part of the Hamiltonian is $$\begin{aligned}
{\cal H}_{int} & =\frac{1}{2N}\sum_{Q,q,p}U(Q,q,p)c_{\frac{Q}{2}+p}^{\dagger}c_{\frac{Q}{2}-p}^{\dagger}c_{\frac{Q}{2}-q}c_{\frac{Q}{2}+q},\end{aligned}$$ where $$\begin{aligned}
U(Q,q,p) & =2({J_{1}}+2J_{2}\cos Q)\sin(q)\sin(p)\\
& +2{J_{2}}\sin(2q)\sin(2p).\end{aligned}$$ The effective attractive interaction between nearest neighbor sites, $J_{1}+2J_{2}\cos Q$, is maximized at $Q=\pi$. Therefore, the lowest energy “two-magnon” bound state is expected to have this momentum for large enough $J_2/|J_1|$.
In the isotropic limit, the nematic phase corresponds to a magnon pair condensate. This state can be approximated by a coherent state built with the two-particle wave function of the bound state. In presence of magnetic anisotropy, the particle number is not conserved because particles can be created or annihilated in pairs. The Hamiltonian eigenstates can then be grouped into two categories based on the particle number parity. The following analysis assumes that the magnetic anisotropy is weak ($\alpha \ll 1$), implying that the finite particle density induced by the spin anisotropy term above and at the critical field, $h_c$ can be made arbitrarily small for small enough values of $\alpha$. This condition also guarantees that two magnon condensation is still the dominant instability for $\alpha \neq 0$. For this reason, we will still refer to the bound state as a “two-magnon” bound states (adiabatic continuation) and we will use the two-particle Green’s function to compute the energy of the two-magnon modes. The two-particle Green’s function is obtained from the two-particle scattering amplitude by solving the corresponding Bethe-Salpeter equation.
Bogoliubov representation
-------------------------
The above non-interacting fermionic Hamiltonian can be diagonalized with a Bogoliubov transformation, $$\begin{aligned}
c_{k} & =\left(u_{k}\alpha_{k}+v_{k}\alpha_{-k}^{\dagger}\right)e^{i\frac{\pi}{4}},\\
c_{k}^{\dagger} & =\left(u_{k}\alpha_{k}^{\dagger}+v_{k}\alpha_{-k}\right)e^{-i\frac{\pi}{4}},\end{aligned}$$ where $$\begin{aligned}
u_{k} & =\sqrt{\frac{1}{2}\left(1+\frac{\epsilon_{k}}{E_{k}}\right)},\\
v_{k} & =-sign(\Delta_{k})\sqrt{\frac{1}{2}\left(1-\frac{\epsilon_{k}}{E_{k}}\right)},\end{aligned}$$ with $E_{k}=\sqrt{\epsilon_{k}^{2}+4\Delta_{k}^{2}}$, $\Delta_{k}=\frac{\alpha}{4}\sin k$ and $$\epsilon_{k}=J_{1}\cos k+J_{2}\cos2k-\left(h+J_{1}+J_{2}\right).$$ The diagonal Hamiltonian $${\cal H}_{0}=\sum_{k}E_{k}\alpha_{k}^{\dagger}\alpha_{k}+E_{0}.$$ leads to the non-interacting Green’s function $$\begin{aligned}
iG_{0}(k,\omega) & =\langle0|{\cal T}\alpha_{k}^{\dagger}\alpha_{k}|0\rangle=\frac{i}{\omega-E_{k}+i0^{+}}.\end{aligned}$$
The next step is to write the interaction vertex in terms of the Bogoliubov quasi-particle operators. The normal interaction term is $$\begin{aligned}
{\cal H}_{int}^{N} & =\frac{1}{4N}\sum_{Q,q,p}\Gamma_{Q}^{(0)N}(q,p)\alpha_{\frac{Q}{2}+p}^{\dagger}\alpha_{\frac{Q}{2}-p}^{\dagger}\alpha_{\frac{Q}{2}-q}\alpha_{\frac{Q}{2}+q},\end{aligned}$$ with a normal interaction vertex
$$\begin{aligned}
\label{eq:normalV}
\Gamma_{Q}^{(0)N}(q,p) & =2U(Q,q,p)\times\left(u_{\frac{Q}{2}+p}u_{\frac{Q}{2}-p}u_{\frac{Q}{2}-q}u_{\frac{Q}{2}+q}+v_{\frac{Q}{2}+p}v_{\frac{Q}{2}-p}v_{\frac{Q}{2}-q}v_{\frac{Q}{2}+q}\right)\\
& -4U(p+q,\frac{Q-p+q}{2},\frac{Q+p-q}{2})\left(u_{\frac{Q}{2}+p}v_{\frac{Q}{2}-p}v_{\frac{Q}{2}-q}u_{\frac{Q}{2}+q}+u_{\frac{Q}{2}-p}v_{\frac{Q}{2}+p}v_{\frac{Q}{2}+q}u_{\frac{Q}{2}-q}\right).\end{aligned}$$
We can verify that $\Gamma_{Q}^{(0)N}(q,p)=\Gamma_{Q}^{(0)N}(-q,-p)$ due to the fermionic statistics. Furthermore, $\Gamma_{Q}^{(0)N}(q,p)=\Gamma_{\bar{Q}}^{(0)N}(-q,-p)=\Gamma_{\bar{Q}}^{(0)N}(q,p)$ because of inversion symmetry. The interaction vertex has been symmetrized with respect to the exchange of external lines. The anomalous interaction terms of the form $\alpha^{\dagger}\alpha^{\dagger}\alpha^{\dagger}\alpha^{\dagger}$ and $\alpha\alpha\alpha\alpha$ are $$\begin{aligned}
{\cal H}_{int}^{A} & =\frac{1}{4!N}\sum_{Q,q,p}\Gamma_{Q}^{(0)A}(q,p)\left(\alpha_{\frac{Q}{2}+p}^{\dagger}\alpha_{\frac{Q}{2}-p}^{\dagger}\alpha_{-\frac{Q}{2}+q}^{\dagger}\alpha_{-\frac{Q}{2}-q}^{\dagger}+h.c.\right),\end{aligned}$$ with an anomalous interaction vertex $$\begin{aligned}
\label{eq:anomalousV}
\Gamma_{Q}^{(0)A}(q,p) & =2U(Q,q,p)\left(u_{\frac{Q}{2}+p}u_{\frac{Q}{2}-p}v_{\frac{Q}{2}-q}v_{\frac{Q}{2}+q}+v_{\frac{Q}{2}+p}v_{\frac{Q}{2}-p}u_{\frac{Q}{2}-q}u_{\frac{Q}{2}+q}\right)\\
& -2U(p+q,\frac{Q-p+q}{2},\frac{Q+p-q}{2})\left(u_{\frac{Q}{2}+p}v_{\frac{Q}{2}-p}u_{\frac{Q}{2}-q}v_{\frac{Q}{2}+q}+v_{\frac{Q}{2}+p}u_{\frac{Q}{2}-p}v_{\frac{Q}{2}-q}u_{\frac{Q}{2}+q}\right)\\
& +2U(p-q,\frac{Q+p+q}{2},\frac{Q-p-q}{2})\left(v_{\frac{Q}{2}+p}u_{\frac{Q}{2}-p}u_{\frac{Q}{2}-q}v_{\frac{Q}{2}+q}+u_{\frac{Q}{2}+p}v_{\frac{Q}{2}-p}v_{\frac{Q}{2}-q}u_{\frac{Q}{2}+q}\right).\end{aligned}$$ We can verify that $\Gamma_{Q}^{(0)A}(q,p)=\Gamma_{\bar{Q}}^{(0)A}(p,q)=\Gamma_{Q}^{(0)A}(-q,-p)$ due to fermionic statistics and $\Gamma_{Q}^{(0)A}(q,p)=\Gamma_{\bar{Q}}^{(0)A}(-q,-p)=\Gamma_{\bar{Q}}^{(0)A}(q,p)$ due to inversion symmetry. This interaction vertex has also been symmetrized with respect to the exchange of external lines.
The remaining anomalous terms ($\alpha^{\dagger}\alpha^{\dagger}\alpha^{\dagger}\alpha$ and $\alpha\alpha\alpha\alpha^{\dagger}$) will not be considered because they give subdominant contributions (in an expansion in powers of $\alpha$) to the effective anomalous interaction vertex shown in Fig. \[fig:BSE\]. We note that the nature of the order parameter is determined by the effective anomalous interaction vertex because the effective normal vertex preserves the U(1) symmetry of ${\cal H}(\alpha=0)$.
Finally, we also note the existence of additional diagrams which are obtained by exchange of the external legs of the diagrams shown in Fig. \[fig:BSE\]. These diagrams are of the same order in an expansion in powers of $\alpha$. However, for the purpose of calculating the two magnon bound state, they can be neglected because they give contributions that remain regular in the proximity of the critical field $h=h_c$.
![Ladder diagrams for the normal and anomalous scattering amplitudes $\Gamma^{N}$ and $\Gamma^{A}$.[]{data-label="fig:BSE"}](BSEs.pdf)
Bethe-Salpeter equation
-----------------------
In the dilute limit, the scattering amplitude can be calculated by summing up the ladder diagrams shown in Fig. \[fig:BSE\]. This sum leads to the Bethe-Salpeter equation [@BSE]: $$\begin{aligned}
\Gamma_{\omega Q}^{N}(q,p) & =\Gamma_{Q}^{(0)N}(q,p)-\frac{1}{2N}\sum_{k}\frac{\Gamma_{Q}^{(0)N}(Q;q,k)\Gamma_{\omega Q}^{N}(k,p)}{E_{\frac{Q}{2}+k}+E_{\frac{Q}{2}-k}-\omega-i0^{+}}-\frac{1}{2N}\sum_{k}\frac{\Gamma_{Q}^{(0)A}(q,k)\Gamma_{\omega Q}^{A}(k,p)}{E_{\frac{Q}{2}+k}+E_{\frac{Q}{2}-k}+\omega-i0^{+}},\label{eq:BSeq1}\\
\Gamma_{\omega Q}^{A}(q,p) & =\Gamma_{Q}^{(0)A}(q,p)-\frac{1}{2N}\sum_{k}\frac{\Gamma_{Q}^{(0)N}(q,k)\Gamma_{\omega Q}^{A}(k,p)}{E_{\frac{Q}{2}+k}+E_{\frac{Q}{2}-k}+\omega-i0^{+}}-\frac{1}{2N}\sum_{k}\frac{\Gamma_{Q}^{(0)A}(q,k)\Gamma_{\omega Q}^{N}(k,p)}{E_{\frac{Q}{2}+k}+E_{\frac{Q}{2}-k}-\omega-i0^{+}},\label{eq:BSeq2}\end{aligned}$$
where $\Gamma_{Q}^{(0)N/A}$ is the bare scattering amplitude of the normal and anomalous type. The energy of the magnon pair bound state can be extracted from the poles of the scattering amplitude. For a wide range of $J_{2}/|J_{1}|$ values, the bound state dispersion has its minimum at $Q=\pi$. The energy of the two-magnon bound state increases with $\alpha$, implying that the critical field for the “two-magnon” condensation decreases relative to the saturation field of the isotropic ($\alpha=0$) Hamiltonian. For the isotropic Heisenberg model, the condensate wave function has a $U(1)$ phase freedom. The spin anisotropy reduces this freedom to $Z_{2}$. The Ginzburg-Landau theory tells us that the phase of the macroscopic wave function can either be real or imaginary depending on the sign of the effective anomalous coupling parameter $u$. In this section we determine the phase of the wave function and also include a microscopic calculation of the parameter $u$. These properties are enclosed in the scattering amplitude obtained from the solution of the Bethe-Salpeter equation. Our analysis shows that wave function of the magnon pair condensate is real, implying that the dominant order parameter is ${\cal O}^{a}$.
### Wave function of the bound state
We start by introducing the two-magnon Green’s function, which can be easily obtained through the scattering amplitude $\Gamma_{\omega Q}^{N/A}(q,p)$: $$\begin{aligned}
G^{(2)}(\omega,Q;q,p) & =G_{0}^{(2)}(\omega,Q;,q,p)+\frac{1}{4}G_{0}^{(2)}(\omega,Q;q,q^{\prime})\nonumber \\
& \times\Gamma_{\omega Q}^{N}(q^{\prime},p^{\prime})G_{0}^{(2)}(\omega,Q;p^{\prime},p),\label{eq:2pGFa}\\
G_{A}^{(2)}(\omega,Q;q,p) & =\frac{1}{4}G_{0}^{(2)}(-\omega,-Q;-q,-q^{\prime})\nonumber \\
& \times\Gamma_{\omega Q}^{A}(q^{\prime},p^{\prime})G_{0}^{(2)}(-\omega,-Q;-p^{\prime},-p),\label{eq:2pGFb}\end{aligned}$$ where $G_{0}^{(2)}$ is the non-interacting two particle Green’s function: $$G_{0}^{(2)}(\omega,Q;q,q^{\prime})=\frac{\delta(q-q^{\prime})-\delta(q+q^{\prime})}{\omega-E_{\frac{Q}{2}+q}-E_{\frac{Q}{2}-q}+i0^{+}}.$$ The Lehmann representation shows explicitly that the two particle Green’s function has the following singular behavior near the pole of bound state, $\omega\sim\omega_{B}$: $$\begin{aligned}
G_{N}^{(2)}(\omega,Q;q,p) & \sim\frac{\psi_{Q}(p)\psi_{Q}^{*}(q)}{\omega-\omega_{B}+i0^{+}}+\text{regular terms},\label{eq:singular1}\\
G_{A}^{(2)}(\omega,Q;q,p) & \sim\frac{\psi_{Q}(p)\phi_{-Q}(q)}{\omega-\omega_{B}+i0^{+}}+\text{regular terms},\label{eq:singular2}\end{aligned}$$ where the regular terms come from higher excited states and $\omega_{B}>0$ is the bound state energy relative to the ground state. The projection of the bound state wave function on the two-magnon sector is obtained from the residue of the pole, $$\begin{aligned}
\psi_{Q}(p) & =\langle G|\alpha_{\frac{Q}{2}+p}\alpha_{\frac{Q}{2}-p}|B\rangle_{Q},\label{eq:bdwf1}\\
\phi_{\bar{Q}}^{*}(p) & =\langle G|\alpha_{-\frac{Q}{2}-p}^{\dagger}\alpha_{-\frac{Q}{2}+p}^{\dagger}|B\rangle_{Q},\label{eq:bdwf2}\end{aligned}$$ where $|B\rangle$ is the ket of the bound state and $|G\rangle$ the ground state. Due to the anomalous interaction arising from the spin anisotropy term, the bound state wave function is a linear combination of states with different particle number. The poles of the two particle Green’s function given by Eqs. (\[eq:2pGFa\]) and (\[eq:2pGFb\]) are obtained after inserting the scattering amplitude, $\Gamma_{\omega Q}^{N/A}(q,p)$, which results from the Bethe-Salpeter equation. According to Eqs. ($\ref{eq:bdwf1}$) and (\[eq:bdwf2\]), the bound state wave function is then obtained by extracting the residue near the pole $\omega_{B}$. For $Q=\pi$, the bound state wave functions, $\psi_{\pi}(p)$ and $\phi_{\pi}(p)$, are even under spatial inversion. This result is found to be correct for all $J_{2}/|J_{1}|$ ratios and for any value of the bound state energy $\omega_{B}$, indicating that the broken symmetry state below the critical field must preserve the spatial inversion symmetry.
![Residue of the two particle normal/anomalous Green’s function near the pole $\omega_{B}$ in frequency space for: (a) $\omega_{B}/|J_{1}|=0.03936$ and (b) $\omega_{B}/|J_{1}|=0.0004$ (close to the “two-magnon” condensation point). The other parameters are taken as $J_{2}/|J_{1}|=1,\alpha=0.2$, momentum $p=0.9\pi$. We find the normal (blue solid line) and anomalous (red dotted line) two-particle Green’s function become the same upon approaching the critical condensation point.[]{data-label="fig:residue"}](residue_GF.pdf)
As the system approaches the critical field corresponding to the onset of the “two-magnon” condensate, $\omega_{B}\rightarrow0$, the normal and the anomalous Green’s function become exactly the same (see Fig. \[fig:residue\]). Consequently, the particle pair wave function, $\psi_{Q}(p)$, is exactly the same as the hole pair wave function $\phi_{\bar{Q}}^{*}(p)$. This is a manifestation of an emergent particle-hole symmetry at zero energy, which sets a constraint on the phase of the bound state wave function. If we describe the condensate state as a coherent state built with the two-body bound state wave function, the phases $\theta_{\pm}$ for the macroscopic components $\Psi_{\pm Q}$ are given by $\psi_{Q}(p)$ and $\phi_{\bar{Q}}(p)$, respectively. The relationship $\psi_{Q}(p)=\phi_{\bar{Q}}^{*}(p)$ indicates that $\theta_{+}+\theta_{-}=0$, which leads to a real order parameter $\langle S_{r}^{+}S_{r+1}^{+}\rangle$ in real space.
Beyond the condensation point, the new ground state, $|G\rangle$, is characterized by the order parameter $\Phi_{Q}(q)=\langle G|\alpha_{\frac{Q}{2}+q}\alpha_{\frac{Q}{2}-q}|G\rangle$. The coherent representation enables us to identify the bound state wave functions $\psi_{Q}(p)$ and $\phi_{\bar{Q}}(p)$ with the two component order parameter, $\Phi_{Q}(p)$ and $\Phi_{\bar{Q}}(p)$, that was discussed in Section \[GG\]. This correspondence leads to the self-consistent equation for the order parameter based on the Bethe-Salpeter equation, from which one can straightforwardly confirm which order is favored. The analysis becomes more transparent by adopting an equivalent but more straightforward approach. We just introduce a small pairing field term into the Hamiltonian, which couples to the order parameter: $$\begin{aligned}
{\cal H}_{paring} & =h_{Q}^{*}\sum_{q}\alpha_{\frac{Q}{2}+q}\alpha_{\frac{Q}{2}-q}+h_{\bar{Q}}^{*}\sum_{q}\alpha_{\frac{\bar{Q}}{2}+q}\alpha_{\frac{\bar{Q}}{2}-q}+h.c..\end{aligned}$$ The renormalized pairing fields $h_{Q}^{R},h_{\bar{Q}}^{R}$ are indicated by the ladder series of vertex corrections in Fig. \[fig:pairing\]. The pairing susceptibility diverges at the condensation point, implying that the order parameter develops spontaneously beyond this point, i.e., in absence of the pairing fields $h_{Q}$ and $h_{\bar{Q}}$. The order parameter in momentum space, $\Phi_{Q}(q)=\langle\alpha_{\frac{Q}{2}+q}\alpha_{\frac{Q}{2}-q}\rangle$, can be calculated as $\Phi_{Q}(q)=-h_{Q}^{R}(q)/\Omega_{q}$. Therefore, the ladder series of vertex corrections in Fig. \[fig:pairing\] leads to the following self-consistent equation:
![Ladder diagrams for the pairing field $h_{\pm Q}^R(q)$. The blue (red) vertex represents the renormalized pairing field at $Q$ ($-Q$). The dashed lines represent the symmetrized interaction vertexes in Eqs. (\[eq:normalV\]) and (\[eq:anomalousV\]).[]{data-label="fig:pairing"}](Pairing1.pdf)
$$\begin{aligned}
\sum_{k}\left(\Omega_{Q}(q)\delta_{q,k}+\frac{1}{2N}\Gamma_{Q}^{(0)N}(q,k)\pm\Gamma_{Q}^{(0)A}(k,q)\right)\Phi_{Q}^{\pm}(k)=0,\label{eq:eom}\end{aligned}$$
where $\Omega_{Q}(q)=E_{{Q\over 2}+q}+E_{{Q\over 2}-q}$ is the energy of a two-magnon excitation and $$\Phi_{Q}^{\pm}(k)=\Phi_{Q}(k)\pm\Phi_{-Q}^{*}(k).$$ The order parameters $\Phi_{Q}^{\pm}(k)$ become finite when the corresponding matrix in Eq. (\[eq:eom\]) is singular. For $Q=\pi$, $\Phi_{\pi}^{+}(k)$ coincides with imaginary part of $\Phi_{\pi}(k)$, while $\Phi_{\pi}^{-}(k)$ is the real part. The numerical calculation shows that the order parameter $\Phi_{\pi}(k)$ is purely imaginary for $Q=\pi$ and that it satisfies $\Phi_{Q}(k)=\Phi_{-Q}^{*}(k)$ for $Q<\pi$.
To understand the meaning of this result in real space, we just need to consider the order parameter $\langle S_{r+1}^{-}S_{r}^{-}\rangle$, which is given by $\langle c_{r+1}c_{r}\rangle$ in terms of the Jordan-Wigner fermionic annihilation operators. By applying Fourier and Bogoliubov transformations, we find $$\begin{aligned}
& \langle S_{r+1}^{-}S_{r}^{-}\rangle=\Psi_{0}-\frac{1}{N}\sum_{q}u_{\frac{Q}{2}+q}u_{\frac{Q}{2}-q}\sin q\nonumber \\
& \times[\Phi_{Q}(q)e^{i\frac{Q}{2}}e^{iQr}+\Phi_{\bar{Q}}(q)e^{i\frac{\bar{Q}}{2}}e^{i\bar{Q}r}]-\frac{1}{N}\sum_{q}v_{\frac{Q}{2}+q}v_{\frac{Q}{2}-q}\nonumber \\
& \times\sin q[\Phi_{\bar{Q}}^{*}(q)e^{i\frac{Q}{2}}e^{iQr}+\Phi_{Q}^{*}(q)e^{i\frac{\bar{Q}}{2}}e^{i\bar{Q}r}].\end{aligned}$$ The relationship $\Phi_{Q}(q)=\Phi_{\bar{Q}}^{*}(q)$ implies that the real space order parameter is $$\begin{aligned}
\langle S_{r+1}^{-}S_{r}^{-}\rangle & =\Psi_{0}-\frac{2}{N}\sum_{q}\left(u_{\frac{Q}{2}+q}u_{\frac{Q}{2}-q}+v_{\frac{Q}{2}+q}v_{\frac{Q}{2}-q}\right)\nonumber \\
& |\Phi_{Q}(q)|\sin q\cos\left(Qr+Q/2+\phi_{+}\right),\end{aligned}$$ where $\phi_{+}$ is the phase of $\Phi_{Q}(q)$. Therefore, the order parameter $\langle S_{r+1}^{-}S_{r}^{-}\rangle$ is real.
### Microscopic calculation of phenomenological parameters
{width="\textwidth"}
To provide a microscopic derivation of the phenomenological parameters $r$ and $u$ we express the Ginzburg-Landau free energy in its diagonal form: $${\cal F}=\frac{r+u}{2}|\Psi_{+}|^{2}+\frac{r-u}{2}|\Psi_{-}|^{2},\label{eq:GLdiagonal}$$ where $\Psi_{\pm }=\Psi_{Q}\pm\Psi_{\bar{Q}}^{*}$ and $\Psi_{Q}$ is the Fourier transform of the bond order parameter $\langle S_{r}^{-}S_{r+1}^{-}\rangle$. The order parameter can be expressed in terms of the fermionic Bogoliubov quasi-particles, $$\Psi_{\pm }=-\frac{e^{iQ/2}}{\sqrt{N}}\sum_{q}\sin(q)(u_{\frac{Q}{2}+q}u_{\frac{Q}{2}-q}\pm v_{\frac{Q}{2}+q}v_{\frac{Q}{2}-q})\Phi_{\pm }(q),\label{eq:orderBogoliubov}$$ where $\Phi_{\pm }(q)=\Phi_{Q}(q)\pm\Phi_{\bar{Q}}^{*}(q)$ and $\Phi_{Q}(q)=\langle\alpha_{\frac{Q}{2}+q}\alpha_{\frac{Q}{2}-q}\rangle$. From Eq. (\[eq:GLdiagonal\]), we can identify the phenomenological parameters with the inverse of the corresponding static susceptibilities: $$r\pm u=\langle \Psi^{\dagger}_{\pm} \Psi_{\pm }\rangle^{-1}=\left(\chi_{\pm }\right)^{-1}.$$ In other words, $\chi_{\pm}$ are the response functions to pairing fields that couple linearly to the $\Psi_{\pm }$ order parameters given in Eq. (\[eq:orderBogoliubov\]): $$\begin{aligned}
\chi_{\pm } & =-\frac{2}{N}\sum_{qp}\frac{\sin q\sin pB_{q}^{\pm}B_{p}^{\pm}}{\Omega_{Q}(q)\Omega_{Q}(p)}\left(\Gamma_{0Q}^{N}(q,p)\pm\Gamma_{0Q}^{A}(q,p)\right)\nonumber \\
& +\frac{2}{N}\sum_{q}\frac{\sin^{2}q(B_{q}^{\pm})^{2}}{\Omega_{Q}(q)},\end{aligned}$$ where $\Gamma_{0Q}^{N/A}(q,p)$ is the scattering amplitude at zero frequency, $\Omega_{Q}(q)=E_{\frac{Q}{2}+q}+E_{\frac{Q}{2}-q}$ and $B_{q}^{\pm}=u_{\frac{Q}{2}+q}u_{\frac{Q}{2}-q}\pm v_{\frac{Q}{2}+q}v_{\frac{Q}{2}-q}$. The second term is the non-interacting susceptibility of the Bogoliubov fermions, which is negligible near the critical point where both $\Gamma_{0Q}^{N}(q,p)$ and $\Gamma_{0Q}^{A}(q,p)$ diverge. The finite $u$ value arises from the non-zero anomalous scattering amplitude, $\Gamma_{0Q}^{A}(q,p)$, which forces the two susceptibilities $\chi_{Q}^{\pm}$ to be different: $$u=\frac{\left(\chi_{Q}^{+}\right)^{-1}-\left(\chi_{Q}^{-}\right)^{-1}}{2}.$$ Numerically, we always find $u<0$ for different ratios of $J_{2}/J_{1}$, in agreement with our previous discussions.
Numerical simulations
=====================
{width="70.00000%"}
The above theoretical analysis for small $\alpha$ indicates that the spin anisotropy stabilizes a dimerized ground state. In this section we present DMRG calculations [@ref:dmrg1; @ref:dmrg2] for the anisotropic 1D spin Hamiltonian ${\cal H}$, which confirm this analysis. The calculations have been done right below the critical field for chains of $L=160$ spins with open boundary conditions. We used up to 400 states and kept the truncation tolerance below $10^{-8}$ throughout the DMRG iterations. We did 6 full sweeps of finite algorithm of DMRG to get well converged observables. Fig. \[fig:dmrg\] shows the correlation functions for the real and imaginary parts of the nematic order parameter, ${\cal O}^{a}$, ${\cal O}^{b}$, for the “pair-density" operator $S_i^z S_{i+1}^z$, and for the bond operators ${\bf S}_i \cdot {\bf S}_{i+1}$, for different values of $\alpha$. The frustration ratio is taken as $J_2/J_1 = 0.9659$, which gives rise to a lowest energy “two-magnon” bound state with center of mass momentum equal to $\pi$. In agreement with the “two-magnon” calculation, the correlation functions of the order parameters ${\cal O}^{a}$ and ${\cal O}^{b}$ oscillate with wave vector $\pi$. The long wave length oscillations are just a consequence of the open boundary conditions and the large spin density wave susceptibility. Note, however, that the incommensurate nature of oscillations precludes the possibility of having long range spin density wave ordering (the incommensurate spin density wave ordering breaks the continuous U(1) symmetry group of translations).
The first column in Fig. \[fig:dmrg\] with $\alpha=0$ corresponds to the U(1) symmetric case for which all the correlators exhibit the expected power law behavior. The order parameters ${\cal O}^{a}$ and ${\cal O}^{b}$ are connected by a $\pi/4$ spin rotation about the $z$-axis, which is a symmetry of the Hamiltonian for $\alpha=0$. Correspondingly, both correlation functions exhibit an identical power-law decay. The bond and the pair density correlators also exhibit a power law decay with long wave length oscillations, which are magnified by the open boundary conditions [@ref:chiral2].
The upper panels of Fig. \[fig:dmrg\] show that the real component of the nematic order parameter dominates over the imaginary part and develops long range ordering for nonzero $\alpha$. The amplitude of the order parameter increases with $\alpha$. As expected from the previous analysis, the pair density, $S_i^z S_{i+1}^z$, and the bond, ${\bf S}_i \cdot {\bf S}_{i+1}$, operators become dimerized as a consequence of the coexistence of uniform and staggered components of the ${\cal O}^a$ order parameter. We recall that the uniform component is directly induced by the $\alpha$ term of the Hamiltonian, while the staggered ($\pi$) component is spontaneously generated. The dimerization is identified by Fourier transforming the real space correlation functions shown in Figs. \[fig:dmrg\](e-h). This information is provided in the insets of Fig. \[fig:pipeak\], which show a clear peak at $Q=\pi$. It is also clear from Figs. \[fig:dmrg\](e-h) that the dimerization order parameter becomes more robust upon increasing $\alpha$. Moreover, the insets of Fig. \[fig:pipeak\] show that the Friedel oscillations of $\langle S_i^z \rangle$ are strongly suppressed for $\alpha = 0.7$, i.e., for a large amplitude of the dimerization order parameter. Consistently with these results, Fig. \[fig:pipeak\] shows that the dimerized phase (shaded region) becomes stable over a larger window of magnetic field values as we increase $\alpha$. The same figure includes the field dependence of the magnetization curve for different values of $\alpha$. As expected, the slope of the magnetization curve above the critical field increases with $\alpha$ and the magnetization saturates asymptotically for $H\rightarrow \infty$.
Discussion and summary
======================
Finally, we discuss the consequences of our theoretical study on the experimental search for nematic phases in various quasi-1D materials. We first note that two previous works studied the role of an XXZ magnetic anisotropy, *which preserves* the $U(1)$ symmetry of the spin Hamiltonian. Both works concluded that the effect of the U(1) invariant spin anisotropy is to stabilize the nematic phase. Similarly, we are finding that the effect of a U(1) symmetry breaking anisotropy is to stabilize the dimerized phase, which is a “direct descendant" of the nematic ordering (real part of the nematic order parameter).
The spin anisotropy of the quasi-1D compound LiCuVO$_4$ is approximately $\alpha = 0.1$. [@ref:Z2material1; @ref:1Dmaterial2; @ref:1Dmaterial3; @ref:1Dmaterial6; @LiCuVO4; @LiCuVO4b; @observation] The results of our simulations for this value of $\alpha$ \[see Figs. \[fig:dmrg\] (b) and (f)\] clearly show that the bond-bond correlation function develops a visible $\pi$ ordering (dimerization). This is a salient experimental signature that can be detected with X-rays. Rb$_2$Cu$_2$Mo$_3$O$_{12}$ is another quasi-1D $J_1$-$J_2$ frustrated magnet [@ref:1DmaterialA1; @ref:1DmaterialA2] with $J_2/\lvert J_1 \rvert \simeq 0.33$. This value falls in the “mutipolar" phase (quasi-condensate of $n$-magnon bound states with $n>2$) of the U(1) invariant spin model. Indeed, the nematic phase becomes unstable for $J_2/\rvert J_1 \rvert <0.37$. Although we have not considered the situation of multipolar orderings of higher order than nematic, our simple GL analysis shows that the $\alpha$-term should once again select the real or the imaginary part of the multipolar order parameter.
A more recent experimental work reports evidence of a gapped “nematic ordering" for the quasi-1D $J_1-J_2$ compound LiCuSbO$_4$. [@grafe2016] Nuclear magnetic resonance (NMR) measurements indicate the opening of a rather large spin gap above a critical magnetic field value $H_{c1} \simeq 13$T. Conventional explanations for the origin of this spin gap, such as saturation of the magnetization or the presence of a staggered Dzyaloshinskii-Moriya (DM) interaction have been ruled out by the experiments. Moreover, the absence of a kink in the low temperature magnetization vs. field curve is consistent with an XYZ magnetic anisotropy. Based on our analysis, this U(1) symmetry breaking anisotropy is responsible for the the spin gap that is inferred from the NMR measurements. According to Ref. [@grafe2016] the value of the spin anisotropy is $\alpha \simeq 0.08$. In addition, a $J_2/\rvert J_1 \rvert \simeq 0.28 $ ratio is obtained from a fit of the magnetization curve with a pure one-dimensional Hamiltonian. This ratio puts the material outside the region of stability of the nematic phase [@ref:chiral2] (higher multipolar orderings are expected for $J_2/\rvert J_1 \lesssim 0.367$). However, Ref. [@LiCusbO4] reports a significantly larger ratio $J_2/\rvert J_1 \rvert \simeq 0.45$, which falls in the nematic phase of the U(1) invariant Hamiltonian. [@ref:chiral2]. If this $J_2/\rvert J_1 \rvert$ ratio is correct, the gapped high-field spin phase of this compound must correspond to a spin dimerized phase. However, further inspection of this material [@grafe2016] shows that the lattice is already dimerized (consecutive bonds are not equivalent) implying a continuous crossover between the dimerized spin phase and the high field paramagnetic state.
In summary, we have demonstrated that the spin-orbit interaction has important consequences for the field-induced spin nematic ordering of U(1) invariant frustrated models. The symmetry reduction of ${\cal H}$ due to the presence of the Ising term renders the nematic order parameter inapplicable. However, the real and imaginary parts of the nematic bond order parameter still break discrete symmetries, which can be directly related with observable quantities. Our analytical and numerical results demonstrate that the spin-orbit interaction stabilizes a bond density wave (bond dimerization for $Q=\pi$) which couples to the lattice via the magneto-elastic effect. These results are confirmed by our DMRG simulations. Given that the spin-orbit interaction is ubiquitous in nature and that continuous symmetries are never strictly present in real magnets, our study indicates that the proposed nematic ordering is likely to be replaced by bond dimerization in systems described by a $J_1$-$J_2$ model with $|J_2|/|J_1| \gtrsim 0.38$. Even in quasi-1D systems, which are approximately described by a U(1) invariant $XXZ$ model, the application of a magnetic field *perpendicular* to the chains should induce the dimerized state that we are proposing here.
We thank Zhentao Wang, Shi-Zeng Lin and Yukitoshi Motome for helpful discussions. S. Z. and C.D.B. are supported by funding from the Lincoln Chair of Excellence in Physics and from the Los Alamos National Laboratory Directed Research and Development program. N. K. was supported by the National Science Foundation, under Grant No. DMR-1404375. E. D. was supported by the U.S. Department of Energy, Office of Basic Energy Sciences, Materials Sciences and Engineering Division.
[1]{}
Vivien Zapf, Marcelo Jaime, and C. D. Batista, Rev. Mod. Phys. [**86**]{}, 563 (2014).
L. Balents, Nature [**464**]{}, 199 (2010).
A. V. Chubukov and O. A. Starykh, Phys. Rev. Lett. [**110**]{}, 217210 (2013).
Toshiya Hikihara, Lars Kecke, Tsutomu Momoi, and Akira Furusaki, Phys. Rev. B [**78**]{}, 144404 (2008).
Gia-Wei Chern, C. J. Fennie, and O. Tchernyshyov, Phys. Rev. B [**74**]{}, 060405 (R) (2006).
P. W. Anderson, Mater. Res. Bull., [**8**]{}, 153 (1973).
Fradkin E., *Field Theories of Condensed Matter Systems* (Addison-Wesley, Reading) 1991.
F. L. Pratt, P. J. Baker, S. J. Blundell, T. Lancaster, S. Ohira-Kawamura, C. Baines, Y. Shimizu, K. Kanoda, I. Watanabe, and G. Saito, Nature. [**471**]{}, 612 (2011).
X. G. Wen, F. Wilczek, and A. Zee, Phys. Rev. B [**39**]{}, 11413 (1989).
M. E. Zhitomirsky and H. Tsunetsugu, Europhys. Lett., [**92**]{}, 37001 (2010).
L. Kecke, T. Momoi, and A. Furusaki, Phys. Rev. B [**76**]{}, 060407 (2007).
T. Hikihara, L. Kecke, T. Momoi, and A Furusaki, Phys. Rev. B [**78**]{}, 144404 (2008).
A. V. Chubukov, Phys. Rev. B **44**, 4693 (1991).
J. Sudan, A. Lüscher, and A. M. Läuchli, Phys. Rev. B [**80**]{}, 140402(R) (2009).
A. Lüscher, J. Sudan, and A. Lüscher, J. Phys. Conf. Ser. [**145**]{}, 012057 (2009).
T. Vekua, A. Honecker, H.-J. Mikeska, and F. Heidrich-Meisner, Phys. Rev. B [**76**]{}, 174420 (2007).
A. V. Syromyatnikov, Phys. Rev. B [**86**]{}, 014423 (2012).
A. Läuchli, J. C. Domenge, C. Lhuillier, P. Sindzingre, and M. Troyer, Phys. Rev. Lett. [**95**]{}, 137206 (2005).
Nic Shannon, Tsutomu Momoi, and Philippe Sindzingre, Phys. Rev. Lett. [**96**]{}, 027213 (2006). G. Marmorini and T. Momoi, Phys. Rev. B [**89**]{}, 134425 (2014).
Ryuichi Shindou and Tsutomu Momoi, Phys. Rev. B [**80**]{}, 064410 (2009).
H. T. Ueda and T. Momoi, Phys. Rev. B [**87**]{}, 144417 (2013).
Emika Takata, Tsutomu Momoi, and Masaki Oshikawa, arXiv: 1510.02373.
M. Hase, H. Kuroe, K. Ozawa, O. Suzuki, H. Kitazawa, G. Kido, and T. Sekine, Phys. Rev. B 70, 104426 2004 .
Y. Yasui, R. Okazaki, I. Terasaki, M. Hase, M. Hagihala, T. Masuda, and T. Sakakibara, JPS Conf. Proc. [**3**]{}, 014014 (2014)
M. Enderle, C. Mukherjee, B. Fák, R. K. Kremer, J.-M. Broto, H. Rosner, S.-L. Drechsler, J. Richter, J. Malek, A. Prokofiev, W. Assmus, S. Pujol, J.-L. Raggazzoni, H. Rakoto, M. Rheinstädter and H. M. Rønnow, Europhys. Lett., [**70**]{} (2), 237 (2005).
J. S. Miller (Ed.): Extended Linear Chain Compounds. Plenum Press, New York and London 1982.
H. A. KrugvonNidda, L. E. Svistov, M. V. Eremin, R. M. Eremina, A. Loidl, V. Kataev, A. Validov, A. Prokofiev, and W. A$\beta$mus, Phys. Rev. B [**65**]{}, 134445 (2002).
N. Büttgen, W. Kraetschmer, L. E. Svistov, L. A. Prozorova, and A. Prokofiev, Phys. Rev. B [**81**]{}, 052403 (2010).
N. Büttgen, H.-A. Krug von Nidda, L. E. Svistov, L. A. Prozorova, A. Prokofiev, and W. A$\beta$mus, Phys. Rev. B [**76**]{}, 014440 (2007).
K. Nawa, M. Takigawa, M. Yoshida, and K. Yoshimura, Journal of the Physical Society of Japan [**82**]{}, 094709 (2013).
N. Büttgen, K. Nawa, T. Fujita, M. Hagiwara, P. Kuhns, A. Prokovief, A. P. Reyes, L. E. Svistov, K. Yoshimura, and M. Takigawa, Phys. Rev. B [**90**]{}, 134401 (2014).
F. Schrettle, S. Krohns, P. Lunkenheimer, J. Hemberger, N. Büttgen, H.-A. Krug von Nidda, A. V. Prokofiev, and A. Loidl, Phys. Rev. B [**77**]{}, 144101 (2008).
M. G. Banks, F. Heidrich-Meisner, A. Honecker, H. Rakoto, J.-M. Broto and R. K. Kremer, J. Phys.: Condens. Matter [**19**]{} 145227 (2007).
B.J. Gibson, R.K. Kremer, A.V. Prokofiev, W. Assmus, G.J. McIntyre, Physica B [**350**]{} e253 (2004).
K. Nawa, Y. Okamoto, A. Matsuo, K. Kindo, Y. Kitahara, S. Yoshida, S. Ikeda, S. Hara, T. Sakurai, S. Okubo, H. Ohta, H. Ohta, and Z. Hiroi, J. Phys. Soc. Jpn. [**83**]{}, 103702 (2014).
K. Nawa, T. Yajima, Y. Okamoto, and Z. Hiroi, Inorg. Chem. [**54**]{}, 5566 (2015).
H. -J. Grafe, S. Nishimoto, M. Iakovleva, E. Vavilova, L. Spillecke, A. Alfonsov, M.-I. Sturza, S. Wurmehl, H. Nojiri, H. Rosner, J. Richter, U.K. Rö$\beta$ler, S.-L. Drechsler, V. Kataev, and B. Büchner, arxiv:1607.05164. S. E. Dutton, M. Kumar, M. Mourigal, Z. Soos, J.-J. Wen, C. Broholm, N. Andersen, Q. Huang, M. Zbiri, R. Toft-Petersen, and R. J. Cava, Phys. Rev. Lett. [**108**]{}, 187206 (2012).
B. Willenberg, M. Schäpers, A.U.B. Wolter, S. L. Drechsler, M. Reehuis, J.-U. Hoffmann, B. Buchner, A. J. Studer, K. C. Rule, B. Ouladdiaf, S. Sullow, and S. Nishimoto, Phys. Rev. Lett. [**116**]{}, 047202 (2016).
H. T. Ueda and K. Totsuka, Phys. Rev. B [**80**]{}, 014417 (2009).
M. Sato, T. Momoi, and A. Furusaki, Phys. Rev. B [**79**]{}, 060406(R).
M. Sato, T. Hikihara, and T. Momoi, Phys. Rev. B [**83**]{}, 064405 (2011).
A. Smerald and N. Shannon, Phys. Rev. B [**93**]{}, 184419 (2016).
G. Jackeli and G. Khaliullin, Phys. Rev. Lett. **102**, 017205 (2009).
M. Sato, T. Hikihara and T. Momoi, Phys. Rev. Lett. [**110**]{}, 077206 (2013).
O. A. Starykh and L. Balents, Phys. Rev. B [**89**]{}, 104407 (2014).
L. E. Svistov, T. Fujita, H. Yamaguchi, S. Kimura, K. Omura, A. Prokofiev, A. I. Smirnov, Z. Honda and M. Hagiwara, Jetp Lett. [**93**]{} 21 (2011).
S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992).
S. R. White, Phys. Rev. B [**48**]{}, 10345 (1993).
H. Onishi, Journal of Physics: Conference Series [**592**]{}, 012109 (2015). S. Nishimoto, S. -L. Drechsler, R. Kuzian, J. Richter and J. vandenBrink, Phys. Rev. B [**92**]{}, 214415 (2015).
M. Mourigal, M. Enderle, B. Fak, R. K. Kremer, J. M. Law, A. Schneidewind, A. Hiess and A. Prokofiev, Phys. Rev. Lett. [**109**]{}, 027203 (2012)
|
{
"pile_set_name": "ArXiv"
}
|
---
address: 'DIFA, University of Bologna, and INFN, Bologna, Italy'
author:
- 'Roberto Zucchini[^1]'
bibliography:
- 'allbibtex.bib'
subtitle: '[LMS/EPSRC Durham Symposium on Higher Structures in M-Theory](http://www.maths.dur.ac.uk/lms/109/index.html)'
title: Wilson Surfaces for Surface Knots
---
Introduction
============
Knots are interesting in topology as well as in gauge theory [@Kauffman:1991ds].
Ordinary knots are embeddings of $S^1$ into a 3–dimensional manifold, say $S^3$ [@adams2004knot; @lickorish1997introduction]. Can one define higher dimensional knots generalizing this simple topological notion? In just one dimension higher there are at least two ways of doing that.
Since $S^1$ is the lowest dimensional non trivial sphere, one may define a 2–dimensional knot as an embedding $S^2$ into $S^4$. This yields the so called 2–knots. Since $S^1$ is also the lowest dimensional non trivial closed oriented manifold, one may define a 2–dimensional knot as an embedding of $S_\ell$ into $S^4$, where $S_\ell$ is a genus $\ell$ closed oriented surface. This leads to genus $\ell$ surface knots. Of course, 2–knots are just genus 0 surface knots. However, they have very special properties which make a separate study meaningful. 2- and surface knots are objects of intense investigation by topologists [@carter1998knotted; @kamada2002braid].
Wilson lines [@Wilson:1974sk] are relevant in the analysis of confinement in quantum chromodynamics, loop formulation of quantum gravity, symmetry breaking in string theory, condensed matter theory and knot topology. As shown in Witten’s seminal work [@Witten:1988hf], one can study knot topology in Chern-Simons theory, an instance of gauge theory, relying on techniques of quantum field theory. With any knot $\xi$, one associates the Wilson line $$W_R(\xi)=\operatorname{tr}_R\!\bigg[\operatorname{Pexp}\bigg(-\int_\xi A\bigg)\bigg].$$ where $R$ is a representation of the gauge group $G$. Chern–Simons correlators of Wilson line operators provide classic knot invariants.
Wilson surfaces [@Chepelev:2001mg; @Alekseev:2015hda] may turn out to be relevant in the study of non perturbative aspects of higher gauge theory, brane theory, quantum gravity and higher knot topology. Following Witten’s paradigm, one can presumably study 2– or surface knot topology computing correlators of knot Wilson surfaces in an appropriate higher version of Chern–Simons theory, an instance of higher gauge theory [@Baez:2002jn; @Baez:2010ya], using again techniques of quantum field theory. To this end, one needs to associate with any surface knot $\varXi$ a Wilson surface $$W(\varXi)=?~,$$ whose expressions is at this point to be found. In this communication, we shall present a proposal for a definition of the Wilson surfaces $W(\varXi)$ in higher gauge theory based mainly on our work [@Zucchini:2015wba; @Zucchini:2015xba].
The problem has two parts:
i) [*define surface knot holonomy*]{};
ii) [*define higher invariant traces*]{}.
Parallel transport and holonomy are related but distinguished, holonomy being a special case of parallel transport.
Earlier endeavours on higher parallel transport includes the work of Caetano and Picken [@Caetano:1993zf], Baez and Schreiber [@Baez:2004in; @Baez:2005qu] Schreiber and Waldorf [@Schreiber:0705.0452; @Schreiber:0802.0663; @Schreiber:2008aa], Faria Martins and Picken [@Martins:2007uki; @Martins:2011:3309], Chatterjee, Lahiri and Sengupta [@Chatterjee:2009ne; @Chatterjee:2014pna; @Chatterjee:2010xa] Soncini and Zucchini [@Soncini:2014zra], Abbaspour and Wagemann [@Abbaspour:1202.2292] and Arias Abad and Schaetz [@Abad:1404.0729; @Abad:1404.0727]. Earlier results on higher holonomy were obtained by Cattaneo and Rossi [@Cattaneo:2002tk] and Faria Martins and Picken [@Martins:2007uki; @Martins:2011:3309].
Following [@Zucchini:2015wba; @Zucchini:2015xba], we shall present a framework for the construction of holonomy invariants of knots and surface knots. In a nutshell, our strategy rests on describing knots by parametrized curves and surface knots by parametrized surfaces. We outline it below, assuming that the reader is familiar with the basic ideas of strict higher gauge theory. In any case, those notion will be reviewed in greater detail in subsequent sections.
In a manifold $M$, a curve $\gamma:p_0\rightarrow p_1$ is a parametrized path joining two points. A homotopy $h:\gamma_0\Rightarrow \gamma_1$ of two curves is a parametrized path joining those curves. Curves can be composed by concatenation and inverted. The resulting operations make curves modulo homotopy a groupoid, the fundamental 1-groupoid $(M,P^0{}_1M)$ of $M$.
In ordinary gauge theory with gauge Lie group $G$, given a flat gauge field $A$ one can construct a gauge covariant and homotopy invariant parallel transport functor $$\begin{aligned}
&F_A:(M,P^0{}_1M)\rightarrow BG,
\\
&\gamma\rightarrow F_A(\gamma),\notag\end{aligned}$$ where $BG$ is the delooping of $G$, that is $G$ seen as the morphism group of a one–object groupoid.
With a knot $\xi$ based at $p$ defined up to ambient isotopy one can associate a curve $\gamma_\xi:p\rightarrow p$ defined up to homotopy and with this the holonomy $$F_A(\xi)=F_A(\gamma_\xi).$$ One can check $F_A(\xi)$ is base point and isotopy invariant and gauge independent up to conjugation. Using invariant traces, one can extract an invariant from the holonomy $F_A(\xi)$.
A ‘gentle’ generalization of the above construction for surface knots is the following.
A curve $\gamma:p_0\rightarrow p_1$ is a parametrized path joining two points. A surface $\varSigma:\gamma_0\Rightarrow \gamma_1$ is a parametrized path joining two curves in a manifold $M$. A thin homotopy $h:\gamma_0\Rightarrow \gamma_1$ of two curves is a parametrized path joining those curves with degenerate (less than two–dimensional) range A homotopy $H:\varSigma_0\Rrightarrow \varSigma_1$ of two surfaces is a parametrized path joining those surfaces.Curves can be composed by concatenation and inverted. Surfaces can be composed by concatenation and inverted in two distinct ways, usually called horizontal and vertical. The resulting operations make curves modulo thin homotopy and surfaces modulo homotopy a 2–groupoid, fundamental 2–groupoid $(M,P_1M,P^0{}_2M)$.
In strict higher gauge theory with gauge Lie crossed module $(G,H)$, given a flat higher gauge field pair $A,B$ one can construct a gauge covariant and (thin) homotopy invariant parallel transport 2–functor $$\begin{aligned}
&F_{A,B}:(M,P_1M,P^0{}_2M)\rightarrow B(G,H),
\\
&\gamma\rightarrow F_A(\gamma), \quad \varSigma\rightarrow F_{A,B}(\varSigma).
\nonumber\end{aligned}$$
With a knot $\xi$ based at $p$ and a surface knot $\varXi$ based at a genus dependent fundamental polygon $\tau$ stemming from cutting the image of $\varXi$ along standard $a$ and $b$ cycles, both defined up to ambient isotopy, one can associate a curve $\gamma_\xi:p\rightarrow p$ and a surface $\varSigma_\varXi:\iota_p\Rightarrow \tau$ up to (thin) homotopy and from this the holonomy $$F_A(\xi)=F_A(\gamma_\xi), \quad F_{A,B}(\varXi)=F_{A,B}(\varSigma_\varXi).$$ One can check that $F_A(\xi)$ is base data and isotopy invariant and gauge independent up to the appropriate form of crossed module conjugation. Using higher invariant traces, one can extract invariants from the holonomy $F_A(\xi)$ and $F_{A,B}(\varXi)$.
There are open issues to be solved. It can be shown that surface knot holonomy necessarily lies in the kernel of the target map $H\longrightarrow\!\!\!\!\!\!\!\!\!\!\!{}^t\,\,\,\,\,\,\,G$ of the Lie crossed module $(G,H)$ and so is central. Thus, for many Lie crossed module this holonomy may turn out to be trivial. The existence of non trivial higher traces on $(G,H)$ is also to be ascertained. This is an problem that can be formalized using higher quandle theory (see Crans [@Crans:2004ve] and Crans and Wagemann[@Crans:1310.4705]).
From a quantum field theoretic point of view, the most delicate question remains obtaining surface knot invariants from a 4–dimensional higher Chern–Simons theory as proposed by Zucchini [@Zucchini:2011aa; @Zucchini:2015ohw] and Soncini and Zucchini [@Soncini:2014ara]. There are problems with the definition of Wilson surface insertions in the quantum theory, which we shall point out in due course.
Curves, surfaces and homotopy {#sec:holo}
=============================
Closed curves and surfaces describe knots and surface knots in an ambient manifold $M$.
Curves and surfaces are smoothly parametrized subsets of $M$. They can be composed and inverted in various ways. In order to preserve smoothness, it is sufficient to require that their parametrization has sitting instants. A smooth map $f:S\times\mathbbm{R}\rightarrow T$, where $S$ and $T$ are manifolds, has sitting instants if $$\begin{aligned}
&f(-,x)=f(-,0) \quad\text{for $x<\epsilon$}, \\
&f(-,x)=f(-,1) \quad \text{for $x>1-\epsilon$}, \nonumber\end{aligned}$$ for some number $\epsilon$ such that $0<\epsilon<1/2$. In what follows, all maps will be tacitly assumed to have sitting instants for each factor $\mathbbm{R}$ of their domains.
Formally, curves and surfaces are defined as follows.
[*For any two points $p_0,p_1\in M$, a curve $\gamma:p_0\rightarrow p_1$ in $M$ is a map $\gamma:\mathbbm{R}\rightarrow M$ such that*]{} $$\begin{aligned}
\gamma(0)=p_0,\quad \gamma(1)=p_1.\end{aligned}$$
[ *For any two points $p_0,p_1\in M$ and any two curves $\gamma_0,\gamma_1:p_0\rightarrow p_1$ of $M$, a surface $\varSigma:\gamma_0\Rightarrow \gamma_1$ of $M$ is a map $\varSigma:$ $\mathbbm{R}^2\rightarrow M$ such that*]{} $$\begin{aligned}
&\varSigma(0,y)=p_0,\quad \varSigma(1,y)=p_1,
\\
&\varSigma(x,0)=\gamma_0(x),\quad \varSigma(x,1)=\gamma_1(x) \nonumber\end{aligned}$$
Curve and surfaces can be combined through a set of natural operations based on the intuitive idea of concatenation. We begin by introducing the basic operations with curves.
[*For a point $p$, the unit curve of $p$ is the curve $\iota_p:p\rightarrow p$ defined by $$\begin{aligned}
\iota_p(x)=p.\end{aligned}$$ For a curve $\gamma:p_0\rightarrow p_1$, the inverse curve of $\gamma$ is the curve $\gamma^{-1_\circ}:p_1\rightarrow p_0$ given by $$\begin{aligned}
\gamma^{-1_\circ}(x)=\gamma(1-x).\end{aligned}$$ For two curves $\gamma_1:p_0\rightarrow p_1$, $\gamma_2:p_1\rightarrow p_2$, the composition of $\gamma_1$, $\gamma_2$ is the curve $\gamma_2\circ\gamma_1:p_0\rightarrow p_2$ piecewise given by*]{} $$\begin{aligned}
&\gamma_2\circ\gamma_1(x)=\gamma_1(2x) \qquad \text{for $x\leq 1/2$},
\\
&\gamma_2\circ\gamma_1(x)=\gamma_2(2x-1) \qquad \text{for $x\geq 1/2$}.\notag
\vphantom{\Big]}\end{aligned}$$
We introduce next the basic operations with surfaces. These turn out to be of two types, called horizontal and vertical.
[*For a curve $\gamma:p_0\rightarrow p_1$, the unit surface of $\gamma$ is the surface $I_\gamma:\gamma\Rightarrow\gamma$ defined by $$\begin{aligned}
I_\gamma(x,y)=\gamma(x).\end{aligned}$$ For a surface $\varSigma:\gamma_0\Rightarrow\gamma_1$, the vertical inverse of $\varSigma$ is the surface $\varSigma^{-1_\bullet}:\gamma_1\Rightarrow\gamma_0$ defined by $$\begin{aligned}
\varSigma^{-1_\bullet}(x,y)=\varSigma(x,1-y).\end{aligned}$$ For two surfaces $\varSigma_1:\gamma_0\Rightarrow\gamma_1$, $\varSigma_2:\gamma_1\Rightarrow\gamma_2$, the vertical composition of $\varSigma_1$, $\varSigma_2$ is the surface $\varSigma_2\bullet\varSigma_1:\gamma_0\Rightarrow \gamma_2$ given by $$\begin{aligned}
&\varSigma_2\bullet\varSigma_1(x,y)=\varSigma_1(x,2y) \qquad \text{for $y\leq 1/2$},
\vphantom{\Big]}
\\
&\varSigma_2\bullet\varSigma_1(x,y)=\varSigma_2(x,2y-1) \qquad \text{for $y\geq 1/2$}.\notag
\vphantom{\Big]}\end{aligned}$$ For a surface $\varSigma:\gamma_0\Rightarrow\gamma_1$, the horizontal inverse of $\varSigma$ is the surface $\varSigma^{-1_\circ}:\gamma_0{}^{-1_\circ}\Rightarrow \gamma_1{}^{-1_\circ}$ defined by $$\begin{aligned}
\varSigma^{-1_\circ}(x,y)=\varSigma(1-x,y).\end{aligned}$$ For two surfaces $\varSigma_1:\gamma_0\Rightarrow\gamma_1$, $\varSigma_2:\gamma_2\Rightarrow\gamma_3$, the horizontal composition of $\varSigma_1$, $\varSigma_2$ is the surface $\varSigma_2\circ\varSigma_1:\gamma_2\circ\gamma_0\Rightarrow \gamma_3\circ\gamma_1$ given by*]{} $$\begin{aligned}
&\varSigma_2\circ\varSigma_1(x,y)=\varSigma_1(2x,y) \qquad \text{for $x\leq 1/2$},
\vphantom{\Big]}
\\
&\varSigma_2\circ\varSigma_1(x,y)=\varSigma_2(2x-1,y) \qquad \text{for $x\geq 1/2$}.\notag
\vphantom{\Big]}\end{aligned}$$
Unfortunately, these operations are not nice enough; associativity and invertibility fail to hold in general. The operation are in fact nice only up to homotopy.
[*A homotopy $h:\gamma_0\Rightarrow \gamma_1$ of two curves $\gamma_0,\gamma_1:p_0\rightarrow p_1$ of $M$ with the same end-points is a map $h:\mathbbm{R}^2\rightarrow M$ of $M$ such that $$\begin{aligned}
&h(0,y)=p_0,\quad h(1,y)=p_1,
\\
&h(x,0)=\gamma_0(x),\quad h(x,1)=\gamma_1(x).
\nonumber\end{aligned}$$ The homotopy is thin if in addition $\operatorname{rank}dh(x,y)<2$. (Thin) homotopy of curves is an equivalence relation.*]{}
[*A homotopy $H:\varSigma_0\Rrightarrow\varSigma_1$ of two surfaces $\varSigma_0:\gamma_0\Rightarrow\gamma_1$, $\varSigma_1:\gamma_2\Rightarrow\gamma_3$, where $\gamma_0,\gamma_1,\gamma_2,\gamma_3:p_0\rightarrow p_1$ are four curves with the same end-points, is a map $H:\mathbbm{R}^3\rightarrow M$ such that $\operatorname{rank}dH(x,0,z)$, $\operatorname{rank}dH(x,1,z)\leq 1$ and $$\begin{aligned}
&H(0,y,z)=p_0,\quad H(1,y,z)=p_1,
\\
&H(x,y,0)=\varSigma_0(x,y), \quad H(x,y,1)=\varSigma_1(x,y).
\nonumber\end{aligned}$$ The homotopy is thin if $\operatorname{rank}dH(x,y,z)<3$. (Thin) homotopy of surfaces is an equivalence relation.*]{}
Let us denote by $\Pi_1M$, $\Pi_2M$ the sets of all curves and surfaces of $M$, by $P_1M$ and $P^0{}_1M$ the sets of thin homotopy and homotopy classes of curves and $P_2M$ and $P^0{}_2M$ the sets of thin homotopy and homotopy classes of surfaces, respectively. The following results are basic.
[*$(M,P_1M)$ and $(M,P^0{}_1M)$ with the operations induced by those of $\Pi_1M$ are groupoids, the path and fundamental groupoids of $M$.*]{}
[*$(M,P_1M,P_2M)$ and $(M,P_1M,P^0{}_2M)$ with the operations induced by those of $\Pi_1M$ and $\Pi_2M$ are 2–groupoids, the path and fundamental 2–groupoids of $M$.*]{}
Higher parallel transport {#sec:partr}
=========================
In gauge theory, holonomy is a special case of parallel transport. Therefore, it is necessary to examine in some detail the definition and the properties of the latter. We begin be reviewing parallel transport in ordinary gauge theory and then we introduce and describe parallel transport in higher gauge theory.
Let $G$ be a Lie group with Lie algebra $\mathfrak{g}$ and let $M$ be a manifold. Consider an ordinary gauge theory on the trivial principal $G$–bundle $M\times G$.
[*A $G$–connection on $M$ is a $\mathfrak{g}$–valued $1$–form $\theta\in\Omega^1(M,\mathfrak{g})$. $\theta$ is flat if*]{} $$\begin{aligned}
{\rm d}\theta+\frac{1}{2}[\theta,\theta]=0.
\label{twoholo7}\end{aligned}$$
Parallel transport requires a $G$–connection on $M$ as input datum.
[*For a curve $\gamma$ of $M$, the parallel transport along $\gamma$*]{} is the element $F_\theta(\gamma)\in G$ defined by $$\begin{aligned}
F_\theta(\gamma)=u(1),\label{twofholo1}\end{aligned}$$ [*where $u:\mathbbm{R}\rightarrow G$ is the unique solution of the differential problem*]{} $$\begin{aligned}
{\rm d}_xu(x)u(x)^{-1}=-\gamma^*\theta_x(x), \quad u(0)=1_G.
\label{twofholo2}\end{aligned}$$
The first relevant property of parallel transport is its consistency with the operations with curves defined in Section \[sec:holo\].
[*For any point $p$ and any curves $\gamma,\gamma_1,\gamma_2$, one has*]{} $$\begin{aligned}
\label{ptc1}
&F_\theta(\iota_p)=1_G,
\\
\label{ptc2}
&F_\theta(\gamma^{-1_\circ})=F_\theta(\gamma)^{-1},
\\
\label{ptc3}
&F_\theta(\gamma_2\circ\gamma_1)=F_\theta(\gamma_2)F_\theta(\gamma_1).\end{aligned}$$ [*whenever defined.*]{}
The second relevant property of parallel transport is its compatibility with homotopy of curves as defined in Section \[sec:holo\].
[*For any two thinly homotopic curves $\gamma_0$, $\gamma_1$, one has $$\begin{aligned}
&F_\theta(\gamma_1)=F_\theta(\gamma_0).
\label{twoholo6}\end{aligned}$$ When $\theta$ is flat, holds also when $\gamma_0$, $\gamma_1$ are homotopic.* ]{}
Parallel transport has an elegant categorical interpretation.
[*Parallel transport yields a functor $\bar F_\theta:(M,P_1M)\rightarrow BG$ from the path groupoid $(M,P_1M)$ of $M$ into $BG$. For flat $\theta$, parallel transport yields a functor $\bar F^0{}_\theta:(M,P^0{}_1M)\rightarrow BG$ from the fundamental groupoid $(M,P^0{}_1M)$ of $M$ into $BG$.*]{}
Any meaningful gauge theoretic construction should be gauge covariant in the appropriate sense. Parallel transport has also this property.
[*A $G$–gauge transformation is just a $G$–valued mapping $g\in\operatorname{Map}(M,G)$.*]{}
Gauge transformations act on connections in the well–known manner.
[*The gauge transform of the $G$–connection $\theta$ is the $G$–connection*]{} $$\begin{aligned}
&{}^g\theta=\operatorname{Ad}g (\theta)-{\rm d}gg^{-1}.
\label{gauholo1}\end{aligned}$$ [*If $\theta$ is flat, ${}^g\theta$ is flat, too.*]{}
Parallel transport has simple covariance properties under gauge transformation.
[*For any curve $\gamma:p_0\rightarrow p_1$ of $M$,* ]{} $$\begin{aligned}
F_{{}^g\theta}(\gamma)=g(p_1)F_\theta(\gamma)g(p_0)^{-1}.
\vphantom{\Big]}
\label{gauholo2}\end{aligned}$$
Gauge transformation of parallel transport also has an elegant categorical interpretation.
[*A gauge transformation $g$ encodes a natural transformation $\bar F_\theta\Rightarrow \bar F_{{}^g\theta}$ of parallel transport functors. When $\theta$ is flat, $g$ encodes a natural transformation $\bar F^0{}_\theta\Rightarrow \bar F^0{}_{{}^g\theta}$ of flat parallel transport functors.*]{}
An appropriate form of parallel transport can be defined also in strict higher gauge theory. The intuitive idea of the construction is still simple, though the technical details are much more involved.
Let $K$ be a strict Lie $2$ group with strict Lie $2$–algebra $\mathfrak{k}$ and let $M$ be a manifold. Consider a higher gauge theory on the trivial principal $K$–$2$–bundle $M\times K$. As it is natural and convenient, we shall view the Lie $2$–group $K$ as a Lie crossed module $H\longrightarrow\!\!\!\!\!\!\!\!\!\!\!{}^t\,\,\,\,\,\,\,G
\longrightarrow\!\!\!\!\!\!\!\!\!\!\!\!\!{}^m\,\,\,\,\operatorname{Aut}(H)$ and the Lie $2$–algebra $\mathfrak{k}$ as the differential Lie crossed module $\mathfrak{h}\longrightarrow\!\!\!\!\!\!\!\!\!\!\!{}^{\dot t}\,\,\,\,\,\,\,\mathfrak{g}
\longrightarrow\!\!\!\!\!\!\!\!\!\!\!\!\!\!{}^{\widehat m}\,\,\,\,\,\,\mathfrak{der}(\mathfrak{h})$ corresponding to it.
[*A $(G,H)$–$2$–connection on $M$ is a pair formed by a $\mathfrak{g}$–valued $1$–form $\theta\in\Omega^1(M,\mathfrak{g})$ and a $\mathfrak{h}$–valued $2$–form $\varUpsilon\in\Omega^2(M,\mathfrak{h})$ such that*]{} $$\begin{aligned}
{\rm d}\theta+\frac{1}{2}[\theta,\theta]-\dot t(\varUpsilon)=0.
\label{twoholo8}\end{aligned}$$ [*(vanishing fake curvature condition). $(\theta,\varUpsilon)$ is flat if*]{} $$\begin{aligned}
{\rm d}\varUpsilon+\widehat{m}(\theta,\varUpsilon)=0.
\label{twoholo27}\end{aligned}$$
Analogously to the ordinary case, higher parallel transport requires a $(G,H)$ connection on $M$ as input datum.
[*For a curve $\gamma$ of $M$, the parallel transport $F_\theta(\gamma)\in G$ is constructed as done earlier for the $G$–connection $\theta$. For a surface $\varSigma$ of $M$, the parallel transport along $\varSigma$ is the element $F_{\theta,\varUpsilon}(\varSigma)\in H$ defined by* ]{} $$\begin{aligned}
F_{\theta,\varUpsilon}(\varSigma)=E(0,1),
\label{twofholo3}\end{aligned}$$ [*where $E:\mathbbm{R}^2\rightarrow H$ is the unique solution of the two step differential problem*]{} $$\begin{aligned}
&\partial_xu(x,y)u(x,y)^{-1}=-\varSigma^*\theta_x(x,y),\quad u(1,y)=1_G,
\label{cycle36}
\\
&\partial_yv(x,y)v(x,y)^{-1}=-\varSigma^*\theta_y(x,y),\,\,\quad v(x,0)=1_G,
\label{cycle37}
\\
&\partial_x(\partial_yE(x,y)E(x,y)^{-1})=
\\
&\qquad\qquad=-\dot m(v(1,y)^{-1}u(x,y)^{-1})(\varSigma^*\varUpsilon_{xy}(x,y))
~~\text{or}
\nonumber
\\
&\partial_y(E(x,y)^{-1}\partial_xE(x,y))=
\nonumber
\\
&\qquad\qquad=-\dot m(u(x,0)^{-1}v(x,y)^{-1})(\varSigma^*\varUpsilon_{xy}(x,y)),
\nonumber
\\
&\hspace{5cm} E(1,y)=E(x,0)=1_H
\nonumber\end{aligned}$$ [*with $u,v:\mathbbm{R}^2\rightarrow G$.*]{}
The two forms of the differential problem for $E$ are equivalent: any solution of one is automatically a solution of the other.
Higher parallel transport has several remarkable properties which extend those of the ordinary case. First, higher parallel transport along surfaces is compatible with that along their end-curves.
[*For a surface $\varSigma:\gamma_0\Rightarrow \gamma_1$ joining the curve $\gamma_0$ to the curve $\gamma_1$,* ]{} $$\begin{aligned}
&F_\theta(\gamma_1)=t(F_{\theta,\varUpsilon}(\varSigma))F_\theta(\gamma_0).
\label{twoholo14}\end{aligned}$$
Second, higher parallel transport is consistent with the operations with curves and surfaces defined in Section \[sec:holo\].
[*For any point $p$, any curves $\gamma,\gamma_1,\gamma_2$ and any surfaces $\varSigma,\varSigma_1,\varSigma_2$, relations – and the further relations*]{} $$\begin{aligned}
&F_{\theta,\varUpsilon}(I_\gamma)=1_H,
\\
&F_{\theta,\varUpsilon}(\varSigma^{-1\bullet})=F_{\theta,\varUpsilon}(\varSigma)^{-1},
\\
&F_{\theta,\varUpsilon}(\varSigma_2\bullet\varSigma_1)
=F_{\theta,\varUpsilon}(\varSigma_2)F_{\theta,\varUpsilon}(\varSigma_1),
\\
&F_{\theta,\varUpsilon}(\varSigma^{-1\circ})
=m(F_\theta(\gamma_0)^{-1})(F_{\theta,\varUpsilon}(\varSigma)^{-1}),
\\
&F_{\theta,\varUpsilon}(\varSigma_2\circ\varSigma_1)
=F_{\theta,\varUpsilon}(\varSigma_2)m(F_\theta(\gamma_2))(F_{\theta,\varUpsilon}(\varSigma_1)),\end{aligned}$$ [*hold whenever defined, where in the last two identities $\varSigma:\gamma_0\Rightarrow\gamma_1$ and $\varSigma_2:\gamma_2\Rightarrow\gamma_3$.*]{}
Third, higher parallel transport is compatible with homotopy of curves and surfaces, as defined again in Section \[sec:holo\], in the following sense.
[*For any two thinly homotopic curves $\gamma_0$, $\gamma_1$*]{} $$\begin{aligned}
\label{twoholo6/2}
&F_\theta(\gamma_1)=F_\theta(\gamma_0). \end{aligned}$$ [*For any two thinly homotopic surfaces $\varSigma_0:\gamma_{00}\Rightarrow\gamma_{01}$, $\varSigma_1:\gamma_{10}\Rightarrow\gamma_{11}$,*]{} $$\begin{aligned}
&F_\theta(\gamma_{10})=F_\theta(\gamma_{00}),
\\
&F_\theta(\gamma_{11})=F_\theta(\gamma_{01}),
\\
&F_{\theta,\varUpsilon}(\varSigma_1)=F_{\theta,\varUpsilon}(\varSigma_0).\end{aligned}$$ [*The same relations hold if $(\theta,\varUpsilon)$ is flat and $\varSigma_0$, $\varSigma_1$ are homotopic.*]{}
Higher parallel transport has an elegant $2$–categorical interpretation.
[*Higher parallel transport is equivalent to a strict $2$–functor $\bar F_{\theta,\varUpsilon}:(M,P_1M,P_2M)\rightarrow B(G,H)$ from the path $2$–groupoid $(M,P_1M,P_2M)$ of $M$ into $B(G,H)$. For a flat $(\theta,\varUpsilon)$, higher parallel transport is likewise equivalent to a strict $2$–functor $\bar F^0{}_{\theta,\varUpsilon}:(M,P_1M,P^0{}_2M)\rightarrow B(G,H)$ from the fundamental $2$–groupoid $(M,P_1M,P^0{}_2M)$ of $M$ into $B(G,H)$.*]{}
Here, with an abuse of notation, $B(G,H)$ stands for the delooping of the strict Lie 2–group corresponding to the Lie crossed module $(G,H)$.
Analogously to ordinary gauge theory, higher parallel transport is gauge covariant in the appropriate higher sense.
[*A $(G,H)$–$1$–gauge transformation is a pair of a $G$–valued map $g\in\operatorname{Map}(M,G)$ and an $\mathfrak{h}$–valued $1$–form $J\in\Omega^1(M,\mathfrak{h})$.*]{}
$1$–gauge transformations act on $2$ connections.
[*The 1–gauge transform of the $(G,H)$–$2$–connection $(\theta,\varUpsilon)$ is the $2$–connection*]{} $$\begin{aligned}
&{}^{g,J}\theta=\operatorname{Ad}g(\theta)-dgg^{-1}-\dot t(J),
\vphantom{\Big]}
\label{gauholo3}
\\
&{}^{g,J}\varUpsilon=\dot m(g)(\varUpsilon)-{\rm d}J-\frac{1}{2}[J,J]\,-
\label{gauholo4}
\\
&\hspace{3cm}-\widehat{m}(\operatorname{Ad}g(\theta) -{\rm d}gg^{-1}-\dot t(J),J).
\vphantom{\Big]}
\nonumber \end{aligned}$$ [*If $(\theta,\varUpsilon)$ is flat, $({}^{g,J}\theta,{}^{g,J}\varUpsilon)$ is flat, too.*]{}
There exits a notion of parallel transport for $1$–gauge transformations similarly to $2$–connections.
[*For a curve $\gamma$ of $M$, the gauge parallel transport along $\gamma$ is the element $G_{g,J;\theta}(\gamma)\in H$ given by*]{} $$\begin{aligned}
G_{g,J;\theta}(\gamma)=\varLambda(0),\end{aligned}$$ [*where $\varLambda:\mathbbm{R}\rightarrow H$ is the unique solution of the two–step differential problem*]{} $$\begin{aligned}
&{\rm d}_xu(x)u(x)^{-1}=-\gamma^*\theta_x(x), \quad u(1)=1_G,
\vphantom{\Big]}
\\
&\varLambda(x)^{-1}{\rm d}_x\varLambda(x)=-\dot m(u(x)^{-1}\gamma^*g(x)^{-1})(\gamma^*J_x(x)),
\\
&\hspace{6cm}\varLambda(1)=1_H.
\nonumber
\vphantom{\Big]}
\nonumber \end{aligned}$$
As ordinary parallel transport, gauge parallel transport is consistent with the operations with curves defined in Section \[sec:holo\].
[*For any point $p$ and any curves $\gamma,\gamma_1,\gamma_2$, one has*]{} $$\begin{aligned}
&G_{g,J;\theta}(\iota_p)=1_H,
\\
&G_{g,J;\theta}(\gamma^{-1_\circ})=m(F_\theta(\gamma)^{-1})(G_{g,J;\theta}(\gamma)^{-1}),
\\
&G_{g,J;\theta}(\gamma_2\circ\gamma_1)
=G_{g,J;\theta}(\gamma_2)m(F_\theta(\gamma_2))(G_{g,J;\theta}(\gamma_1)).\end{aligned}$$ [*whenever defined.*]{}
Again as ordinary parallel transport, gauge parallel transport is compatible with homotopy of curves as defined in Section \[sec:holo\].
[*For any two thinly homotopic curves $\gamma_0$, $\gamma_1$, one has* ]{} $$\begin{aligned}
G_{g,J;\theta}(\gamma_1)=G_{g,J;\theta}(\gamma_0).
\label{gauholo6/2}\end{aligned}$$
The reason why we introduced gauge parallel transport is that it enters in the $1$–gauge covariance relation of higher parallel transport in a non trivial manner.
[*For any curve $\gamma:p_0\rightarrow p_1$, one has*]{} $$\begin{aligned}
\label{gauholo16/1}
F_{{}^{g,J}\theta}(\gamma)
=g(p_1)t(G_{g,J;\theta}(\gamma))F_\theta(\gamma)g(p_0)^{-1}.\end{aligned}$$ [*For any two curves $\gamma_0,\gamma_1:p_0\rightarrow p_1$ and any surface $\varSigma:\gamma_0\Rightarrow\gamma_1$, one has*]{} $$\begin{aligned}
\label{gauholo7}
&F_{{}^{g,J}\theta,{}^{g,J}\varUpsilon}(\varSigma)=
\\
&\hspace{1cm}=m(g(p_1))\big(G_{g,J;\theta}(\gamma_1)
F_{\theta,\varUpsilon}(\varSigma)G_{g,J;\theta}(\gamma_0)^{-1}\big).
\nonumber\end{aligned}$$
Gauge parallel transport has as expected a categorical interpretation.
[*Gauge parallel transport defines a pseudonatural transformation $\bar G_{g,J;\theta}:\bar F_{\theta,\varUpsilon}\Rightarrow \bar F_{{}^{g,J}\theta,{}^{g,J}\varUpsilon}$ of parallel transport $2$–functors. If $(\theta,\varUpsilon)$ is a flat $2$–connection, gauge parallel transport defines a pseudonatural transformation $\bar G^0{}_{g,J;\theta}:\bar F^0{}_{\theta,\varUpsilon}\Rightarrow \bar F^0{}_{{}^{g,J}\theta,{}^{g,J}\varUpsilon}$ of flat parallel transport $2$–functors.*]{}
Higher gauge theory is characterized also by gauge for gauge symmetry.
[*A $(G,H)$–$2$–gauge transformation is just a mapping $\varOmega\in\operatorname{Map}(M,H)$.*]{}
$(G,H)$–$2$–gauge transformations describe gauge transformations of $(G,H)$–$1$– gauge transformations depending on an assigned $(G,H)$–$2$–connection $(\theta,\varUpsilon)$. They encode modifications $\bar G_{g,J;\theta}\Rrightarrow\bar G_{{}^{\tilde \varOmega} g_{|\theta},{}^{\tilde \varOmega}J_{|\theta};\theta}$ of gauge pseudonatural transformations of parallel transport functors. The apparently have no role in knot holonomy.
$C$– and $S$–knots {#sec:csknots}
==================
Knots are embeddings of a fixed closed model manifold into an ambient manifold $M$. Thus, knots are not simply subsets of $M$ but mappings into $M$. Knots differing by an ambient isotopy are identified.
The simplest closed model manifold is the oriented circle $C$.
[*A $C$–marking of $C$ is a pointing $p_C\in C$ of $C$.*]{}
[*A $C$–marking of an oriented manifold $M$ is a pointing $p_M\in M$ of $M$.*]{}
$C$–knots are circles embedded in $M$.
[*A marked $C$–knot of $M$ is embedding $\xi:C\rightarrow M$ of the circle $C$ into $M$ such that*]{} $$\begin{aligned}
&\xi(p_C)=p_M.\end{aligned}$$
Ambient isotopy is the natural notion of mutual deformability of marked $C$–knots.
[*Two marked $C$–knots $\xi_0,\xi_1$ are ambient isotopic if there is a smooth family $F_z\in\operatorname{Diff}_+(M)$, $z\in\mathbbm{R}$, of orientation preserving diffeomorphisms such that*]{} $$\begin{aligned}
&F_0=\operatorname{id}_M,\quad \xi_1=F_1\circ \xi_0,
\\
&F_z(p_M)=p_M.\end{aligned}$$
In order to compute $C$–knot holonomy, we need parametrized $C$–knots. This is achieved by assigning a curve to any marked $C$–knot as detailed next.
*A compatible curve in $C$ is a curve $\gamma_C:p_C\rightarrow p_C$ in $C$ such that*
i) $I_C=\gamma_C{}^{-1}(C\setminus p_C)$ is an open interval in $\mathbbm{R}$;
ii) $\gamma_C|_{I_C}:I_C\rightarrow C\setminus p_C$ is an orientation preserving diffeomorphism.
[*Example*]{}. Let $C=S^1$ be the circle standardly embedded in $\mathbbm{R}^2$ through $$s_{S^1}(\vartheta)=(\cos\vartheta,\sin\vartheta),$$ where $\vartheta\in [0,2\pi)$, with the $C$–marking $p_{S^1}=(1,0)$. A compatible curve $\gamma_{S^1}:\mathbbm{R}\rightarrow S^1$ is given by $$\gamma_{S^1}(x)=s_{S^1}(2\pi\alpha(x))
\label{gammas1}$$ where $\alpha:\mathbbm{R}\rightarrow [0,1]$ is a function such that $d_x\alpha(x)\geq 0$ and $\alpha(x)=0$ for $x<\epsilon$ and $\alpha(x)=1$ for $x>1-\epsilon$.
A curve furnishing a natural parametrization of a given marked $C$–knot can now be constructed.
[*With every marked $C$–knot $\xi$ there is associated a curve $\gamma_\xi:p_M\rightarrow p_M$ given by*]{} $$\begin{aligned}
\gamma_\xi=\xi\circ \gamma_C.\end{aligned}$$
The curve $\gamma_\xi$ has a number of nice properties.
[*$\gamma_\xi$ is independent of the choice of the compatible curve $\gamma_C$ up to thin homotopy.*]{}
Note that the $C$–marking $p_C$ of $C$ is fixed here.
Ambient isotopic marked $C$–knots have homotopic curves
[*If $\xi_0$, $\xi_1$ are ambient isotopic marked $C$–knots, then $\gamma_{\xi_0}$, $\gamma_{\xi_1}$ are homotopic in the sense of Section \[sec:holo\].*]{}
It should be possible to alter marking data changing $\gamma_\xi$ at most by a (thin) homotopy.
[*Two marked $C$–knots $\xi_0$, $\xi_1$ with respect to two distinct $C$–marking $p_{M0}$, $p_{M1}$ of $M$ are freely ambient isotopic if there is a smooth family $F_z\in\operatorname{Diff}_+(M)$, $z\in\mathbbm{R}$, of orientation preserving diffeomorphisms such that*]{} $$\begin{aligned}
&F_0=\operatorname{id}_M,\quad \xi_1=F_1\circ \xi_0.\end{aligned}$$
Again, the $C$–marking $p_C$ of $C$ is fixed.
Freely ambient isotopic marked $C$–knots have homotopic curves up to conjugation.
[*If $\xi_0$, $\xi_1$ two freely ambient isotopic marked $C$–knots, there is a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ such that $\gamma_{\xi_0}$, $\gamma_1{}^{-1_\circ }\circ\gamma_{\xi_1}\circ\gamma_1$ are homotopic.*]{}
Notice that the “compose rightmost first” convention is used here and in the following for curve composition.
The same embedding of $C$ into $M$ can be a marked $C$–knot in more than one way. The corresponding curves are related in the expected manner.
[*If the embedding $\xi:C\rightarrow M$ is a marked $C$–knot with respect to two distinct $C$–markings $p_{C0}$, $p_{M0}$ and $p_{C1}$, $p_{M1}$ of $C$ and $M$, there exists a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ in $\xi(C)$ such that $\gamma_{\xi|0}$, $\gamma_1{}^{-1_\circ}\circ\gamma_{\xi|1}\circ\gamma_1$ are thinly homotopic.*]{}
The results just expounded are standard. Our aim is finding their generalization to surface knots. As we shall see, this task is not completely straightforward. Problems occur for higher genus knots. We shall propose a solution in due course. To this end, we need to introduce further notions.
To construct higher genus $S$–knot holonomy, it will be necessary to cut the model manifold $S$ along its standard $a$– and $b$–cycles. The cuts are the images of spiky $C$–knots, generalized $C$–knots which are continuous but not smooth at the marked point.
[*A spiky $C$–knot is an embedding $\xi:C\rightarrow M$ that obeys $$\begin{aligned}
\xi(p_C)=p_M\end{aligned}$$ and is smooth on $C\setminus p_C$ with finite derivatives and non zero first derivatives at both ends of $C\setminus p_C$.*]{}
Note that spiky $C$–knots are marked.
With any spiky $C$–knot $\xi$, one can associate a curve $\gamma_\xi$ defined in the same way as above and smooth anyway.
[*For every spiky marked $C$–knot $\xi$, the curve $\gamma_\xi$ is smooth.*]{}
We can now introduce $S$–knots. The next to simplest closed manifold is a genus $\ell_S$ closed oriented surface $S$.
*An $S$–marking of $M$ consists of the following elements:*
i) a $C$–marking $p_M$ of $M$;
ii) a set of spiky $C$–knots $\zeta_{Mi}$ of $M$, $i=1,\dots,2\ell_S$, such that:
iii) the $\zeta_{Mi}(C)$ intersect only at $p_M$;
iv) there is an embedding $\varPhi:S\rightarrow M$ with the property that $\varPhi(p_S)=p_M$, $\varPhi\circ\zeta_{Si}=\zeta_{Mi}$.
Note that the notion of $S$–marking of $S$ is compatible with that of $S$–marking of $M$ when $M=S$.
$S$–knots are surfaces embedded in $M$.
[*A marked $S$–knot of $M$ an is embedding $\varXi:S\rightarrow M$ of the surface $S$ into $M$ such that*]{} $$\begin{aligned}
&\varXi(p_S)=p_M,
\\
&\varXi\circ\zeta_{Si}=\zeta_{Mi}.\end{aligned}$$
Ambient isotopy is the natural notion of mutual deformability also of marked $S$–knots.
[*Two marked $S$–knots $\varXi_0,\varXi_1$ are ambient isotopic if there is a smooth family $F_z\in\operatorname{Diff}_+(M)$, $z\in\mathbbm{R}$, of orientation preserving diffeomorphisms such that*]{} $$\begin{aligned}
&F_0=\operatorname{id}_M,\quad \varXi_1=F_1\circ \varXi_0,
\nonumber
\\
&F_z(p_M)=p_M,~~F_z\circ\zeta_{Mi}=\zeta_{Mi}.
\nonumber\end{aligned}$$
Analogously to $C$–knots, to compute $S$–knot holonomy we need parametrized $S$–knots. This is achieved by assigning a surface to any marked $S$–knot.
The fundamental polygon of $S$ is the boundary of the simply connected open $2$–fold that results cutting $S$ along the standard $a$– and $b$–cycles. It plays a basic role in the subsequent constructions.
[*View $S$ as a $C$–marked manifold and let*]{} $$\begin{aligned}
\gamma_{Si}=\gamma_{\zeta_{Si}} ~~\text{that is}~~
\alpha_{Sr}=\gamma_{\xi_{Sr}}, ~ \beta_{Sr}=\gamma_{\eta_{Sr}},\end{aligned}$$ [*Then the fundamental polygon of $S$ is the curve given by*]{} $$\begin{aligned}
&\tau_S=\beta_{S\ell_S}{}^{-1_\circ}\circ\alpha_{S\ell_S}{}^{-1_\circ}\circ\beta_{S\ell_S}\circ
\alpha_{S\ell_S}\circ
\\
&\hspace{3cm}
\cdots\circ \beta_{S1}{}^{-1_\circ}\circ\alpha_{S1}{}^{-1_\circ}\circ\beta_{S1}\circ\alpha_{S1}
\nonumber\end{aligned}$$ [*if $\ell_S=0$, $\tau_S=\iota_{p_S}$*]{}.
As a compatible curve in $C$ is required in order to associate a curve to each marked $C$–knot, a compatible surface in $S$ is required in order to associate a surface to each marked $C$–knot.
*A compatible surface in $S$ is a surface $\varSigma_S:\iota_{p_S}\rightarrow \tau_S$ such that*
i) $D_S=\varSigma_S{}^{-1}(S\setminus \cup_i\zeta_{Si}(C))$ is an open simply connected domain in $\mathbbm{R}^2$;
ii) $\varSigma_S|_{D_S}:D_S\rightarrow S\setminus \cup_i\zeta_{Si}(C)$ is an orientation preserving diffeomorphism.
[*Example*]{}. Let $S=S^2$ be the sphere embedded in $\mathbbm{R}^3$ as $$\begin{aligned}
&S_{S^2}(\vartheta,\varphi)=(\cos\vartheta\sin\vartheta(1-\cos\varphi),
$$ $$\begin{aligned}
&\hspace{3cm}-\sin\vartheta\sin\varphi,1-\sin^2\vartheta(1-\cos\varphi)),
\nonumber\end{aligned}$$ where $\vartheta\in(0,\pi)$, $\varphi\in[0,2\pi)$, with the $S$–marking $p_{S^2}=(0,0,1)$. A compatible surface $\varSigma_{S^2}:\mathbbm{R}^2\rightarrow S^2$ is given by $$\begin{aligned}
\varSigma_{S^2}(x,y)=S_{S^2}(\pi\alpha(y),2\pi\alpha(x)),\end{aligned}$$ where $\alpha:\mathbbm{R}\rightarrow [0,1]$ is a function enjoying the properties listed below Equation .
The surface $\varSigma_{S^2}$ describes a parametrized family of circles on $S^2$ which spring from the north pole on one side of it, sweep $S^2$ dilating, reaching the south pole and then contracting and finally converge to the north pole on the other side.
[*Example*]{}. Let $S=T^2$ be the torus embedded in $\mathbbm{R}^3$ as $$\begin{aligned}
&S_{T^2}(\vartheta_1,\vartheta_2)=(\cos\vartheta_1(1+r\cos\vartheta_2),
\\
&\hspace{3.5cm}\sin\vartheta_1(1+r\cos\vartheta_2),r\sin\vartheta_2),
\nonumber\end{aligned}$$ where $r<1$ is fixed and $\vartheta_1,\vartheta_2\in[0,2\pi)$, with the $S$–marking $p_{T^2}=(1+r,0,0)$ and $$\begin{aligned}
\xi_{T^2}(\vartheta)&=((1+r)\cos\vartheta,(1+r)\sin\vartheta,0),
\vphantom{\Big]}
\nonumber \\
\eta_{T^2}(\vartheta)&=(1+r\cos\vartheta,0,r\sin\vartheta),
\vphantom{\Big]}
\nonumber $$ where $\vartheta\in[0,2\pi)$. A compatible surface $\varSigma_{T^2}:\mathbbm{R}^2\rightarrow T^2$ is $$\begin{aligned}
&\varSigma_{T^2}(x,y)=S_{T^2}(2\pi c_1(x,y)),2\pi c_2(x,y)))
\vphantom{\Big]}
\nonumber \\
&c_1(x,y)=\varrho(4\alpha(x),\alpha(y))-\varrho(4\alpha(x)-2,\alpha(y)),
\vphantom{\Big]}
\nonumber \\
&c_2(x,y)=\varrho(4\alpha(x)-1,\alpha(y))-\varrho(4\alpha(x)-3,\alpha(y)),
\vphantom{\Big]}
\nonumber $$ where $\alpha:\mathbbm{R}\rightarrow [0,1]$ is a function with the same properties as before and $\varrho:\mathbbm{R}\times[0,1]\rightarrow [0,1]$ is the function given by $$\varrho(s,t)=tg_\beta\bigg(\frac{1-2s}{(1+s-t)(2-s-t)}\bigg),
\label{sample10}$$ where $g_\beta(w)=1/(\exp(\beta w)+1)$ with $\beta>0$ is the Fermi–Dirac function.
Upon unfolding the torus $T^2$ into a square $I^2$ by cutting it along the $a$– and $b$–cycle, the surface $\varSigma_{T^2}$ describes a parametrized family of closed curves on $I^2$ which spring from one corner of the square and sweep it all eventually approximating the square’s boundary.
A surface furnishing a natural parametrization of a given marked $S$–knot can now be constructed. To this end, we need to identify a curve in $M$ that matches the fundamental polygon of $S$.
[*View $M$ as a $C$–marked manifold and let*]{} $$\begin{aligned}
\gamma_{Mi}=\gamma_{\zeta_{Mi}} ~~\text{that is}~~
\alpha_{Mr}=\gamma_{\xi_{Mr}}, ~ \beta_{Mr}=\gamma_{\eta_{Mr}},\end{aligned}$$ [*Then, the fundamental polygon of the marking of $M$ is the curve*]{} $$\begin{aligned}
&\tau_M=\beta_{M\ell_S}{}^{\!-1_\circ}\circ\alpha_{M\ell_S}{}^{\!-1_\circ}\circ\beta_{M\ell_S}\circ\alpha_{M\ell_S}
\circ\\
&\hspace{2cm}\cdots\circ
\beta_{M1}{}^{\!-1_\circ}\circ\alpha_{M1}{}^{\!-1_\circ}\circ\beta_{M1}\circ\alpha_{M1}
\nonumber\end{aligned}$$ [*if $\ell_S=0$, $\tau_M=\iota_{p_M}$.*]{}
A surface furnishing a natural parametrization of a given marked $S$–knot can now be constructed.
[*With every marked $S$–knot $\varXi$, there is associated a surface $\varSigma_\varXi:\iota_{p_M}\Rightarrow \tau_M$ given by*]{} $$\begin{aligned}
\varSigma_\varXi=\varXi\circ \varSigma_S.\end{aligned}$$
Note that $\tau_M=\varXi\circ\tau_S$.
For a marked $C$–knot $\xi$, the source and target of the associated curve $\gamma_\xi:p_M\rightarrow p_M$ are equal. In gauge theory, this ensures nice ambient isotopy and gauge covariance properties of $C$–knot holonomy. For a genus $\ell_S=0$ marked $S$–knot $\varXi$, the source and target of the associated surface $\varSigma_\varXi:\iota_{p_M}\Rightarrow \iota_{p_M}$ are equal as well. In higher gauge theory, this also ensures nice ambient isotopy and gauge covariance properties of $S$–knot holonomy. However, for a genus $\ell_S>0$ marked $S$–knot $\varXi$, the source and target of the associated surface $\varSigma_\varXi:\iota_{p_M}\Rightarrow \tau_M\not=\iota_{p_M}$ are different. This is likely to be a problem for ambient isotopy and gauge covariance properties of holonomy. We have a proposal for the solution of this difficulty.
For given $\ell_S$ and $C$–marking of $M$, pick a reference marked $S$–knot $\varDelta_M$ (e. g. Hosokawa’s and Kawauchi’s surface unknots in $S^4$ [@hosokawa1979proposals]).
[*The normalized surface of a marked $S$–knot $\varXi$ is the surface $\varSigma^\sharp{}_{\varXi}:\iota_{p_M}\Rightarrow\iota_{p_M}$ given by*]{} $$\begin{aligned}
\varSigma^\sharp{}_{\varXi}=\varSigma_M{}^{-1_\bullet}\bullet\varSigma_{\varXi},\end{aligned}$$ [*where $\varSigma_M:=\varSigma_{\varDelta_M}$ and $\bullet$ denotes vertical surface composition (cf. Section \[sec:holo\]).*]{}
An intuitive way of thinking of the normalized surface of $\varXi$ is as a surface characterizing the $S$–knot “ratio” of $\varXi$ to $\varDelta_M$, with $\varDelta_M$ acting as a normalizing knot.
The normalized surface of a marked $S$–knot has nice properties.
[*$\varSigma^\sharp{}_{\varXi}$ is independent from the choice of $\varSigma_S$ and $\gamma_C$ up to thin homotopy.*]{}
Note that the markings $p_C$ and $(p_S,\zeta_{Si})$ are fixed.
Ambient isotopic reference $S$–knots yield homotopic normalized marked $S$–knot surfaces.
[*If the reference marked $S$–knots $\varDelta_{M0}$, $\varDelta_{M1}$ are ambient isotopic, then for every marked $S$–knot $\varXi$ the normalized surfaces $\varSigma^\sharp{}_{\varXi|0}$, $\varSigma^\sharp{}_{\varXi|1}$ are homotopic.*]{}
Ambient isotopic marked $S$–knots have homotopic normalized surfaces.
[*If $\varXi_0$, $\varXi_1$ are ambient isotopic marked $S$–knots, then the normalized surfaces $\varSigma^\sharp{}_{\varXi_0}$, $\varSigma^\sharp{}_{\varXi_1}$ are homotopic.* ]{}
As for $C$–knots, it should be possible to alter the marking changing $\varSigma^\sharp{}_\varXi$ by a (thin) homotopy.
[*Two marked $S$–knots $\varXi_0$, $\varXi_1$ with respect to distinct $S$–markings $(p_{M0},\zeta_{M0i})$, $(p_{M1},\zeta_{M1i})$ of $M$ are said to be freely ambient isotopic if there is a smooth family $F_z\in\operatorname{Diff}_+(M)$, $z\in\mathbbm{R}$, of orientation preserving diffeomorphisms such that*]{} $$\begin{aligned}
F_0=\operatorname{id}_M,\quad \varXi_1=F_1\circ \varXi_0.\end{aligned}$$
Notice that above the $S$–marking $(p_S,\zeta_{Si})$ of $S$ is kept fixed.
[*Two pairs $\varXi_0$, $\varXi_1$ and $\varXi_0{}'$, $\varXi_1{}'$ of freely ambient isotopic marked $S$–knots are called concordant if there exist ambient isotopies $F_z$ of $\varXi_0$, $\varXi_1$ and $F'{}_z$ of $\varXi_0{}'$, $\varXi_1{}'$ s. t. $F_z(p_{M0})=F'{}_z(p_{M0})$, $F_z\circ\zeta_{M0i}=F'{}_z\circ\zeta_{M0i}$.*]{}
Freely ambient isotopic marked $S$–knots have homotopic normalized surfaces up to conjugation under concordance with reference knots.
[*Suppose the reference marked $S$–knots $\varDelta_{M0}$, $\varDelta_{M1}$ are freely ambient isotopic. If the marked $S$–knots $\varXi_0$, $\varXi_1$ are freely ambient isotopic concordantly with $\varDelta_{M0}$, $\varDelta_{M1}$, then there is curve a $\gamma_1:p_{M0}\rightarrow p_{M1}$ such that $\varSigma^\sharp{}_{\varXi_0|0}$, $I_{\gamma_1}{}^{-1_\circ}
\circ\varSigma^\sharp{}_{\varXi_1|1}\circ I_{\gamma_1}$ are homotopic.*]{}
Before stating the next result, we recall the following property. For two $S$–markings $(p_{S0},\zeta_{S0i})$, $(p_{S1},\zeta_{S1i})$ of $S$, there is an orientation preserving ambient isotopy $k_z$ of $S$ such that $k_1(p_{S0})=p_{S1}$, $k_1\circ \zeta_{S0i}=\zeta_{S1i}$.
[*If the embeddings $\varDelta_M,\varXi:S\rightarrow M$ are simultaneously the reference and considered marked $S$–knot with respect to two distinct $S$–markings $(p_{S0},\zeta_{S0i})$, $(p_{M0},\zeta_{M0i})$ and $(p_{S1},\zeta_{S1i})$, $(p_{M1},\zeta_{M1i})$ of $S$ and $M$ and there is an ambient isotopy $k_z$ of $S$ shifting $\{p_{S0},\zeta_{S0i}\}$ to $\{p_{S1},\zeta_{S1i}\}$ such that $\varXi\circ k_z(p_{S0})=\varDelta_M\circ k_z(p_{S0})$ and $\varXi\circ k_z\circ \zeta_{S0i}=\varDelta_M\circ k_z\circ \zeta_{S0i}$, then there is a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ lying in the image $\varXi(S)$ such that $\varSigma^\sharp{}_{\varXi|0}$, $I_{\gamma_1}{}^{-1_\circ}\circ\varSigma^\sharp{}_{\varXi|1}\circ I_{\gamma_1}$ are thinly homotopic.*]{}
Relying on the above results, we can now tackle the task of constructing higher knot holonomy.
$C$– and $S$–knot holonomy {#sec:csknothol}
==========================
Our aim is constructing holonomy invariants of knots up to conjugation. We begin with reviewing how this is done for $C$–knots
We let $G$ be a Lie group and $\theta$ be a flat $G$–connection on $M$. Further, we fix $C$–markings $p_C$ and $p_M$ of $C$ and $M$, respectively.
The holonomy of a marked $C$–knot is built out of the curve associated to the knot.
[*The holonomy of a marked $C$–knot $\xi$ is the element $F_\theta(\xi)\in G$ given by* ]{} $$\begin{aligned}
F_\theta(\xi)=F_\theta(\gamma_\xi),\end{aligned}$$ [*where $\gamma_\xi:p_M\rightarrow p_M$ curve of $\xi$ (cf. Section \[sec:csknots\]) and $F_\theta$ is the parallel transport functor (cf. Section \[sec:partr\]).*]{}
$C$–knot holonomy is independent from the choice of parametrization.
[*For any marked $C$–knot $\xi$, $F_\theta(\xi)$ is independent of the choice of the compatible curve $\gamma_C$ of $C$.*]{}
$C$–knot holonomy is further invariant under ambient isotopy.
[*If $\xi_0$, $\xi_1$ are ambient isotopic marked $C$–knots of $M$, then* ]{} $$\begin{aligned}
F_\theta(\xi_1)=F_\theta(\xi_0). \end{aligned}$$
This property generalizes as follows. Fix the $C$–marking $p_C$ of $C$ but allow two distinct $C$–markings $p_{M0}$, $p_{M1}$ of $M$.
[*If $\xi_0$, $\xi_1$ are freely ambient isotopic marked $C$–knots, then there exists a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ of $M$ such that* ]{} $$\begin{aligned}
F_\theta(\xi_1)=F_\theta(\gamma_1)F_\theta(\xi_0)F_\theta(\gamma_1)^{-1}.
\nonumber $$
$C$–knot holonomy is independent of the way a given $C$–knot is marked up to conjugation.
[*If $\xi$ is a marked $C$–knot with respect to two distinct $C$–markings $p_{C0}$, $p_{M0}$ and $p_{C1}$, $p_{M1}$ of $C$ and $M$, then there is a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ lying in $\xi(C)$ such that* ]{} $$\begin{aligned}
F_{\theta|1}(\xi)=F_\theta(\gamma_1)F_{\theta|0}(\xi)F_\theta(\gamma_1)^{-1}.
\nonumber $$
$C$–knot holonomy is also gauge covariant as desired.
[*Let $\xi$ be a marked $C$–knot of $M$. Then, for any $G$–gauge transformation $g$, one has*]{} $$\begin{aligned}
F_{{}^g\theta}(\xi)=g(p_M)F_\theta(\xi)g(p_M)^{-1}.
\nonumber $$
In summary, $C$–knot holonomy is $C$–marking and gauge independent and isotopy invariant up to $G$–conjugation.
Next, using the treatment of $C$–knot holonomy presented above as a model, we illustrate the construction of $S$–knot holonomy.
We let $(G,H)$ be a Lie crossed module and $(\theta,\varUpsilon)$ be a flat $(G,H)$–$2$–connection pair on $M$. Furthermore, we fix $S$-markings $(p_S,\zeta_{Si})$ and $(p_M,\zeta_{SMi})$ of $S$ and $M$, respectively.
[*The holonomy of a marked $S$–knot $\varXi$ is the element $F_\theta(\varXi)$ $\in H$ given by*]{} $$\begin{aligned}
F_{\theta,\varUpsilon}(\varXi)=F_{\theta,\varUpsilon}(\varSigma^\sharp{}_\varXi)
=F_{\theta,\varUpsilon}(\varSigma_M)^{-1}F_{\theta,\varUpsilon}(\varSigma_\varXi),\end{aligned}$$ [*where $\varSigma^\sharp{}_\varXi:\iota_{p_M}\Rightarrow \iota_{p_M}$ is the normalized surface of $\varXi$ and $F_{\theta,\varUpsilon}$ is the parallel transport 2–functor.*]{}
The fact that $\varSigma^\sharp{}_\varXi:\iota_{p_M}\Rightarrow \iota_{p_M}$ has the following crucial consequence.
[*For a marked $S$–knot $\varXi$,* ]{} $$\begin{aligned}
t(F_{\theta,\varUpsilon}(\varXi))=1_G.\end{aligned}$$ [*Thus, $F_{\theta,\varUpsilon}(\varXi)=1_H$ unless $\ker t\not= \{1_H\}$. Further, $F_{\theta,\varUpsilon}(\varXi)\in Z_H$.*]{}
Thus, unlike $C$–knot holonomy, $S$–knot holonomy is fundamentally Abelian and non trivial only for crossed modules whose target map has non trivial kernel.
$S$–knot holonomy is independent from the choice of parametrization.
[*For every marked $S$–knot $\varXi$, $F_{\theta,\varUpsilon}(\varXi)$ is independent from the choice of the compatible surface $\varSigma_S$ of $S$ and curve $\gamma_C$ of $C$.*]{}
$S$–knot holonomy is invariant under a change of the reference marked $S$–knots in the following sense.
[*If the reference marked $S$–knots $\varDelta_{M0}$, $\varDelta_{M1}$ are ambient isotopic, then for any marked $S$–knot $\varXi$*]{} $$\begin{aligned}
F_{\theta,\varUpsilon|0}(\varXi)=F_{\theta,\varUpsilon|1}(\varXi).\end{aligned}$$
$S$–knot holonomy is further invariant under ambient isotopy.
[*If $\varXi_0$, $\varXi_1$ are ambient isotopic marked $S$–knots, then* ]{} $$\begin{aligned}
F_{{\theta,\varUpsilon}}(\varXi_1)=F_{\theta,\varUpsilon}(\varXi_0).\end{aligned}$$
This property generalizes as follows. Fix the $S$-markings $(p_S,\zeta_{Si})$ of $S$ but allow distinct $S$–marking $(p_{M0},\zeta_{M0i})$, $(p_{M1},\zeta_{M1i})$ of $M$.
[*Suppose $\varDelta_{M0}$, $\varDelta_{M1}$ are freely ambient isotopic reference marked $S$–knots. If the marked $S$–knots $\varXi_0$, $\varXi_1$ are freely ambient isotopic concordantly with $\varDelta_{M0}$, $\varDelta_{M1}$, then there is a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ such that*]{} $$\begin{aligned}
F_{\theta,\varUpsilon|1}(\varXi_1)=m(F_\theta(\gamma_1))(F_{\theta,\varUpsilon|0}(\varXi_0)).\end{aligned}$$
$S$–knot holonomy is independent of the way a given knot $S$–knot is marked up to conjugation.
[*If the embeddings $\varDelta_M,\varXi:S\rightarrow M$ are simultaneously the reference and considered marked $S$–knot with respect to two distinct $S$–markings $(p_{S0},\zeta_{S0i})$, $(p_{M0},\zeta_{M0i})$ and $(p_{S1},\zeta_{S1i})$, $(p_{M1},\zeta_{M1i})$ of $S$ and $M$ and there is an ambient isotopy $k_z$ of $S$ shifting $\{p_{S0},\zeta_{S0i}\}$ to $\{p_{S1},\zeta_{S1i}\}$ such that $\varXi\circ k_z(p_{S0})=\varDelta_M\circ k_z(p_{S0})$ and $\varXi\circ k_z\circ \zeta_{S0i}=\varDelta_M\circ k_z\circ \zeta_{S0i}$, then there is a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ lying in the image $\varXi(S)$ such that*]{} $$\begin{aligned}
F_{\theta,\varUpsilon|1}(\varXi)=m(F_\theta(\gamma_1))(F_{\theta,\varUpsilon|0}(\varXi)).\end{aligned}$$
In this higher gauge theoretic set–up, one can define also $C$–knot holonomy in the same way as before.
[*The holonomy of a marked $C$–knot $\xi$ is the element $F_\theta(\xi)\in G$ given by* ]{} $$\begin{aligned}
F_\theta(\xi)=F_\theta(\gamma_\xi)\end{aligned}$$ [*is defined.*]{}
This $C$–knot holonomy has however weaker properties than in ordinary gauge theory.
$C$–knot holonomy is still independent of the choice of parametrization.
[*For any marked $C$–knot $\xi$, $F_\theta(\xi)$ is independent of the choice of the compatible curve $\gamma_C$ of $C$.*]{}
Since, however, $\theta$ is not flat unless $\dot t(\varUpsilon)=0$, $F_\theta(\xi)$ is not ambient isotopy invariant.
[*If $\xi_0$, $\xi_1$ are two ambient isotopic marked $C$–knots of $M$, then there is a surface $\varSigma:\gamma_{\xi_0}\Rightarrow \gamma_{\xi_1}$ of $M$ such that* ]{} $$\begin{aligned}
F_\theta(\xi_1)=t(F_{\theta,\varUpsilon}(\varSigma))F_\theta(\xi_0). \end{aligned}$$
This property generalizes as follows. Fix the $C$–marking $p_C$ of $C$ but allow two distinct $C$–markings $p_{M0}$, $p_{M1}$ of $M$.
[*If $\xi_0$, $\xi_1$ are freely ambient isotopic marked $C$–knots, then there exist a curve $\gamma_1:p_{M0}\rightarrow p_{M1}$ and a surface $\varSigma:\gamma_{\xi_0}\Rightarrow\gamma_1{}^{-1_\circ }\circ\gamma_{\xi_1}\circ\gamma_1$ of $M$ such that*]{} $$\begin{aligned}
F_\theta(\xi_1)=F_\theta(\gamma_1)t(F_{\theta,\varUpsilon}(\varSigma))
F_\theta(\xi_0)F_\theta(\gamma_1)^{-1}.\end{aligned}$$
$C$–knot holonomy is again independent of the way a given $C$–knot is marked up to conjugation.
[*If $\xi$ is a marked $C$–knot with respect to two distinct $C$–markings $p_{C0}$, $p_{M0}$ and $p_{C1}$, $p_{M1}$ of $C$ and $M$, then there is $\gamma_1:p_{M0}\rightarrow p_{M1}$ curve in $\xi(C)$ such that* ]{} $$\begin{aligned}
F_{\theta|1}(\xi)=F_\theta(\gamma_1)F_{\theta|0}(\xi)F_\theta(\gamma_1)^{-1}.
\label{xiholo2}\end{aligned}$$
Also in higher gauge theory, $S$– and $C$–knot holonomy is gauge covariant in the appropriate sense.
[*Let $\varXi$ be a marked $S$–knot and $\xi$ a marked $C$–knot. If $(g,J)$ is a $(G,H)$–$1$–gauge transformation, then*]{} $$\begin{aligned}
F_{{}^{g,J}\theta,{}^{g,J}\varUpsilon}(\varXi)=m(g(p_M))(F_{\theta,\varUpsilon}(\varXi))\end{aligned}$$ [*and* ]{} $$\begin{aligned}
F_{{}^{g,J}\theta}(\xi)=g(p_M)t(G_{g,J;\theta}(\gamma_\xi))F_\theta(\gamma)g(p_M)^{-1},\end{aligned}$$ [*where $G_{g,J;\theta}(\gamma_\xi)$ is the gauge parallel transport along $\gamma_\xi$ defined in Section \[sec:partr\].*]{}
To summarize, $C$– and $S$–knot holonomy are $C$–marking and gauge independent and isotopy invariant up to the appropriate form of crossed module conjugation. We shall analyze this point in greater depth in the next section.
Invariant traces {#sec:invtr}
================
Having applications to knot topology in mind, we aim at a construction of holonomy invariants. This requires working out invariant traces.
We let $G$ again be a Lie group and $\theta$ be a flat $G$–connection on $M$. Further, we let $C$–markings $p_C$ and $p_M$ of $C$ and $M$, respectively, be given.
We have seen in Section \[sec:csknothol\] that for a $C$–knot $\xi$, its holonomy $F_\theta(\xi)$ is $C$–marking and isotopy invariant and gauge independent up to $G$–conjugation, that is $$\begin{aligned}
F_\theta(\xi) \equiv a F_\theta(\xi) a^{-1}\end{aligned}$$ for $a\in G$.
There is a well established way of extracting $C$–knot invariants from knot holonomy.
[ *The Wilson line*]{} $$\begin{aligned}
W_{R,\theta}(\xi)=\operatorname{tr}_R(F_\theta(\xi)),\end{aligned}$$ [*with $R$ a representation of $G$ provides a $C$–knot invariant.*]{}
Next, taking the procedure just reviewed to construct $C$–knot holonomy invariants as a model, we propose a systematic way to build $S$–knot holonomy invariants.
We let $(G,H)$ be a Lie crossed module and $(\theta,\varUpsilon)$ be a flat $(G,H)$–$2$–connection pair on $M$. Furthermore, we let $S$-markings $(p_S,\zeta_{Si})$ and $(p_M,\zeta_{SMi})$ of $S$ and $M$, respectively, be given.
In Section \[sec:csknothol\], we have also seen that for a $C$–knot $\xi$ and an $S$–knot $\varXi$, the holonomy $F_\theta(\xi)$ and $F_{\theta,\Upsilon}(\varXi)$ is $C$– and $S$–marking and isotopy invariant and gauge independent up to $(G,H)$–conjugation $$\begin{aligned}
&F_\theta(\xi) \equiv a F_\theta(\xi) a^{-1}t(A),
\\
&F_{\theta,\Upsilon}(\varXi)\equiv m(a)(F_{\theta,\Upsilon}(\varXi))\end{aligned}$$ with $(a,A)\in G\times H$. $(G,H)$–conjugation is defined by $$\begin{aligned}
u'=aua^{-1}t(A),\qquad U'=m(a)(U)\end{aligned}$$ with $(u,U),(u',U'),(a,A)\in G\times H$ and is an equivalence relation.
To obtain knot invariants, one needs traces invariant under $(G,H)$–conjugation. To this end, one could proceed as follows.
Assume $G$, $H$ are compact with bi-invariant Haar measures $\mu_G$, $\mu_H$. Pick representations $R$, $S$ of $G$, $H$. Set $$\begin{aligned}
&\operatorname{tr}_{R,S|b}(u)=\int_Hd\mu_H(X)\operatorname{tr}_R(ut(X)),
\label{trb}
\\
&\operatorname{tr}_{R,S|f}(U)=\int_Gd\mu_G(x)\operatorname{tr}_S(m(x)(U)),
\label{trf}\end{aligned}$$ $(u,U)\in G\times H$.
[*The traces $\operatorname{tr}_{R,S|b}$, $\operatorname{tr}_{R,S|f}(U)$ are invariant under $(G,H)$ conjugation*]{}, $$\begin{aligned}
&\operatorname{tr}_{R,S|b}(aua^{-1}t(A))=\operatorname{tr}_{R,S|b}(u),
\\
&\operatorname{tr}_{R,S|f}(m(a)(U))=\operatorname{tr}_{R,S|f}(U)\end{aligned}$$ [*for $(u,U), (a,A)\in G\times H$*]{}.
These invariant traces can be used to extract $C$– and $S$–knot invariants from knot holonomy as in the ordinary case.
[*The Wilson line and surface*]{} $$\begin{aligned}
&W_{R,S,\theta|b}(\xi)=\operatorname{tr}_{R,S|b}(F_\theta(\xi)),
\\
&W_{R,S,\theta,\Upsilon|f}(\varXi)=\operatorname{tr}_{R,S|f}(F_{\theta,\Upsilon}(\varXi)). \end{aligned}$$ [*provide a $C$– and $S$–knot invariant.*]{}
There is a problem with this way of proceeding. The traces may be trivial. For instance, if $t(H)=G$, $\operatorname{tr}_{R,Sb}(u)$ does not depend on $u$ and $\operatorname{tr}_{R,Sf}(U)=\operatorname{tr}_S(U)$ for $U\in \ker t$ (the case of interest for surface knots).
In ordinary gauge theory with gauge group $G$, a trace is a map $\operatorname{tr}:G\rightarrow \mathbbm{C}$ invariant under the action $$\begin{aligned}
a{{\,\vartriangleright\,}}u:=aua^{-1}\end{aligned}$$ with $a,u\in G$, that is $$\begin{aligned}
\operatorname{tr}(a{{\,\vartriangleright\,}}u)=\operatorname{tr}(u). \end{aligned}$$ If $G$ is a compact Lie group, then $\operatorname{tr}$ reduces to a linear combination of ordinary traces $\operatorname{tr}_R$ associated with the irreducible representations $R$ of $G$.
What matters is not the group structure of $G$ but its conjugation structure codified in the conjugation pointed quandle of $G$.
[*A pointed quandle is a set $G$ with a binary operation ${{\,\vartriangleright\,}}:G\times G\rightarrow G$ and a distinguished element $1_G\in G$ such that*]{} $$\begin{aligned}
&a{{\,\vartriangleright\,}}a=a,
\\
&a{{\,\vartriangleright\,}}(b{{\,\vartriangleright\,}}c)=(a{{\,\vartriangleright\,}}b){{\,\vartriangleright\,}}(a{{\,\vartriangleright\,}}c)\end{aligned}$$ [*with $a,b,c\in G$. Further, the map $a{{\,\vartriangleright\,}}\cdot:G\rightarrow G$ is invertible for any $a\in G$ and*]{} $$\begin{aligned}
a{{\,\vartriangleright\,}}1_G=1_G,\quad 1_G{{\,\vartriangleright\,}}a=a\end{aligned}$$ [*for $a\in G$*]{}.
In higher gauge theory with gauge crossed module $(G,H)$, a similar point of view is appropriate. A trace pair is a pair of maps $\operatorname{tr}_b:G\rightarrow \mathbbm{C}$, $\operatorname{tr}_f:H\rightarrow \mathbbm{C}$ invariant under the action $$\begin{aligned}
&a{{\,\vartriangleright\,}}u:=aua^{-1}~,
\\
&A{{\,\succ\,}}u:=ut(A)~,
\\
&a{{\,\vartriangleright\,}}U:=m(a)(U)\end{aligned}$$ with $a,u\in G$, $A,U\in H$, that is $$\begin{aligned}
&\operatorname{tr}_b(a{{\,\vartriangleright\,}}u)=\operatorname{tr}_b(u),
\\
&\operatorname{tr}_b(A{{\,\succ\,}}u)=\operatorname{tr}_b(u),
\\
&\operatorname{tr}_f(a{{\,\vartriangleright\,}}U)=\operatorname{tr}_f(U).\end{aligned}$$
What matters is not $(G,H)$ itself but its conjugation augmented pointed quandle crossed module $(G,H)$ [@Crans:2004ve; @Crans:1310.4705; @Zucchini:2015xba]:
*An augmented pointed quandle crossed module is a pair of sets $G$, $H$ endowed with three operations ${{\,\vartriangleright\,}}:G\times G\rightarrow G$, $H\times H\rightarrow H$, $G\times H\rightarrow H$ and distinguished elements $1_G\in G$, $1_H\in H$ such that*
i) $G$ is a pointed quandle,
ii) $H$ is a pointed quandle
and the following requirements are satisfied.
The relations $$\begin{aligned}
&a{{\,\vartriangleright\,}}(b{{\,\vartriangleright\,}}A)=(a{{\,\vartriangleright\,}}b){{\,\vartriangleright\,}}(a{{\,\vartriangleright\,}}A),
\\
&a{{\,\vartriangleright\,}}(A{{\,\vartriangleright\,}}B)=(a{{\,\vartriangleright\,}}A){{\,\vartriangleright\,}}(a{{\,\vartriangleright\,}}B)\end{aligned}$$ with $a,b\in G$, $A,B\in H$ hold.
For any $a\in G$, the map $a{{\,\vartriangleright\,}}\cdot:H\rightarrow H$ is invertible.
For $a\in G$, $A\in H$, the relations $$\begin{aligned}
&1_G{{\,\vartriangleright\,}}A=A,
\\
&a{{\,\vartriangleright\,}}1_H=1_H\end{aligned}$$ are fulfilled.
Further, a quandle morphism $\alpha:H\rightarrow G$ (a map respecting ${{\,\vartriangleright\,}}$ and $1$) is given such that $$\begin{aligned}
&\alpha(a{{\,\vartriangleright\,}}A)=a{{\,\vartriangleright\,}}\alpha(A)~,
\label{trg}
\\
&\alpha(A){{\,\vartriangleright\,}}B=A{{\,\vartriangleright\,}}B
\label{peif}\end{aligned}$$ with $a\in G$, $A,B\in H$.
Finally, an augmentation map ${{\,\succ\,}}:H\times G\rightarrow G$ is given with the following properties.
For $a,b\in G$, $A\in H$, $$\begin{aligned}
a{{\,\vartriangleright\,}}(A{{\,\succ\,}}b)=(a{{\,\vartriangleright\,}}A){{\,\succ\,}}(a{{\,\vartriangleright\,}}b)~.\end{aligned}$$
For $A\in H$, $A{{\,\succ\,}}\cdot:G\rightarrow G$ is invertible.
For $a\in G,~A\in H$ $$\begin{aligned}
&A{{\,\succ\,}}1_G=\alpha(A),
\\
&1_H{{\,\succ\,}}a=a.\end{aligned}$$
Notice that $\alpha$ is the quandle crossed module analog of the group crossed module target morphisms. In particular is the quandle counterpart of the Peiffer identity.
The following question is still open. If $G$, $H$ are compact, does a trace pair $\operatorname{tr}_b$, $\operatorname{tr}_f$ reduce to linear combinations of traces $\operatorname{tr}_{R,S|b}$, $\operatorname{tr}_{R,S|f}$ of the form , with $R$, $S$ irreducible representations of $G$, $H$, respectively?
Higher Chern–Simons theory
==========================
To compute knot invariants in quantum field theory, one needs Chern–Simons theory. This has been known for a long time since Witten’s 1988 paper [@Witten:1988hf].
Chern–Simons theory is a Schwarz type topological gauge theory on a closed 3-dimensional manifold $M_3$. Suppose that $G$ is the gauge group and $\mathfrak{g}$ is its Lie algebra. Suppose further that $\mathfrak{g}$ is equipped with a properly normalized invariant non singular bilinear form $(\cdot,\cdot):\mathfrak{g}\times\mathfrak{g}\rightarrow \mathfrak{g}$ so that $$\begin{aligned}
([z,x],y)+x,[z,y])=0\end{aligned}$$ with $x,y,z\in \mathfrak{g}$.
[*The Chern–Simons action is given by*]{} $$\begin{aligned}
\mathrm{CS}(\theta)=\frac{k}{4\pi}\int_{M_3}\bigg(\theta,{\rm d}\theta+\frac{1}{3}[\theta,\theta]\bigg)\end{aligned}$$ [*with $\theta$ a $G$–connection.*]{} The coefficient $k$ is called level.
[*The Chern–Simons field equations are equivalent to the flatness condition of $\theta$ (cf. Equation ):*]{} $$\begin{aligned}
{\rm d}\theta+\frac{1}{2}[\theta,\theta]=0.\end{aligned}$$
[*The Chern–Simons action is invariant under a $G$–gauge transformations $g$ only modulo $2\pi\mathbbm{Z}$:*]{} $$\begin{aligned}
\mathrm{CS}({}^g\theta)=\mathrm{CS}(\theta)-2\pi k\cdot\mathrm{wn}(g), \end{aligned}$$ [*where $\mathrm{wn}(g)$ is the winding number of $g$.*]{}
[*Quantum gauge invariance holds if the level $k$ is integer.*]{}
Chern–Simons correlators of Wilson loop $W_{R,\theta}(\xi)$ yield knot invariants, for instance:
$G=\operatorname{SU}(2)$, $R=F$ $\Rightarrow$ Jones polynomial;
$G=\operatorname{SU}(n)$, $R=F$ $\Rightarrow$ HOMFLY polynomial;
$G=\operatorname{SO}(n)$, $R=F$ $\Rightarrow$ Kauffman polynomial...
In the Chern–Simons path integral, $\theta$ is not flat and consequently $W_{R,\theta}(\xi)$ is not ambient isotopy invariant. However, the theory somehow localizes on the moduli space of flat connections even though it is not a cohomological topological field theory. This has been proven by Beasley and Witten for $M_3$ Seifert, e. g. $S^1\times S^2$, $S^3$, …. Therefore, Chern–Simons Wilson loop correlators $W_{R,\theta}(\xi)$ furnish genuine knot invariants.
In order to compute surface knots invariants in quantum field theory, one needs a higher version of Chern–Simons theory, 2-Chern–Simons theory. We have a proposal for such a model. There are however unsolved problems to be discussed below.
2–Chern–Simons theory is a Schwarz type topological gauge theory on a closed 4-dimensional manifold $M_4$. Assume that $H\longrightarrow\!\!\!\!\!\!\!\!\!\!\!{}^t\,\,\,\,\,\,\,G
\longrightarrow\!\!\!\!\!\!\!\!\!\!\!\!\!{}^m\,\,\,\,\operatorname{Aut}(H)$ is the gauge Lie crossed module and that $\mathfrak{h}\longrightarrow\!\!\!\!\!\!\!\!\!\!\!{}^{\dot t}\,\,\,\,\,\,\,\mathfrak{g}
\longrightarrow\!\!\!\!\!\!\!\!\!\!\!\!\!\!{}^{\widehat m}\,\,\,\,\,\,\mathfrak{der}(\mathfrak{h})$ is its differential Lie crossed module. Assume further $(\mathfrak{g},\mathfrak{h})$ is equipped with a properly normalized invariant non singular bilinear pairing $(\cdot,\cdot):
\mathfrak{g}\times\mathfrak{h}\rightarrow \mathbbm{R}$ such that $$\begin{aligned}
&(\dot t(X),Y)-(\dot t(Y),X)=0,
\nonumber
\\
&([y,x],X)+(x,\widehat{m}(y)(X))=0
\nonumber\end{aligned}$$ with $x,y\in\mathfrak{g}$, $X,Y\in\mathfrak{h}$. Note that this requires that the crossed module is balanced, that is $\dim\mathfrak{g}=\dim\mathfrak{h}$.
[*The 2–Chern–Simons action is given by*]{} $$\begin{aligned}
\mathrm{CS}_2(\theta,\Upsilon)
=\kappa_2\int_{M_4} \bigg({\rm d}\theta+\frac{1}{2}[\theta,\theta]-\frac{1}{2}\dot t(\Upsilon),\Upsilon\bigg),\end{aligned}$$ [*with $(\theta,\Upsilon)$ a $(G,H)$ $2$–preconnection [@Soncini:2014ara; @Zucchini:2015ohw]. $\kappa_2$ is a coefficient analog to level.*]{}
A $(G,H)$ $2$–preconnection is just a pair $(\theta,\Upsilon)\in\Omega^1(M_4,\mathfrak{g})\times\Omega^2(M_4,\mathfrak{h})$. $(\theta,\Upsilon)$ a $(G,H)$–2–connection if in addition it satisfies the vanishing fake curvature condition .
[*The $2$–Chern–Simons field equations are equivalent to $(\theta,\Upsilon)$ being a flat $(G,H)$–2–connection (cf. Equation ), ):*]{} $$\begin{aligned}
&{\rm d}\theta+\frac{1}{2}[\theta,\theta]-\dot t(\varUpsilon)=0,
\label{}
\\
&{\rm d}\varUpsilon+\widehat{m}(\theta,\varUpsilon)=0.
\label{}\end{aligned}$$ [*Thus, at once, $(\theta,\Upsilon)$ satisfies the vanishing fake curvature condition, which makes it a genuine $(G,H)$–2–connection, and the vanishing curvature condition, which characterizes as a flat one*]{}.
This is quite nice, but it signals a potential problem for the construction of 2–Chern–Simons theory as a full quantum field theory. Since $(\theta,\Upsilon)$ does not obey the zero fake curvature condition in the 2–Chern–Simons path integral, the insertion of Wilson surfaces $W_{R,S,\theta,\Upsilon}(\varXi)$ of surface knots $\varXi$ in the path integral is problematic, as the definition of the $W_{R,S,\theta,\Upsilon}(\varXi)$ requires that condition in a basic way.
Another unexpected feature of the model concerns $1$–gauge invariance.
[*The 2-Chern–Simons action is invariant under $(G,H)$–1–gauge transformation $(g,J)$,*]{} $$\begin{aligned}
\mathrm{CS}_2({}^{g,J}\theta,{}^{g,J}\theta\Upsilon)=\mathrm{CS}_2(\theta,\Upsilon).
\label{}\end{aligned}$$
In $2$–Chern–Simons theory, there is no shift by some kind of higher winding number such to cause level quantization as in the ordinary Chern–Simons model. This surprising and somewhat disappointing finding can be explained by hypothesising that either all $(G,H)$–1–gauge transformations $(g,J)$ are small unlike ordinary gauge transformation or that we are missing all the topologically non trivial $(G,H)$–1–gauge transformations. This is still an open problem.
In spite of these open issues, the possibility of obtaining surface knot invariants as correlators of Wilson surface insertion in 2–Chern–Simons theory remains an intriguing possibility. Here are further reasons for this.
Studying pull–backs of knots may be interesting.
All orientation preserving diffeomorphisms $f\in\operatorname{Diff}_+(C)$ of the circle $C$ are homotopic to $\operatorname{id}_C$. Consequently, for a $C$–knot $\xi$, the curves $\gamma_\xi$, $\gamma_{f^*\xi}$ are thinly homotopicand the $C$–knots $\xi$ and $f^*\xi$ have the same holonomy.
Conversely, for a higher genus surface $S$, not all orientation preserving diffeomorphisms $f\in\operatorname{Diff}_+(S)$ of $S$ are homotopic to $\operatorname{id}_S$. Consequently, for a $S$–knot $\varXi$, the normalized surfaces $\varSigma^\sharp{}_\varXi$, $\varSigma^\sharp{}_{f^*\varXi}$ are not thinly homotopic and the $S$–knots $\varXi$ and $f^*\varXi$ do not have the same holonomy in general
This suggests that $S$–knot invariants computed using higher gauge theory may have interesting covariance properties under the mapping class group $$\begin{aligned}
\text{MCG}_+(S)=\operatorname{Diff}_+(S)/\operatorname{Diff}_0(S),\end{aligned}$$ about which there exists a well–developed mathematical theory.
We thank R. Picken for his interest in the subject and for many useful discussions.
[^1]: Corresponding author e-mail: [](mailto:[email protected])
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We classify the distance-regular Cayley graphs with least eigenvalue $-2$ and diameter at most three. Besides sporadic examples, these comprise of the lattice graphs, certain triangular graphs, and line graphs of incidence graphs of certain projective planes. In addition, we classify the possible connection sets for the lattice graphs and obtain some results on the structure of distance-regular Cayley line graphs of incidence graphs of generalized polygons.'
author:
- Alireza Abdollahi
- 'Edwin R. van Dam'
- Mojtaba Jazaeri
date: 'Received: date / Accepted: date'
title: 'Distance-regular Cayley graphs with least eigenvalue $-2$'
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction
============
Distance-regular graphs form an important class of graphs in the area of algebraic graph theory. Originally, they were defined as a generalization of distance-transitive graphs, and many of them are not even vertex-transitive. For background on distance-regular graphs, we refer to the monograph by Brouwer, Cohen, and Neumaier [@BCN] and the recent survey by Van Dam, Koolen, and Tanaka [@DKT]. Here we study the question which distance-regular graphs are Cayley graphs. This question has been well-studied for distance-regular graphs with diameter two, that is, for strongly regular graphs, see the survey paper on partial difference sets by Ma [@Ma]. Miklavič and Potočnik [@MP1; @MP2] classified the distance-regular circulant graphs and distance-regular Cayley graphs on dihedral groups, whereas Miklavič and Šparl [@MS] studied a particular class of distance-regular Cayley graphs on abelian groups. See also the monograph by Konstantinova [@EK] for some basic facts and problems on Cayley graphs and distance-regular graphs.
It is well-known that graphs with least eigenvalue $-2$ have been classified by using root lattices, see [@BCN §3.12]. In particular, it follows that a distance-regular graph with least eigenvalue $-2$ is strongly regular or the line graph of a regular graph with girth at least five. The strongly regular graphs with least eigenvalue $-2$ have been classified by Seidel [@Se]. We will give an overview of which of these graphs is a Cayley graph and in particular, we will classify the possible connection sets for the lattice graphs, using some general results that we obtain for the distance-regular line graphs of incidence graphs of generalized polygons. We will also classify the Cayley graphs with diameter three among the distance-regular line graphs, in particular the line graphs of Moore graphs and the line graphs of incidence graphs of projective planes. What remains open is to classify which line graphs of incidence graphs of generalized quadrangles and hexagons are Cayley graphs.
Preliminaries
=============
Let $G$ be a finite group with identity element $e$ and $S\subseteq G\setminus\{e\}$ be a set such that $S=S^{-1}$ (we call $S$ inverse-closed). An (undirected) Cayley graph $\operatorname{Cay}(G,S)$ with connection set $S$ is the graph whose vertex set is $G$ and where two vertices $a$ and $b$ are adjacent (denoted by $a \sim b$) whenever $ab^{-1}\in S$. The Cayley graph $\operatorname{Cay}(G,S)$ is connected if and only if the subgroup $\langle S \rangle$ generated by $S$ is equal to $G$. In the literature, it is sometimes assumed explicitly that a Cayley graph is connected. In this case, the connection set is also called a generating set. Here we follow the terminology used by Alspach [@BW]. We denote the order of an element $a \in G$ by $O(a)$, the subgroup generated by $a$ by ${\langle a\rangle}$ and the cyclic group of order $n$ by $\mathbb{Z}_{n}$. Furthermore, the cycle graph of order $m$ is denoted by $C_{m}$ and the line graph of a graph $\Gamma$ by $L(\Gamma)$.
Distance-regular graphs
-----------------------
A strongly regular graph with parameters $(v,k,\lambda,\mu)$ is a $k$-regular graph with $v$ vertices such that every pair of adjacent vertices has $\lambda$ common neighbors and every pair of non-adjacent vertices has $\mu$ common neighbors. Here we exclude disjoint unions of complete graphs and edgeless graphs, and therefore strongly regular graphs are connected with diameter two.
A connected graph with diameter $d$ is distance-regular whenever for all vertices $x$ and $y$, and all integers $i,j\leq d$, the number of vertices at distance $i$ from $x$ and distance $j$ from $y$ depends only on $i$, $j$, and the distance between $x$ and $y$. A distance-regular graph with diameter two is the same as a strongly regular graph.
A generalized $d$-gon is a point-line incidence structure whose (bipartite) incidence graph has diameter $d$ and girth $2d$. It is of order $(s,t)$ if every line contains $s+1$ points, and every point is on $t+1$ lines. For $s=t$, both the incidence graph and its line graph are distance-regular. This line graph can also be viewed as the point graph of a generalized $2d$-gon of order $(s,1)$. For some basic background on generalized polygons, we refer to the monographs by Godsil and Royle [@Go §5.6] and Brouwer, Cohen, and Neumaier [@BCN §6.5].
The (adjacency) spectrum of a graph is the multiset of eigenvalues of its adjacency matrix. As mentioned in the introduction, distance-regular graphs with least eigenvalue $-2$ can be classified. In particular, we have the following.
\[thm:drg-2\] [@BCN Thm. 3.12.4 and 4.2.16] Let $\G$ be a distance-regular graph with least eigenvalue $-2$. Then $\G$ is a cycle of even length, or its diameter $d$ equals $2,3,4,$ or $6$. Moreover,
- If $d=2$, then $\G$ is a cocktail party graph, a triangular graph, a lattice graph, the Petersen graph, the Clebsch graph, the Shrikhande graph, the Schläfli graph, or one of the three Chang graphs,
- If $d=3$, then $\G$ is the line graph of the Petersen graph, the line graph of the Hoffman-Singleton graph, the line graph of a strongly regular graph with parameters $(3250,57,0,1)$, or the line graph of the incidence graph of a projective plane,
- If $d=4$, then $\G$ is the line graph of the incidence graph of a generalized quadrangle of order $(q,q)$,
- If $d=6$, then $\G$ is the line graph of the incidence graph of a generalized hexagon of order $(q,q)$.
Recall that the triangular graph $T(n)$ is the line graph of the complete graph $K_{n}$, the lattice graph $L_{2}(n)$ is the line graph of the complete bipartite graph $K_{n,n}$ (a generalized $2$-gon), and the cocktail party graph $CP(n)$ is the complete multipartite graph with $n$ parts of size two. Note also that a projective plane is a generalized $3$-gon.
We note that the distance-regular graphs with least eigenvalue [*larger*]{} than $-2$ are also known. Besides the complete graphs (with least eigenvalue $-1$), there are the cycles of odd length, and these are clearly Cayley graphs.
Vertex-transitivity and edge-transitivity
-----------------------------------------
Recall that a graph $\Gamma$ is vertex-transitive whenever the automorphism group of $\Gamma$ acts transitivity on the vertex set of $\Gamma$, i.e. if $x$ is a fixed vertex of $\Gamma$, then $\{x^{\sigma}|\sigma \in \operatorname{Aut}(\Gamma)\}$ is equal to the set of vertices of $\Gamma$. It is clear that Cayley graphs are vertex-transitive. In fact, a graph $\Gamma$ is a Cayley graph if and only if the automorphism group $\operatorname{Aut}(\Gamma)$ of $\Gamma$ contains a regular subgroup, see [@BW Thm. 2.2].
A graph $\Gamma$ is called edge-transitive whenever the automorphism group of $\Gamma$ acts transitivity on the edge set of the graph. Because line graphs play an important role in this paper, also the concept of edge-transitivity is relevant. Indeed, the following result provides us with a connection between the vertex-transitivity of the line graph of a graph $\Gamma$ and the edge-transitivity of $\Gamma$.
[@Sa Thm. 5.3] \[line isomorphism\] Let $\Gamma$ be a connected graph which is not isomorphic to the complete graphs $K_{2}$, $K_{4}$, a triangle with an extra edge attached, and two triangles sharing an edge. Then the automorphism group of $\Gamma$ and its line graph are isomorphic, with the natural group isomorphism $\varphi : \operatorname{Aut}(\Gamma) \rightarrow \operatorname{Aut}(L(\Gamma))$, defined by $\varphi(\sigma)=\widetilde{\sigma}$ for $\sigma \in \operatorname{Aut}(\Gamma)$, where $\widetilde{\sigma}$ acts on the line graph of $\Gamma$ such that $\widetilde{\sigma}(\{v,w\})=\{\sigma(v),\sigma(w)\}$, where $v$ and $w$ are adjacent in $\Gamma$.
\[lem:edgevertextransitive\] A connected regular graph is edge-transitive if and only if its line graph is vertex-transitive.
Let $\Gamma$ be connected and regular. If $\Gamma$ is isomorphic to $K_{2}$ or $K_{4}$, then $\Gamma$ is edge-transitive and the line graph of $\Gamma$ is vertex-transitive. On the other hand, if $\Gamma$ is not isomorphic to $K_{2}$ or $K_{4}$, then the automorphism group of $\Gamma$ and its line graph are isomorphic with the natural group isomorphism by Theorem \[line isomorphism\], which completes the proof.
Groups and products
-------------------
Two subgroups $H$ and $K$ in $G$ are conjugate whenever there exists an element $g \in G$ such that $K=g^{-1}Hg$. The semidirect product $G$ of a group $N$ by a group $H$ is denoted by $H\ltimes N$ or $N\rtimes H$. It has the property that it contains a normal subgroup $N_1$ isomorphic to $N$ and a subgroup $H_1$ isomorphic to $H$ such that $G=N_1H_1$ and $N_1\cap H_1=\{e\}$.
Let $G$ be a finite group with subgroups $H$ and $K$ such that $G=HK$ and the intersection of $H$ and $K$ is the identity of $G$. Then $G$ is called a general product of $H$ and $K$ (see [@Co]).
Finally we mention a result that we will use in Section \[sec:projplane\].
[@R Thm. 9.1.2] \[Hall\] Let $N$ be a normal subgroup of a finite group $G$, and let $n=|N|$, and $m=[G:N]$. Suppose that $n$ and $m$ are relatively prime. Then $G$ contains subgroups of order $m$ and any two of them are conjugate in $G$.
Some results on generalized polygons {#sec:genpol}
====================================
Let $\G$ be a distance-regular line graph of the incidence graph of a generalized $d$-gon of order $(q,q)$. Then $\G$ can also be seen as the point graph of a generalized $2d$-gon of order $(q,1)$. It follows that each vertex of $\G$ is contained in two maximal cliques, of size $q+1$, and every edge of $\G$ is contained in a unique maximal clique. Thus, $\G$ does not have $K_{1,3}$ nor $K_{1,2,1}$ as an induced subgraph. Moreover, $\G$ has diameter $d$ and every induced cycle in $\G$ is either a $3$-cycle or a $2d$-cycle. We will use these properties to derive some general results on the structure of the connection set in case $\G$ is a Cayley graph.
\[Cayley line structure\] Let $d \geq 2$, let $\Gamma$ be the line graph of the incidence graph of a generalized $d$-gon of order $(q,q)$, and suppose that $\Gamma$ is a Cayley graph $\operatorname{Cay}(G,S)$. Then there exist two subgroups $H$ and $K$ of $G$ such that $S=(H \cup K) \setminus \{e\}$, with $|H|=|K|=q+1$ and $H \cap K=\{e\}$ if and only if ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ of order $2d$ in $S$.
One direction is clear: if there are subgroups $H$ and $K$ of $G$ such that $S=(H \cup K) \setminus \{e\}$, then ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ in $S$. To show the other direction, assume that ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ of order $2d$ in $S$.
We first claim that ${\langle a\rangle} \subseteq S \cup \{e\}$ for [*all*]{} $a \in S$. In order to prove this, let $a\in S$ and $n=O(a) \neq 2d$. If $n=2$ or $3$, then ${\langle a\rangle} \subseteq S \cup \{e\}$ since $S=S^{-1}$. If $n \geq 4$, then it is clear that the induced subgraph $\G_{{\langle a\rangle}}$ of $\G$ on ${\langle a\rangle}$ contains a cycle $e \sim a \sim a^2 \sim \cdots \sim a^{n-1} \sim e$ of length $n$ (see also [@AV Lemma 2.6]). Because $n \neq 3$ and $n\neq 2d$, it follows that this cycle is not an induced cycle. Thus, there must be an extra edge in $\G_{{\langle a\rangle}}$, that is, an edge that is not generated by $a$ or $a^{-1}$, and hence $a^i \in S$ for some $i$ with $1<i<n-1$. Now $e$ is adjacent to $a,a^{-1}$, and $a^i$, and because $\G$ does not contain an induced subgraph $K_{1,3}$, it follows that $a^2 \in S$, or $a^{i-1}\in S$, or $a^{i+1}\in S$. Let us consider the case that $a^{i-1} \in S$, with $i>2$. By considering the induced subgraph on $\{e,a,a^{i-1},a^{i}\}$, it follows that $a^{i-2}\in S$ because $\G$ does not contain an induced subgraph $K_{1,2,1}$. Similarly, by considering the induced subgraph on $\{e,a^{-1},a^{i-1},a^{i}\}$, it follows that $a^{i+1}\in S$. By repeating this argument, it follows that ${\langle a\rangle} \subseteq S \cup \{e\}$. The other cases go similarly, which proves our claim.
Let $H$ and $K$ be the two cliques of size $q+1$ that contain $e$. Then $S=(H \cup K) \setminus \{e\}$ and $H \cap K=\{e\}$. What remains to be shown is that $H$ and $K$ are subgroups of $G$.
Let $a\in H \setminus \{e\}$. Because the graph induced on ${\langle a\rangle}$ is a clique, and there are no edges between $H \setminus \{e\}$ and $K \setminus \{e\}$, it follows that ${\langle a\rangle} \subseteq H$, In particular, $a^{-1}\in H$.
Now let $a,b \in H$, and let us show that $ba^{-1} \in H$, thus showing that $H$ is a subgroup of $G$. If $a=b$, $a=e$, or $b=e$, then this clearly implies that $ba^{-1} \in H$. In the other cases, we have that $b \sim a$, so $ba^{-1} \in S$. Because $ba^{-1} \sim a^{-1}$, and there are no edges between $H \setminus \{e\}$ and $K \setminus \{e\}$, it follows that $ba^{-1} \in H$. Thus, $H$ — and similarly $K$ — is a subgroup of $G$.
The condition that ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ of order $2d$ in $S$ is not redundant. Indeed, the lattice graph $L_2(2)$, which is the line graph of $K_{2,2}$ (the incidence graph of a generalized $2$-gon of order $(1,1)$) is isomorphic to the Cayley graph $\operatorname{Cay}(\mathbb{Z}_4,\{\pm1\})$. Both elements $a$ in $S =\{\pm1\}$ have order $4$, but $a^2 \notin S$, and indeed $S \cup \{e\}$ does not contain a nontrivial subgroup of $\mathbb{Z}_4$.
The proof of Theorem \[Cayley line structure\] indicates that the condition ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ of order $2d$ in $S$ can be replaced by the condition that $a^2 \in S$ for every element $a$ of order $2d$ in $S$. We can in fact generalize this as follows.
Let $d \geq 2$, let $\Gamma$ be the line graph of the incidence graph of a generalized $d$-gon of order $(q,q)$, and suppose that $\Gamma$ is a Cayley graph $\operatorname{Cay}(G,S)$. Then there exist two subgroups $H$ and $K$ of $G$ such that $S=(H \cup K) \setminus \{e\}$, with $|H|=|K|=q+1$ and $H \cap K=\{e\}$ if and only if for every element $a$ of order $2d$, there exists an element $s \in S$ such that $s \neq a$, $a^{-1}$ and $sas^{-1} \in S$.
Let $a$ be of order $2d$ in $S$, and assume that there exists an element $s \in S$ such that $s \neq a$, $a^{-1}$ and $sas^{-1} \in S$. By Theorem \[Cayley line structure\], it suffices to prove that ${\langle a\rangle} \subseteq S \cup \{e\}$. Because $e$ is adjacent to $a$, $a^{-1}$, and $s$, and $G$ has no induced subgraph $K_{1,3}$, it follows that there is at least one edge within $\{a,a^{-1},s\}$. If $a$ and $a^{-1}$ are adjacent, then $a^2 \in S$. Because $\G$ does not contain an induced subgraph $K_{1,2,1}$, it then follows by induction and by considering the induced subgraph on $\{e,a,a^{i},a^{i+1}\}$ (for $i \geq 2$) that ${\langle a\rangle} \subseteq S \cup \{e\}$. So let us assume that $a$ and $a^{-1}$ are not adjacent. Without loss of generality, we may thus assume that $s$ is adjacent $a^{-1}$, and hence that $sa \in S$. Now $e$ is adjacent to $sa$, $a$ and $s$. Furthermore, $sa$ is adjacent to $a$ and $s$ since $sas^{-1} \in S$. It follows, again because $\G$ does not contain an induced subgraph $K_{1,2,1}$, that $a$ is adjacent to $s$. Using the same argument once more gives that $a$ and $a^{-1}$ are adjacent, which is a contradiction that finishes the proof.
In view of the above, if there exists an element $a \in S$ of order $2d$ such that ${\langle a\rangle} \nsubseteq S \cup \{e\}$, then $a$ and $a^{-1}$ are not adjacent. We may therefore assume that $a \in H$ and $a^{-1} \in K$, where $H$ and $K$ are the two maximal cliques (but not subgroups) containing $e$. But clearly the set $Ka$ is a maximal clique containing $a$ and $e$. Because every edge is in a unique maximal clique, it follows that $Ka = H$. Therefore, in this case, $S=(K \cup Ka)\setminus \{e\}$. In the case of the Cayley graph $\operatorname{Cay}(\mathbb{Z}_4,\{\pm1\})$, we indeed have $K=\{-1,0\}$ and $H=K+1$.
As a first application of the above, we obtain that the (distance-regular) line graph of the Tutte-Coxeter graph is not a Cayley graph.
\[tuttecoxeter\] The line graph of the Tutte-Coxeter graph is not a Cayley graph.
The Tutte-Coxeter graph is the incidence graph of a generalized quadrangle ($4$-gon) of order $(2,2)$. It has $30$ vertices and $45$ edges. If its line graph is a Cayley graph $Cay(G,S)$, then $|G|=45$ and $|S|=4$. Because $G$ has no element of order $8$, it follows from Theorem \[Cayley line structure\] that there exist two subgroups $H$ and $K$ of $G$ such that $S=(H \cup K) \setminus \{e\}$, where $|H|=|K|=3$ and $H \cap K=\{e\}$. Furthermore, the group $G$ is an abelian group isomorphic to $\mathbb{Z}_{3} \times \mathbb{Z}_{3} \times \mathbb{Z}_{5}$ or $\mathbb{Z}_{9} \times \mathbb{Z}_{5}$ since $G$ has only one subgroup of order $9$ and one subgroup of order $5$ by Sylow’s theorems. By the structure of the connection set $S$, it now follows that $G$ must be the abelian group isomorphic to $\mathbb{Z}_{3} \times \mathbb{Z}_{3} \times \mathbb{Z}_{5}$ but in this case the Cayley graph $Cay(G,S)$ is not connected, a contradiction. Therefore the line graph of the Tutte-Coxeter graph is not a Cayley graph.
We finish this section with a result that shows that the obtained structure of $S$ in the above fits naturally with line graphs of bipartite graphs.
\[lem:line\] Let $\G$ be a Cayley graph $\operatorname{Cay}(G,S)$, where $S=(H \cup K) \setminus \{e\}$ for nontrivial subgroups $H$ and $K$ of $G$ such that $H \cap K=\{e\}$. Then $\G$ is the line graph of a bipartite graph.
From the structure of $S$, it follows that each vertex is in two maximal cliques, and every edge is in a unique maximal clique. By a result of Krausz [@Kr] (see [@W Thm. 7.1.16]) it follows that $\G$ is a line graph of a graph, $\G'$, say. The graph $\G'$ has the maximal cliques of $\G$ as vertices, and two such cliques are adjacent in $\G'$ if and only if they intersect; the corresponding edge in $\G'$ is the vertex in $\G$ that is contained in both cliques.
Because $S=(H \cup K) \setminus \{e\}$ and $H \cap K=\{e\}$, we can distinguish between two kinds of maximal cliques. We call such a clique an $H$-clique if the edges in the clique are generated by an element in $H$, and the other cliques are similarly called $K$-cliques. Now it is clear that every edge in $\G'$ has one vertex in the set of $H$-cliques and the other vertex in the set of $K$-cliques. Thus $\G'$ is bipartite.
Strongly regular graphs {#sec:srg}
=======================
In this section, we will determine which strongly regular graphs with least eigenvalue $-2$ are Cayley graphs, using the case of diameter $d=2$ in the classification given in Theorem \[thm:drg-2\].
The sporadic graphs
-------------------
Besides the three infinite families of strongly regular graphs with least eigenvalue $-2$, we have to consider the Petersen graph, the Clebsch graph, the Shrikhande graph, the Schläfli graph, and the Chang graphs. The Petersen graph is the unique strongly regular graph with parameters $(10,3,0,1)$. It is the complement of the line graph of the complete graph $K_{5}$, and therefore it is not a Cayley graph, by Corollary \[triangular\] below (see also [@Go Lemma 3.1.3]).
(Folklore). \[Petersen\_no\] The Petersen graph is not a Cayley graph.
It is well-known that the complement of the Clebsch graph is the folded $5$-cube, which is strongly regular with parameters $(16,5,0,2)$ (see [@BH p. 119]). The $d$-dimensional cube $Q_{d}$ is the distance-regular graph whose vertex set can be labeled with the $2^{d}$ binary $d$-tuples such that two vertices are adjacent whenever their labels differ in exactly one position (clearly this is a Cayley graph). The folded $d$-cube is the distance-regular graph that can be obtained from the cube $Q_{d-1}$ by adding a perfect matching that connects vertices at distance $d-1$ (see [@van; @Bon]). It is evident that the folded $d$-cube is the Cayley graph $\operatorname{Cay}(G,S)$, where $G$ is the elementary abelian $2$-group of order $2^{d-1}$ and $$S=\{(1,0,0,\ldots,0), (0,1,0,\ldots,0), \ldots , (0,0,0,\ldots,0,1), (1,1,\ldots,1)\}.$$ Thus, the Clebsch graph is a Cayley graph.
The Shrikhande graph is a strongly regular graph with the same parameters as the lattice graph $L_{2}(4)$ and can be constructed as a Cayley graph $$\operatorname{Cay}(\mathbb{Z}_{4} \times \mathbb{Z}_{4}, \{\pm (0,1), \pm (1,0), \pm (1,-1)\}).$$ This construction ‘on the torus’ is accredited to Biggs [@Biggs] by Gol’fand, Ivanov, and Klin [@GIK p. 182].
The Schläfli graph is the unique strongly regular graphs with parameters $(27,16,10,8)$. It follows from the work by Liebeck, Praeger, and Saxl [@LPS] (see also [@LPM Lemma 2.6]) that it is a Cayley graph over the semidirect product $\mathbb{Z}_{9} \rtimes \mathbb{Z}_{3}$. Using [GAP]{} [@GAP], we checked that with $G=\mathbb{Z}_{9} \rtimes \mathbb{Z}_{3}={\langle a,b|a^9=b^3=1,b^{-1}ab=a^7\rangle}$ and $S=\{a,a^8,a^3,a^6,b,b^2,a^7b,a^5b^2,a^2b,a^4b^2\}$, the Cayley graph $\G=\operatorname{Cay}(G,S)$ indeed is the complement of the Schläfli graph. Note that $\G$ is also the point graph of the unique generalized quadrangle of order $(2,4)$, with lines thus being the triangles in $\G$. Therefore, these lines can be obtained as the right ‘cosets’ of the five triangles $\{e,a,a^2b\}$, $\{e,a^8,a^7b\}$,$\{e,a^3,a^6\}$,$\{e,b,b^2\}$,$\{e,a^5b^2,a^4b^2\}$ through $e$.
The Schläfli graph can also be constructed as a Cayley graph over $(\mathbb{Z}_{3} \times \mathbb{Z}_{3}) \rtimes \mathbb{Z}_{3}$, the other nonabelian group of order 27. Indeed, we again checked with [GAP]{} [@GAP] that $\G=\operatorname{Cay}(G',S')$ for $G'= {\langle a,b,c|a^3=b^3=c^3=e, abc=ba,ac=ca,bc=cb\rangle}$ and $S'=\{a,a^2,b,b^2,c,c^2,cba,a^2b^2c^2,aba,bab\}$. In this case all nonidentity elements of the group have order $3$, and hence the triangles through $e$ are subgroups $H_1,\dots,H_5$ of $G'$, with trivial intersection and $S'=(H_1\cup\cdots\cup H_5) \setminus \{e\}$ (cf. Theorem \[Cayley line structure\]). Again, the cosets of these subgroups give the lines of the generalized quadrangle of order $(2,4)$. From the above, we conclude the following.
\[sporadicsrg\] The Clebsch graph, the Shrikhande graph, and the Schläfli graph are Cayley graphs.
The Chang graphs are strongly regular graphs with the same parameters as the line graph of the complete graph $K_{8}$. These three graphs can be obtained by Seidel switching in $L(K_{8})$. According to [@Math2], the orders of the automorphism groups of these graphs are $384$, $360$, and $96$, respectively.
\[chang\] The three Chang graphs are not Cayley graphs.
Let $\Gamma$ be one of the Chang graphs, and suppose on the contrary that it is a Cayley graph, and hence that it is vertex-transitive. Let $x$ be a fixed vertex in $\Gamma$. Then the order of $\{x^{\sigma}| \sigma \in \operatorname{Aut}(\Gamma)\}$ is equal to $28$ since $\Gamma$ is vertex-transitive. It follows that the index of $\operatorname{Aut}(\Gamma)$ over the stabilizer of $x$ is $28$. Therefore $28$ must divide the order of $\operatorname{Aut}(\Gamma)$, which is a contradiction.
The infinite families
---------------------
A cocktail party graph is a complete multipartite graph with parts of size two, and clearly such a graph is a Cayley graph. By [@AJ1 Prop. 2.6], we obtain the following result.
\[cocktail\] A Cayley graph $\operatorname{Cay}(G,S)$ is a cocktail party graph if and only if $G$ has an element $a$ of order $2$ and $S=G \setminus {\langle a\rangle}$ .
Consider a set $X$ of size $n$ and let $V$ be the collection of all subsets of size $m$ in $X$, with $m \geq 2$ and $n \geq 2m+1$. The Kneser graph $K(n,m)$ is the graph with vertex set $V$ such that two vertices $A$ and $B$ in $V$ are adjacent whenever $|A \cap B|=0$. The Kneser graph $K(n,2)$ is the complement of the triangular graph $T(n)$. Godsil [@G] characterized the Cayley graphs among the Kneser graphs.
[@G] Except in the following cases, the Kneser graph $K(n,m)$ is not a Cayley graph.
- $m=2$, $n$ is a prime power and $n\equiv 3 \text{\em{ (mod 4)}}$,
- $m=3$, $n=8$ or $n=32$.
As a corollary, we obtain a result first obtained by Sabidussi [@SaVTG]. Note that the triangular graphs $T(2)$ and $T(3)$ are complete graphs, and that $T(4)$ is isomorphic to the cocktail party graph $CP(3)$.
[@SaVTG] \[triangular\] The triangular graph $T(n)$ is a Cayley graph if and only if $n=2,3,4$ or $n\equiv 3 \text{\em{ (mod 4)}}$ and $n$ is a prime power.
Godsil [@mathoverflow] gave the following construction of the triangular graph $T(n)$ as a Cayley graph $\operatorname{Cay}(G,S)$ for prime powers $n\equiv 3$ (mod 4). Let $\mathbb{F}$ be the field of order $n$. For $a,b \in \mathbb{F}$, let the map $T_{a,b}: \mathbb{F} \rightarrow \mathbb{F}$ be defined by $T_{a,b}(x)=ax+b$. Let $G$ be the group of maps $T_{a,b}$, with $a$ a non-zero square and $b$ arbitrary. It is not hard to see that $G$ acts regularity on the edges of the complete graph $K_{n}$ (with vertex set $\mathbb{F}$), using that $-1$ is a non-square (whence the assumption that $n\equiv 3$ (mod 4)). As connection set $S$ one can take the set of maps $T_{a,b} \in G$ such that either $T_{a,b}(0) \in \{0,1\}$ or $T_{a,b}(1) \in \{0,1\}$ (thus mapping the vertex $\{0,1\}$ of the triangular graph to an adjacent vertex).
As a final family of strongly regular graphs with least eigenvalue $-2$, we consider the lattice graphs. Let $n \geq 2$. The lattice graph $L_2(n)$ is the line graph of the complete bipartite graph $K_{n,n}$. It is isomorphic to the Cartesian product of two complete graphs $K_n$, and hence to the Cayley graph $\operatorname{Cay}(\mathbb{Z}_{n} \times \mathbb{Z}_{n},\{(0,1),\ldots,(0,n-1),(1,0),\ldots,(n-1,0)\})$. Because $K_{n,n}$ is the incidence graph of a generalized $2$-gon, we can apply the results of Section \[sec:genpol\]. We will use these to give a characterization of the lattice graphs as Cayley graphs.
\[lattice\] Let $n \geq 2$, let $G$ be a finite group, $S$ be an inverse-closed subset of $G$, and let $\G=\operatorname{Cay}(G,S)$. Then the following hold:
- If $G$ is a general product of two of its subgroups $H$ and $K$ of order $n$ and $S=(H \cup K)\setminus \{e\}$, then $\G$ is isomorphic to the lattice graph $L_{2}(n)$,
- If $\G$ is isomorphic to the lattice graph $L_{2}(n)$ and ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ of order $4$ in $S$, then $G$ is a general product of two of its subgroups $H$ and $K$ of order $n$ and $S=(H \cup K)\setminus \{e\}$.
Let $G$ be a general product of two of its subgroups $H$ and $K$ of order $n$ and let $S=(H \cup K)\setminus \{e\}$. Then $|G|=n^2$. By using the results in Section \[sec:genpol\], we know that every vertex in $\G$ is in two maximal cliques of size $n$. A simple counting argument shows that there are $2|G|/n=2n$ maximal cliques in $\G$. By Lemma \[lem:line\] and its proof, it follows that $\G$ is the line graph of a bipartite graph $\G'$ on the $2n$ maximal cliques of $\G$, and that each clique is on $n$ edges in $\G'$. This implies that $\G'$ is the complete bipartite graph $K_{n,n}$, and hence $\G$ is the lattice graph $L_2(n)$.
To prove the second item, suppose that $\G$ is isomorphic to the lattice graph $L_{2}(n)$ and ${\langle a\rangle} \subseteq S \cup \{e\}$ for every element $a$ of order $4$ in $S$. It follows by Theorem \[Cayley line structure\] that there are subgroups $H$ and $K$ of order $n$ in $G$ such that $H \cap K=\{e\}$ and $S=(H \cup K)\setminus \{e\}$. Now $K$ is a maximal clique in $\G$. Let $g$ be a vertex not in $K$. Then the structure of the lattice graph implies that $g$ is adjacent to precisely one vertex $k \in K$. Thus $gk^{-1} \in S$, and hence it follows that $gk^{-1} \in H$ (because if it were in $K$, then so would $g$), so $g=hk$ for some $h \in H$. Therefore $G$ is the general product of $H$ and $K$, which completes the proof.
We recall from Section \[sec:genpol\] that the lattice graph $L_2(2)$ is isomorphic to the Cayley graph $\operatorname{Cay}(\mathbb{Z}_4,\{\pm1\})$, which is an example such that $G=\mathbb{Z}_4$ cannot be written as a general product $HK$ with inverse-closed sets $H$ and $K$ of size $2$.
We now conclude this section by giving the classification of all strongly regular Cayley graphs with least eigenvalue at least $-2$ (which follows from the above). Recall that the only strongly regular graph with least eigenvalue larger than $-2$ is the $5$-cycle.
A graph $\Gamma$ is a strongly regular Cayley graph with least eigenvalue at least $-2$ if and only if $\Gamma$ is isomorphic to one of the following graphs.
- The cycle $C_{5}$, the Clebsch graph, the Shrikhande graph, or the Schläfli graph,
- The cocktail party graph $CP(n)$, with $n\geq 2$,
- The triangular graph $T(n)$, with $n=4$, or $n\equiv 3 \text{\em{ (mod 4)}}$ and $n$ a prime power, $n>4$,
- The lattice graph $L_{2}(n)$, with $n\geq 2$.
Distance regular graphs with diameter three {#sec:diameter3}
===========================================
In this section, we will determine which distance-regular graphs with least eigenvalue $-2$ and diameter three are Cayley graphs. By the classification given in Theorem \[thm:drg-2\], we again have to consider a few sporadic examples and an infinite family.
The line graphs of Moore graphs
-------------------------------
\[Petersen\] The line graph of the Petersen graph is not a Cayley graph.
Let $\G$ be the line graph of the Petersen graph, and suppose that $\G \cong \operatorname{Cay}(G,S)$, hence $|G|=15$ and $|S|=4$. Therefore there exists a subgroup of order $15$ of the automorphism group of $\G$ which acts transitively on the edges of the Petersen graph. By Sylow’s theorems, it is easy to see that the only group of order $15$ is the cyclic group $\mathbb{Z}_{15}$. This abelian group $G$ acts transitively on the edges of the Petersen graph, and because this graph is not bipartite, it follows that $G$ acts transitively on the vertices of the Petersen graph (cf. [@Go Lemma 3.2.1]). But every transitive abelian group acts regularly (cf. [@Biggs Prop. 16.5]), which gives a contradiction because the Petersen graph does not have 15 vertices.
\[Hoffman\] The line graph of the Hoffman-Singleton graph is not a Cayley graph.
Let $\G$ be the line graph of the Hoffman-Singleton graph, and suppose that $\G \cong \operatorname{Cay}(G,S)$, hence $|G|=175$ and $|S|=12$. It is easy to see that there exist only two groups of order $175$ by Sylow’s theorems, which are the abelian groups $\mathbb{Z}_{175}$ and $\mathbb{Z}_{35} \times \mathbb{Z}_{5}$. The result now follows similarly as in Proposition \[Petersen\]
The final case in this section is the line graph of a putative Moore graph on $3250$ vertices.
\[unknown\] The line graph of a strongly regular graph with parameters $(3250,57,0,1)$ is not a Cayley graph.
Let $\Gamma$ be a strongly regular graph with parameters $(3250,57,0,1)$ and suppose that the line graph of $\Gamma$ is a Cayley graph. Then $L(\Gamma)$ is vertex-transitive and therefore $\Gamma$ is edge-transitive by Lemma \[lem:edgevertextransitive\]. On the other hand, it is known that $\Gamma$ is not vertex-transitive, see [@BH Prop. 11.2], and therefore $\Gamma$ must be bipartite by [@Go Lemma 3.2.1], which is a contradiction.
The line graphs of the incidence graphs of projective planes {#sec:projplane}
------------------------------------------------------------
Recall that a projective plane of order $q$ is a point-line incidence structure such that each line has $q+1$ points, each point is on $q+1$ lines, and every pair of points in on a unique line. It is the same as a generalized $3$-gon of order $(q,q)$ and a $2$-$(q^2+q+1,q+1,1)$ design. Currently, projective planes of order $q$ are only known to exist for prime powers $q$, and for $q=1$. For $q>1$, the classical construction of a projective plane of order $q$ uses the finite field GF($q$) and gives the so-called Desarguesian plane of order $q$. We note that Loz, Mačaj, Miller, Šiagiová, Širáň, and Tomanová [@LMM] showed that the (distance-regular) incidence graph of a Desarguesian plane is a Cayley graph. Here we will consider the line graph, however. For $q=1$, the line graph of the incidence graph is a $6$-cycle, which is a Cayley graph. We therefore assume from now on that $q>1$. We note that the dual incidence structure of a projective plane is also a projective plane; if a projective plane is isomorphic to its dual, then we say it is self-dual.
Consider now a projective plane $\pi$ of order $q$, and let $\G_{\pi}$ be the incidence graph of $\pi$. Recall from Theorem \[line isomorphism\] that the automorphism group of $\G_{\pi}$ and its line graph $L(\G_{\pi})$ are isomorphic. A collineation (automorphism) of $\pi$ is a permutation of the points and lines that maps points to points, lines to lines, and that preserves incidence. If $\pi$ is not self-dual, then an automorphism of the incidence graph $\G_{\pi}$ must be a collineation. Additionally, if the projective plane is self-dual, then the automorphism group of $\G_{\pi}$ has index $2$ over the automorphism group of $\pi$; in this case the plane has so-called correlations (isomorphisms between the plane and its dual; see also [@M]) on top of collineations.
By construction, a vertex in $L(\G_{\pi})$ corresponds to an incident point-line pair — also called flag – of $\pi$. If $L(\G_{\pi})$ is a Cayley graph (or more generally, is vertex-transitive), then we have a group of collineations and correlations of $\pi$ that is transitive on flags. In particular, we have the following lemma.
Let $\pi$ be a projective plane of order $q$, with $q$ even. If $L(\G_{\pi})$ is a Cayley graph, then $\pi$ has a collineation group acting regularly on its flags.
If $L(\G_{\pi})$ is a Cayley graph, then there must be a group $G$ of automorphisms of $\G_{\pi}$ acting regularly on the edges of $\G_{\pi}$. The group $G$ therefore has order $(q+1)(q^2+q+1)$. Moreover, $G$ is (isomorphic to) a group of collineations and correlations of $\pi$ that acts regularly on its flags. If this group contains correlations, then it has an index 2 subgroup of collineations, but this is impossible because the order of $G$ is odd. Hence $\pi$ has a collineation group acting regularly on its flags.
For $q$ even, we can therefore use the following characterization by Kantor [@K].
[@K Thm. A] \[Kantor\] Let $q \geq 2$, let $\pi$ be a projective plane of order $q$, and let $F$ be a collineation group of $\pi$ that is transitive on flags. Then either
- $PSL(3,q)$ is contained in $F$ and $\pi$ is Desarguesian, or
- $F$ is a Frobenius group of odd order $(q+1)(q^2+q+1)$, and $q^2+q+1$ is prime.
Recall that $PSL(3,q)$ is the projective special linear group, which has order $$\frac{q^3(q^{3}-1)(q^{2}-1)}{\gcd(3,q-1)}.$$ If $L(\G_{\pi})$ is a Cayley graph $\operatorname{Cay}(G,S)$, then $|G|=(q^2+q+1)(q+1)$, and the action of $G$ on the flags of $\pi$ must be regular. Because the order of $PSL(3,q)$ is larger than $|G|$, it follows that $G$ is a Frobenius group of odd order $(q^2+q+1)(q+1)$, and that $q^2+q+1$ is prime. Recall that a Frobenius group is a group $F$ which has a non-trivial subgroup $H$ such that $H \cap x^{-1}Hx = \{e\}$ for all $x \in F \setminus H$. Furthermore, $N= F \setminus \bigcup_{x \in F}(x^{-1}Hx\setminus \{e\})$ is a normal subgroup of $F$ such that $F=HN$ and $H \cap K = \{e\}$, i.e. $F$ is the semidirect product $N \rtimes H$ (see [@R]).
\[incidence\] If the line graph of the incidence graph of a projective plane $\pi$ of order $q$ is a Cayley graph $\operatorname{Cay}(G,S)$, where $G$ corresponds to a group of collineations of $\pi$, then $G$ is $N \rtimes H$ in which $N$ is a normal subgroup of prime order $q^2+q+1$ and $H$ is a subgroup of odd order $q+1$.
It follows from the above that $G$ is a Frobenius group of odd order $(q^2+q+1)(q+1)$, and $q^2+q+1$ is a prime number. It follows that $G$ has a normal $(q^2+q+1)$-Sylow subgroup $N$ of order $q^2+q+1$ by Sylow’s theorems. On the other hand, there exists a subgroup $H$ of order $q+1$ in $G$ by Theorem \[Hall\], and the intersection of $N$ and $H$ is the identity element of $G$. Therefore $G$ is $N \rtimes H$.
It is widely believed that there is no non-Desarguesian plane admitting a collineation group acting transitively on flags. Thas and Zagier [@TZ] showed that if such a plane exists, then its order is at least $2 \times 10^{11}$.
On the other hand, Higman and McLaughlin [@HM] showed that the only Desarguesian planes admitting a collineation group acting regularly on flags are those of order $2$ and $8$. Indeed, the line graphs of the incidence graphs of these projective planes can be constructed as Cayley graphs as follows:
The Heawood graph is the incidence graph of the Fano plane; its line graph is the unique graph with spectrum $\{4^1,(1+\sqrt{2})^6,(1-\sqrt{2})^6,-2^8\}$ (see [@VH]). Let $G=\mathbb{Z}_{7} \rtimes \mathbb{Z}_{3}={\langle a,b|a^{7}=b^{3}=e,b^{-1}ab=a^2\rangle}$. Let $H={\langle b\rangle}$, $K={\langle a^{-1}ba\rangle}$ and $S=(H \cup K) \setminus \{e\}$ (cf. Theorem \[Cayley line structure\]). By using [GAP]{} [@GAP] and similar codes as in [@AJ2 p. 4], it is checked that the Cayley graph $\operatorname{Cay}(G,S)$ is indeed the line graph of the Heawood graph.
Similarly the line graph of the incidence graph of the (unique) projective plane of order $8$ is obtained by taking $G=\mathbb{Z}_{73} \rtimes \mathbb{Z}_{9}={\langle a,b|a^{73}=b^{9}=e,b^{-1}ab=a^2\rangle}$, $H={\langle b\rangle}$, $K={\langle a^{-1}ba\rangle}$, and $S=(H \cup K) \setminus \{e\}$.
We may thus conclude the following.
\[cayley3\] Let $\Gamma$ be a distance-regular Cayley graph with diameter three and least eigenvalue at least $-2$. Then $\Gamma$ is isomorphic to one of the following graphs.
- The cycle $C_6$ or $C_7$,
- The line graph of the incidence graph of the Desarguesian projective plane of order $2$ or $8$,
- The line graph of the incidence graph of a non-Desarguesian projective plane of order $q$, where $q^2+q+1$ is prime and $q$ is even and at least $2 \times 10^{11}$,
- The line graph of the incidence graph of a projective plane of odd order with a group of collineations and correlations acting regularly on its flags.
It would be interesting to find out whether any of the results on collineations of projective planes can be extended to groups of collineations and correlations, and thus rule out the final case of Theorem \[cayley3\]. We could not find any such results in the literature.
Besides the line graph of the Tutte-Coxeter graph (see Proposition \[tuttecoxeter\]) we leave the case of the line graphs of incidence graphs of generalized quadrangles and hexagons open (cf. Theorem \[thm:drg-2\]). For some results on flag-transitive generalized quadrangles, we refer to Bamberg, Giudici, Morris, Royle, and Spiga [@flagquad]; for flag-transitive generalized hexagons, we refer to Schneider and Van Maldeghem [@flaghexa].
The authors thank Brendan McKay for pointing to [@Sa Thm. 5.3] in order to prove Lemma \[lem:edgevertextransitive\]. Mojtaba Jazaeri thanks the Graduate Studies of University of Isfahan and Tilburg University since this paper was partly written during his visit at Tilburg University as part of his PhD program in Isfahan. The research of Alireza Abdollahi was in part supported by a grant from School of Mathematics, Institute for Research in Fundamental Sciences (IPM) (No. 94050219). Alireza Abdollahi is also supported financially by the Center of Excellence, University of Isfahan. The research of Mojtaba Jazaeri was in part supported by a grant from School of Mathematics, Institute for Research in Fundamental Sciences (IPM) (No. 94050039).
A. Abdollahi and M. Jazaeri, On groups admitting no integral Cayley graphs besides complete multipartite graphs, Appl. Anal. Discrete Math. 7 (2013) 119–128.
A. Abdollahi and M. Jazaeri, Groups all of whose undirected Cayley graphs are integral, European J. Combin. 38 (2014) 102–109.
A. Abdollahi and E. Vatandoost, Which Cayley graphs are integral?, Electron. J. Combin. 16 (2009), \#R 122.
B. Alspach, Cayley graphs, Chapter 6 in: Topics in algebraic graph theory, L.W. Beineke and R.J. Wilson, eds., Cambridge University Press, Cambridge, 2007.
J. Bamberg, M. Giudici, J. Morris, G. F. Royle, and P. Spiga, Generalised quadrangles with a group of automorphisms acting primitively on points and lines, J. Combin. Theory Ser. A 119 (2012) 1479–1499.
N. Biggs, Algebraic graph theory, Cambridge University Press, Cambridge, 1974.
J. van Bon, Finite primitive distance-transitive graphs, European J. Combin. 28 (2007) 517–532.
A. E. Brouwer, Chang graphs, <http://www.win.tue.nl/~aeb/graphs/Chang.html> (February 2015).
A. E. Brouwer, A. M. Cohen, and A. Neumaier, Distance-regular graphs, Springer-Verlag Berlin Heidelberg, New York, 1989.
A. E. Brouwer and W. H. Haemers, Spectra of graphs, Springer, New York, 2012.
P. M. Cohn, A remark on the general product of two infinite cyclic groups, Arch. Math. 7 (1956) 94–99.
E. R. van Dam and W. H. Haemers, Which graphs are determined by their spectrum?, Linear Algebra Appl. 373 (2003) 241–272.
E. R. van Dam, J. H. Koolen, and H. Tanaka, Distance-regular graphs, Electron. J. Combin. (2016), \#DS22.
C. Godsil, More odd graph theory, Discrete Mathematics 32 (1980) 205–207.
C. Godsil, Cayley graph which is isomorphic to the line graph of a complete graph, \[Online discussion group\], [http://mathoverflow.net/questions/164500/Cayley graph which is isomorphic to the line graph of a complete graph](http://mathoverflow.net/questions/164500/Cayley graph which is isomorphic to the line graph of a complete graph) (2014)
C. Godsil and G. Royle, Algebraic graph theory, Springer, New York, 2001.
Ja. Ju. Gol’fand, A. V. Ivanov, and M. Klin, Amorphic cellular rings, in [*Investigations in Algebraic Theory of Combinatorial Objects*]{} (I.A. Faradžev et al., eds.), Kluwer, Dordrecht, 1994, pp. 167–186.
D. G. Higman and J. E. McLaughlin, Geometric $ABA$-groups, Illinois J. Math. 5 (1961) 382–397.
W. M. Kantor, Primitive permutation groups of odd degree, and an application to finite projective planes, J. Algebra 106 (1987) 15–45.
E. Konstantinova, Some problems on Cayley graphs, University of Primorska Press, Koper, <http://www.hippocampus.si/ISBN/978-961-6832-51-9/index.html>, 2013.
J. Krausz, Démonstration nouvelle d’une théorème de Whitney sur les réseaux. Matematikai és Fizikai Lapok 50 (1943) 75–89.
C. H. Li, J. Pan, and L. Ma, Locally primitive graphs of prime-power order, J. Aust. Math. Soc. 86 (2009) 111–122.
M. W. Liebeck, C. E. Praeger, and J. Saxl, Regular subgroups of primitive permutation groups, Mem. Amer. Math. Soc. 203 (2009).
E. Loz, M. Mačaj, M. Miller, J. Šiagiová, J. Širáň, and J. Tomanová, Small vertex-transitive and Cayley graphs of girth six and given degree: An algebraic approach, J. Graph Theory, 68 (2011) 265–284.
S. L. Ma, A survey of partial difference sets, Des. Codes Cryptogr. 4 (1994), 221–261.
Š. Miklavič and P. Potočnik, Distance-regular circulants, European J. Combin. 24 (2003) 777–784.
Š. Miklavič and P. Potočnik, Distance-regular Cayley graphs on dihedral groups, J. Combin. Theory Ser. B 97 (2007) 14–33.
Š. Miklavič and P. Šparl, On distance-regular Cayley graphs on abelian groups, J. Combin. Theory Ser. B 108 (2014) 102–122.
G. E. Moorhouse, Incidence Geometry, <http://www.uwyo.edu/moorhouse/handouts/incidence_geometry.pdf> (2007).
D. J. S. Robinson, A course in the theory of groups, Second Edition, Springer, New York, 1995.
G. Sabidussi, Graph derivatives, Math. Zeitschr. 76 (1961) 385–401.
G. Sabidussi, Vertex-transitive graphs, Monatsh. Math. 68 (1964) 426–438.
C. Schneider and H. Van Maldeghem, Primitive flag-transitive generalized hexagons and octagons, J. Combin. Theory Ser. A 115 (2008) 1436–1455.
J. J. Seidel, Strongly regular graphs with $(-1, 1, 0))$ adjacency matrix having eigenvalue $3$, Linear Algebra Appl. 1 (1968) 281–298.
K. Thas and D. Zagier, Finite projective planes, Fermat curves, and Gaussian periods, J. Eur. Math. Soc. 10 (2008) 173–190.
The GAP Group, *GAP – Groups, Algorithms, and Programming, Version 4.6.4*; 2013, <http://www.gap-system.org>.
D. B. West, Introduction to graph theory, Second Edition, Pearson Education, India, 2002.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'By exploring a spinor space whose elements carry a spin 1/2 representation of the Lorentz group and satisfy the the Fierz-Pauli-Kofink identities we show that certain symmetries operations form a Lie group. Moreover, we discuss the reflex of the Dirac dynamics in the spinor space. In particular, we show that the usual dynamics for massless spinors in the spacetime is related to an incompressible fluid behavior in the spinor space.'
author:
- 'J. M. Hoff da Silva'
- 'R. T. Cavalcanti'
- 'D. Beghetto'
- 'R. da Rocha'
title: Spinor symmetries and underlying properties
---
Introduction
============
Spinors have constituted a comprehensive mathematical object of study, with a variety of applications in Physics. In particular, spinors are the main ingredient in the description of fermionic particles, that encode ordinary matter in the Universe. Fundamentally constructed upon the Lorentz group, a lot have been looked at the spacetime symmetries, underlying the Lounesto classification, rather than the symmetries of the spinor space itself. The Lounesto classification of spinor fields allocates classical spinors into six disjoint classes of regular and singular spinors. Regular spinors encompass Dirac spinors, and the singular ones, constituting flag-dipole, flagpole and dipole structures, comprehend the Majorana and Weyl spinors. Refs. [@Cavalcanti:2014wia; @Fabbri:2016msm] constructed a reciprocal classification for spinors, including gauge field theoretical aspects. Astonishingly, spinors satisfying the Dirac equation in several backgrounds have been found into five out of the six Lounesto’s classes. However, the subclasses of spinors, in each spinor class, whose equations of motion have been already stablished, are not fully determined. In other words, each spinor class has subclasses with precise dynamics, but the problem of categorizing the dynamics of all spinors in each class is intricate. Several approaches were scrutinized in Refs. [@Bonora:2017oyb; @Fabbri:2017lvu; @Fabbri:2016msm; @daRocha:2016bil; @daRocha:2005ti; @daRocha:2013qhu; @Ablamowicz:2014rpa; @5]. However, for example the fifth class of singular spinors has at least three subclasses: neutral spinors satisfying Majorana equation, eigenspinors of the charge conjugation operator with dual helicity, satisfying Elko coupled first order equations of motion, and also charged spinors satisfying the Dirac equation that induce an underlying fluid flow structure in some background spacetimes [@daRocha:2016bil]. The Lounesto spinor classification, encompassing classes of charged and neutral spinors under the U(1) gauge symmetry, was extended in Ref. [@Fabbri:2017lvu] to non-Abelian gauge symmetries. A second-quantized field theoretical approach [@Bonora:2017oyb] poses a similar classification in the framework of second quantization. Unexpected tensorial objetcs emerging from the spinor dynamics were also found in Ref. [@Fabbri:2018crr]. Therefore, it is clear that the question regarding the dynamics and kinematics of all spinors in Lounesto classification lacks still.
Within this motivation, a hybrid paradigm, uniting the symmetries on the spinor space and the spinors as representatives of the spinor classes in Lounesto classification, was previously proposed [@propo]. The ideia was to use this spinor representation space to envisage physical characteristics as an output of geometric, algebraic and topological properties of the constructed space, by exploring the point of view of spinors completely characterized by its bilinear covariants. In this work we continue exploring such a space, this time further exploring symmetries acting upon spinors themselves. Naturally, these symmetries should preserve each one of the spinor classes in Lounesto classification. Starting from its definition we investigate symmetries properties in the spinor space. It is shown that (invertible) symmetries transformations are rescaling for every bilinear covariants components constituting a subgroup of $GL(4,\mathbb{C})$. Also, the possibility of projective representations is explored, where we highlight an algebraic parallel of a superselection rule. These results are presented in Sec. III which is preceded by a review about the Lounesto spinor classification.
Section IV is reserved to the investigation of the usual Dirac dynamics in the spinor space $\Sigma$. Assuming the existence of a homomorphism between $\Sigma$ and an open set of ${\bf P}_{Spin_{1,3}^e}\times_\tau \mathbb{C}^4$ (see further specifications of this bundle in Sec. II) it is shown, under certain general conditions, that for massless spinors an analog of the Liouville theorem may be set. This section is finished contrasting such a result for the case of exotic spinors. In the final section we conclude.
Lounesto classification {#II}
=======================
Let $M$ denote the Minkowski spacetime. Spinors are objects in the spinor bundle, ${\bf P}_{Spin_{1,3}^e}\times_\tau \mathbb{C}^4$ associated to $M$, carrying the so-called $\tau={\left(1/2, 0\right)}\oplus{\left(0, 1/2\right)}$ representations of the Lorentz group [@1; @2; @3]. Due to several applications, arbitrary bases $\{\upgamma^\mu\}\subset \Omega(M)=\oplus_{i=0}^4\Omega^i(M)$ of the exterior bundle may be adopted. The bilinear covariants are the following exterior bundle sections [@TAKA],
$$\begin{aligned}
\textcolor{black}{\upsigma}&=&\textcolor{black}{\bar{\psi}\psi}\in\Upomega^0(M),
\label{upsig}\\
\mathbb{J}&=&\bar{\psi}\upgamma _{\mu }\psi\,\upgamma^\mu\in\Upomega^1(M),\label{jj}\\
\mathbb{S}&=&\frac{i}{2}\bar{\psi}[\upgamma _{\mu},\upgamma_{
\nu }]\psi\,\upgamma^\mu\wedge\upgamma^\nu\in\Upomega^2(M),\label{ss}\\
\mathbb{K}&=&\bar{\psi}\upgamma _{\mu }\upgamma^5\psi\,\upgamma^\mu\in\Upomega^3(M),\label{kk}\\
\upomega&=&{i\bar{\psi}\upgamma_{5}\psi\,,}\in\Upomega^4(M).\label{upom}\end{aligned}$$
The generators $\{\upgamma^\mu\}$ also satisfy the Clifford-Dirac algebra, $\upgamma_{\mu }\upgamma _{\nu
}+\upgamma _{\nu }\upgamma_{\mu }=2\eta_{\mu \nu }\mathbf{1}$. Besides, $\upgamma_5=i\upgamma_0\upgamma_1\upgamma_2\upgamma_3$. The spinor conjugation is denoted by $\bar\psi=\psi^\dagger\upgamma_0$.
Lounesto classification [@4; @5] allocates spinors into classes according to their bilinear covariants, categorizing and organizing the physically relevant spinorial space which can, therefore, be faced as composed by these six distinct pieces. In fact, the Lounesto classification split off the following:
$$\begin{aligned}
&&(1)\;\;\;\mathbb{K}\neq 0, \;\;\;\mathbb{S}\neq0,\;\;\;\upomega\neq0,\;\;\; \upsigma\neq0,\;\;\text{}
\label{tipo1}\\
&&(2)\;\;\;\mathbb{K}\neq 0, \;\;\;\mathbb{S}\neq0,\;\;\;\upomega\neq0,\;\;\; \upsigma=0,\;\;\label{tipo2}\\
&&(3)\;\;\;\mathbb{K}\neq 0, \;\;\;\mathbb{S}\neq0,\;\;\;\upomega=0,\;\;\; \upsigma\neq0,\label{tipo3}\\
&&(4)\;\;\;\mathbb{K}\neq 0, \;\;\;\mathbb{S}\neq0,\;\;\;\upomega=0=\upsigma, \;\;\text{}\quad\qquad\label{tipo4}\\
&&(5) \;\;\;\mathbb{K}=0, \;\;\;\mathbb{S}\neq0,\;\;\;\upomega=0=\upsigma,\text{}\quad\qquad\label{tipo5}\\
&&(6)\;\;\;\mathbb{K}\neq0, \;\;\;\mathbb{S}=0,\;\;\;\upomega=0=\upsigma.\;\;\;\text{}\quad\qquad\label{tipo6}\end{aligned}$$
Some physically important observables, such as the current and spin densities in the Dirac’s electron theory, are respectively identified to $\mathbb{J}$ and $\mathbb{K}$, while $\mathbb{S}$ is the spin density itself. The scalar $\sigma$ is the mass term in Lagrangians, whereas the pseudoscalar $\omega$ bilinears can reveal CP violations. The sum $\sigma^2+\omega^2$ corresponds to a probability density. Some regular spinors and most of the singular ones in the above classes are not supported by the same physical interpretation given to the electron. Besides, particular subclasses of the Lounesto’s classification satisfy the Fierz–Pauli–Kofink (FPK) relations [@TAKA]:
$$\begin{aligned}
\label{fifi1}
S^{\rho\sigma}\epsilon_{\;\;\;\alpha\beta}^{\rho\sigma}\upsigma-{S}_{\alpha\beta}\upomega&=&\epsilon_{\alpha\beta\rho\sigma}J^\alpha {K}^\beta,\\\eta_{\alpha\beta}(J^\alpha J^\beta+K^\alpha K^\beta)&=&0=\eta_{\alpha\beta}J^\alpha K^\beta,\label{fifi2}\\ \eta_{\alpha\beta}J^\alpha J^\beta&=&\upsigma^{2}+\upomega^{2}\,.\label{fifi3} \end{aligned}$$
Symmetries in the spinorial space {#II}
=================================
Let us denote a given bilinear by $\bar{\psi}\Gamma\psi$, for which $\Gamma$ is any element of the set $\{\mathbb{I},\gamma_5,\gamma_\mu,\gamma_5\gamma_\mu,\gamma_{\mu\nu}\}$, where $\mathbb{I}$ stands for the identity matrix, and the spinor dual is the usual (Dirac) one. Lounesto classification depends whether a given subset of bilinear is null or not, respecting the FPK identities. The relevant aspect to be emphasized here is that, concerning classical spinors in physics, the Lounesto classification is based in the physical observables and, hence, the belonging to a given type is by itself a physical information. This remark motivates the following definition.
Thus any phase multiplying a spinor is also a symmetry and one is facing a ray representation of spinors, very much like the use of Hilbert space vectors in quantum mechanics, representing physical states. Denoting by $\Sigma_i$ the part of the spinor space encompassing type-$i$ spinors $(i\in\{1, 2, \ldots,6\})$, with $\Sigma=\cup_{i=1}^6\Sigma_i$, a certain spinor $\psi$ is better characterized by an equivalence class representing its ray, denoted by $R$. Hence $\psi \in R \subset \Sigma_i \subset \Sigma$.
We shall now explore symmetry transformations. Let $S$ be a transformation leading rays into symmetry-preserving rays, that is $$\begin{aligned}
S_i: R\subset\Sigma_i&\to& R'\subset\Sigma_i, \forall i\in\{1, 2, \ldots,6\} \nonumber\\
\left[\psi\right] &\mapsto& [\psi ']=S_i[\psi], \label{pri}\end{aligned}$$ where $[\psi]$ denotes the equivalence class to which $\psi$ belongs. Therefore, if $\psi'$ is a spinor different of $\psi$, namely a modified spinor, a symmetry means $S_i(\Sigma_i) \subset \Sigma_i$. More explicitly, a symmetry should obey $$\begin{aligned}
[\bar{\psi}]\gamma^0S^\dagger\gamma^0\Gamma S[\psi]=\beta_{\Gamma} [\bar{\psi}]\Gamma[\psi],\label{se}\end{aligned}$$ where $\beta_{\Gamma} \in \mathbb{R}$ is the shift resulting from the transformation action. This shift will be non null, in general, and different from one[^1]. Depending on the bilinear dealt with, $\beta_\Gamma$ must be replaced by an array, or disposed into a matrix structure, albeit this is not important now. In the case of the $\Sigma_6$ space, the representation of a general symmetry $S_6$ can be straightforwardly displayed by a block diagonal matrix of one of the two following forms:[^2] $$\begin{aligned}
\label{stipo6}
\left( \begin{array}{cc}
A & \mathbb{O} \\
\mathbb{O} & B
\end{array} \right) \qquad \text{or} \qquad \left( \begin{array}{cc}
\mathbb{O} & A \\
B & \mathbb{O}
\end{array} \right),
\end{aligned}$$ where $A$ and $B$ are $2\times 2$ matrices with only $A$ or $B$ necessarily non-null. In fact, for a general singular spinor $\psi=(a,b,c,d)^\intercal$, the transformation $S_i$, $i=4,5,6$, must preserve the algebraic relation $a=\frac{bcd^*}{\Vert c\Vert^2}$. On the other hand, the opposite relation $a\neq \frac{bcd^*}{\Vert c\Vert^2}$, must be preserved for regular spinors [@Cavalcanti:2014wia].
Being $\chi$ a mapping between $\Sigma$ and the dual space $\bar{\Sigma}$, which is defined in a quite similar manner to $\Sigma$, we will restrict our analysis to the case in which $\chi$ is one-to-one. The reason is simple: being $\chi$ not one-to-one, then to an element of a given $R$, say $e^{i\alpha}\psi$ with $\alpha \in \mathbb{R}$, it would correspond $e^{-i\beta}\bar{\psi}$ in $\bar{\Sigma}$, with $\beta$ possibly different of $\alpha$. Then $\bar{\psi}'\Gamma\psi'=e^{i(\alpha-\beta)}\bar{\psi}\Gamma\psi$, breaking the symmetry.
In the following we are interested in symmetries such that, for every $S$ leading from a ray to another one, there should exist an inverse mapping, $S^{-1}$, pulling the transformation back. Besides, if $S_1$ transforms a ray $R$ into $R'$ and $S_2$ leads $R'$ into $R''$, then the acting of $S_1$ followed by $S_2$ should have the same effect of an unique transformation, say $S_3$, going directly from $R$ to $R''$. Taking all into account, provided associativity, symmetries transformations, if allowed, may form a group. We stress that the existence of symmetry transformations without inverse for all spinor types is, in principle, not forbidden. Not invertible symmetry transformations may also be physically relevant. However, we concentrate in the invertible case, since we are interested in a possible group structure.[^3]
[**Proof:**]{} Being the scalar and pseudo-scalar bilinear covariants non null, one may write $$\begin{aligned}
\gamma^0S^\dagger\gamma^0 S&=&\alpha\mathbb{I},\label{l21}\\
\gamma^0S^\dagger\gamma^0\gamma^5 S&=&\beta \gamma^5.\label{l22}\end{aligned}$$ These equations combine into $$\alpha S^{-1}\gamma^5 S=\beta \gamma^5. \label{l23}$$ Taking the determinant of both sides of (\[l23\]) yields $\alpha=\pm \beta$. A similar reasoning may be straigtforwardly extended to all components of the bilinear covariants, covering all the possible types. Some remarks, nevertheless, are in order before concluding. First, the proportionality between a transformed tensorial bilinear may be performed by a tensorial quantity, as to allow – respecting symmetry – the vanishing of some given component and the raising of another one. In any case, the final value of the tensorial quantity components are subject to the analysis above. Finally, the possible change of sign must, obviously, respect the constraints coming from FPK identities. $\Box$ With these results we are able to enunciate the next theorem.
[**Proof:**]{} Let $\{X,Y,S,\ldots\}$ be a set of symmetry transformations for type-$1$ spinors in which every element belongs to $\mathbb{M}(4,\mathbb{C})$. Suppose $X$ and $Y$ both satisfying (\[se\]) for, say, $\beta_{\Gamma X}$ and $\beta_{\Gamma Y}$, respectively. Hence $$\begin{aligned}
\gamma^0(XY)^\dagger\gamma^0\Gamma(XY),\label{pre}\end{aligned}$$ shall also satisfy (\[se\]) for $\beta_{\Gamma }=\beta_{\Gamma X}\beta_{\Gamma Y}$. Besides, it is fairly simple to see that the inverse transformation respects $\gamma^0(S^{-1})^\dagger\gamma^0\Gamma S^{-1}=\beta^{-1}_{\Gamma }\Gamma$. Once again we remark that when necessary the proportionality, and its inverse, must be engendered by a tensorial object. $\Box$
We finalize this section by stating some facts about the representation of the symmetry group found in the spinor space. As a matter of fact, while symmetry transformations act upon rays, the operators representing the above group transform spinors itself. In this regard, the representation will inherit most of the group properties. Denoting by $O(S)$ the operator representing the symmetry action in the spinor space, the resulting state $O(S_1)O(S_2)\psi$ differs from $O(S_1 S_2)\psi$, as usual, by a phase at most. This is, of course, the indication of a possible projective representation. At this point we have not enough information about the topology of the subgroup referred in the above theorem, although the elimination of its central charge seems to be reachable. Hence we willl postpone the elimination, so to speak, of the projective representation for the future. Instead we would like to point out an interesting peculiarity of the representation.
When dealing with representation up to a phase $O(S_1)O(S_2)\psi_k=e^{\phi_k}O(S_1 S_2)\psi_k$ the usual approach to quantum states yields a phase that does not depend on the state (here evinced by the label $k$) upon which the operators act, exception made to forbidden states, whose existence is precluded by means of a superselection rule [@WIWI]. The general picture may be straightforwardly recalled as follows: taking the sum of two spinors, say $\psi_m$ and $\psi_n$, and representing the transformation we have $O(S_1)O(S_2)(\psi_m+\psi_n)=e^{i\phi_{mn}}O(S_1 S_2)(\psi_m+\psi_n)$. After working out the right-hand side and acting with $O^{-1}(S_1S_2)$, we are left with $$e^{i\phi_m}\psi_m+e^{i\phi_n}\psi_n=e^{i\phi_{mn}}(\psi_m+\psi_n),\label{phase}$$ where $O$ is assumed unitary, for simplicity. Clearly, a solution for the above equation is $\phi_{mn}=\phi_m=\phi_n$ pointing to a phase independent to the state, but as symmetries transformations are allowed in this space, $\psi_m$ and $\psi_n$ may well be connected, and therefore it is hard to accept that the independence of the phases is reached by chance. In this regard, the very existence of the symmetry may be faced as the analogue of the superselection rule. As a final comment, we remark that the reasoning just outlined cannot be applied to type-$i$ spinors as a whole, as the type is not necessarily preserved by the sum of spinors [@4]. Regarding representations in the sector of $\Sigma_i$, for which the type is not preserved by the sum, the situation is quite unclear so far.
Dynamics avatar
===============
The group of transformations regarding type-$i$ spinors may be faced as an additional step towards the continuity of such a sector of $\Sigma$. However, any spinor in this space may be endowed of a dynamics inherited from the dynamics in spacetime. In this section we will investigate the behavior of spinors as elements in $\Sigma$. Let $\bar{\varphi}$ be a one-to-one, linear, and invertible mapping from $\Sigma$ to sections of ${\bf P}_{Spin_{1,3}^e}\times_\tau \mathbb{C}^4$, i. e. $$\begin{aligned}
\bar{\varphi} &: \Sigma \rightarrow {\bf P}_{Spin_{1,3}^e}\times_\tau \mathbb{C}^4 \nonumber\\ & \psi \mapsto \bar{\varphi}[\psi]=\Psi(\vec{x},t). \label{var} \end{aligned}$$ We shall restrict ourselves to the subset $U\subset {\bf P}_{Spin_{1,3}^e}\times_\tau \mathbb{C}^4$ such that the spinors $\Psi(\vec{x},t) \in U$ are subjected to the usual dynamics dictated by the Dirac operator $\mathcal{D}$, i. e. $\mathcal{D}\Psi(\vec{x},t)=0$. In addition, we are going to restrict $\bar{\varphi}$ to ${\varphi}=\bar{\varphi}\mid_{\bar{\varphi}^{-1}(U)}$, namely, the domain of ${\varphi}$ shall be the preimage of $U$, denoted by $\bar{\varphi}^{-1}(U)$. In analogy to the Dirac operator $\mathcal{D}$, let $\nabla$ be a “dynamical” operator (an automorphism) in $\Sigma$ such that $$\begin{aligned}
\nabla: \Sigma &\rightarrow& \Sigma \nonumber\\ \psi&\mapsto& \nabla\psi, \label{nab} \end{aligned}$$ whose relation with the dynamical operators be simply given by $\mathcal{D}=\varphi\circ\nabla\circ\varphi^{-1}$. As $\varphi^{-1}\circ \varphi=Id_{\Sigma}$ one has $\nabla=\varphi^{-1}\circ \mathcal{D}\circ \varphi$. Notice, in particular, that the algebraic zero resulting from the action of the Dirac operator is mapped into the null spinor in $\Sigma$. In fact, $$\begin{aligned}
\nabla\psi=\varphi^{-1}\circ \mathcal{D}\circ \varphi[\psi]=\varphi^{-1}\circ \mathcal{D}\Psi(\vec{x},t),\label{pre}\end{aligned}$$ and $\mathcal{D}\Psi(\vec{x},t)=0$ yields $\nabla\psi=0_{\Sigma}$. That is the alluded dynamical avatar which, despite have been straightforwardly obtained, leads to interesting consequences.
For free fermionic particles in the spacetime, the Dirac operator is usually expressed as $\mathcal{D}=i\gamma^{\mu}\partial_{\mu}-m\mathbb{I}$, where $\gamma^\mu$ are the Dirac matrices and $m$ the mass parameter. Therefore $$\begin{aligned}
\nabla\psi=\varphi^{-1}\circ (i\gamma^{\mu}\partial_{\mu}-m\mathbb{I})\Psi(\vec{x},t), \label{indo}\end{aligned}$$ which, by means of the map linearity, leads to $$\begin{aligned}
i\varphi^{-1}\circ (\gamma^\mu\partial_{\mu}\Psi(\vec{x},t))-m\psi=0_\Sigma. \label{nov}\end{aligned}$$ At first sight, one might speculate that the matrix representation of $\varphi$ commutes with gamma matrices. However, it would imply that $\varphi$ is proportional to the identity [@3]. This scenario is too restrictive. Here we will require something less limiting, by demanding the commutation of $\varphi$ only with $\gamma_0$. This requirement will be useful in what follows. Let us denote, then, the pullback of the spinor $\partial_t\Psi(\vec{x},t)$ by $\delta_t\psi:=\varphi^{-1}\circ (\partial_t\Psi(\vec{x},t))=\varphi^{-1}\circ \partial_t \circ \varphi[\psi]$, and then write $$\begin{aligned}
i\gamma^0\delta_t\psi+i\varphi^{-1}\circ (\vec{\gamma}\cdot \partial_{\vec{x}}\Psi(\vec{x},t))-m\psi=0_{\Sigma}.\label{primeiro}\end{aligned}$$
Eq. (\[primeiro\]) performs a shadow, so to speak, of the spacetime dynamics respected by the physical spinor. It may be applied to every sector of $\varphi^{-1}(U)\subset \Sigma$ and in this space as a whole. However, it is not completely clear so far which connections may be reached inside the spinor space, see for instance [@jmp]. Therefore we will assume, in a first moment, a conservative approach adopting the physically sound particularization that the spinor type is not changed by the dynamics and study its consequences for each $\Sigma_i$ separately.
The spinors belonging to $\Sigma_i$ are called physical, in the sense that they satisfy the FPK identities. The attribute “physical" in dealing with spinors, however, must be used with a great care. In fact, a spinor alone describing a fermion cannot be detected. Its dual – and the correspondent theory – must be taken into account. Despite of these important matters, if the elements of $\Sigma_i$ are representatives of physical states, then they have to be conserved. Consider a macroscopically dense set of spinors in $\mathcal{F}\subset\Sigma_i$ and suppose that the surface $\partial\mathcal{F}$ is orientable. A conservation law will encounter an analogue within $\mathcal{F}$. Hence, being the density $\rho$ of spinor states in this region characterized by $\rho(\psi,t)$ one may be able to write $$\frac{\partial \rho}{\partial t}=-\frac{\delta (\rho \delta_t \psi)}{\delta \psi}. \label{cons}$$ Some considerations concerning Eq. (\[cons\]) are in order. As to represent a conservation law, $\delta_t \psi$ denotes a generalized velocity in $\mathcal{F}$ leading, then, the term $\rho\delta_t \psi$ to express a current of states. Therefore, a decreasing \[increasing\] in the density of states is taken due to an output \[input\] current. Besides, the functional derivative present in Eq. (\[cons\]) may be taken in exact same footing as its counterpart in classical and quantum field theory. We are now in position to assert the following result, in close analogy to the Statistical Mechanics Liouville theorem for physical states in the phase space [@pat].
[**Proof:**]{} The time variation of the representative density reads $$\frac{d\rho(\psi,t)}{dt}=\frac{\delta \rho(\psi,t)}{\delta \psi}\delta_t\psi+\frac{\partial \rho(\psi,t)}{\partial t}$$ and taking (\[cons\]) into account yields $$\frac{d\rho(\psi,t)}{dt}=-\rho \frac{\delta (\delta_t \psi)}{\delta \psi}.\label{dois}$$ The equation governing the behavior in the spinor space, (\[primeiro\]), for massless spinors may be recast into the form $$\mathbb{I}\delta_t \psi=-\gamma^0\varphi^{-1}\circ (\vec{\gamma}\cdot\partial_{\vec{x}}\Psi(\vec{x},t)).\label{sei}$$ From Eqs. (\[dois\]) and (\[sei\]) it is fairly simple to see that $$\frac{d(\mathbb{I} \rho(\psi,t))}{dt}=0,\label{ies}$$ culminating in four identical equations satisfied by a constant density. $\Box$
The massive case is just inconsistent. We are currently investigating this case, for which, we speculate, none conservative equation analogue can be stated, but have not a satisfactory interpretation for that so far. Before concluding this section, we would like to contrast our results with the case concerning exotic spinors.
It is well known that when the base manifold, $M$, is not simply connected there is not only one spinorial structure [@exo1]. This fact is traduced by the non triviality of the (first) cohomology group $H^1(M, \mathbb{Z}_2)$. The non trivial topology is then reflected in the dynamics [@exo2; @exo21], by means of an additional term in the Dirac operator now reading $\tilde{\mathcal{D}}=i\gamma^{\mu}\partial_{\mu}+i\gamma^\mu\partial_\mu\theta(\vec{x},t)-m\mathbb{I}=\mathcal{D}+i\gamma^\mu\partial_\mu\theta(\vec{x},t)$, where $\theta(\vec{x},t)$ is a real scalar function (for mathematical details see, for instance, [@exo3]). Concerning the topics approached in Sect. \[II\], there are very few substantial differences in dealing with exotic spinors instead of usual spinors. As a matter of fact, it is still possible to define a spinor space for exotic spinors and, as before, we also will have a categorization of the spinorial space into types. The unique novelty is that there are three more possible types of spinors, but the results of the previous section remain valid. A noteworthy difference occurs in the appreciation of the proposition above to exotic spinors. Denoting exotic spinors in the exotic spinorial space ($\tilde{\Sigma}$) by $\tilde{\psi}$ (and its spacetime counterpart by $\tilde{\Psi}(\vec{x},t)$), it is fairly direct to see that the analogue of Eq. (\[primeiro\]) for the case at hands reads $$\begin{aligned}
i\gamma^0\delta_t\tilde{\psi}-m\tilde{\psi}+i\gamma^0\dot{\theta}(\vec{x},t)\tilde{\psi}+i\varphi^{-1}\circ \Big(\vec{\gamma}\cdot\{\partial_{\vec{x}}\tilde{\Psi}(\vec{x},t)+\partial_{\vec{x}}\theta(\vec{x},t)\tilde{\Psi}(\vec{x},t)\}\Big)=0_{\tilde{\Sigma}},\label{penult}\end{aligned}$$ where $\dot{\theta}=\partial_t\theta$. In this vein, massless exotic spinors shall obey $$\begin{aligned}
\mathbb{I}\delta_t\tilde{\psi}=-\mathbb{I}\dot{\theta}(\vec{x},t)\tilde{\psi}-\gamma^0\varphi^{-1}\circ \Big(\vec{\gamma}\cdot\{\partial_{\vec{x}}\tilde{\Psi}(\vec{x},t)+\partial_{\vec{x}}\theta(\vec{x},t)\tilde{\Psi}(\vec{x},t)\}\Big).\label{ult}\end{aligned}$$ Though the form of Eq. (\[ult\]) is not particularly clear from the physical point of view, the investigation of the exotic spinors behavior in $\tilde{\Sigma}$ leads to the fact that the spinorial density, provided conservation, is given in terms of the exotic additional term $\rho(\psi,t)=\rho_0\exp(\theta(\vec{x},t))$. Of course, $\rho_0$ is constant in such a way that if $\theta=0$ (the usual case of trivial topology) the proposition result is recovered, as expected.
Concluding remarks
==================
In Sec. III we show the possibility of symmetries transformations in the spinor space as elements of a subgroup of $GL(4,\mathbb{C})$. These symmetries respect the Lounesto classification and so do not accross the spinor type. While relevant results on their own, we would like here to give a comprehensive account on results. In Ref. [@nove] it was proposed an interpolation between sectors of a given representation, encompassing spinors satisfying the Heisenberg equation of motion, which could lead to the neutrino oscillation even in the massless case. All these spinors was shown to belong to Lounesto type-1 case [@ult]. The results here explored may serve as a first step towards the mathematical investigation of such an interpolation, in the sense that it was conjectured to be performed by an unitary operator [@nove] whose action preserves the spinor type [@ult].
In Sec. IV, we explore the interelationship between the dynamics occurring in the spacetime and its reflex in the spinor space. The interplay between spinors, bilinear covariants and hydrodynamics was implemented in Refs. [@daRocha:2016bil; @Bonora:2015ppa], in the context of the Lounesto spinor classification. In Ref. [@Bonora:2015ppa] suitable black hole backgrounds were considered, having a current density that interpolates between a timelike Killing vector field at the spatial infinity and the null Killing vector field on the black hole event horizon. This current density was identified to a spinor fluid flow. In Ref. [@daRocha:2016bil], flag-dipole spinors, satisfying the Dirac equation in another black hole background was shown to induce an underlying fluid flow structure in the background spacetime. These two results are quite particular, relating fluid mechanics to the Lounesto classification. On the other hand, the results in the Proposition here presented are universal, relating the dynamics of certain spinors with the equations of motion of incompressible fluids. The investigation of this result in the context of exotic spinors was presented. It was shown that unusual topology in the spacetime leads to a modification in the spinor space dynamics. While some modification is generically expected, since the connection is changed, we emphasize that the dynamical interplay was strong enough to reveal that unusual topology forbids the perfect fluid behavior. We are currently investigating additional developments of this interplay, as well its limitations.
Acknowledgments {#acknowledgments .unnumbered}
===============
JMHS thanks to CNPq (Grant No. 303561/2018-1) for partial financial support. RdR is grateful to FAPESP (Grant No. 2017/18897-8), to CNPq (Grants No. 406134/2018-9 and No. 303293/2015-2) and to HECAP - ICTP, Trieste, for partial financial support, and this last one also for the hospitality.
[99]{}
P. Lounesto, [*Clifford Algebras and Spinors*]{}, Cambridge Univ. Press, Cambridge, 2001. R. T. Cavalcanti, Int. J. Mod. Phys. D [**23**]{} (2014) 1444002 \[arXiv:1408.0720 \[hep-th\]\].
L. Fabbri, Int. J. Geom. Meth. Mod. Phys. [**13**]{} (2016) 1650078 \[arXiv:1603.02554 \[gr-qc\]\].
L. Bonora, J. M. Hoff da Silva and R. da Rocha, Eur. Phys. J. C [**78**]{} (2018) 157 \[arXiv:1711.00544 \[hep-th\]\]. L. Fabbri and R. da Rocha, Phys. Lett. B [**780**]{} (2018) 427 \[arXiv:1711.07873 \[hep-th\]\].
R. da Rocha and R. T. Cavalcanti, Phys. Atom. Nucl. [**80**]{} (2017) 329 \[arXiv:1602.02441 \[hep-th\]\]. R. da Rocha and W. A. Rodrigues, Jr., Mod. Phys. Lett. A [**21**]{} (2006) 65 \[math-ph/0506075\].
R. da Rocha, L. Fabbri, J. M. Hoff da Silva, R. T. Cavalcanti and J. A. Silva-Neto, J. Math. Phys. [**54**]{} (2013) 102505 \[arXiv:1302.2262 \[gr-qc\]\].
R. Abłamowicz, I. Gonçalves and R. da Rocha, J. Math. Phys. [**55**]{} (2014) 103501 \[arXiv:1409.4550 \[math-ph\]\].
J. M. Hoff da Silva and R. T. Cavalcanti, Mod. Phys. Lett. A [**32**]{} (2017) 1730032 \[arXiv:1708.06222 \[physics.gen-ph\]\].
L. Fabbri, Eur. Phys. J. C [**78**]{}, no. 9, 783 (2018) \[arXiv:1810.01290 \[physics.gen-ph\]\].
J. M. Hoff da Silva, C.H. Coronado Villalobos, R. da Rocha, and R.J. Bueno Rogerio, Eur. Phys. J. C [**77**]{} (2017) 487 \[arXiv:1702.05034 \[math-ph\]\].
R. A. Mosna and W. A. Rodrigues Jr., J. Math. Phys. [**45**]{} (2004) 2945 \[arXiv:math-ph/0212033\].
W. A. Rodrigues Jr., J. Math. Phys. [**45**]{} (2004) 2908 \[arXiv:math-ph/0212030\].
J. Vaz Jr. and R. da Rocha, [*An Introduction to Clifford Algebras and Spinors*]{}, Oxford Univ. Press, Oxford, 2016.
J. P. Crawford, J. Math. Phys. [**26**]{} (1985) 1439. Y. Takahashi, Phys. Rev. D [**26**]{} (1982) 2169.
G. C. Wick, A. S. Wightman, and E. P. Wigner, Phys. Rev. [**88**]{} (1952) 101.
R. da Rocha, and J. M. Hoff da Silva, J. Math. Phys. [**48**]{} (2007) 123517 \[arXiv:0711.1103 \[math-ph\]\].
R. K. Pathria, [*Statistical Mechanics*]{}, Butterworth-Heinemann, Oxford, 1996.
R. P. Geroch, J. Math. Phys. [**9**]{} (1968) 1739.
S. J. Avis and C. J. Isham, Nucl. Phys. B [**156**]{} (1979) 441 S. J. Avis and C. J. Isham, Commun. Math. Phys. [**72**]{} (1980) 103.
R. da Rocha, A. E. Bernardini and J. M. Hoff da Silva, JHEP [**04**]{} (2011) 110 \[arXiv:1103.4759 \[hep-th\]\].
L. Bonora and R. da Rocha, JHEP [**1601**]{} (2016) 133 \[arXiv:1508.01357 \[hep-th\]\].
M. Novello, Europhys. Lett. [**80**]{} (2007) 41001 \[arXiv:0705.2692 \[astro-ph\]\].
D. Beghetto and J. M. Hoff da Silva, Europhys. Lett. [**119**]{} (2017) 40006 \[arXiv:1710.07086 \[math-ph\]\].
[^1]: If $\beta_{\Gamma}=1$, for any $\Gamma$, the inversion theorem [@TAKA; @TAKA1] yields necessarily $S=\mathbb{I}$.
[^2]: As the apparatus is representation-independent, we are adopting the Weyl representation for the $\gamma$ matrices.
[^3]: The particular case of type-6 spinors does not require the whole matrix $S_6$ being invertible for having the group structure. It is sufficient being the non-null block invertible.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose new concepts of statistical depth, multivariate quantiles, vector quantiles and ranks, ranks, and signs, based on canonical transportation maps between a distribution of interest on ${{\rm I}\kern-0.18em{\rm R}}^d$ and a reference distribution on the $d$-dimensional unit ball. The new depth concept, called [*Monge-Kantorovich depth*]{}, specializes to halfspace depth for $d=1$ and in the case of spherical distributions, but, for more general distributions, differs from the latter in the ability for its contours to account for non convex features of the distribution of interest. We propose empirical counterparts to the population versions of those Monge-Kantorovich depth contours, quantiles, ranks, signs, and vector quantiles and ranks, and show their consistency by establishing a uniform convergence property for empirical (forward and reverse) transport maps, which is the main theoretical result of this paper.'
address:
- |
Department of Economics and Center for Statistics\
Massachusetts Institute of Technology\
Cambridge, MA 02139, USA\
- |
Economics Department and Courant Institute,\
New York University\
New York, NY 10012, USA\
and\
Sciences Po, Economics Department\
75007 Paris, France\
- |
ECARES\
Universit' e libre de Bruxelles\
CP 114 Brussels, Belgium\
and\
ORFE\
Princeton University\
Princeton, NJ 08540, USA\
- |
Department of Economics\
The Pennsylvania State University\
University Park, PA 16802, USA\
author:
- '[^1],'
- '[^2],'
- '[^3],'
- '[^4]'
title: |
Monge-Kantorovich depth,\
quantiles, ranks, and signs
---
\
Introduction {#intro}
============
The concept of statistical depth was introduced in order to overcome the lack of a canonical ordering in ${{\rm I}\kern-0.18em{\rm R}}^d$ for $d>1$, hence the absence of the related notions of quantile and distribution functions, ranks, and signs. The earliest and most popular depth concept is halfspace depth, the definition of which goes back to Tukey [@Tukey:75]. Since then, many other concepts have been considered: [simplicial depth]{} [@Liu:90], majority depth ([@Singh:91] and [@LS:93]), [projection depth]{} ([@Liu:92], building on [@Stahel:81] and [@Donoho:82], [@Zuo:2003]), [Mahalanobis depth]{} ([@Mahalanobis:36], [@Liu:92], [@LS:93]), [Oja depth]{} [@Oja:83], [zonoid depth]{} ([@KM:97] and [@Koshevoy:2002]), [spatial depth]{} ([@DK:92], [@MO:95], [@Chaudhuri:96], [@VZ:2000]), $L^p$ depth [@ZS:2000], among many others. An axiomatic approach, aiming at unifying all those concepts, was initiated by Liu [@Liu:90] and Zuo and Serfling [@ZS:2000], who list four properties that are generally considered desirable for any statistical depth function, namely affine invariance, maximality at the center, linear monotonicity relative to the deepest points, and vanishing at infinity (see Section \[sec:LZS\] for details). Halfspace depth is the prototype of a depth concept satisfying the Liu-Zuo-Serfling axioms for the family $\mathcal P$ of all absolutely continuous distributions on ${{\rm I}\kern-0.18em{\rm R}}^d$.
An important feature of halfspace depth is the convexity of its contours, which thus satisfy the star-convexity requirement embodied in the linear monotonicity axiom. That feature is shared by most existing depth concepts and might be considered undesirable for distributions with non convex supports or level contours, and multimodal ones. Proposals have been made, under the name of [ local depths]{}, to deal with this, while retaining the spirit of the Liu-Zuo-Serfling axioms: see [@CDPB:2009], [@HKV:2010], [@AR:2011], and [@PB:2013] who provide an in-depth discussion of those various attempts. In this paper, we take a totally different and more agnostic approach, on the model of the discussion by Serfling in [@Serfling:2002]: if the ultimate purpose of statistical depth is to provide, for each distribution $ P$, a $ P$-related ordering of ${{\rm I}\kern-0.18em{\rm R}}^d$ producing adequate concepts of quantile and distribution functions, ranks and signs, the relevance of a given depth function should be evaluated in terms of the relevance of the resulting ordering, and the quantiles, ranks and signs it produces.
Now, the concepts of quantiles, ranks and signs are well understood in two particular cases, essentially, that should serve as benchmarks. The first case is that of the family $\mathcal{P}^1$ of all distributions with nonvanishing Lebesgue densities over convex support sets. Here, the concepts of quantile and distribution functions, ranks, and signs are related to the “classical" univariate ones. The second case is that of the family $\mathcal{P}^d_{\mbox{\scriptsize ell}}$ of all full-rank elliptical distributions over ${{\rm I}\kern-0.18em{\rm R}}^d$ ($d>1$) with radial densities over elliptical support sets. Recall that the family $\mathcal{P}^d_{\mbox{\scriptsize ell}; g}=\{P_{\mu ,\Sigma , g}\}$ of elliptical distributions with given [ radial density $g$ and distribution function $G$]{} is a parametric family indexed by a location parameter $\mu$ and a [ scatter]{} parameter $\Sigma$ (a symmetric positive definite real matrix) such that a random vector $X$ has distribution $P_{\mu ,\Sigma , g}$ iff the residual $Y:= \Sigma^{-1/2}(X-\mu)$, which results from transforming $X$ into isotropic position, has spherical distribution $P_{0 ,I , g}$. Further, this is equivalent to ${{\rm R}}_P (Y) = (Y/\|Y\|)G(\|Y\|)$ having the spherical uniform distribution $U_d$ on the unit ball $\mathbb{S}^d$ in ${{\rm I}\kern-0.18em{\rm R}}^d$. By [*spherical uniform*]{}, we mean the distribution of a random vector $r{\varphi}$, where $r$ is uniform on $[0,1]$, ${\varphi}$ is uniform on the unit sphere $\mathcal S^{d-1}$, and $r$ and ${\varphi}$ are mutually independent. There, spherical contours with $ P_{\mu ,I , g}$-probability contents $\tau$ coincide with the halfspace depth contours, and provide a natural definition of [ $\tau$-quantile contours]{} for $Y$, while ${{\rm R}}_P (Y)$, ${{\rm R}}_P (Y)/\|{{\rm R}}_P (Y)\|$ and $\|{{\rm R}}_P(Y)\|$ play the roles of vector ranks, signs, and ranks, respectively ([@HP:2002; @HP:2004; @HP:2005; @HP:2006; @HP:2008]): we call them spherical [*vector ranks*]{}, [ *signs*]{}, and [*ranks*]{}. On the other hand, we call the inverse map $u \longmapsto {{\rm Q}}_P(u)$ of the vector rank map $y \longmapsto {{\rm R}}_P (y) = (y/\|y\|)G(\|y\|)$ the [*vector quantile*]{} map. In both cases, the relevance of ranks and signs, whether traditional or spherical, is related to their role as maximal invariants under groups of transformations minimally generating $\mathcal{P}^1$ or the family $\mathcal{P}^d_{\mbox{\scriptsize sph}}=\{P_{0,I,f}\}$ of [ spherical]{} distributions, of which distribution-freeness of ${{\rm R}}_P$ is just a by-product, as explained in [@HW:2003]. We argue that an adequate depth function, when restricted to those two particular cases, should lead to the same well-established concepts—classical quantiles, ranks and signs for $\mathcal{P}^1$, and spherical ones for $\mathcal P^d_{\mbox{\scriptsize sph}}$—hence should coincide with halfspace depth.
Now, a closer look at those two particular cases reveals that halfspace depth contours, in $\mathcal{P}^1$ and $\mathcal P^d_{\mbox{\scriptsize sph}}$, are the images, by the vector quantile map ${{\rm Q}}_P$, of the hyperspheres $\mathcal{S}(\tau)$ with radii $\tau\in[0,1)$ centered at the origin. The map ${{\rm Q}}_P$ is the gradient of a convex function and it transports the spherical uniform distribution $U_d$ on the unit ball $\mathbb S^d$ of ${{\rm I}\kern-0.18em{\rm R}}^d$ into the univariate distribution $P~\!\in~\!\mathcal{P}^1$ or into the spherical distribution $P=P_{0 ,I , f}$ of interest.
For the case of general distributions $P$, we proceed similarly, and define the map ${{\rm Q}}_P$ as a gradient of a convex function that transform the spherical uniform distribution $U_d$ into the target distribution, namely if $U \sim U_d$ then $Y = {{\rm Q}}_P(U) \sim P$. It follows by McCann’s [@McCann1995] extension of Brenier’s celebrated Polar Factorization Theorem [@brenier] that, for any distribution $P$ on ${{\rm I}\kern-0.18em{\rm R}}^d$, such a gradient ${{\rm Q}}_P$ exists, and is essentially unique. Moreover, when $P$ has finite moments of order two, that mapping ${{\rm Q}}_P$ is the Monge-Kantorovich optimal transport map that transfers the spherical uniform distribution $U_d$ to $P$, where optimality is the sense of minimizing the expected quadratic cost $\min_{{\rm Q}}{{\rm I}\kern-0.18em{\rm E}}_{U} ( {{\rm Q}}(U) -U)^2$ subject to $U \sim U_d$ and ${{\rm Q}}(U) \sim P$.
This suggests a new concept of statistical depth, which we call the [*Monge-Kantorovich (or MK) depth*]{} ${{\rm D}}^{\scriptsize{\text{MK}}}$, the contours of which are obtained as the images by ${{\rm Q}}_P$ of the hyperspheres with radius $\tau\in[0,1]$. When restricted to $\mathcal{P}^1$ or $\mathcal P^d_{\mbox{\scriptsize sph}}$, Monge-Kantorovich and halfspace depths coincide. Under suitable regularity conditions due to Caffarelli (see [@villani1], Section 4.2.2), ${{\rm Q}}_P$ is a homeomorphism, and its inverse ${{\rm R}}_P:={{\rm Q}}_P^{-1}$ is also the gradient of a convex function; the Monge-Kantorovich depth contours are continuous and the corresponding depth regions are nested, so that Monge-Kantorovich depth indeed provides a center-outward ordering of ${{\rm I}\kern-0.18em{\rm R}}^d$, namely, $$\label{MKorder}
y_2 \geq_{{{\rm D}}_P^{\scriptsize{\text{MK}}}} y_1 \mbox{ if and only if } \Vert{{\rm R}}_P(y_2)\Vert \leq \Vert{{\rm R}}_P(y_1)\Vert .$$ Thus, our approach based on the theory of measure transportation allows us to define
(a) an [*MK vector quantile*]{} map ${{\rm Q}}_P$, and the associated [*MK quantile*]{} correspondence, which maps $\tau\in[0,1]$ to ${{\rm Q}}_P(\mathcal S(\tau))$,
(b) an [*MK vector rank*]{} (or [*MK signed rank*]{}) function ${{\rm R}}_P$, which can be decomposed into an [*MK rank*]{} function $r_P$ from ${{\rm I}\kern-0.18em{\rm R}}^d$ to $[0,1]$, with $r_P(x):=\|{{\rm R}}_P(x)\|$, and an [*MK sign*]{} function $u_P$, mapping $x\in{{\rm I}\kern-0.18em{\rm R}}^d$ to $u_P(x):={{\rm R}}_P(x)/\|{{\rm R}}_P(x)\|\in\mathcal S^{d-1}$.
To the best of our knowledge, this is the first proposal of a depth concept based on the Monge-Kantorovich theory of measure transportation —hence the first attempt to provide a measure-driven ordering of ${{\rm I}\kern-0.18em{\rm R}}^d$ based on measure transportation theory. Previous proposals have been made, however, of measure transportation-based vector quantile functions in Ekeland, Galichon and Henry [@EGH] and Galichon and Henry [@GH:2012] (with moment conditions) and Carlier, Chernozhukov and Galichon [@CCG] (dropping moment conditions) who also extended the notion to vector quantile regression, creating a vector analogue of Koenker and Basset’s [@KB:78] scalar quantile regression. More recently, Decurninge [@Decurninge] proposed a new concept of multivariate $L^p$ moments based upon a similar notion. In these contributions, however, the focus is not statistical depth and the associated ranks and quantiles, and the leading case for the reference distribution is uniform on the unit hypercube in ${{\rm I}\kern-0.18em{\rm R}}^d$, as opposed to the spherical uniform distribution $U_d$ we adopt here as leading case, while pointing out that other reference distributions may be entertained, such as the standard Gaussian distribution on ${{\rm I}\kern-0.18em{\rm R}}^d$ or the uniform on the hypercube $[0,1]^d$ as mentioned above.
We then proceed to define the empirical notions corresponding to the concepts given above. We define the empirical MK vector quantiles and ranks as the essentially unique gradients $\hat{{\rm Q}}_n$ and $\hat {{\rm R}}_n$ of a pair of convex functions solving the Kantorovich dual problem for the Monge optimal transport with quadratic costs. Using the plug-in principle, we then define the empirical rank and sign maps as $\|\hat {{\rm R}}_n\|$ and $\hat {{\rm R}}_n/\|\hat {{\rm R}}_n\|$ and the empirical $\tau$-quantile sets and contours as $\hat {{\rm Q}}_n(\mathbb{S}(\tau))$ and $\hat {{\rm Q}}_n(\mathcal{S}(\tau))$. We establish the uniform convergence of these quantities to their theoretical counterparts. We derive these results as a consequence of the uniform convergence of empirical transport (vector quantile and rank) maps $\hat {{\rm Q}}_n$ and $\hat {{\rm R}}_n$ to their theoretical counterparts ${{\rm Q}}_P$ and ${{\rm R}}_P$ on compact subsets of the domain’s interior. This is the main theoretical result of the paper presented in Theorem 3.1. This result in turn is derived through an application of the extended continuous mapping theorem and a set of new theorems on stability of transport under deterministic perturbations of the source and target measures, given as Theorems A.1 and A.2 in the Appendix, which are new results of independent interest. Application of the extended continuous mapping theorem allows to us then to replace the deterministic perturbations by stochastic perturbations of measures and obtain the stochastic uniform convergence of the empirical transport maps.
Notation, conventions and preliminaries {#notation-conventions-and-preliminaries .unnumbered}
---------------------------------------
Let $(\Omega, \mathcal{A}, {{\rm I}\kern-0.18em{\rm P}})$ be some probability space. Throughout, $\mathcal P$ denotes a class of probability distributions over ${{\rm I}\kern-0.18em{\rm R}}^d$—unless otherwise specified, the class of all Borel probability measures on ${{\rm I}\kern-0.18em{\rm R}}^d$. Denote by $\mathbb S^d:=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;\|x\|\leq1\}$ the unit ball, and by $\mathcal S^{d-1}:=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;\|x\|=1\}$ the unit sphere, in ${{\rm I}\kern-0.18em{\rm R}}^d$. For $\tau\in(0,1]$, $\mathbb{S}(\tau ):=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;\|x\|\leq\tau\}$ is the ball, and $\mathcal{S}(\tau ):=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;\|x\|=\tau\}$ the sphere, of radius $\tau$. Let $P_X$ stand for the distribution of the random vector $X$. The symbol $\partial$ will denote either the boundary of a set or the subdifferential, as will be clear from the context. Following Villani [@villani1], we denote by $g\#\mu$ the [*image measure*]{} (or [*push-forward*]{}) of a measure $\mu\in\mathcal P$ by a measurable map $g:{{\rm I}\kern-0.18em{\rm R}}^d\rightarrow{{\rm I}\kern-0.18em{\rm R}}^d$. Explicitly, for any Borel set $A$, $g\#\mu(A):=\mu(g^{-1}(A))$. For a Borel subset $\mathbb{D}$ of a vector space equipped with the norm $\|\cdot\|$ and $f: \mathbb{D} \mapsto {{\rm I}\kern-0.18em{\rm R}}$, let $$\|f\|_{\mathrm{BL}(\mathbb{D})} := \sup_{x} |f(x)| \vee \sup_{x \neq x'} | f(x) - f(x')|\|x - x'\|^{-1} .$$ For two probability distributions $P$ and $P'$ on a measurable space $\mathbb{D}$, define the bounded Lipschitz metric as $$d_{{{\rm BL}}} (P, P') := \| P - P'\|_{{{\rm BL}}} := \sup_{\|f\|_{\mathrm{BL}(\mathbb{D})} \leq 1} \int f d (P - P'),$$ which metrizes the topology of weak convergence. Throughout the paper, we let $\mathcal{U}$ and $\mathcal{Y}$ be convex subsets of ${{\rm I}\kern-0.18em{\rm R}}^d$ with non-empty interiors. A [*convex*]{} function $\psi$ on ${\mathcal{U}}$ refers to a function $\psi:{\mathcal{U}}\rightarrow{{\rm I}\kern-0.18em{\rm R}}\cup\{+\infty\}$ for which $\psi((1-t)x+tx')\leq (1-t)\psi(x)+t\psi(x')$ for any $(x,x')$ such that $\psi(x)$ and $\psi(x')$ are finite and for any $t\in(0,1)$. Such a function is continuous on the interior of the convex set dom $\psi:=\{x\in{\mathcal{U}}: \psi(x)<\infty\}$, and differentiable Lebesgue-almost everywhere in dom $\psi$, by Rademacher’s theorem. Write $\nabla\psi$ for the gradient of $\psi$. For any function $\psi: {\mathcal{U}}\mapsto {{\rm I}\kern-0.18em{\rm R}}\cup \{+\infty\}$, the [*conjugate*]{} $\psi^*: {\mathcal{Y}}\mapsto {{\rm I}\kern-0.18em{\rm R}}\cup \{+\infty\}$ of $\psi$ is defined for each $y \in {\mathcal{Y}}$ by $$\psi^*(y) := \sup_{z \in {\mathcal{U}}} [y^\top z - \psi(z)].$$ The conjugate $\psi^\ast$ of $\psi$ is a convex lower-semi-continuous function on ${\mathcal{Y}}$. We shall call a [*conjugate pair of potentials*]{} over $({\mathcal{U}}, {\mathcal{Y}})$ any pair of lower-semi-continuous convex functions $(\psi, \psi^*)$ that are conjugates of each other. The transpose of a matrix $A$ is denoted $A^\top$. Finally, we call [*weak order*]{} a complete reflexive and transitive binary relation. Finally, recall the definition of Hausdorff distance between two non-empty sets $A$ and $B$ in ${{\rm I}\kern-0.18em{\rm R}}^d$: $$d_H(A,B) := \sup_{b \in B}\inf_{a \in A}\| a- b \| \vee \sup_{a \in A}\inf_{b \in B}\| a- b \|.$$
Outline of the paper {#outline-of-the-paper .unnumbered}
--------------------
Section \[sec:MK\] introduces and motivates the concepts of statistical depth, vector quantiles and vector ranks based on optimal transport maps. Section \[sec:emp\] describes estimators of depth contours, quantiles and ranks, and proves consistency of these estimators. Section \[sec:comp\] describes computational characterizations. The appendix presents additional theoretical results and proofs.
Statistical depth and vector ranks and quantiles {#sec:MK}
================================================
Statistical depth, regions and contours
---------------------------------------
The notion of statistical depth serves to define a center-outward ordering of points in the support of a distribution on ${{\rm I}\kern-0.18em{\rm R}}^d$, for $d>1$. As such, it emulates the notion of quantile for distributions on the real line. We define it as a real-valued index on ${{\rm I}\kern-0.18em{\rm R}}^d$ as follows.
\[def:depth\] A depth function is an upper-semi-continuous mapping ${{\rm D}}: {{\rm I}\kern-0.18em{\rm R}}^d \longmapsto {{\rm I}\kern-0.18em{\rm R}}$. In our context these functions will be indexed by a distribution $P$. The quantity ${{\rm D}}_P(x)$ is called the [*depth of*]{} $x$ [*relative to*]{} $P$. For each $P\in~\!\mathcal P$, the [*depth ordering*]{} $\geq_{{{\rm D}}_P}$ [*associated with*]{} ${{\rm D}}_P$ is the weak order on ${{\rm I}\kern-0.18em{\rm R}}^d$ defined, for $(y_1,y_2)\in~\!{{\rm I}\kern-0.18em{\rm R}}^{2d}$, by $$y_1\geq_{{{\rm D}}_P}y_2 \mbox{ if and only if } {{\rm D}}_P(y_1)\geq {{\rm D}}_P(y_2),$$ in which case $y_1$ is said to be [*deeper*]{} than $y_2$ relative to $P$.
The depth function thus defined allows graphical representations of the distribution $P$ through depth contours, which are collections of points of equal depth relative to $P$.
\[def:rc\] Let ${{\rm D}}_P$ be a depth function relative to distribution $P$ on ${{\rm I}\kern-0.18em{\rm R}}^d$. The [*region of depth*]{} $d$ is the upper contour set of level $d$ of ${{\rm D}}_P$, namely $\mathbb C_P(d)=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;{{\rm D}}_P(x)\geq d\}$; the [*contour of depth*]{} $d$ is the boundary $\mathcal C_P(d)= \partial \mathbb C_P(d)$.
By construction, the depth regions are nested: $$\forall (d,d')\in{{\rm I}\kern-0.18em{\rm R}}_+^2,\;
d'\geq d \;\Longrightarrow\;\mathbb C_P(d')\subseteq\mathbb C_P(d).$$ Hence, the depth ordering qualifies as a [*center-outward ordering*]{} of points in ${{\rm I}\kern-0.18em{\rm R}}^d$ relative to the center given by the set of the deepest points, $\arg\sup_{ x \in {{\rm I}\kern-0.18em{\rm R}}^d} {{\rm D}}_P(x).$
It is often convenient to work with depth regions indexed by their probability content.
\[def:rcq\]For $\tau \in [0,1]$, the depth region with probability content at least $\tau$ is $$\mathbb K_P(\tau):=\mathbb{C}_P(d(\tau)), \quad d(\tau):= \inf\{d \in {{\rm I}\kern-0.18em{\rm R}}:
{{\rm I}\kern-0.18em{\rm P}}_P (\mathbb{C}(d)) \geq \tau \};$$ the corresponding contour region is the boundary $\mathcal K_P(\tau):= \partial \mathbb K_P(d)$.
Liu-Zuo-Serfling axioms and Tukey’s halfspace depth {#sec:LZS}
---------------------------------------------------
\[fig:Tban\]
![Tukey halfspace depth contours for a banana-shaped distribution, produced with the algorithm of Paindaveine and Šiman [@PS:2012] from a sample of 9999 observations. The banana-like geometry of the data cloud is not picked by the convex contours, and the deepest point is close to the boundary of the support.](Banana_N=9999.eps){width="5in" height="5in"}
The four axioms proposed by Liu [@Liu:90] and Zuo and Serfling [@ZS:2000] to unify the diverse depth functions proposed in the literature are the following.
1. (Affine invariance) ${{\rm D}}_{ P_{AX+b}}(Ax+b)={{\rm D}}_{{ P}_X}(x)$ for any $x\in\mathbb{R}^d$, any full-rank $d\times d$ matrix $A$, and any $b\in{{\rm I}\kern-0.18em{\rm R}}^d$.
2. (Maximality at the center) If $x_0$ is a center of symmetry for ${ P}$ (symmetry here can be either [*central*]{}, [*angular*]{} or [*halfspace*]{} symmetry), it is [*deepest*]{}, that is, ${{\rm D}}_{ P}(x_0)=\max_{x\in{{{\rm I}\kern-0.18em{\rm R}}^d}}{{\rm D}}_{ P}(x)$.
3. (Linear monotonicity relative to the deepest points) If ${{\rm D}}_{ P}(x_0)=\max_{x\in{{{\rm I}\kern-0.18em{\rm R}}^d}}{{\rm D}}_{ P}(x)$, then ${{\rm D}}_{ P}(x) \leq {{\rm D}}_{ P}((1-\alpha) x_0 + \alpha x)$ for all $\alpha\in[0,1]$ and $x\in{{\rm I}\kern-0.18em{\rm R}}^d$: depth is monotonically decreasing along any straight line running through a deepest point.
4. (Vanishing at infinity) $\lim_{\Vert x\Vert \to\infty}{{\rm D}}_{ P}(x) = 0$.
The earliest and most popular depth function is [*halfspace depth*]{} proposed by Tukey [@Tukey:75]:
\[def:half\] The halfspace depth ${{\rm D}}^{\scriptsize{\text{Tukey}}}_{ P}(x)$ of a point $x\in~\!{{\rm I}\kern-0.18em{\rm R}}^d$ with respect to the distribution $ P_X$ of a random vector $X$ on ${{\rm I}\kern-0.18em{\rm R}}^d$ is defined as $${{\rm D}}^{\scriptsize{\text{Tukey}}}_{ P_X}(x):=\min_{{\varphi}\in\mathcal S^{d-1}} {{\rm I}\kern-0.18em{\rm P}}[(X-x)^\top {\varphi}\geq 0].$$
Halfspace depth relative to any distribution with nonvanishing density on ${{\rm I}\kern-0.18em{\rm R}}^d$ satisfies (A1)-(A4). The appealing properties of halfspace depth are well known and well documented: see Donoho and Gasko [@DG:92], Mosler [@Mosler:2002], Koshevoy [@Koshevoy:2002], Ghosh and Chaudhuri [@GC:2005], Cuestas-Albertos and Nieto-Reyes [@CN:2008], Hassairi and Regaieg [@HR:2008], to cite only a few. Halfspace depth takes values in $[0, 1/2]$, and its contours are continuous and convex; the corresponding regions are closed, convex, and nested as $d$ decreases. Under very mild conditions, halfspace depth moreover fully characterizes the distribution $P$. For somewhat less satisfactory features, however, see Dutta et al. [@DGC:2011]. An important feature of halfspace depth is the convexity of its contours, which implies that halfspace depth contours cannot pick non convex features in the geometry of the underlying distribution, as illustrated in Figure 1.
We shall propose below a new depth concept, the Monge-Kantorovich (MK) depth, that relinquishes the affine equivariance and star convexity of contours imposed by Axioms (A1) and (A3) and recovers non convex features of the underlying distribution. As a preview of the concept, without going through any definition, we illustrate in Figure 2 (using the same banana-shaped distribution as in Figure 1) the ability of the MK depth to capture non-convexities. In what follows, we characterize these abilities more formally. We shall emphasize that this notion comes in a package with new, interesting notions of vector ranks and quantiles, based on optimal transport, which reduce to classical notions in the univariate and multivariate spherical cases.
Monge-Kantorovich depth
-----------------------
The principle behind the notion of depth we define here is to map the depth regions and contours relative to a well-chosen reference distribution $F$, into depth contours and regions relative to a distribution of interest $P$ on ${{\rm I}\kern-0.18em{\rm R}}^d$, using a well-chosen mapping. The mapping proposed here is the gradient of a convex function $\nabla \psi$ such that if $U$ has distribution $F$, then $Y = \nabla \psi (U)$ has distribution $P$, or, in terms of measures, $\nabla \psi \# F = P$. The gradient $\nabla \psi$ is said to to push $F$ forward to $P$, which is conventionally denoted by the push-forward notation, $\nabla \psi \# F = P$, which is defined in the notation section.
The gradient of a convex function property is a generalization of monotonicity in the one-dimensional case. When $F$ and $P$ have finite second-order moments, these maps are the optimal Monge-Kantorovich transport maps from $F$ to $P$ for the quadratic cost, as explained below. In the unidimensional case, when $F$ is the standard uniform, the gradient/optimal transport map $\nabla \psi$ coincides with the classical quantile function.
\[fig:MKban\]
{width="4in" height="5in"}
The following theorem, due to Brenier [@brenier] and McCann [@McCann1995], establishes existence of gradients of convex functions with the required properties.
\[thm:polar\] Let $P$ and $F$ be two distributions on ${{\rm I}\kern-0.18em{\rm R}}^d$. (1) If $F$ is absolutely continuous with respect to the Lebesgue measure on $\Bbb{R}^d$, with support contained in a convex set ${\mathcal{U}}$, the following holds: there exists a convex function $\psi:{\mathcal{U}}\rightarrow{{\rm I}\kern-0.18em{\rm R}}\cup\{+\infty\}$ such that $\nabla\psi\# F=P$. The function $\nabla\psi$ exists and is unique, $F$-almost everywhere. (2) If, in addition, $P$ is absolutely continuous on $\Bbb{R}^d$ with support contained in a convex set ${\mathcal{Y}}$, the following holds: there exists a convex function $\psi^*:{\mathcal{Y}}\rightarrow{{\rm I}\kern-0.18em{\rm R}}\cup\{+\infty\}$ such that $\nabla\psi^*\# P=F$. The function $\nabla\psi^*$ exists, is unique and equal to $\nabla \psi^{-1}$, $P$-almost everywhere.
If $P$ and $F$ have finite second moments, ${{\rm Q}}$ is $F$-almost everywhere equal to the [*optimal transport plan*]{} $\nabla \psi$ from $F$ to $P$ for quadratic cost: namely, the map ${{\rm Q}}: {{\rm I}\kern-0.18em{\rm R}}^d\longrightarrow{{\rm I}\kern-0.18em{\rm R}}^d$ solves the problem $$\begin{aligned}
\inf_{Q} \int (u - Q(u))^2 d F(u) : \quad Q\#F=P,\end{aligned}$$ or, equivalently, $$\begin{aligned}
\sup_{Q} \int u^\top Q(u)\;dF(u) : \quad Q\#F=P.\label{eq:OT}\end{aligned}$$ This definition has a classical counterpart in the case of univariate distributions. When $d=1$ and $F$ is uniform on $[0,1]$, the optimal transport $u\mapsto {{\rm Q}}(u)$ is the classical quantile function for distribution $P$.
We now state a fundamental duality result due to Kantorovich and Brenier, which we explicitly rely on in Section \[sec:emp\].
\[dual\] Suppose hypothesis (1) of Theorem \[thm:polar\] holds and $P$ and $F$ have finite second moments, then the function $\psi$, or optimal potential, solves the optimization problem $$\begin{aligned}
\label{eq:Kdual}
\int\psi dF+\int\psi^\ast dP=\inf_{(\varphi, \varphi^*)} \left( \int\varphi dF+\int\varphi^\ast dP \right),\end{aligned}$$ where the infimum is taken over the class of conjugate pairs of potentials $(\varphi, \varphi^*)$ over $( {\mathcal{U}}, {\mathcal{Y}})$.
This problem is dual to the optimal transport problem (\[eq:OT\]). Moreover, under the hypotheses of Theorem \[dual\], $\nabla\psi$ is the unique optimal transport map from $F$ to $P$ for quadratic cost, in the sense that any other optimal transport coincides with $\nabla \psi$ on a set of $F$-measure 1 (see [@villani1]). Under the hypotheses of Theorem \[dual\] and hypothesis (2) of Theorem \[thm:polar\], $\nabla\psi^*$ is the unique optimal (reverse) transport map from $P$ to $F$ for quadratic cost, in the sense that any other optimal transport coincides with $\nabla \psi^*$ on a set of $P$-measure one (see [@villani1]).
Next we use Theorem \[thm:polar\] to define a natural notion of *vector quantiles* and *vector ranks*.
\[def:MK\] Let $F$ be an absolutely continuous reference distribution with support in a convex set ${\mathcal{U}}\subseteq {{\rm I}\kern-0.18em{\rm R}}^d$, and let $P$ be an arbitrary distribution with support in a convex set ${\mathcal{Y}}\subseteq {{\rm I}\kern-0.18em{\rm R}}^d$. Let $\nabla \psi$ be the $F$-almost surely unique gradient of a convex function $\psi$ of Theorem \[thm:polar\] and let $\psi^\ast$ be the conjugate of $\psi$ over $({\mathcal{U}}, {\mathcal{Y}})$. Vector quantiles and ranks are defined as follows: $${{\rm Q}}_P(u) \in \arg \sup_{y \in {\mathcal{Y}}}[ y^\top u - \psi^*(y)] , \ \ { u \in {\mathcal{U}}}; \quad
{{\rm R}}_P(y) \in \arg \sup_{u \in {\mathcal{U}}} [y^\top u - \psi(u)], \ \ { y \in {{\rm I}\kern-0.18em{\rm R}}^d}.$$
Thus we define the MK vector quantiles ${{\rm Q}}_P$ and ranks ${{\rm R}}_P$ as any solutions of the optimization problems in the display above. Our definition here does not impose any moment condition and ensures that the quantities are defined for every value of the argument in the appropriate domains. By the envelope theorem and Rademacher’s theorem ([@villani1]), the maps ${{\rm Q}}_P$ and ${{\rm R}}_P$ essentially coincide with the gradients $\nabla \psi$ and $\nabla \psi^*$ of conjugate potentials $\psi$ and $\psi^*$, namely $${{\rm Q}}_P = \nabla \psi \text{ a.e. on } \ \mathcal{U}, \quad {{\rm R}}_P = \nabla \psi^* \text{ a.e. on } \ \mathcal{Y},$$ where “a.e." abbreviates “almost everywhere with respect to the Lebesgue measure". In the fact, the equality holds everywhere on certain domains under condition (C) stated below. Under the conditions of Theorem \[dual\], the pair $(\psi, \psi^*)$ has the variational characterization given in (\[eq:Kdual\]).
When requiring regularity of vector quantiles and ranks, we shall impose the following condition on the conjugate pair of optimal potentials $(\psi,\psi^\ast)$ over $({\mathcal{U}}, {\mathcal{Y}})$.
- Let $\mathcal{U}$ and $\mathcal{Y}$ be closed, convex subsets of ${{\rm I}\kern-0.18em{\rm R}}^d$, and ${\mathcal{U}}_0 \subset {\mathcal{U}}$ and ${\mathcal{Y}}_0 \subset {\mathcal{Y}}$ be open, non-empty sets in ${{\rm I}\kern-0.18em{\rm R}}^d$. Let $\psi: \mathcal{U} \mapsto {{\rm I}\kern-0.18em{\rm R}}$ and $\psi^*: \mathcal{Y} \mapsto {{\rm I}\kern-0.18em{\rm R}}$ form a conjugate pair over $({\mathcal{U}}, {\mathcal{Y}})$ and possess gradients $\nabla \psi(u)$ for all $u \in {\mathcal{U}}_0$, and $\nabla \psi^*(y)$ for all $y \in {\mathcal{Y}}_0$. The gradients $\nabla \psi |_{{\mathcal{U}}_0}: {\mathcal{U}}_0 \mapsto {\mathcal{Y}}_0$ and $\nabla \psi^* |_{{\mathcal{Y}}_0}: {\mathcal{Y}}_0 \mapsto {\mathcal{U}}_0$ are homeomorphisms and $\nabla \psi|_{{\mathcal{U}}_0} = (\nabla \psi^* |_{{\mathcal{Y}}_0})^{-1}$.
Under Condition (C), we have: $$\label{as gradients}
{{\rm Q}}_P(u) = \nabla\psi(u) \text{ for all } u \in {\mathcal{U}}_0, \quad
{{\rm R}}_P(y) = \nabla\psi^\ast(y)=\left( \nabla\psi \right)^{-1}(y) \text{ for all } y \in {\mathcal{Y}}_0,$$ that is, vector ranks and quantiles are defined as gradients of conjugate potentials for each (as opposed to almost every) value in the indicated sets, and inverse functions of each other.
Sufficient conditions for Condition (C) in the context of Definition \[def:MK\] are provided by Caffarelli’s regularity theory (Villani [@villani1], Theorem 4.14). One set of sufficient conditions is as follows.
\[prop:C\] Suppose that $P$ and $F$ admit densities, which are of smoothness class $C^{\beta}$ for $\beta>0$ on convex, compact support sets ${\rm cl}({\mathcal{Y}}_0)$ and ${\rm cl}({\mathcal{U}}_0)$, and the densities are bounded away from zero and above uniformly on the support sets. Then Condition (C) is satisfied for the conjugate pair $(\psi,\psi^\ast)$ such that $\nabla\psi\#F=P$ and $\nabla\psi^\ast\#P=F.$
We now can give our [*main*]{} definition – that of multivariate notions of quantiles and ranks, through which a depth function will be inherited from the reference distribution $F=U_d$.
\[def:MK\] Let $F$ be the spherical uniform distribution $U_d$ on a unit ball ${\mathcal{U}}= \mathbb{S}^d$, and $P$ be an arbitrary distribution with support in a convex region ${\mathcal{Y}}\subseteq {{\rm I}\kern-0.18em{\rm R}}^d$. MK quantiles, ranks, signs and depth are defined as follows.
1. The [*MK rank*]{} of $y \in {{\rm I}\kern-0.18em{\rm R}}^d$ is $\|{{\rm R}}_P(y)\|$ and the [*MK sign*]{} is ${{\rm R}}_P(y)/\|{{\rm R}}_P(y)\|$.
2. The [*MK*]{} $\tau$-[*quantile contour*]{} is the set ${{\rm Q}}_P(\mathcal S(\tau))$ and the [*MK depth region*]{} with probability content $\tau$ is ${{\rm Q}}_P(\mathbb S(\tau))$.
3. The [*MK depth*]{} of $y \in{{\rm I}\kern-0.18em{\rm R}}^d$ is the depth of ${{\rm R}}_P(y)$ under ${{\rm D}}^{\scriptsize{\text{Tukey}}}_{U_d}$: $${{\rm D}}^{\rm\scriptsize MK}_P(y):={{\rm D}}^{\scriptsize{\text{Tukey}}}_{U_d}({{\rm R}}_P(y)).$$
The notion of depth proposed in Definition \[def:MK\] is based on an optimal transport map from the reference spherical uniform distribution $F=U_d$ to the distribution of interest $P$. Under Condition (C), ${{\rm Q}}_P$ and ${{\rm R}}_P$ are continuous and are mutual inverse maps, so that the MK $\tau$-quantile contours are continuously deformable into spheres and the MK depth regions with probability content $\tau$ are nested.
By choosing other reference distributions $F$, such as the uniform distribution on a unit hypercube, or the standard Gaussian distribution, we can give a more general definition of MK ranks, quantiles, and signs, which may be of interest.
\[def:MK2\]Let $F$ be an absolutely continuous reference distribution with support contained in a convex region ${\mathcal{U}}\subseteq \Bbb{R}^d$, and let $\| \cdot \|$ be a norm on ${\mathcal{U}}$. Let ${{\rm D}}_F: {{\rm I}\kern-0.18em{\rm R}}^d \to \Bbb{R}_+$ be an associated reference depth function and ${\mathcal{K}}(\tau)$ the associated $\tau$-quantile contour and ${\mathbb{K}}(\tau)$ the associated depth region with probability content $\tau$. The MK quantiles, ranks, signs and depth are defined as follows.
1. The [*MK rank*]{} of $y \in {{\rm I}\kern-0.18em{\rm R}}^d$ is $\|{{\rm R}}_P(y)\|$ and the [*MK sign*]{} is ${{\rm R}}_P(y)/\|{{\rm R}}_P(y)\|$.
2. The [*MK*]{} $\tau$-[*quantile*]{} is the set ${{\rm Q}}_P({\mathcal{K}}(\tau))$ and the [*MK depth region*]{} with probability mass $\tau$ is ${{\rm Q}}_P({\mathbb{K}}(\tau))$.
3. The [*MK depth*]{} of $y \in{{\rm I}\kern-0.18em{\rm R}}^d$ is the depth of ${{\rm R}}_P(y)$ under ${{\rm D}}_F$: $${{\rm D}}^{\rm\scriptsize MK}_P(y):={{\rm D}}_F({{\rm R}}_P(y)).$$
Of course, all the quantities thus defined depend on the choice of the reference distribution $F$ and the depth function ${{\rm D}}_F$.
When the reference distribution $F$ is spherical, it is natural to use Tukey’s depth function ${{\rm D}}_F = {{\rm D}}^{\scriptsize{\text{Tukey}}}_{F}$ to define the MK depth of $y \in{{\rm I}\kern-0.18em{\rm R}}^d$ relative to $P$ as the halfspace depth of ${{\rm R}}_P(y)$ relative to the reference distribution $F$, namely $${{\rm D}}^{\rm\scriptsize MK}_P(y):={{\rm D}}^{\scriptsize{\text{Tukey}}}_{F}({{\rm R}}_P(y)).$$ The choice of halfspace depth may be less natural for non-spherical reference distributions. One example is where $F$ is the standard uniform distribution $U[0,1]^d$ on the unit cube $[0,1]^d$. Then it seems natural to use the sup norm $\|\cdot\|_\infty$ as the norm $\| \cdot \|$ and the depth function ${{\rm D}}_{U[0,1]^d}(y) = 1/2 - \| y - \mathbf{1}/2\|_\infty$, where $\mathbf{1} = (1,\dots,1)'$, in which case ${\mathbb{K}}(\tau)$ is a cube of diameter $\tau^{1/d}$ centered at $\mathbf{1}/2$. In this case, the MK depth is $${{\rm D}}^{\rm\scriptsize MK}_P(y):={{\rm D}}_{U[0,1]^d}({{\rm R}}_P(y)).$$
Monge-Kantorovich depth with spherical uniform reference distribution
---------------------------------------------------------------------
Here we consider in more detail the Monge-Kantorovich depth defined from a baseline spherical uniform distribution $U_d$ supported on the unit ball $\mathbb S^d$ of ${{\rm I}\kern-0.18em{\rm R}}^d$. Recall that this distribution is that of a random vector $r{\varphi}$, where $r$ is uniform on $[0,1]$, ${\varphi}$ is uniform on the unit sphere $\mathcal S^{d-1}$, and $r$ and ${\varphi}$ are mutually independent.
The spherical symmetry of distribution $U_d$ produces halfspace depth contours that are concentric spheres, the deepest point being the origin. The radius $\tau $ of the ball $\mathbb{S}(\tau )=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;\|x\|\leq\tau\}$ is also its $U_d$-probability contents, that is, $\tau =U_d({{\mathbb{S}}}(\tau ))$. Letting $\theta :=\arccos \tau $, the halfspace depth with respect to $U_d$ of a point $\tau u\in\mathcal{S}(\tau ):=\{x\in{{\rm I}\kern-0.18em{\rm R}}^d:\;\|x\|=\tau\}$, where $\tau \in (0,1]$ and $u\in\mathbb S^d$, is $$\label{hlfd} {{\rm D}}_U(\tau u)=\left\{
\begin{array}{ll} \pi^{-1}[\theta -\cos\theta \log \vert \sec\theta + \tan\theta\vert ] &
d\geq 2\\(1-\tau )/2 & d=1.\end{array}\right.$$ Note that for $d=1$, $u$ takes values $\pm 1$ and, in agreement with rotational symmetry of $U_d$, that depth does not depend on $u$.
The principle behind the notion of depth we investigate further here is to map the depth regions and contours relative to the spherical uniform distribution $U_d$, namely, the concentric spheres, into depth contours and regions relative to a distribution of interest $P$ on ${{\rm I}\kern-0.18em{\rm R}}^d$ using the optimal transport plan from $U_d$ to $P$. Under the sufficient conditions for Condition (C) provided in Lemma \[prop:C\] (note that the conditions on $F$ are automatically satisfied in case $F=U_d$), ${{\rm Q}}_P$ and ${{\rm R}}_P$ are continuous and are inverse maps of each other, so that the MK depth contours are continuously deformable into spheres, the MK depth regions are nested, and regions and contours, when indexed by probability content, take the respective forms $${{\rm Q}}_P\left( \mathbb S(\tau)\right) \mbox{ and } {{\rm Q}}_P\left( \mathcal S(\tau)\right),\mbox{ for }\tau\in(0,1].$$
### MK depth is halfspace depth in dimension 1 {#mk-depth-is-halfspace-depth-in-dimension-1 .unnumbered}
The halfspace depth of a point $x\in{{\rm I}\kern-0.18em{\rm R}}$ relative to a distribution $P$ over ${{\rm I}\kern-0.18em{\rm R}}$ takes the very simple form $$D^{\scriptsize{\text{Tukey}}}_P(x)=\min(P(x), 1-P(x)),$$ where, by abuse of notation, $P$ stands for both distribution and distribution function. The nondecreasing map defined for each $x\in{{\rm I}\kern-0.18em{\rm R}}$ by $x\mapsto {{\rm R}}_P(x)=2P(x) -1$ is the derivative of a convex function and it transports distribution $P$ to $U_1$, which is uniform on $[-1,1]$, i.e., ${{\rm R}}_P\#P=U_1$. Hence ${{\rm R}}_P$ coincides with the MK vector rank of Definition \[def:MK\]. Therefore, for each $x\in{{\rm I}\kern-0.18em{\rm R}}$, $$D_P(x)=D^{\scriptsize{\text{Tukey}}}_{U_d}({{\rm R}}_P(x))=\min(P(x), 1-P(x))$$ and MK depth coincides with Tukey depth in case of all distributions with nonvanishing densities on the real line.
More generally (still in the univariate case), denoting by $F_1$ and $F_2$ the distribution functions associated with two absolutely continuous distributions $P_1$ and $P_2$, the mapping $F_2^{-1}\circ F_1$, being monotone increasing, is also the optimal transport from $P_1$ to $P_2$. The same transformation has been studied, in a different context, by Doksum [@Doksum:74] and Doksum and Sievers [@DS:76]; see also the concept of convex ordering proposed by van Zwet [@VZ:1964].
### MK depth is halfspace depth for elliptical families {#mk-depth-is-halfspace-depth-for-elliptical-families .unnumbered}
As explained in the introduction, a $d$-dimensional random vector $X$ has elliptical distribution $P_{\mu,\Sigma,g}$ with location $\mu\in~\!{{\rm I}\kern-0.18em{\rm R}}^d$, positive definite symmetric $d\times d$ scatter matrix $\Sigma$ and radial density function $g$ (radial distribution function $G$) if and only if, denoting by $\Sigma^{1/2}$ the symmetric root of $\Sigma$, $Y:=\Sigma^{-1/2}(X-\mu)$ has spherical distribution $P= P_{0,I,g}$ (hence $\Vert Y\Vert $ has density $f$), which holds if and only if $$\label{eq:ell}
{{\rm R}}_P (Y):=\frac{Y}{\Vert Y\Vert }G\big(\Vert Y\Vert\big)
\mbox{ is distributed according to } U_d.$$ Let $\Psi(t) = \int_{-\infty}^t G(r) d r$, and note that the map $z\mapsto {{\rm R}}_P(z)$ is the gradient of $\psi^\ast(z):= \Psi(\Vert z \Vert)$ so that, from (\[eq:ell\]), $\nabla\psi^\ast\#P=U_d$ as Definition \[def:MK\] requires. That $\psi^\ast$ is convex follows from Theorem 5.1 of [@RockafellarConvex] by noting that $\psi^\ast$ is a composition of $\Psi: {{\rm I}\kern-0.18em{\rm R}}\to {{\rm I}\kern-0.18em{\rm R}}$, a convex, non-decreasing map, and $\|\cdot \|: {{\rm I}\kern-0.18em{\rm R}}^d \to R$, a convex function by definition of the norm. As a consequence, the mapping ${{\rm R}}_P$ in (\[eq:ell\]) is the MK vector rank function associated with $P=P_{0,I,f}$; and, the MK depth contours (with probability content $\tau$) of $P$ are spheres with radii $G^{-1}(\tau)$ centered at the origin: $$D_P(x) = \{ y \in {{\rm I}\kern-0.18em{\rm R}}^d: \| y \| \leq G^{-1} (\tau) \}.$$ These spheres are halfspace depth contours for $P$. This is the precise sense in which MK depth reduces to halfspace for elliptical families.
It should be noted above, that we treat location and scatter parameters as known, and transform $X$ to a vector $Y$ in isotropic position. This transformation ensures basic invariance properties of the resulting depth, ranks, and quantiles with respect to affine transformations. When those parameters are unknown, they will have to be replaced with by affine-equivariant estimators, as in the usual definition of elliptical ranks and signs (see,e.g., [@HP:2002]) in order to insure similar invariance properties for the empirical analogs. Without the aforementioned transformation, however, the invariance properties are not guaranteed, owing to the fact that composition of two gradients of convex functions is not necessarily the gradient of a convex function, unless the composition has a specific structure, as is the case above.
Empirical depth, ranks and quantiles {#sec:emp}
====================================
Having defined Monge-Kantorovich vector quantiles, ranks and depth relative to a distribution $P$ based on reference distribution $F$ on ${{\rm I}\kern-0.18em{\rm R}}^d$, we now turn to the estimation of these quantities. Hereafter, we shall assume that Condition (C) holds. Then, the MK vector quantiles and ranks of Definition \[def:MK\] are $${{\rm Q}}_P(u) := \nabla\psi(u) , \quad
{{\rm R}}_P(y) := \nabla\psi^\ast(y)=\left( \nabla\psi \right)^{-1}(y),$$ for each $u \in {\mathcal{U}}_0$ and $y \in {\mathcal{Y}}_0$, respectively. We define $\Phi_0({\mathcal{U}}, {\mathcal{Y}})$ as a collection of conjugate potentials $({\varphi}, {\varphi}^*)$ on $({\mathcal{U}}, {\mathcal{Y}})$ such that ${\varphi}(u_0)= 0$ for some fixed point $u_0 \in {\mathcal{U}}_0$. Under the conditions of Theorem \[dual\], the potentials $(\psi, \psi^*)$ solve the dual problem $$\label{kantorovich 2}
\int \psi dF+ \int \psi^* d P = \inf_{ (\varphi, \varphi^*) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}}) } \int \varphi dF + \int \varphi^* dP .$$ Constraining the conjugate pair to lie in $\Phi_0({\mathcal{U}}, {\mathcal{Y}})$ is a normalization that (without any loss of generality) pins down the constant, so that $(\psi,\psi^\ast)$ are uniquely determined, as argued in the proof.
We propose empirical versions of MK quantiles and ranks based on estimators $\hat P$ of $P$. The typical case is when the reference measure $F$ is known. However, our theory allows us to handle the case where $F$ is itself unknown, and so it is estimated by some $\hat F$. This is indeed useful for at least two reasons. First, we may be interested in a classical problem of comparing one distribution $P$ to a reference distribution $F$, both of which are known only up to a random sample available from each of them. Second, we may be interested in discretizing $F$ for computational reasons, as we discuss in Section \[sec:comp\], in which case the discretized $F$ is the estimator of $F$.
Conditions on estimators of $P$ and $F$
---------------------------------------
Suppose that $\{\hat P_n\}_{n=1}^\infty$ and $\{\hat F_n\}^\infty_{n=1}$ are sequences of random measures on ${\mathcal{Y}}$ and ${\mathcal{U}}$, with finite total mass, that are consistent for $P$ and $F$, in the sense that $$\label{empirical KR}
d_{{{\rm BL}}}(\hat P_n, P) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \quad d_{{{\rm BL}}}(\hat F_n, F) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0,$$ where $\to_{{{\rm I}\kern-0.18em{\rm P}}^*}$ denotes convergence in (outer) probability under probability measure ${{\rm I}\kern-0.18em{\rm P}}$, see van der Vaart and Wellner [@VvW]. A basic example is where $\hat P_n$ is the empirical distribution of a random sample $(Y_i)_{i=1}^n$ drawn from $P$ and $\hat F_n$ is the empirical distribution of a random sample $(U_i)_{i=1}^n$ drawn from $F$. Other, much more complicated examples, including smoothed empirical measures and data originating from dependent processes, satisfy sufficient conditions for (\[empirical KR\]) that we now give. In order to develop some examples, we introduce an ergodicity condition:
- Let $\mathcal{W}$ be a measurable subset of ${{\rm I}\kern-0.18em{\rm R}}^d$. A data stream $\{(W_{t,n})_{t=1}^n\}_{n=1}^\infty$, with $W_{t,n} \in \mathcal{W} \subseteq {{\rm I}\kern-0.18em{\rm R}}^d$ for each $t$ and $n$, is ergodic for the probability law $P_W$ on $\mathcal{W}$ if for each $g: \mathcal{W} \mapsto {{\rm I}\kern-0.18em{\rm R}}$ such that $\|g\|_{{{\rm BL}}(\mathcal{W})} < \infty$, the law of large numbers holds:
$$\label{erg}
\frac{1}{n} \sum_{t=1}^n g(W_{t,n}) \to_{{{\rm I}\kern-0.18em{\rm P}}} \int g(w) d P_W(w).$$
The class of ergodic processes is extremely rich, including in particular the following cases:
- $W_{t,n}=W_t$, where $(W_t)_{t=1}^\infty$ are independent, identically distributed random vectors with distribution $P_W$;
- $W_{t,n}=W_t$, where $(W_t)_{t=1}^\infty$ is stationary strongly mixing process with marginal distribution $P_W$;
- $W_{t,n}=W_{t}$, where $(W_t)_{t=1}^\infty$ is an irreducible and aperiodic Markov chain with invariant distribution $P_W$;
- $W_{t,n}=w_{t,n}$, where $(w_{t,n})_{t=1}^n$ is a deterministic sequence of points such that (\[erg\]) holds deterministically.
For a detailed motivation and discussion of the use of deterministic sequences such as, for example, the so-called [*low-discrepancy sequences*]{}: see, e.g., Chapter 9 and, more particularly, page 314 of [@ken:judd].
Thus, if the data stream $\{(W_{t,n})_{t=1}^n\}_{n=1}^\infty$ is ergodic for $P_W$, we can estimate $P_W$ by the empirical and smoothed empirical measures $$\hat P_W(A) = \frac{1}{n} \sum_{t=1}^n 1\{W_{t,n} \in A \}, \ \quad
\ \tilde P_W(A) = \frac{1}{n} \sum_{t=1}^n \int_{{{\rm I}\kern-0.18em{\rm R}}^d} 1\{W_{t,n} + h_n {\varepsilon}\in A \cap \mathcal{W} \} d \Phi ({\varepsilon}),$$ where $\Phi $ is the probability law of the standard $d$-dimensional Gaussian vector, $N(0,I_d)$, and $h_n \geq 0$ a semi-positive-definite matrix of bandwidths such that $\|h_n\| \to 0$ as $n \to \infty$. Note that $\tilde P_W$ may not integrate to 1, since we are forcing it to have support in $\mathcal{W}$.
\[ergodic\] Suppose that $P_W$ is absolutely continuous with support in the compact set $\mathcal{W} \subset \Bbb{R}^d$. If $\{(W_{t,n})_{t=1}^n\}_{n=1}^\infty$ is ergodic for $P_W$ on $\mathcal{W}$, then $$d_{{{\rm BL}}} (\hat P_W, P_W) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \quad d_{{{\rm BL}}} (\tilde P_W, P_W) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0.$$ Thus, if $P_Y:=P$ and $P_U:=F$ are absolutely continuous with support sets contained in compact sets ${\mathcal{Y}}$ and ${\mathcal{U}}$, and if $\{(Y_{t,n})_{t=1}^n\}_{n=1}^\infty$ is ergodic for $P_Y$ on ${\mathcal{Y}}$ and $\{(U_{t,n})_{t=1}^n\}_{n=1}^\infty$ is ergodic for $P_U$ on ${\mathcal{U}}$, then $\hat P_n= \hat P_W$ or $\tilde P_W$ and $\hat F_n = \hat P_U$ or $\tilde P_U$ obey condition (\[empirical KR\]).
Absolute continuity of $P_W$ in Lemma \[ergodic\] is invoked to show that the smoothed estimator $\tilde P_W$ is asymptotically non-defective.
Empirical vector quantiles and ranks
------------------------------------
We base empirical versions of MK quantiles, ranks and depth on estimators $\hat P_n$ for $P$ and $\hat F_n$ for $F$ satisfying (\[empirical KR\]). This includes cases where the reference measure $F$ is known, i.e. $\hat F_n=F$. Recall Assumption (C) is maintained throughout this section.
Empirical vector quantile $\hat{{\rm Q}}_n$ and vector rank $\hat{{\rm R}}_n$ are any pair of functions satisfying, for each $u \in {\mathcal{U}}$ and $y \in {\mathcal{Y}}$, $$\label{eq: empirical Q and R}
\hat {{\rm Q}}_n(u) \in \arg \sup_{y \in {\mathcal{Y}}}[ y^\top u - \hat \psi_n^*(y)] , \quad
\hat {{\rm R}}_n(y) \in \arg \sup_{u \in {\mathcal{U}}} [y^\top u - \hat \psi_n(u)],$$ where $(\hat \psi_n, \hat \psi^*_n) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}})$ is such that $$\label{empirical kantorovich 2}
\int \hat \psi_n d\hat F_n + \int \hat \psi_n^* d \hat P_n =\inf_{ ({\varphi}, {\varphi}^*) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}}) } \int \varphi d\hat F_n + \int \varphi^* d\hat P_n.$$
We now state the main result of Section \[sec:emp\].
\[theorem: empirical\] Suppose that the sets ${\mathcal{U}}$ and ${\mathcal{Y}}$ are compact subsets of ${{\rm I}\kern-0.18em{\rm R}}^d$, and that the probability measures $P$ and $F$ are absolutely continuous with respect to the Lebesgue measure, with ${\rm support }(P) \subseteq {\mathcal{Y}}$ and ${\rm support }(F) \subseteq {\mathcal{U}}$. Suppose that $\{\hat P_n\}$ and $\{\hat F_n\}$ are sequences of random measures on ${\mathcal{Y}}$ and ${\mathcal{U}}$, with finite total mass, that are consistent for $P$ and $F$ in the sense of (\[empirical KR\]). Suppose that Condition (C) holds for the solution of (\[kantorovich 2\]) for ${\mathcal{Y}}_0 := {\rm int}({\rm support}(P))$ and ${\mathcal{U}}_0:= {\rm int}({\rm support}(F))$. Then, as $n \to \infty$, for any closed set $K \subset \ {\mathcal{U}}_0$ and any closed set $K' \subset {\mathcal{Y}}_0$, $$\sup_{ u \in K} \| \hat {{\rm Q}}_n(u) - {{\rm Q}}_P(u) \| \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \quad
\sup_{ y \in K'} \| \hat {{\rm R}}_n(y) - {{\rm R}}_P(y) \| \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0,$$ and $$\sup_{A \subseteq K }d_H(\hat {{\rm Q}}_{n}(A),{{\rm Q}}_P(A)) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \quad
\sup_{A' \subseteq K'} d_H(\hat {{\rm R}}_{n}(A'), {{\rm R}}_P(A')) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0,$$ where the suprema are taken over nonempty subsets.
The first result establishes the uniform consistency of empirical vector quantile and rank maps, hence also of empirical ranks and signs. The set ${{\rm Q}}_P(K)$ with $K = \mathbb{K}(\tau)$ is the statistical depth contour with probability content $\tau$. The second result, therefore, establishes consistency of the approximation $\hat{{\rm Q}}_n(K)$ to the theoretical depth region ${{\rm Q}}_P(K)$.
Empirical MK quantiles, ranks, and signs and their convergence
--------------------------------------------------------------
We work with the conditions of the previous theorem, but here, for the sake of simplicity, we first consider the lead case where $F$ is known, i.e. $\hat F_n = F$.
\[def:EMK1\]Let $F$ be an absolutely continuous reference distribution with support contained in a convex region ${\mathcal{U}}\subseteq \Bbb{R}^d$, and let $\| \cdot \|$ be a norm on ${\mathcal{U}}$. The MK empirical quantiles, ranks, signs and depth are defined as follows.
1. The [*MK empirical rank*]{} and [*sign*]{} of $y \in {{\rm I}\kern-0.18em{\rm R}}^d$ are $\|\hat {{\rm R}}_n(y)\|$ and $ \hat {{\rm R}}_n (y)/\|\hat {{\rm R}}_P(y)\|$.
2. The [*MK empirical*]{} $\tau$-[*quantile contour*]{} is the set $\hat {{\rm Q}}_n({\mathcal{K}}(\tau))$ and the [*MK empirical depth region*]{} with probability mass $\tau$ is $\hat {{\rm Q}}_n({\mathbb{K}}(\tau))$.
3. The [*MK empirical depth*]{} of $y \in{{\rm I}\kern-0.18em{\rm R}}^d$ is the depth of $\hat {{\rm R}}_n(y)$ under ${{\rm D}}_F$: $$\hat{{\rm D}}^{\rm\scriptsize MK}_{P,n}(y):={{\rm D}}_F(\hat {{\rm R}}_n(y)).$$
Uniform convergence of empirical MK rank, signs and depth to their theoretical counterparts follows by an application of the Extended Continuous Mapping Theorem.
\[cor:ERS\] Work with the assumptions of Theorem \[theorem: empirical\], and assume that $D_F$ is continuous on ${\mathcal{U}}_0$. As $n \to \infty$, for any closed set $K' \subset {\mathcal{Y}}_0$, $$\begin{aligned}
\sup_{ y \in K'} | \|\hat {{\rm R}}_n(y)\| - \|{{\rm R}}_P(y)\| | \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \\
\sup_{ y \in K'} | \hat {{\rm R}}_n (y)/\|\hat {{\rm R}}_n(y)\| - {{\rm R}}_n (y)/\| {{\rm R}}_P(y) \| \big |
\to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \\
\sup_{ y \in K'} | \hat{{\rm D}}^{\rm\scriptsize MK}_{P,n}(y) - {{\rm D}}^{\rm\scriptsize MK}_{P}(y) | \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0. \\
\end{aligned}$$
Uniform convergence of MK empirical $\tau$-quantile contours and MK empirical depth regions with probability content $\tau$ follows also through an application of the Extended Continuous Mapping Theorem.
\[cor:eqd\] Work with the assumptions of Theorem \[theorem: empirical\]. Consider $\mathcal{T} \subset (0,1)$ such that $\mathrm{cl}(\cup_{\tau \in \mathcal{T}} \mathbb{K}(\tau)) \subset {\mathcal{U}}_0$, then $$\begin{aligned}
\quad \sup_{\tau\in \mathcal{T}} d_H(\hat {{\rm Q}}_{n}(\mathbb K(\tau)),{{\rm Q}}_P(\mathbb K(\tau))) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \ \ \sup_{\tau\in \mathcal{T}}d_H(\hat {{\rm Q}}_{n}( \mathcal K(\tau)),{{\rm Q}}_P(\mathcal K(\tau))) & \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0.\end{aligned}$$
The main results are derived assuming we know the reference distribution $F$ and the associated depth function ${{\rm D}}_F$ as well as depth regions $\mathbb{K}(\tau)$ and quantile contours $\mathcal{K}(\tau)$. There are cases where these will be approximated numerically or using data. The same definitions and results extend naturally where these quantities are replaced by uniformly consistent estimators $\hat {{\rm D}}_{F,n}$, $\hat {\mathbb{K}}_n(\tau)$, and $\hat {\mathcal{K}}_n(\tau)$: $$\label{depth}
\begin{array}{rc}
\sup_{ u \in K } | \hat D_{F,n}(u) - D_F(u) | & \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \\
\sup_{\tau\in \mathcal{T}}d_H( \hat {\mathbb K}_n(\tau), \mathbb K(\tau)) & \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \\
\sup_{\tau\in \mathcal{T}}d_H( \hat {\mathcal K}_n(\tau), \mathcal K(\tau)) & \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0,
\end{array}$$ where $K$ is any closed subset of ${\mathcal{U}}_0$. These high-level conditions hold trivially for the numerical approximations we use in Section \[sec:comp\]. They also hold, for example, for Tukey’s halfspace depth under regularity conditions. We will not discuss these conditions here.
\[def:EMK2\]Let $F$ be an absolutely continuous reference distribution with support contained in a convex and compact region ${\mathcal{U}}\subset \Bbb{R}^d$, and let $\| \cdot \|$ be a norm on ${\mathcal{U}}$. Given estimators $\hat {{\rm D}}_{F,n}$, $\hat {\mathbb{K}}_n(\tau)$ and $\hat {\mathcal{K}}_n(\tau)$ satisfying (\[depth\]), the MK empirical quantiles, ranks, signs and depth are defined as follows.
1. The [*MK empirical rank and sign*]{} of $y \in {{\rm I}\kern-0.18em{\rm R}}^d$ are $\|\hat {{\rm R}}_n(y)\|$ and $ \hat {{\rm R}}_n (y)/\|\hat {{\rm R}}_P(y)\|$.
2. The [*MK empirical*]{} $\tau$-[*quantile contour*]{} is the set $\hat {{\rm Q}}_n(\hat{{\mathcal{K}}}_n(\tau))$ and the [*MK empirical depth region*]{} with probability mass $\tau$ is $\hat {{\rm Q}}_n(\hat{{\mathbb{K}}}_n(\tau))$.
3. The [*MK empirical depth*]{} of $y \in{{\rm I}\kern-0.18em{\rm R}}^d$ is the depth of $\hat {{\rm R}}_n(y)$ under $\hat {{{\rm D}}}_{F,n}$: $$\hat{{\rm D}}^{\rm\scriptsize MK}_{P,n}(y):=\hat{{{\rm D}}}_{F,n}(\hat {{\rm R}}_n(y)).$$
\[cor:eqd2\] Work with conditions of the previous corollary and suppose that Conditions (\[depth\]) hold. Then the conclusions of Corollary 3.1 hold and the conclusions of Corollary 3.2 hold in the following form: $$\begin{aligned}
\quad \sup_{\tau\in \mathcal{T}} d_H(\hat {{\rm Q}}_{n}(\hat{{\mathbb{K}}}_n(\tau)),{{\rm Q}}_P(\mathbb K(\tau))) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0, \ \
\sup_{\tau\in \mathcal{T}}d_H(\hat {{\rm Q}}_{n}( \hat{{\mathcal{K}}}_n(\tau)),{{\rm Q}}_P(\mathcal K(\tau))) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0.\end{aligned}$$
Computing Empirical Quantiles and Depth Regions {#sec:comp}
===============================================
Here we provide computational characterizations of the empirical quantiles, ranks, and depth regions for various cases of interest.
Smooth $\hat P_n$ and $\hat F_n$ {#smooth-hat-p_n-and-hat-f_n .unnumbered}
--------------------------------
Suppose $\hat P_n$ and $\hat F_n$ satisfy Caffarelli regularity conditions, so that $\hat{{\rm Q}}_n=\nabla\hat\psi_n$ and $\hat{{\rm R}}_n=\nabla\hat\psi^\ast_n$, with $(\hat\psi_n,\hat\psi^\ast_n)$ satisfying (C). The MK empirical vector quantile maps $\hat{{\rm Q}}_n$ and $\hat{{\rm R}}_n$ can then be computed with the algorithm of Benamou and Brenier [@BB:2000].
Discrete $\hat P_n$ and smooth $\hat F_n$ {#discrete-hat-p_n-and-smooth-hat-f_n .unnumbered}
-----------------------------------------
Suppose now $\hat P_n$ is a discrete estimator of $P$ and $\hat F_n$ an absolutely continuous distribution with convex compact support ${\mathcal{U}}\subset{{\rm I}\kern-0.18em{\rm R}}^d\!$. Let $\hat P_n$ be of the form $\hat P_n\!=\!\sum_{k=1}^{K_n}p_{k,n}\delta_{y_{k,n}}$ for some integer $K_n$, some nonnegative weights $p_{1,n},\ldots,p_{K_n,n}$ such that $\sum_{k=1}^{K_n}p_{k,n}=1$, and $y_{1,n},\ldots,y_{K_n,n}\in{{\rm I}\kern-0.18em{\rm R}}^d$ . The leading example is when $\hat P_n$ is the empirical distribution of a random sample $(Y_i)_{i=1}^n$ drawn from $P$.
The MK empirical vector quantile map $\hat{{\rm Q}}_n$ is then equal (almost everywhere) to the gradient of a convex map $\hat\psi_n$ such that $\nabla\hat\psi_n\#\hat F_n=\hat P_n$, i.e., the $\hat F_n$-almost surely unique map $\hat{{\rm Q}}_n=\nabla\hat\psi_n$ satisfying the following:
1. $\nabla\hat\psi_n(u)\in\{y_{1,n},\ldots,y_{K_n,n}\}$, for Lebesgue-almost all $u\in{\mathcal{U}}$,
2. $\hat F_n\left( \{u\in{\mathcal{U}}:\;\nabla\hat\psi_n(u)=y_{k,n} \} \right)=p_{k,n}$, for each $k\in\{1,\ldots,K_n\}$,
3. $\hat\psi_n$ is a convex function.
The following characterization of $\hat\psi_n$ specializes Kantorovich duality to this discrete-continuous case (see, e.g.,[@EGH]).
There exist unique (up to an additive constant) weights $\{v_1^\ast,\ldots,v_n^\ast\}$ such that $\hat\psi_n(u)=\max_{1\leq k\leq K_n}\{u^\top y_{k,n}-v_k^\ast\}$ satisfies conditions (1), (2) and (3). The function $
v\mapsto\int
\hat\psi_nd\hat F_n+\sum_{k=1}^{K_n}p_{k,n}v_k
$ is convex and minimized at $v^\ast=\{v_1^\ast,\ldots,v_n^\ast\}$.
This lemma allows efficient computation of $\hat{{\rm Q}}_n$ using a gradient algorithm proposed in [@AHA:98]. The map $\hat\psi_n$ is piecewise affine and the empirical vector quantile $\hat{{\rm Q}}_n$ is piecewise constant. The correspondence $\hat{{\rm Q}}_n^{-1}$ defined for each $k\leq K_n$ by $$y_{k,n}\mapsto\hat{{\rm Q}}^{-1}_n(y_{k,n}):=\{u\in{\mathcal{U}}:\;\nabla\hat\psi_n(u)=y_{k,n} \}$$ maps $\{y_{1,n},\ldots,y_{K_n,n}\}$ into $K_n$ regions of a partition of ${\mathcal{U}}$, called a [*power diagram*]{}. The estimator $\hat{{\rm R}}_n$ of the MK vector rank can be computed according to formula (\[eq: empirical Q and R\]) after computing the conjugate $\hat \psi^*_n$ of $\hat \psi_n$ via: $\hat \psi^*_n(y) = \sup_{u \in {\mathcal{U}}} \{ u^\top y - \hat \psi_n(u) \}.$ The empirical depth, depth regions, and quantiles can be computed using the depth function, according to their theoretical definitions.
Discrete $\hat P_n$ and $\hat F_n$ {#discrete-hat-p_n-and-hat-f_n .unnumbered}
----------------------------------
Particularly amenable to computation is the case when both distribution estimators $\hat P_n$ and $\hat F_n$ are discrete with uniformly distributed mass on sets of points of the same cardinality. Let $\hat P_n=\sum_{j=1}^n\delta_{y_j}/n$ for a set $\mathcal Y_n=\{y_1,\ldots,y_n\}$ of points in ${{\rm I}\kern-0.18em{\rm R}}^d$ and $\hat F_n=\sum_{j=1}^n\delta_{u_j}/n$, for a set $\mathcal U_n=\{u_1,\ldots,u_n\}$ of points in ${{\rm I}\kern-0.18em{\rm R}}^d$. The restriction of the quantile map $\hat{{\rm Q}}_n$ to $\mathcal U_n$ is the bijection $u \longmapsto y=\hat{{\rm Q}}_n|_{\mathcal U_n}(u)$ from $\mathcal U_n$ onto $\mathcal Y_n$ and $\hat{{\rm R}}_n|_{\mathcal Y_n}$ is its inverse. The solutions $\hat{{\rm Q}}_n$ and $\hat{{\rm R}}_n$ can be computed with any optimal assignment algorithm. More generally, in the case of any two discrete estimators $\hat P_n$ and $\hat F_n$, the problem of finding $\hat{{\rm Q}}_n$ or $\hat{{\rm R}}_n$ is a linear programming problem.
Visualization of Empirical MK Depth and Quantile Contours {#visualization-of-empirical-mk-depth-and-quantile-contours .unnumbered}
---------------------------------------------------------
Whenever $\hat P_n$ is finitely discrete, then the MK empirical depth regions and quantile contours are finite sets of points. For visualization purposes it may be helpful to transform them into nicer looking objects which are close to the original objects in terms of Hausdorff distance. In the example below we used $\alpha$-hulls to create approximations to the depth regions and took the boundaries of the set as a numerical approximation to the quantile contours. It may also be possible to use polygonization methods such as those in [@DS:88] for $d=2$ and [@Gruenbaum:94] for $d=3$.
In the example illustrated in Figure 2, we use a discrete approximation $\hat F_n$ to the spherical uniform reference distribution. Figure 2 shows the MK empirical depth contours for the same banana-shaped distribution as in Figure 1. The specific construction to produce Figure 2 is the following: $\hat P_n$ is the empirical distribution of a random sample $\mathcal Y_n$ drawn from the banana-shaped distribution in ${{\rm I}\kern-0.18em{\rm R}}^2$, with $n=9999$; $\hat F_n$ is a discrete approximation to $F$ with mass $1/n$ on each of the points in $\mathcal U_n$. The latter is a collection of $99$ evenly spaced points on each of $101$ circles, of evenly spaced radii in $(0,1]$. The sets $\mathcal Y_n$ and $\mathcal U_n$ are matched optimally with the assignment algorithm of the [*adagio*]{} package in R. MK empirical depth regions are $\alpha$-hulls of $\hat{{\rm Q}}_n(\mathcal U_n\cap\mathbb S(\tau))$ for $11$ values of $\tau\in(0,1)$ (see [@EKDS:83] for a definition of $\alpha$-hulls). The $\alpha$-hulls are computed using the [*alphahull*]{} package in R, with $\alpha=0.3$. The banana-shaped distribution considered is the distribution of the vector $(X+R\cos\Phi,X^2+R\sin\Phi)$, where $X$ is uniform on $[-1,1]$, $\Phi$ is uniform on $[0,2\pi]$, $Z$ is uniform on $[0,1]$, $X$, $Z$ and $\Phi$ are independent, and $R=0.2Z(1+(1-|X|)/2)$.
Uniform Convergence of Subdifferentials and Transport Maps
==========================================================
Uniform Convergence of Subdifferentials
---------------------------------------
Let $\mathcal{U}$ and $\mathcal{Y}$ be convex, closed subsets of ${{\rm I}\kern-0.18em{\rm R}}^d$. A pair of convex potentials $\psi: \mathcal{U} \mapsto {{\rm I}\kern-0.18em{\rm R}}\cup \{\infty\}$ and $\psi^*: \mathcal{Y} \mapsto {{\rm I}\kern-0.18em{\rm R}}\cup \{\infty\}$ is a conjugate pair over $({\mathcal{U}}, {\mathcal{Y}})$ if, for each $u \in {\mathcal{U}}$ and $y \in {\mathcal{Y}}$, $$\psi(u) = \sup_{y \in {\mathcal{Y}}} [y^\top u - \psi^*(y)], \quad
\psi^*(y) = \sup_{u \in {\mathcal{U}}} [y^\top u - \psi(u)].$$
In the sequel, we consider a fixed pair $(\psi, \psi^*)$ obeying the following condition.
- Let $\mathcal{U}$ and $\mathcal{Y}$ be closed, convex subsets of ${{\rm I}\kern-0.18em{\rm R}}^d$, and ${\mathcal{U}}_0 \subset {\mathcal{U}}$ and ${\mathcal{Y}}_0 \subset {\mathcal{Y}}$ some open, non-empty sets in ${{\rm I}\kern-0.18em{\rm R}}^d$. Let $\psi: \mathcal{U} \mapsto {{\rm I}\kern-0.18em{\rm R}}$ and $\psi^*: \mathcal{Y} \mapsto {{\rm I}\kern-0.18em{\rm R}}$ form a conjugate pair over $({\mathcal{U}}, {\mathcal{Y}})$ and possess gradients $\nabla \psi(u)$ for all $u \in {\mathcal{U}}_0$, and $\nabla \psi^*(y)$ for all $y \in {\mathcal{Y}}_0$. The gradients $\nabla \psi |_{{\mathcal{U}}_0}: {\mathcal{U}}_0 \mapsto {\mathcal{Y}}_0$ and $\nabla \psi^* |_{{\mathcal{Y}}_0}: {\mathcal{Y}}_0 \mapsto {\mathcal{U}}_0$ are homeomorphisms, and $\nabla \psi|_{{\mathcal{U}}_0} = (\nabla \psi^* |_{{\mathcal{Y}}_0})^{-1}$.
We also consider a sequence $(\psi_n, \psi^*_n)$ of conjugate potentials approaching $(\psi, \psi^*)$.
- A sequence of conjugate potentials $(\psi_n, \psi^*_n)$ over $({\mathcal{U}}, {\mathcal{Y}})$, with $n \in \mathbb{N}$, is such that: $\psi_n(u) \to \psi(u)$ in ${{\rm I}\kern-0.18em{\rm R}}\cup \{\infty\}$ pointwise in $u$ in a dense subset of ${\mathcal{U}}$ and $\psi^*_n(y) \to \psi^*(y)$ in ${{\rm I}\kern-0.18em{\rm R}}\cup \{\infty\}$ pointwise in $y$ in a dense subset of ${\mathcal{Y}}$, as $n \to \infty$.
Condition (A) is equivalent to requiring that either $\psi_n$ or $\psi^*_n$ converge pointwise over dense subsets. There is no loss of generality in stating that both converge.
Define the maps $${{\rm Q}}(u) := \arg \sup_{y \in {\mathcal{Y}}}[ y^\top u - \psi^*(y)] , \quad
{{\rm R}}(y) := \arg \sup_{u \in {\mathcal{U}}} [y^\top u - \psi(u)],$$ for each $u \in {\mathcal{U}}_0$ and $y \in {\mathcal{Y}}_0$. By the envelope theorem, $${{\rm R}}(y) = \nabla \psi^*(y), \text{ for } y \in \ {\mathcal{Y}}_0; \ \ \
{{\rm Q}}(u) = \nabla \psi(u), \text{ for } u \in \ {\mathcal{U}}_0.$$ Let us define, for each $u \in {\mathcal{U}}$ and $y \in {\mathcal{Y}}$, $$\label{define Qn and Rn}
{{\rm Q}}_n(u) \in \arg \sup_{y \in {\mathcal{Y}}} [y^\top u - \psi_n^*(y)] , \quad
{{\rm R}}_n(y) \in \arg \sup_{u \in {\mathcal{U}}} [y^\top u - \psi_n(u)].$$
It is useful to note that $${{\rm R}}_n(y) \in \partial \psi^*_n(y) \text{ for } y \in {\mathcal{Y}}; \ \ {{\rm Q}}_n(u) \in \partial \psi_n(u) \text{ for } u \in {\mathcal{U}},$$ where $\partial$ denotes the sub-differential of a convex function; conversely, any pair of elements of $\partial \psi^*_n(y)$ and $\partial \psi_n(u)$, respectively, could be taken as solutions to the problem (\[define Qn and Rn\]) (by Proposition 2.4 in Villani [@villani1]). Hence, the problem of convergence of ${{\rm Q}}_n$ and ${{\rm R}}_n$ to ${{\rm Q}}$ and ${{\rm R}}$ is equivalent to the problem of convergence of subdifferentials. Moreover, by Rademacher’s theorem, $\partial \psi^*_n(y) = \{ \nabla \psi_n^*(y) \} $ and $\partial \psi_n(u) = \{ \nabla \psi_n(u) \}$ almost everywhere with respect to the Lebesgue measure (see, e.g., [@villani1]), so the solutions to (\[define Qn and Rn\]) are unique almost everywhere on $u \in {\mathcal{U}}$ and $y \in {\mathcal{Y}}$.
\[lemma: ucg\] Suppose that Conditions (A) and (C) hold. Then, as $n \to \infty$, for any compact set $K \subset {\mathcal{U}}_0$ and any compact set $K' \subset {\mathcal{Y}}_0$, $$\sup_{ u \in K} \| {{\rm Q}}_n(u) - {{\rm Q}}(u) \| \to 0, \quad
\sup_{ y \in K'} \| {{\rm R}}_n(y) - {{\rm R}}(y) \| \to 0.$$
This result appears to be new. It complements the result stated in Lemma 5.4 in Villani [@villani:stability] for the case ${\mathcal{U}}_0 = {\mathcal{U}}={\mathcal{Y}}_0 = {\mathcal{Y}}= {{\rm I}\kern-0.18em{\rm R}}^d$. This result also trivially implies convergence in $L^p$ norms, $1 \leq p < \infty$: $$\int_{{\mathcal{U}}} \| {{\rm Q}}_n(u) - {{\rm Q}}(u)\|^p d F(u) \to 0, \quad \int_{{\mathcal{Y}}} \| {{\rm R}}_n(y) - {{\rm R}}(y)\|^p d P (y) \to 0,$$ for probability laws $F$ on ${\mathcal{U}}$ and $P$ on ${\mathcal{Y}}$, whenever, for some $\bar p>p$, $$\sup_{n \in \mathbb{N}} \int_{{\mathcal{U}}} \| {{\rm Q}}_n(u)\|^{\bar p} +\|{{\rm Q}}(u)\|^{p} d F(u) < \infty, \quad
\sup_{n \in \mathbb{N}} \int_{{\mathcal{Y}}} \| {{\rm R}}_n(y)\|^{\bar p} +\|{{\rm R}}(y)\|^{p} d P(y) < \infty.$$ Hence, the new result is stronger than available results on convergence in measure (including $L^p$ convergence results) in the optimal transport literature (see, e.g., Villani [@villani1; @villani2]).
The following example also shows that, in general, our result cannot be strengthened to the uniform convergence over entire sets ${\mathcal{U}}$ and ${\mathcal{Y}}$. Consider the sequence of potential maps $\psi_n: {\mathcal{U}}= [0,1] \mapsto {{\rm I}\kern-0.18em{\rm R}}$: $$\psi_n(u) = \int_0^u {{\rm Q}}_n(t) dt, \ \ {{\rm Q}}_n(t) = t \cdot 1( t \leq 1-1/n) + 10 \cdot 1( t > 1 -1/n).$$ Then $
\psi_n(u) = 2^{-1}u^2 1( u \leq 1-1/n) + \big\{
10(u -(1-1/n)) + 2^{-1}(1-1/n)^2 \big\}1(u> 1- 1/n) $ converges uniformly on $[0,1]$ to $
\varphi(u) = 2^{-1} u^2.$ The latter potential has the gradient map ${{\rm Q}}: [0,1] \mapsto {\mathcal{Y}}_0 = [0,1]$ defined by ${{\rm Q}}(t) = t$. We have that $
\sup_{t \in K}|{{\rm Q}}_n(t) - {{\rm Q}}(t)| \to 0
$ for any compact subset $K$ of $(0,1)$. However, the uniform convergence over the entire region $[0,1]$ fails, since $\sup_{t \in [0,1]}|{{\rm Q}}_n(t) - {{\rm Q}}(t)| \geq 9$ for all $n$. Therefore, the theorem cannot be strengthened in general.
We next consider the behavior of image sets of gradients defined as follows: $${{\rm Q}}_{n}(A) := \{{{\rm Q}}_{n}(u): u \in A\}, \quad {{\rm Q}}(A) := \{{{\rm Q}}(u): u \in A\}, \quad A \subseteq K,$$$$\mathrm{R}_{n}(A') := \{ {{\rm R}}_{n}(y): y \in A'\}, \quad {{\rm R}}(A') := \{ {{\rm R}}(y): y \in A'\}, \quad A' \subseteq K',$$ where $K \subset \ {\mathcal{U}}_0$ and $K' \subset \ {\mathcal{Y}}_0$ are compact sets, and the subsets $A$ and $A'$ are understood to be non-empty.
\[cor: csg\] Under the conditions of the previous theorem, we have that $$\sup_{A \subseteq K}d_H({{\rm Q}}_{n}(A),{{\rm Q}}(A)) \to 0, \quad
\sup_{A ' \subseteq K'} d_H({{\rm R}}_{n}(A'), {{\rm R}}(A')) \to 0.$$
\[cor: csg2\] Assume the conditions of the previous theorem. For any sequence of sets $\{ A_n\} \subseteq K$ and $\{A'_n \} \subseteq K'$ such that $d_H(A_n, A) \to 0$ and $d_H(A_n', A')\to 0$ for some sets $A$ and $A'$, we have $$d_H({{\rm Q}}_{n}(A_n),{{\rm Q}}(A)) \to 0, \quad
d_H({{\rm R}}_{n}(A'_n), {{\rm R}}(A')) \to 0.$$
Uniform Convergence of Transport Maps
-------------------------------------
We next consider the problem of convergence for potentials and transport (vector quantile and rank) maps arising from the Kantorovich dual optimal transport problem.
Equip ${\mathcal{Y}}$ and ${\mathcal{U}}$ with absolutely continuous probability measures $P$ and $F$, respectively, and let $${\mathcal{Y}}_0 := \text{int}(\text{support}(P)),\qquad {\mathcal{U}}_0:= \text{int}(\text{support}(F)).$$ We consider sequences of measures $P_n$ and $F_n$ approximating $P$ and $F$:
- There are sequences of measures $\{P_n\}_{n \in \mathbb{N}}$ on ${\mathcal{Y}}$ and $\{F_n\}_{n \in \mathbb{N}}$ on ${\mathcal{U}}$, with finite total mass, that converge to $P$ and $F$, respectively, in the topology of weak convergence: $$d_{{{\rm BL}}} (P_n, P) \to 0, \ \quad d_{{{\rm BL}}} (F_n, F) \to 0 .$$
Recall that we defined $\Phi_0({\mathcal{U}}, {\mathcal{Y}})$ as a collection of conjugate potentials $({\varphi}, {\varphi}^*)$ on $({\mathcal{U}}, {\mathcal{Y}})$ such that ${\varphi}(u_0)= 0$ for some fixed point $u_0 \in {\mathcal{U}}_0$. Let $(\psi_n, \psi^*_n) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}})$ solve the Kantorovich problem for the pair $(P_n,F_n)$: $$\label{empirical kantorovich}
\int \psi_n dF_n + \int \psi_n^* d P_n= \inf_{ ({\varphi}, {\varphi}^*) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}}) } \int \varphi dF_n + \int \varphi^* dP_n.$$ Also, let $(\psi, \psi^*) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}})$ solve the Kantorovich problem for the pair $(P,F)$: $$\label{kantorovich}
\int \psi dF + \int \psi^* d P = \inf_{ ({\varphi}, {\varphi}^*) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}}) } \int \varphi dF + \int \varphi^* dP .$$ It is known that solutions to these problems exist; see, e.g., Villani [@villani1]. Recall also that we imposed the normalization condition in the definition of $\Phi_0({\mathcal{U}}, {\mathcal{Y}})$ to pin down the constants.
\[theorem: uct\] Suppose that the sets ${\mathcal{U}}$ and ${\mathcal{Y}}$ are compact subsets of ${{\rm I}\kern-0.18em{\rm R}}^d$, and that the probability measures $P$ and $F$ are absolutely continuous with respect to the Lebesgue measure, with ${\rm support }(P) \subseteq {\mathcal{Y}}$ and ${\rm support }(F) \subseteq {\mathcal{U}}$. Suppose that Condition (W) holds, and that Condition (C) holds for a solution $(\psi, \psi^*)$ of (\[kantorovich\]) for the sets ${\mathcal{U}}_0$ and ${\mathcal{Y}}_0$ defined as above. Then the conclusions of Theorem \[lemma: ucg\] and Corollary \[cor: csg\] and Corollary \[cor: csg2\]. hold.
Proofs
======
Proof of Theorem \[lemma: ucg\]
-------------------------------
The proof relies on the equivalence of the uniform and continuous convergence.
\[Lemma: Resnick\] Let $\mathbb{D}$ and $\mathbb{E}$ be complete separable metric spaces, with $\mathbb{D}$ compact. Suppose $f: \mathbb{D}
\mapsto\mathbb{E}$ is continuous. Then a sequence of functions $f_{n}:
\mathbb{D} \mapsto\mathbb{E}$ converges to $f$ uniformly on $\mathbb{D}$ if and only if, for any convergent sequence $x_{n} \to x$ in $\mathbb{D}$, we have that $f_{n}(x_{n})\to f(x)$.
For the proof, see, e.g., Rockafellar and Wets [@RW]. The proof also relies on the following convergence result, which is a consequence of Theorem 7.17 in Rockafellar and Wets [@RW]. For a point $a$ and a non-empty set $A$ in $\Bbb{R}^d$, define $d(a,A) := \inf_{a' \in A} \| a- a'\|$.
\[lemma: convex\] Suppose that $g$ is a lower-semi-continuous convex function mapping ${{\rm I}\kern-0.18em{\rm R}}^d$ to ${{\rm I}\kern-0.18em{\rm R}}\cup \{+ \infty\}$ that attains a minimum on the set ${\mathcal{X}}_0 = \arg\inf_{x \in {{\rm I}\kern-0.18em{\rm R}}^d} g(x)\subset \mathrm{int}({\mathcal{D}}_0)$, where ${\mathcal{D}}_0 = \{x \in {{\rm I}\kern-0.18em{\rm R}}^d: g(x) < \infty \}$, and $ \mathrm{int}({\mathcal{D}}_0)$ is a non-empty, open set in ${{\rm I}\kern-0.18em{\rm R}}^d$. Let $\{g_n\}$ be a sequence of convex, lower-semi-continuous functions mapping ${{\rm I}\kern-0.18em{\rm R}}^d$ to ${{\rm I}\kern-0.18em{\rm R}}\cup \{+ \infty\}$ and such that $g_n(x) \to g(x)$ pointwise in $x \in {{\rm I}\kern-0.18em{\rm R}}_0^d$, where ${{\rm I}\kern-0.18em{\rm R}}_0^d$ is a countable dense subset of ${{\rm I}\kern-0.18em{\rm R}}^d$. Then any $x_n \in \arg\inf_{x \in {{\rm I}\kern-0.18em{\rm R}}^d} g_n(x)$ obeys $$d(x_n, {\mathcal{X}}_0) \to 0,$$ and, in particular, if ${\mathcal{X}}_0$ is a singleton $\{x_0\}$, $x_n \to x_0$.
The proof of this lemma is given below, immediately after the conclusion of the proof of this theorem.
We define the extension maps $y \mapsto g_{n,u}(y)$ and $u \mapsto \bar g_{n,y}(u)$ mapping ${{\rm I}\kern-0.18em{\rm R}}^d$ to ${{\rm I}\kern-0.18em{\rm R}}\cup \{-\infty\}$ $$g_{n,u} (y) := \left \{ \begin{array}{cc} y^\top u - \psi^*_n(y) & \text{ if }
y \in {\mathcal{Y}}\\
- \infty & \text{ if }
y \not \in {\mathcal{Y}}\end{array} \right., \quad
\bar g_{n,y} (u) := \left \{ \begin{array}{cc} y^\top u - \psi_n(u) & \text{ if }
u \in {\mathcal{U}}\\
- \infty & \text{ if }
u \not \in {\mathcal{U}}.
\end{array} \right.$$ By the convexity of $\psi_n$ and $\psi^*_n$ over convex, closed sets ${\mathcal{Y}}$ and ${\mathcal{U}}$, we have that the functions are proper upper-semi-continuous concave functions. Define the extension maps $y \mapsto g_{u}(y)$ and $u \mapsto \bar g_{y}(u)$ mapping ${{\rm I}\kern-0.18em{\rm R}}^d$ to ${{\rm I}\kern-0.18em{\rm R}}\cup \{-\infty\}$ analogously, by removing the index $n$ above.
Condition (A) assumes pointwise convergence of $\psi^*_n$ to $\psi^*$ on a dense subset of ${\mathcal{Y}}$. By Theorem 7.17 in Rockafellar and Wets [@RW], this implies the uniform convergence of $\psi^*_n$ to $\psi^*$ on any compact set $K' \subset \text{ int } {\mathcal{Y}}$ that does not overlap with the boundary of the set ${\mathcal{D}}_1 = \{ y \in {\mathcal{Y}}: \psi^*(y) < + \infty\}$. Hence, for any sequence $\{u_n\}$ such that $u_n \to u \in K$, a compact subset of $ {\mathcal{U}}_0$, and any $y \in (\text{int } {\mathcal{Y}}) \setminus \partial {\mathcal{D}}_1$, $$g_{n,u_n} (y) = y^\top u_n - \psi^*_n(y) \to g_{u}(y) = y^\top u - \psi^*(y).$$ Next, consider any $y \not \in {\mathcal{Y}}$, in which case, $
g_{n,u_n}(y) = -\infty \to g_u(y) = - \infty.
$ Hence, $$g_{n,u_n}(y) \to g_u(y) \text{ in } {{\rm I}\kern-0.18em{\rm R}}\cup \{ - \infty \} , \text{ for all } y \in {{\rm I}\kern-0.18em{\rm R}}^d_1 = {{\rm I}\kern-0.18em{\rm R}}^d \setminus (\partial {\mathcal{Y}}\cup \partial {\mathcal{D}}_1),$$ where ${{\rm I}\kern-0.18em{\rm R}}^d_1$ is a dense subset of ${{\rm I}\kern-0.18em{\rm R}}^d$. We apply Lemma \[lemma: convex\] to conclude that $$\arg\sup_{y \in {{\rm I}\kern-0.18em{\rm R}}^d} g_{n,u_n}(y) \ni {{\rm Q}}_n(u_n) \to {{\rm Q}}(u) \in \arg\sup_{y \in {{\rm I}\kern-0.18em{\rm R}}^d} g_{u}(y) = \{ \nabla \psi(u) \}.$$
Take $K$ as any compact subset of $ {\mathcal{U}}_0$. The above argument applies for every point $u \in~\!K$ and every convergent sequence $u_n \to u$. Therefore, since by Assumption (C) the map $u \mapsto Q(u) = \nabla \psi (u)$ is continuous in $u \in K$, we conclude by the equivalence of the continuous and uniform convergence, Lemma \[Lemma: Resnick\], that $${{\rm Q}}_n(u) \to {{\rm Q}}(u) \text{ uniformly in } u \in K.$$
By symmetry, the proof of the second claim is identical to the proof of the first one.
Proof of Lemma \[lemma: convex\]
--------------------------------
By assumption, ${\mathcal{X}}_0 = \arg \min g \subset \mathrm{int}({\mathcal{D}}_0)$, and ${\mathcal{X}}_0$ is convex and closed. Let $x_0$ be an element of ${\mathcal{X}}_0$. We have that, for all $ 0< {\varepsilon}\leq {\varepsilon}_0$ with ${\varepsilon}_0$ such that $B_{{\varepsilon}_0}({\mathcal{X}}_0) \subset \mathrm{int}({\mathcal{D}}_0)$, $$\label{g bound}
g(x_0) < \inf_{x \in \partial B_{{\varepsilon}}({\mathcal{X}}_0)} g(x),$$ where $B_{{\varepsilon}} ({\mathcal{X}}_0) := \{x \in {{\rm I}\kern-0.18em{\rm R}}^d: d(x, {\mathcal{X}}_0) \leq {\varepsilon}\}$ is convex and closed.
Fix an ${\varepsilon}\in (0, {\varepsilon}_0]$. By convexity of $g$ and $g_n$ and by Theorem 7.17 in Rockafellar and Wets [@RW], the pointwise convergence of $g_n$ to $g$ on a dense subset of ${{\rm I}\kern-0.18em{\rm R}}^d$ is equivalent to the uniform convergence of $g_n$ to $g$ on any compact set $K$ that does not overlap with $\partial {\mathcal{D}}_0$, i.e. $K \cap \partial {\mathcal{D}}_0 = \emptyset$. Hence, $g_n \to g$ uniformly on $B_{{\varepsilon}_0}({\mathcal{X}}_0)$. This and (\[g bound\]) imply that eventually, i.e. for all $n \geq n_{\varepsilon}$, $$g_n(x_0) < \inf_{x \in \partial B_{\varepsilon}({\mathcal{X}}_0)} g_n(x).$$ By convexity of $g_n$, this implies that $g_n(x_0) < \inf_{x \not \in B_{\varepsilon}({\mathcal{X}}_0)} g_n(x)$ for all $n \geq n_{\varepsilon}$, which is to say that, for all $n \geq n_{\varepsilon}$, $$\arg\inf g_n = \arg\min g_n \subset B_{\varepsilon}({\mathcal{X}}_0).$$ Since ${\varepsilon}>0$ can be set as small as desired, it follows that any $x_n \in \arg \inf g_n$ is such that $d(x_n, {\mathcal{X}}_0) \to 0$.
Proof of Corollary \[cor: csg\]
-------------------------------
By Theorem \[lemma: ucg\] and the definition of Hausdorff distance, for $A$ denoting non-empty subsets, $$\begin{aligned}
&& \sup_{A \subseteq K}d_H({{\rm Q}}_{n}(A), {{\rm Q}}(A)) \\
&& = \sup_{A \subseteq K} \left ( \sup_{u \in A}\inf_{\bar u \in A}\| {{\rm Q}}_n(\bar u) - {{\rm Q}}(u) \|
\vee \sup_{\bar u\in A} \inf_{u \in A} \| {{\rm Q}}_n(\bar u) - {{\rm Q}}(u) \| \right) \\
&& \leq \sup_{A \subseteq K} \left(
\sup_{u \in A} \| {{\rm Q}}_n(u) - {{\rm Q}}(u) \| \vee \sup_{\bar u \in A} \| {{\rm Q}}_n(\bar u) - {{\rm Q}}(\bar u) \|
\right) \\
&& = \sup_{ u \in K} \| {{\rm Q}}_n(u) - {{\rm Q}}(u) \| \to 0.
\end{aligned}$$ The proof of the second claim is identical.
Proof of Corollary \[cor: csg2\]
--------------------------------
We have that $$\begin{aligned}
d_H({{\rm Q}}_n(A_n), {{\rm Q}}(A)) \!\!\!\! \!&\!\leq\!&\!\! \! d_H({{\rm Q}}_n(A_n), {{\rm Q}}(A_n)) + d_H({{\rm Q}}(A_n), {{\rm Q}}(A))\\ \! \!&\leq\!&\! \!\! \sup_{A \subseteq K} \!d_H({{\rm Q}}_n(A), Q(A)) +\! \sup_{\bar u, u \in K}\! \{{{\rm Q}}(\bar u) - {{\rm Q}}(u)\!: \|\bar u - u\| \leq d_H(A_n, A) \!\} \\ &\to& 0,\end{aligned}$$ where the first inequality holds by the triangle inequality, the second inequality holds by definition and by $A_n, A \subseteq K$, and the last conclusion follows by Corollary \[cor: csg\] and continuity of the map $u \longmapsto Q(u)$ on $u \in K$.
The proof of the second claim is identical.
Proof of Theorem \[theorem: uct\]
---------------------------------
<span style="font-variant:small-caps;">Step 1.</span> Here we show that the set of conjugate pairs is compact in the topology of uniform convergence. First we notice that, for any pair $({\varphi}, {\varphi}^*) \in \Phi_0({\mathcal{U}}, {\mathcal{Y}})$, $$\begin{aligned}
\| \varphi \|_{{{\rm BL}}({\mathcal{U}})} & \leq & (2 \|{\mathcal{Y}}\| \|{\mathcal{U}}\|) \vee \|{\mathcal{Y}}\| < \infty, \ \
\| \varphi^* \|_{{{\rm BL}}({\mathcal{Y}})} \leq (2 \|{\mathcal{Y}}\| \|{\mathcal{U}}\|) \vee \|{\mathcal{U}}\| < \infty,
\end{aligned}$$ with $\|A\| := \sup_{a \in A} \|a\|$ for $A \subseteq {{\rm I}\kern-0.18em{\rm R}}^d$, where we have used the fact that ${\varphi}(u_0) = 0$ for some $u_0 \in {\mathcal{U}}$ as well as compactness of ${\mathcal{Y}}$ and ${\mathcal{U}}$.
The Arzela-Ascoli Theorem implies that $\Phi_0({\mathcal{U}}, {\mathcal{Y}})$ is relatively compact in the topology of the uniform convergence. We want to show compactness, namely that this set is also closed. For this we need to show that all uniformly convergent subsequences $({\varphi}_n, {\varphi}_n^*)_{n \in \mathbb{N}'}$ (where $\mathbb{N}' \subseteq \mathbb{N}$) have the limit point in this set: $$({\varphi}, {\varphi}^*) := \lim_{n \in \mathbb{N}'} ({\varphi}_n, {\varphi}_n^*) \in \Phi_0( {\mathcal{U}}, {\mathcal{Y}}).$$ This is true, since uniform limits of convex functions are necessarily convex (see [@RW]), and since $$\begin{aligned}
{\varphi}(u) & = & \lim_{n \in \mathbb{N}'} \left [ \sup_{y \in {\mathcal{Y}}} [u^\top y - {\varphi}^*_n(y)] \right ] \\
& \leq & \limsup_{n \in \mathbb{N}'} \left [ \sup_{y \in {\mathcal{Y}}} [u^\top y - {\varphi}^*(y)] + \sup_{y \in {\mathcal{Y}}} | {\varphi}^*_n(y) - {\varphi}^*(y) | \right ] = \sup_{y \in {\mathcal{Y}}} [u^\top y - \varphi^*(y)];\end{aligned}$$ and $$\begin{aligned}
{\varphi}(u) & = & \lim_{n \in \mathbb{N}'} \left [ \sup_{y \in {\mathcal{Y}}} [u^\top y - {\varphi}^*_n(y)] \right ] \\
& \geq & \liminf_{n \in \mathbb{N}'} \left [ \sup_{y \in {\mathcal{Y}}} [u^\top y - {\varphi}^*(y)] - \sup_{y \in {\mathcal{Y}}} | {\varphi}^*_n(y) - {\varphi}^*(y) | \right ] = \sup_{y \in {\mathcal{Y}}} [u^\top y - \varphi^*(y)].\end{aligned}$$ Analogously, $\varphi^*(y) = \sup_{u \in {\mathcal{U}}} [u^\top y - \varphi(y)]$.
<span style="font-variant:small-caps;">Step 2.</span> The claim here is that $$\label{value}
I_n: = \int \psi_n d F_n + \int \psi_n^* d P_n \to_{n \in \mathbb{N}} \int \psi dF + \int \psi^* dP =: I_0.$$ Indeed, $$I_n \leq \int \psi d F_n + \int \psi^* d P_n \to_{n \in \mathbb{N}} I_0,$$ where the inequality holds by definition, and the convergence holds by $$\left | \int \psi d (F_n - F) \right | + \left |\int \psi^* d(P_n - P) \right | \lesssim d_{{{\rm BL}}} (F_n, F) + d_{{{\rm BL}}} (P_n, P) \to 0,$$ where $x \lesssim y$ means $x \leq A y$, for some constant $A$ that does not depend on $n$. Moreover, by definition, $$II_n:= \int \psi_n dF + \int \psi^*_n d P \geq I_0,$$ but $$\left | I_n - II_n\right | \leq \left | \int \psi_n d (F_n - F)\right | + \left | \int \psi_n^* d(P_n - P) \right |
\lesssim d_{{{\rm BL}}} (F_n, F) + d_{{{\rm BL}}} (P_n, P) \to 0.$$
<span style="font-variant:small-caps;">Step 3.</span> Here we conclude.
First, we observe that the solution pair $(\psi, \psi^*)$ to the limit Kantorovich problem is unique on ${\mathcal{U}}_0\times{\mathcal{Y}}_0$ in the sense that any other solution $({\varphi}, {\varphi}^*)$ agrees with $(\psi, \psi^*)$ on ${\mathcal{U}}_0\times{\mathcal{Y}}_0$. Indeed, suppose that ${\varphi}(u_1) \neq \psi(u_1)$ for some $ u_1 \in {\mathcal{U}}_0$. By the uniform continuity of elements of $\Phi_0({\mathcal{U}}, {\mathcal{Y}})$ and openness of ${\mathcal{U}}_0$, there exists a ball $B_{\varepsilon}(u_1) \subset {\mathcal{U}}_0$ such that $\psi(u) \neq {\varphi}(u)$ for all $ u \in B_{\varepsilon}(u_1)$. By the normalization assumption ${\varphi}(u_0) = \psi(u_0)=0$, there does not exist a constant $c \neq 0$ such that $\psi(u) = {\varphi}(u) + c$ for all $u \in {\mathcal{U}}_0$, so this must imply that $\nabla \psi (u) \neq \nabla {\varphi}(u)$ on a set $K \subset {\mathcal{U}}_0$ of positive measure (otherwise, if they disagree only on a set of measure zero, we would have $\psi(u) - \psi(u_0) = \int_0^1 \nabla \psi(u_0 + v^{\top}(u-u_0))^\top (u-u_0) d v = \int_0^1 \nabla {\varphi}(u_0 + v^{\top}(u-u_0)) ^\top(u-u_0) d v
= {\varphi}(u)- {\varphi}(u_0)$ for almost all $u \in B_{\varepsilon}(u_1)$, which is a contradiction). However, the statement $\nabla \psi \neq \nabla {\varphi}$ on a set $K \subset {\mathcal{U}}_0$ of positive Lebesgue measure would contradict the fact that any solution $\psi$ or ${\varphi}$ of the Kantorovich problem must obey $$\int h \circ \nabla {\varphi}d F = \int h \circ \nabla \psi d F = \int h dP ,$$ for each bounded continuous $h$, i.e. that $\nabla {\varphi}\# F = \nabla \psi \# F = P$, established on p.72 in Villani [@villani1]. Analogous arguments apply to establish uniqueness of $\psi^*$ on the set ${\mathcal{Y}}_0$.
Second, we can split $\mathbb{N}$ into subsequences $\mathbb{N} = \cup_{j=1}^\infty \mathbb{N}_{j}$ such that, for each $j$, $$\label{unif}
(\psi_n, \psi^*_n) \to_{n \in \mathbb{N}_j} (\varphi_j, \varphi^*_j) \in \Phi_0( {\mathcal{U}}, {\mathcal{Y}}), \ \text{ uniformly on ${\mathcal{U}}\times {\mathcal{Y}}$. }$$ But by Step 2 this means that $$\int \varphi_j d F + \int \varphi^*_j dP = \int \psi dF + \int \psi^* dP.$$ It must be that each pair $({\varphi}_j, {\varphi}_j^*)$ is the solution to the limit Kantorovich problem, and by the uniqueness established above we have that $$({\varphi}_j, {\varphi}_j^*) = (\psi, \psi^*) \text{ on } {\mathcal{U}}_0 \times {\mathcal{Y}}_0.$$ By Condition (C) we have that, for $u \in {\mathcal{U}}_0$ and $y \in {\mathcal{Y}}_0$: $${{\rm Q}}(u) = \nabla \psi(u) = \nabla {\varphi}_j(u), \quad {{\rm R}}(u) = \nabla \psi^*(u) = \nabla \varphi^*_j(u).$$ By (\[unif\]) and Condition (C) we can invoke Theorem \[lemma: ucg\] to conclude that ${{\rm Q}}_n \to {{\rm Q}}$ uniformly on compact subsets of ${\mathcal{U}}_0$ and ${{\rm R}}_n \to {{\rm R}}$ uniformly on compact subsets of ${\mathcal{Y}}_0$.
Proof of Lemma \[ergodic\]
--------------------------
The proof is a variant of standard arguments, for example, those given in ([@PRW:1999], proof of Theorem 2.1), so is delegated to the Supplemental Appendix.
Proof of Theorem \[theorem: empirical\]
---------------------------------------
The proof is an immediate consequence of the Extended Continuous Mapping Theorem, as given in van der Vaart and Wellner [@VvW], Theorem \[lemma: ucg\] and Corollary \[cor: csg\].
The theorem, specialized to our context, reads as follows: Let $\mathbb{D}$ and $\mathbb{E} $ be normed spaces and let $x \in \mathbb{D}$. Let $\mathbb{D}_n \subseteq \mathbb{D}$ be arbitrary subsets and $g_n: \mathbb{D}_n \mapsto \mathbb{E}$ be arbitrary maps ($n \geq 0$), such that for every sequence $x_n \in
\mathbb{D}_n$ such that $x_{n} \to x$, along a subsequence, we have that $g_{n}(x_{n}) \to g_0(x)$, along the same subsequence. Then, for arbitrary (i.e. possibly non-measurable) maps $X_n: \Omega \mapsto \mathbb{D}_n$ such that $X_n \to_{{{\rm I}\kern-0.18em{\rm P}}^*} x$, we have that $g_n(X_n) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} g_0(x)$.
In our case $X_n = (\hat P_n, \hat F_n)$ is a stochastic element of $\mathbb{D}$, viewed as an arbitrary map from $\Omega$ to $\mathbb{D}$, and $x = (P, F)$ is a non-stochastic element of $\mathbb{D}$, where $\mathbb{D}$ is the space of linear operators $\mathbb{D}$ acting on the space of bounded Lipschitz functions. This space can be equipped with the norm (see notation section): $$\|\cdot\|_{\mathbb{D}}:
\| (x_1, x_2) \|_{\mathbb{D}} = \| x_1 \|_{{{\rm BL}}({\mathcal{Y}})} \vee \| x_2 \|_{{{\rm BL}}({\mathcal{U}})}.$$ Moreover, $X_n \to_{{{\rm I}\kern-0.18em{\rm P}}^*} x$ with respect to this norm, i.e. $$\| X_n - x\|_{\mathbb{D}}:= \| \hat P_n - P \|_{{{\rm BL}}({\mathcal{Y}})} \vee \|\hat F_n-F\|_{{{\rm BL}}({\mathcal{U}})} \to_{{{\rm I}\kern-0.18em{\rm P}}^*} 0.$$ Then $g_n(X_n) := (\hat {{\rm Q}}_n, \hat {{\rm R}}_n)$ and $g (x) := ({{\rm Q}}, {{\rm R}})$ are viewed as elements of the space$\mathbb{E}= \ell^\infty (K\times K', {{\rm I}\kern-0.18em{\rm R}}^d \times {{\rm I}\kern-0.18em{\rm R}}^d)$ of bounded functions mapping $K\times K'$ to ${{\rm I}\kern-0.18em{\rm R}}^d \times {{\rm I}\kern-0.18em{\rm R}}^d$, equipped with the supremum norm. The maps have the continuity property: if $\|x_n - x\|_{\mathbb{D}} \to 0$ along a subsequence, then $\|g_n(x_n) - g(x)\|_{\mathbb{E}} \to 0$ along the same subsequence, as established by Theorem \[lemma: ucg\]. Hence conclude that $g_n(X_n) \to_{{{\rm I}\kern-0.18em{\rm P}}^*} g(x)$.
The second claim follows by the Extended Continuous Mapping Theorem and Corollary \[cor: csg\].
Proof of Corollaries \[cor:ERS\], \[cor:eqd\], and \[cor:eqd2\]
---------------------------------------------------------------
Corollaries \[cor:ERS\] and \[cor:eqd\] follow by Theorem \[theorem: empirical\] and the Extended Continuous Mapping Theorem; and Corollary \[cor:eqd2\] follows by Theorem \[theorem: empirical\], the Extended Continuous Mapping Theorem, and Corollary \[cor: csg2\].
[99]{}
Agostinelli, C., and Romanazzi, M. (2011). Local depth, [*Journal of Statistical Planning and Inference*]{} [**141**]{}, 817-–830.
Aurenhammer, F., Hoffmann, F., and Aronov, B. (1998). Minkowski-type theorems and mean-square clustering, *Algorithmica* [**20**]{}, 61–76.
Benamou, J.-D., Brenier, Y. (2000). A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem. [*Numerische Mathematik*]{} [**84**]{}, 375–393.
Brenier, Y. (1991). Polar factorization and monotone rearrangement of vector-valued functions, [*Communications in Pure and Applied Mathematics*]{} [**44**]{}, 375–417.
Carlier, G., Chernozhukov, V., and Galichon, A. (2014). Vector quantile regression. ArXiv preprint arXiv:1406.4643.
Chaudhuri, P. (1996). On a geometric notion of quantiles for multivariate data, [*Journal of the American Statistical Association*]{} [**91**]{}, 862–-872.
Chen, Y., Dang, X., Peng, H., and Bart, H. L. J. (2009). Outlier detection with the kernelized spatial depth function, [*IEEE Transactions on Pattern Analysis and Machine Intelligence*]{} [**31**]{}, 288-–305.
Cuesta-Albertos, J., and Nieto-Reyes, A. (2008). The random Tukey depth, [*Computational Statistics and Data Analysis*]{} [**52**]{}, 4979–-4988.
Decurninge, A. (2014). Multivariate quantiles and multivariate $L$-moments, ArXiv preprint arXiv:1409.6013
Deneen, L., and Shute, G. (1988). Polygonization of point sets in the plane, *Discrete and Computational Geometry* [**3**]{}, 77-87.
Doksum, K.A. (1974). Empirical probability plots and statistical inference for nonlinear models in the two-sample case, [*Annals of Statistics*]{} [**2**]{}, 267-277.
Doksum, K.A. and Sievers, G.L. (1976). Plotting with confidence: Graphical comparison of two populations, [*Biometrika*]{} [**63**]{}, 421-434.
Donoho, D. L. (1982). Breakdown properties of multivariate location estimators, Qualifying Paper, Harvard University.
Donoho, D. L., and Gasko, M. (1992). Breakdown properties of location estimates based on halfspace depth and projected outlyingness, [*The Annals of Statistics*]{} [**20**]{}, 1803–1827.
Dutta, S., Ghosh, A. K., and Chaudhuri, P. (2011). Some intriguing properties of Tukey’s halfspace depth, [*Bernoulli*]{} [**17**]{}, 1420–-1434.
Edelsbrunner, H., Kirkpatrick, D., and Seidel, R. (1983). On the shape of a set of points in the plane, [*IEEE Transactions on Information Theory*]{} [**29**]{}, 551–-559.
Ekeland, I., Galichon, A., and Henry, M. (2012). Comonotonic measures of multivariate risks, [*Mathematical Finance*]{} [**22**]{}, 109–132.
Galichon, A., and Henry, M. (2012). Dual theory of choice under multivariate risk, [*Journal of Economic Theory*]{} [**147**]{}, 1501–1516.
Ghosh, A. K., and Chaudhuri, P. (2005). On maximum depth and related classifiers, [*Scandinavian Journal of Statistics*]{} [**32**]{}, 327-–350.
Gruenbaum, B. (1994). Hamiltonian polygons and polyhedra, *Geombinatorics* [**3**]{}, 83–89.
Hallin, M., and Paindaveine, D. (2002). Optimal tests for multivariate location based on interdirections and pseudo-Mahalanobis ranks, [*The Annals of Statistics*]{} [**30**]{}, 1103–1133.
Hallin, M., and Paindaveine, D. (2004). Rank-based optimal tests of the adequacy of an elliptic VARMA model, [*The Annals of Statistics*]{} [**32**]{}, 2642–-2678.
Hallin, M., and Paindaveine, D. (2005). Affine-invariant aligned rank tests for the multivariate general linear model with VARMA errors, [*Journal of Multivariate Analysis*]{} [**93**]{}, 122-–163.
Hallin, M., and Paindaveine, D. (2006). Semiparametrically efficient rank-based inference for shape. I. Optimal rank-based tests for sphericity, [*The Annals of Statistics*]{} [**34**]{}, 2707-–2756.
Hallin, M., and Paindaveine, D. (2008). Optimal rank-based tests for homogeneity of scatter, [*The Annals of Statistics*]{} [**36**]{}, 1261-–1298.
Hallin, M., and Werker, B. J. M. (2003). Semiparametric efficiency, distribution-freeness, and invariance, [*Bernoulli*]{} [**9**]{}, 137–-165.
Hallin, M., Paindaveine, D., and [Š]{}iman, M. (2010). Multivariate quantiles and multiple-output regression quantiles: from $L^1$ optimization to halfspace depth (with discussion), [*The Annals of Statistics*]{} [**38**]{}, 635-–669.
Hardy, G., Littlewood, J., and Pólya, G. (1952). *Inequalities*. Cambridge: Cambridge University Press.
Hassairi, A., and Regaieg, O. (2008). On the Tukey depth of a continuous probability distribution, [*Statistics and Probability Letters*]{} [**78**]{}, 2308–2313.
Hlubinka, D., Kotík, L., and Vencálek, O. (2010). Weighted halfspace depth, [*Kybernetika*]{} [**46**]{}, 125–-148.
Judd, K.L. (1998). [*Numerical Methods in Economics*]{}, Cambridge: MIT Press.
Koenker, R., and Bassett, G., Jr. (1987). Regression quantiles, [*Econometrica*]{} [**46**]{}, 33–50.
Koshevoy, G. (2002). The Tukey depth characterizes the atomic measure, [*Journal of Multivariate Analysis*]{} [**83**]{}, 360-–364.
Koshevoy, G., and Mosler, K. (1997). Zonoid trimming for multivariate distributions, [*The Annals of Statistics*]{} [**25**]{}, 1998–-2017.
Koltchinskii, V. (1997). M-estimation, convexity and quantiles, *Annals of Statistics* [**25**]{}, 435–477.
Koltchinskii, V., and Dudley, R. (1992) On spatial quantiles, unpublished manuscript.
Liu, R. Y. (1990). On a notion of data depth based on random simplices, [*The Annals of Statistics*]{} [**18**]{}, 405–-414.
Liu, R. Y. (1992). Data depth and multivariate rank tests, in [*$L^1$-Statistics and Related Methods*]{} (Y. Dodge, ed.) 279–294. North-Holland, Amsterdam.
Liu, R. Y., Parelius, J. M., and Singh, K. (1999). Multivariate analysis by data depth: descriptive statistics, graphics and inference (with discussion), [*The Annals of Statistics*]{} [**27**]{}, 783-–858.
Liu, R., and Singh, K. (1993). A quality index based on data depth and multivariate rank tests, [*Journal of the American Statistical Association*]{} [**88**]{}, 257–260.
Mahalanobis, P. C. (1936). On the generalized distance in statistics, [*Proceedings of the National Academy of Sciences of India*]{} [**12**]{}, 49–55.
McCann, R. J. (1995). Existence and uniqueness of monotone measure-preserving maps, *Duke Mathematical Journal* [**80**]{}, 309–324
Mosler, K. (2002). [*Multivariate Dispersion, Central Regions and Depth: the Lift Zonoid Approach*]{}, New York: Springer.
M\^\^22ott\^\^22onen, J., and Oja, H. (1995). Multivariate sign and rank methods, [*Journal of Nonparametric Statistics*]{} [**5**]{}, 201–213.
Oja, H. (1983). Descriptive statistics for multivariate distributions, [*Statistics and Probability Letters*]{} [**1**]{}, 327–-332.
Paindaveine, D., and van Bever, G. (2013). From depth to local depth, [*Journal of the American Statistical Association*]{} [**108**]{}, 1105–1119.
Paindaveine, D., and [Š]{}iman, M. (2012). Computing multiple-output regression quantile regions, [*Computational Statistics and Data Analysis*]{} [**56**]{}, 841–853.
Politis, D., Romano, J., and Wolf, M. (1999). Weak convergence of dependent empirical measures with application to subsampling in function spaces, [*Journal of Statistical Planning and Inference*]{} [**79**]{}, 179-190.
Rockafellar, R. (1997). [*Convex Analysis*]{}, Princeton University Press.
Rockafellar, R. T., and Wets, R. J.-B. (1998). *Variational Analysis,* Berlin: Springer.
Serfling, R. (2002). Quantile functions for multivariate analysis: approaches and applications, [*Statistica Neerlandica*]{} [**56**]{}, 214–232.
Singh, K. (1991). Majority depth, unpublished manuscript.
Stahel, W. (1981). Robuste Sch\^\^22atzungen : infinitesimale Optimalit\^\^22at und Sch\^\^22atzungen von Kovarianzmatrizen, PhD Thesis, University of Z\^\^22urich.
Tukey, J. W. (1975). Mathematics and the picturing of data, in [*Proceedings of the International Congress of Mathematicians*]{} (Vancouver, B. C., 1974), Vol. 2, Montreal: Canadian Mathematical Congress, pp. 523–-531.
van der Vaart, A. W., and Wellner, J. A. (1996). *Weak Convergence*. Springer New York.
van Zwet, W.R. (1964). *Convex Transformations of Random Variables*. Mathematical Centre Tracts, Amsterdam: Mathematisch Centrum. Vardi, Y., and Zhang, C.-H. (2000). The multivariate $L^1$-median and associated data depth, [*Proceedings of the National Academy of Sciences*]{} [**97**]{}, 1423–1426.
Villani, C. (2003). *Topics in Optimal Transportation*. Providence: American Mathematical Society.
Villani, C. (2009). *Optimal transport: Old and New*. Grundlehren der Mathematischen Wissenschaften, Heidelberg: Springer-Verlag.
Villani, C. (2008). Stability of a 4th-order curvature condition arising in optimal transport theory, *Journal of Functional Analysis* [**255**]{}, 2683–2708.
Zuo, Y. (2003). Projection-based depth functions and associated medians, [*The Annals of Statistics*]{} [**31**]{}, 1460–-1490.
Zuo, Y., and Serfling, R. (2000). General notions of statistical depth function, [*The Annals of Statistics*]{} [**28**]{}, 461–-482.
[^1]: Research supported by the NSF Grant SES 1061841.
[^2]: Research funded by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013), ERC grant agreement \#313699.
[^3]: Research supported by the IAP research network grant P7/06 of the Belgian government (Belgian Science Policy), a Cr' edit aux Chercheurs of the Fonds National de la Recherche Scientifique, and the Discovery grant DP150100210 of the Australian Research Council.
[^4]: Research supported by the SSHRC Grant 435-2013-0292 and the NSERC Grant 356491-2013.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The Hodge correlators ${\text{\rm Cor}}_{\mathcal{H}}(z_0,z_1,\dots,z_n)$ are functions of several complex variables, defined by (Goncharov, 2008) by an explicit integral formula. They satisfy some linear relations: dihedral symmetry relations, distribution relations, and the shuffle relations.
We found new *second shuffle relations*. When $z_i\in{\left\{0\right\}}\cup\mu_N$, where $\mu_N$ are the $N$-th roots of unity, they are expected to give almost all relations.
When $z_i$ run through a finite subset $S$ of ${\mathbb{C}}$, the Hodge correlators describe the real mixed Hodge-Tate structure on the pronilpotent completion of the fundamental group $\pi_1^{\text{\rm nil}}({\mathbb{C}}{\mathbb{P}}^1{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$. The latter is a Lie algebra in the category of mixed ${\mathbb{Q}}$-Hodge-Tate structures. The Hodge correlators are lifted to canonical elements ${\text{\rm Cor}}_{\text{\rm Hod}}(z_0,\dots,z_n)$ in the Tannakian Lie coalgebra ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ of this category. We prove that these elements satisfy the second shuffle relations.
Let $S\subset{\overline{{\mathbb{Q}}}}$. The pronilpotent fundamental group is the Betti realization of the motivic fundamental group, which is a Lie algebra in the category of mixed Tate motives over ${\overline{{\mathbb{Q}}}}$. The Hodge correlators are lifted to elements ${\text{\rm Cor}}_{\text{\rm Mot}}(z_0,\dots,z_n)$ in the Tannakian Lie coalgebra ${\text{\rm Lie}}_{\text{\rm MT}}^\vee$ of the category of mixed Tate motives. We prove the second shuffle relations for these motivic elements.
The universal enveloping algebra of ${\text{\rm Lie}}_{\text{\rm MT}}^\vee$ was described by Goncharov via motivic multiple polylogarithms, which obey a similar yet different set of double shuffle relations. Motivic correlators have several advantages: they obey dihedral symmetry relations at all points, not only at roots of unity; they are defined for any curve, and the double shuffle relations admit a generalization to elliptic curve; and they describe elements of the motivic Lie coalgebra rather than its universal enveloping algebra.
author:
- Nikolay Malkin
bibliography:
- 'bibliography.bib'
title: Shuffle relations for Hodge and motivic correlators
---
Introduction and main results {#sec:intro}
=============================
Summary
-------
The Hodge correlators ${\text{\rm Cor}}_{\mathcal{H}}(z_0,z_1,\dots,z_n)$ are functions of several complex variables, defined by an explicit integral formula in [@goncharov-hodge-correlators]. They satisfy some linear relations: the dihedral symmetry relations, the distribution relations, and the shuffle relations.
We found new relations, called *second shuffle relations*. When $z_i\in{\left\{0\right\}}\cup\mu_N$, where $\mu_N$ are the $N$-th roots of unity, they should give almost all relations: the results of [@goncharov-motivic-modular] suggest that the other relations are sporadic, i.e., cannot be described by universal formulae.
When $z_i$ run through a finite subset $S$ of ${\mathbb{C}}$, the Hodge correlators are the canonical real periods of the mixed Hodge-Tate structures on the pronilpotent completion of the fundamental group $\pi_1^{\text{\rm nil}}({\mathbb{C}}{\mathbb{P}}^1{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$, with the tangential base point at $\infty$. The latter is a Lie algebra in the category of mixed ${\mathbb{Q}}$-Hodge-Tate structures. The Hodge correlators describe the real mixed Hodge structure on this Lie algebra tensored over ${\mathbb{Q}}$ by ${\mathbb{R}}$.
The category of mixed ${\mathbb{Q}}$-Hodge-Tate structures is canonically equivalent to the category of representations of a graded Lie algebra over ${\mathbb{Q}}$. Let us take its image in the representation defining $\pi_1^{\text{\rm nil}}({\mathbb{C}}{\mathbb{P}}^1{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$, and consider the graded dual Lie coalgebra ${\text{\rm Lie}}_{\text{\rm HT}}^\vee(S)$. The Hodge correlators were lifted in [@goncharov-hodge-correlators] to canonical elements $${\text{\rm Cor}}_{\text{\rm Hod}}(z_0,\dots,z_n)\in{\text{\rm Lie}}_{\text{\rm HT}}^\vee(S).\label{eqn:cor_hod}$$ The real numbers ${\text{\rm Cor}}_{\mathcal{H}}$ are the canonical real periods of these elements. We prove that our new relations can be lifted to relations on the elements (\[eqn:cor\_hod\]).
Let $S\subset{\overline{{\mathbb{Q}}}}\subset{\mathbb{C}}$. The Lie algebra $\pi_1^{\text{\rm nil}}({\mathbb{C}}{\mathbb{P}}^1{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$ is the Betti realization of the motivic fundamental group $\pi_1^{\text{\rm Mot}}({\mathbb{P}}^1{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$. The latter is a Lie algebra in the category of mixed Tate motives over ${\overline{{\mathbb{Q}}}}$, defined in [@deligne-goncharov]. This category is identified with the category of representations of the motivic Galois Lie algebra. Just like in the Hodge case, we take the image of this Lie algebra in the representation provided by the motivic fundamental group, and consider the graded dual Lie coalgebra ${\text{\rm Lie}}_{\text{\rm MT}}^\vee(S)$. In [@goncharov-hodge-correlators], the elements (\[eqn:cor\_hod\]) were lifted to elements $${\text{\rm Cor}}_{\text{\rm Mot}}(z_0,\dots,z_n)\in{\text{\rm Lie}}_{\text{\rm MT}}^\vee(S).\label{eqn:cor_mot}$$ We prove that our relations can be upgraded to linear relations on these elements.
The universal enveloping algebra for the Lie coalgebra ${\text{\rm Lie}}_{\text{\rm MT}}^\vee(S)$ was described in [@goncharov-polylogs-tate] via motivic multiple polylogarithms. The motivic double shuffle relations for them were proved in [@goncharov-periods-mm]. The explicit relation between motivic correlators and multiple polylogarithms is an interesting open problem.
The multiple polylogarithms obey a similar system of double shuffle relations, but the dihedral symmetry relation holds only at roots of unity. The combinatorics of those relations, originally described by [@goncharov-polylogs-modular]-[@goncharov-polylogs-tate], were studied further by [@racinet].
The motivic correlator description of $\pi_1^{\text{\rm Mot}}({\mathbb{P}}^1{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$ has several advantages. Most importantly, motivic correlators are defined for any algebraic curve, not only ${\mathbb{A}}^1{\setminus}S$, and the double shuffle relations admit a generalization to elliptic curves [@malkin-ec]. The motivic correlators obey double shuffle and cyclic symmetry relations at all points. Motivic correlators describe elements of the Lie coalgebra rather than its universal enveloping algebra. Finally, they give the best way to describe the mysterious connection between the Lie coalgebra ${\text{\rm Lie}}_{\text{\rm MT}}^\vee({\left\{0\right\}}\cup\mu_N)$ and modular manifolds [@goncharov-motivic-modular].
### Acknowledgements {#acknowledgements .unnumbered}
I am grateful to A.B. Goncharov for introducing me to this problem, for many helpful discussions and explanations, and for comments on a draft of this paper.
This material is based upon work supported by the National Science Foundation under grant DMS-1440140 while the author was in residence at the Mathematical Sciences Research Institute in Berkeley, California, during the Fall 2019 semester. The author also acknowledges support from NSF grants DMS-1107452, 1107263, 1107367 “RNMS: Geometric Structures and Representation Varieties” (the GEAR Network).
Hodge correlators and shuffle relations
---------------------------------------
We describe a family of functions of several complex variables, the Hodge correlators ([@goncharov-hodge-correlators]).[^1] Our main result is a set of functional equations on the Hodge correlators and the Hodge-theoretic and motivic upgrades of these relations.
### Definition
Let $z_0,\dots,z_n\in{\mathbb{C}}$. We define the Hodge correlator of weight $n$, ${\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)$.
Draw a disc in the plane with a sequence of points $V^{\partial}={\left\{v_0,\dots,v_n\right\}}$ placed counterclockwise around the boundary, and label $v_i$ by the value $z_i$. Choose a plane trivalent tree $T$ inside the disc with leaves at the labeled boundary vertices. Such a tree has $n-1$ interior vertices $V^\circ$ and $2n-1$ edges $E={\left\{E_0,\dots,E_{2n-2}\right\}}$. The embedding into the plane gives a canonical orientation $\rm{Or}_T\in{\left\{\pm1\right\}}$ (a choice of component of ${\mathbb{R}}^{\wedge E}$, i.e., ordering of the edges up to even permutation).
Let us assign to each edge $E_j$ a function $f_j$ on $${\mathbf X}:={\mathbb{C}}^{V^\circ}\times {\mathbb{C}}^{V^{\partial}}.$$ Precisely, to an edge $E_i=(u,v)$, assign $f_i=(2\pi i){^{-1}}\log{\left|x_u-x_v\right|}$, where $x_u$ is the coordinate on $\mathbf{X}$ corresponding to a vertex $u$. Then fix the coordinate at each boundary vertex $v_i$ to be $z_i$. Abusing notation, also denote by $f_j$ the restriction of $f_j$ to ${\mathbb{C}}^{V^\circ}$ with the boundary coordinates fixed.
Setting $d^{\mathbb{C}}={\partial}-{\overline{{\partial}}}$, we define: $$c_T(z_0,\dots,z_n)=(-4)^{n-1}\binom{2n-2}{n-1}{^{-1}}\mathrm{Or}_T\int_{{\mathbb{C}}^{V^\circ}} f_0\,d^{\mathbb{C}}f_1\wedge\dots\wedge d^{\mathbb{C}}f_{2n-2},
\label{eqn:integral_onetree_p1}$$ This expression is independent of the numbering of the edges. The Hodge correlator is defined as the sum of these integrals over all plane trivalent trees $T$: $${\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)
=\sum_Tc_T(z_0,\dots,z_n).$$ It takes values in $(2\pi i)^{-n}{\mathbb{R}}$. The simplest example, in weight 1, is shown in Fig. \[fig:w1\_hc\].
[1emFigure \[fig:w1\_hc\]. ${\text{\rm Cor}}_{\mathcal{H}}(z_0,z_1)=(2\pi i)^{-1}\log{\left|z_0-z_1\right|}$.\[fig:w1\_hc\]]{}
In weight 2, the Hodge correlators are given by $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(z_0,z_1,z_2)&=-{\frac{1}{8}}\int_x(2\pi i)^{-3}\log{\left|x-z_0\right|}\,d^{\mathbb{C}}\log{\left|x-z_1\right|}\wedge d^{\mathbb{C}}\log{\left|x-z_2\right|}.\end{aligned}$$ This integral is described by the Feynman diagram in Fig. \[fig:w2\_hc\].
[1emFigure \[fig:w2\_hc\]. \[fig:w2\_hc\]]{}
Recall the single-valued version of the dilogarithm, called the Bloch-Wigner function: $${\mathcal{L}}_2(z)=\Im{\left({\text{\rm Li}}_2(z)\right)}+\log{\left|z\right|}\arg(1-z),\quad\Im(a+bi):=b.$$ The weight 2 Hodge correlator integral can be calculated explicitly as $${\text{\rm Cor}}_{\mathcal{H}}(z_0,z_1,z_2)=(2\pi i)^{-2}{\mathcal{L}}_2{\left({\frac{z_1-z_0}{z_2-z_0}}\right)}.
\label{eqn:h_w2_l2}$$
### Properties
The Hodge correlators satisfy *dihedral symmetry* relations: $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(z_0,z_1,\dots,z_n)&={\text{\rm Cor}}_{\mathcal{H}}(z_1,\dots,z_n,z_0)\\&=(-1)^{n+1}{\text{\rm Cor}}_{\mathcal{H}}(z_n,\dots,z_1,z_0).\end{aligned}$$ One can show using (\[eqn:integral\_onetree\_p1\]) that the Hodge correlators are invariant under an additive shift of the arguments. In weight $>1$, they are also invariant under a multiplicative shift: $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)&={\text{\rm Cor}}_{\mathcal{H}}(z_0+a,\dots,z_n+a),\\
{\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)&={\text{\rm Cor}}_{\mathcal{H}}(az_0,\dots,az_n)\quad(a\in{\mathbb{C}}{^{*}}, n>1).\end{aligned}$$
Furthermore, the Hodge correlators satisfy *shuffle relations*: for $r,s\geq1$ and $z_0,\dots,z_{r+s}\in{\mathbb{C}}$, $$\sum_{\sigma\in\Sigma_{r,s}}{\text{\rm Cor}}_{\mathcal{H}}(z_0,z_{\sigma{^{-1}}(1)},z_{\sigma{^{-1}}(2)},\dots,z_{\sigma{^{-1}}(r+s)})=0,
\label{eqn:first_shuffle_cyc}$$ where $\Sigma_{r,s}\subset S_{r+s}$ is the set of *$(r,s)$-shuffles*, consisting of the permutations $\sigma$ such that $$\sigma(1)<\dots<\sigma(r),\quad\sigma(r+1)<\dots<\sigma(r+s).$$
For example, the $(1,1)$-shuffle relation states: $${\text{\rm Cor}}_{\mathcal{H}}(z_0,{\textcolor{blue}{z_1}},{\textcolor{red}{z_2}})+{\text{\rm Cor}}_{\mathcal{H}}(z_0,{\textcolor{red}{z_2}},{\textcolor{blue}{z_1}})=0;$$ the $(2,1)$-shuffle relation is:
$${\text{\rm Cor}}_{\mathcal{H}}(z_0,{\textcolor{blue}{z_1}},{\textcolor{blue}{z_2}},{\textcolor{red}{z_3}})+{\text{\rm Cor}}_{\mathcal{H}}(z_0,{\textcolor{blue}{z_1}},{\textcolor{red}{z_3}},{\textcolor{blue}{z_2}})+{\text{\rm Cor}}_{\mathcal{H}}(z_0,{\textcolor{red}{z_3}},{\textcolor{blue}{z_1}},{\textcolor{blue}{z_2}})=0.$$ The shuffle relations may be considered “easy” because they hold on the level of the sum over trees of the *integrands* in (\[eqn:integral\_onetree\_p1\]).
### Second shuffle relation
We found another relation on the Hodge correlators. Together, the two relations form the *double shuffle relations*. To state the new relations, we must introduce some notation.
Because of the multiplicative invariance (in weight $>1$) of Hodge correlators, it is possible and convenient to introduce an inhomogeneous notation for them, where the arguments are represented by the quotients between successive nonzero values and the number of 0s between them. Precisely, given $w_0,\dots,w_k\in{\mathbb{C}}{^{*}}$ such that $w_0w_1\dots w_k=1$, define $$\begin{aligned}
&{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_0|n_0,w_1|n_1,\dots,w_k|n_k):=\\&={\text{\rm Cor}}_{\mathcal{H}}(\underbrace{0,\dots,0}_{n_0},1,\underbrace{0,\dots,0}_{n_1},w_1,\underbrace{0,\dots,0}_{n_2},w_1w_2,\dots,\underbrace{0,\dots,0}_{n_k},w_1\dots w_k).\end{aligned}$$ This definition is illustrated in Fig. \[fig:homo\].
----------------------------------------- ----------------------------------------- --
{width="30.00000%"} {width="30.00000%"}
----------------------------------------- ----------------------------------------- --
[1emFigure \[fig:homo\]. ${\text{\rm Cor}}_{\mathcal{H}}(z_0,\underbrace{0,\dots,0}_{n_1},z_1,\underbrace{0,\dots,0}_{n_2},z_2,\underbrace{0,\dots,0}_{n_0})\equiv{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,w_2|n_2,w_3|n_3)$.\[fig:homo\]]{}
Define the *depth* of an expression ${\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)$ to be one less than the number of arguments in the multiplicative notation, that is, $k$ in the formula above.
Our new shuffle relation states: $$\sum_{\sigma\in\Sigma_{r,s}}{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_{\sigma{^{-1}}(1)}|n_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(r+s)}|n_{\sigma{^{-1}}(r+s)},w_0|n_0)+\text{lower-depth terms}=0.
\label{eqn:second_shuffle_cyc}$$ That is, we shuffle two ordered sets of expressions $(w_i|n_i)$, while leaving the segment $(w_0|n_0)$ fixed. For example the $(1,1)$-shuffle relation begins:
[cc]{}  &\
\
${\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,w_2|n_2,w_0|n_0)$&$+\:{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_2|n_2,w_1|n_1,w_0|n_0)$\
To describe the lower-depth terms, we need the notion of *quasishuffle*. Let $A={\left\{a_1<\dots<a_r\right\}}$ and $B={\left\{b_1<\dots<b_s\right\}}$ be two ordered sets. A quasishuffle of $A$ and $B$ is a sequence of slots ${\left\{1,\dots,M\right\}}$ and a placement of each element of $A\cup B$ in a slot, such that each slot is filled with one of:
- some $a_i\in A$,
- some $b_j\in B$,
- a pair ${\left\{a_i,b_j\right\}}$,
and the sequence of slots containing the $a_1,\dots,a_r$ and the sequence of slots containing the $b_1,\dots,b_s$ are ordered left to right. If $a_i$ and $b_j$ share a slot, they are said to *collide*. If no elements collide, the quasishuffle is said to be a shuffle.
Let $A={\left\{1,\dots,r\right\}}$ and $B={\left\{r+1,\dots,r+s\right\}}$ with the natural orders. Then, equivalently, the quasishuffles are the surjective maps ${\left\{1,\dots,r+s\right\}}{\xrightarrow{\sigma}}{\left\{1,\dots,M_\sigma\right\}}$ that are strictly increasing on $1,\dots,r$ and $r+1,\dots,r+s$.
Indices $i\in{\left\{1,\dots,r\right\}}$ collide with indices $j\in{\left\{r+1,\dots,r+s\right\}}$ whenever $\sigma(i)=\sigma(j)$. Let ${\overline{\Sigma}}_{r,s}$ be the set of such quasishuffles.
A quasishuffle $\sigma$ is a shuffle if $M_\sigma=r+s$. Recall the set of $(r,s)$-shuffles $\Sigma_{r,s}\subset S_{r+s}$. We naturally identify $\Sigma_{r,s}$ with the subset of the shuffles in ${\overline{\Sigma}}_{r,s}$.
The lower-depth terms in (\[eqn:second\_shuffle\_cyc\]) come in two kinds:
(1) Terms coming from the $(r,s)$-*quasishuffles* that are not proper shuffles. Whenever the segments $(w_i|n_j)$ and $(w_j|n_j)$ collide, we get a new segment $(w_iw_j|n_i+n_j+1)$ in their place – a 0 is inserted – and the term picks up a negative sign.
For the $(1,1)$-shuffle relation, there is only one quasishuffle that is not a shuffle. In this quasishuffle, the two segments $(w_1|n_1)$ and $(w_2|n_2)$ collide:
[c]{} \
\
$-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1w_2|n_1+n_2+1,w_0|n_0)$\
(2) Two extra terms: one where the segments $w_1,\dots,w_r$ appear in order and the remaining segments $w_{r+1},\dots,w_{r+s},w_0$ collapse; another where the segments $w_{r+1},\dots,w_{r+s}$ appear in order and $w_1,\dots,w_r,w_0$ collapse. These terms come with a negative sign.
For the $(1,1)$-shuffle relation:
[cc]{}  &\
\
$-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,w_2w_0|n_2+n_0+1)$&$-\:{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_2|n_2,w_1w_0|n_1+n_0+1)$\
In summary, the $(1,1)$-shuffle relation states, for $w_0,w_1,w_2\in{\mathbb{C}}{^{*}}$ and $w_0w_1w_2=1$, $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,w_2|n_2,w_0|n_0)&+{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_2|n_1,w_1|n_1,w_0|n_0)\\&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1w_2|n_1+n_2+1,w_0|n_0)\\&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,w_2w_0|n_2+n_0+1)\\&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_2|n_2,w_1w_0|n_1+n_0+1)&=0.\end{aligned}$$ It is already a nontrivial relation, which is not easy to prove from the definition (\[eqn:integral\_onetree\_p1\]) even for $n_0=n_1=n_2=0$.
By formula (\[eqn:h\_w2\_l2\]), Hodge correlators in weight 2 are expressed in a simple way in terms of the Bloch-Wigner function ${\mathcal{L}}_2$. The $(1,1)$-shuffle relation with $n_0=n_1=n_2=0$ is equivalent to the five-term relation, $${\mathcal{L}}_2{\left({\frac{1-w_1}{1-w_1w_2}}\right)}+{\mathcal{L}}_2{\left({\frac{1-w_2}{1-w_1w_2}}\right)}+{\mathcal{L}}_2(1-w_1w_2)+{\mathcal{L}}_2(w_1)+{\mathcal{L}}_2(w_2)=0.$$ According to [@bloch-irvine], this is essentially the only functional equation for ${\mathcal{L}}_2$. It follows that the dihedral symmetry and shuffle relations are the *only* relations between the Hodge correlators in weight 2.
For further illustration, let us write out the $(2,1)$-shuffle relation for the Hodge correlator $${\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|0,w_2|1|w_3|1,w_4|0),$$ where ${\color{blue}{w_1}}$ and ${\color{blue}{w_2}}$ will be shuffled with ${\color{red}{w_3}}$:
(1) There are three terms from the shuffles:
---------------------------------------------------------------- ---------------------------------------------------------------- ----------------------------------------------------------------
${\color{blue}{w_1}}\,{\color{blue}{w_2}}\,{\color{red}{w_3}}$ ${\color{blue}{w_1}}\,{\color{red}{w_3}}\,{\color{blue}{w_2}}$ ${\color{red}{w_3}}\,{\color{blue}{w_1}}\,{\color{blue}{w_2}}$
  
---------------------------------------------------------------- ---------------------------------------------------------------- ----------------------------------------------------------------
(2) There are two terms from the quasishuffles that are not shuffles:
------------------------------------------------------------------------------ ------------------------------------------------------------------------------ --
${\begin{bmatrix}{\color{blue}{w_1}}\\#2\end{bmatrix}}\,{\color{blue}{w_2}}$ ${\color{blue}{w_1}}\,{\begin{bmatrix}{\color{blue}{w_2}}\\#2\end{bmatrix}}$
 
------------------------------------------------------------------------------ ------------------------------------------------------------------------------ --
(3) There are two additional terms:
----------------------- ----------------------- --
 
----------------------- ----------------------- --
The full relation is then $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|0,w_2|1,w_3|1,w_4|0)&+{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|0,w_3|1,w_2|1,w_4|0)+{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_3|1,w_1|0,w_2|1,w_4|0)\\
&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1w_3|2,w_2|1,w_4|0)-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|0,w_2w_3|3,w_4|0)\\
&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|0,w_2|1,(w_1w_2){^{-1}}|2)-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_3|1,w_3{^{-1}}|3)&=0,\end{aligned}$$ where the $3+2+2$ terms in the three rows match the $3+2+2$ pictures above.
We now write out the general relation:
(a) Suppose that $r,s>1$ and that not all $n_i=0$ or not all $w_i=1$. Then the Hodge correlators satisfy the relation: $$\begin{aligned}
&\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_{\sigma{^{-1}}(1)}|n_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)}|n_{\sigma{^{-1}}(M_\sigma)},w_0|n_0)\\
&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,\dots,w_r|n_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}|n_{{\left\{r+1,\dots,r+s,0\right\}}})\\
&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_{r+1}|n_{r+1},\dots,w_{r+s}|n_{r+s},w_{{\left\{1,\dots,r,0\right\}}}|n_{{\left\{1,\dots,r,0\right\}}})&=0,
\end{aligned}$$ where $$n_S=\sum_{i\in S}(n_i+1)-1,\quad w_S=\prod_{i\in S}w_i.$$
(b) The Hodge correlators satisfy all specializations of this relation as any subset of the $w_i$ $(1\leq i\leq n)$ approaches 0.
\[thm:period\_main\]
### Applications
Theorem \[thm:period\_main\] gives simple proofs of certain results of [@goncharov-rudenko].
For $n>2$, every Hodge correlator of weight $n$ is a linear combination of Hodge correlators of weight $n$ and depth at most $n-2$.
Precisely, for $z_1,\dots,z_n\in{\mathbb{C}}{^{*}}$, we have $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(z_1,\dots,z_n,0)
&=\sum_{i=1}^n{\text{\rm Cor}}_{\mathcal{H}}{\left(z_1,\dots,z_{i-1},z_i,z_i{\frac{z_1}{z_n}},\dots,z_{n-1}{\frac{z_1}{z_n}},z_n{\frac{z_1}{z_n}}\right)}\nonumber\\
&\quad-\sum_{i=2}^n{\text{\rm Cor}}_{\mathcal{H}}{\left(z_1,\dots,z_{i-1},0,z_i{\frac{z_1}{z_n}},\dots,z_{n-1}{\frac{z_1}{z_n}},z_n{\frac{z_1}{z_n}}\right)}\nonumber\\
&\quad-{\text{\rm Cor}}_{\mathcal{H}}{\left(z_1,z_1{\frac{z_1}{z_n}},0,\dots,0\right)}.\label{eqn:lower_depth_reduction}
\end{aligned}$$ \[cly:gr28\]
In weight 3, we deduce the Hodge correlator version of relations (27) and (29) from [@goncharov-rudenko].
The Hodge correlators in weight 3 satisfy the relations: $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,x)&+{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-x)+{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-x{^{-1}})={\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1),\\
{\text{\rm Cor}}_{\mathcal{H}}(0,x,1,y)&=
-{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-x{^{-1}})
-{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-y{^{-1}})
-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{y}{x}}\right)}\nonumber\\\label{eqn:corr_29}
&\quad-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-y}{1-x}}\right)}
+{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-y{^{-1}}}{1-x{^{-1}}}}\right)}
+{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1).\end{aligned}$$ \[cly:gr2729\]
We have noted that the double shuffle and dihedral symmetry relations give all relations between Hodge correlators in weight 2.
In weight 3, the Hodge correlators of depth 1 are expressed in terms of the single-valued trilogarithm ${\mathcal{L}}_3$ (see §\[sec:rel\_hodge\_polylog\]). By the results of [@goncharov-rudenko], the relations (\[eqn:corr\_29\]) imply the general functional equation for ${\mathcal{L}}_3$ ([@goncharov-conf-polylogs]). We conclude that the double shuffle relations for Hodge correlators imply all functional equations for ${\mathcal{L}}_2$ and ${\mathcal{L}}_3$.
Quasidihedral Lie coalgebras
----------------------------
Let $G$ be an abelian group. We use the multiplicative notation for $G$; the identity element is $1\in G$. Typically, $G$ will be the multiplicative group of a field $F^\times$ or the group of $N$-th roots of unity $\mu_N$. We adjoin to $G$ a formal element 0, where $0\cdot g=0$ for $g\in G\cup{\left\{0\right\}}$.
We define the *quasidihedral Lie coalgebra* ${\mathcal{D}}(G)$. It generalizes the dihedral Lie coalgebra of [@goncharov-dihedral]; the latter is the associated graded for the depth filtration of of ${\mathcal{D}}(G)$. The aim of the construction of ${\mathcal{D}}(G)$ is twofold:
(1) It is the main combinatorial ingredient in the proof of the double shuffle relations for correlators.
(2) The Lie coalgebra ${\mathcal{D}}(G)$ describes the coproduct of motivic correlators.
### Cyclic Lie coalgebra
Let $V$ be the ${\mathbb{Q}}$-vector space with basis indexed by $G\cup{\left\{0\right\}}$
Let $T(V)=\bigoplus_{n\geq0}V^{\otimes n}$ be the tensor algebra of $V$ over ${\mathbb{Q}}$. We impose a grading by weight, where $V^{\otimes n}$ has weight $n-1$. Then define the cyclic Lie coalgebra, as a vector space, by
$${\mathcal{C}}(G)={\frac{T(V)}{\text{cyclic symmetry}}}.$$ It is positively graded and generated in weight $n$ by elements $x_0\otimes\dots\otimes x_n$ modulo the relation $x_0\otimes\dots\otimes x_n=x_1\otimes\dots\otimes x_n\otimes x_0$. We can represent these elements by elements of $G\cup{\left\{0\right\}}$ written counterclockwise at marked points on a circle.
The coproduct on ${\mathcal{C}}(G)$ is defined on such a generator by splitting the circle into two arcs that share exactly one point. That is, consider a line inside the circle, starting at a marked point and ending between two marked points. It splits the circle into two parts, representing generators $x'$ and $x''$, and the coproduct of $x_0\otimes\dots\otimes x_n$ is the sum of $x'\wedge x''$ over all such cuts.
Precisely, the coproduct is defined by $$\delta{\left(x_0\otimes\dots\otimes x_n\right)}=\sum_{\rm cyc}\sum_{i=1}^{n-1}{\left(x_0\otimes x_1\otimes\dots\otimes x_i\right)}\wedge{\left(x_0\otimes x_{i+1}\otimes\dots\otimes x_n\right)}.
\label{eqn:cyc_coproduct}$$ It respects the weight grading and satisfies the co-Jacobi identity.
We will write elements of ${\mathcal{C}}(G)$ as $$C(x_0,\dots,x_n)=x_0\otimes\dots\otimes x_n.$$ Also introduce a notation, analogous to that for Hodge correlators, for $w_0,\dots,w_k\in G$ with $w_0\dots w_k=1$: $$\begin{aligned}
&C{^{*}}(w_0|n_0,w_1|n_1,\dots,w_k|n_k):=\\&=C(\underbrace{0,\dots,0}_{n_0},1,\underbrace{0,\dots,0}_{n_1},w_1,\underbrace{0,\dots,0}_{n_2},\dots,w_1\dots w_{k-1},\underbrace{0,\dots,0}_{n_k},w_1\dots w_k).\end{aligned}$$
### Relations
A *first shuffle* in ${\mathcal{C}}(G)$ is an element of the form $$\sum_{\sigma\in\Sigma_{r,s}}C(x_0,x_{\sigma{^{-1}}(1)},x_{\sigma{^{-1}}(2)},\dots,x_{\sigma{^{-1}}(r+s)}).$$ Define $$\widetilde{\mathcal{D}}(G)={\frac{{\mathcal{C}}(G)}{\text{first shuffles, scaling relations, distribution relations}}}.$$ The scaling relations we impose are:
(1) In weight 1, we have $C(0,0)=0$ and $C(ab,ac)=C(0,a)+C(b,c)$ for $a\in G$.
(2) In weight $>1$, multiplicative invariance: $$C(x_0,\dots,x_n)=C(ax_0,\dots,ax_n),\quad a\in G.$$
The distribution relations are the following. For $l\in{\mathbb{Z}}_{>0}$, let $G_l$ denote the $l$-torsion of $G$. Suppose that $G_l$ is finite and $l$ divides ${\left|G_l\right|}$, and suppose $x_0,\dots,x_n\in G\cup{\left\{0\right\}}$ are divisible by $l$ (note $0$ is always divisible by $l$). Let $m$ be the number of 0s among the $x_i$. Then the relation is $$C(x_0,\dots,x_n)={\frac{l^m}{{\left|G_l\right|}}}\sum_{y_i^l=x_i}C(y_0,\dots,y_n),
\label{eqn:distr_rel}$$ except in the case that $n=1$ and $x_0=x_1$.
The following is immediate from the constructions of [@goncharov-dihedral] (Theorem 4.3).
The first shuffles, scaling relations, and distribution relations generate a coideal in ${\mathcal{C}}(G)$. The coproduct on ${\mathcal{C}}(G)$ descends to a well-defined coproduct on $\widetilde{\mathcal{D}}^\vee(G)$.
Abusing notation, denote also by $C$ and $C{^{*}}$ the images in $\widetilde{\mathcal{D}}(G)$ of the elements $C,C{^{*}}$ in ${\mathcal{C}}(G)$.
A *second shuffle* in ${\mathcal{C}}(G)$ is an element of the form suggested by Theorem \[thm:period\_main\]: $$\begin{aligned}
&\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}C{^{*}}(w_{\sigma{^{-1}}(1)}|n_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)}|n_{\sigma{^{-1}}(M_\sigma)},w_0|n_0)\\
&-C{^{*}}(w_1|n_1,\dots,w_r|n_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}|n_{{\left\{r+1,\dots,r+s,0\right\}}})\\
&-C{^{*}}(w_{r+1}|n_{r+1},\dots,w_{r+s}|n_{r+s},w_{{\left\{1,\dots,r,0\right\}}}|n_{{\left\{1,\dots,r,0\right\}}}),\end{aligned}$$ where $$n_S=\sum_{i\in S}(n_i+1)-1,\quad w_S=\prod_{i\in S}w_i.$$ Define the *quasidihedral Lie coalgebra* $${\mathcal{D}}(G)={\frac{\widetilde{\mathcal{D}}(G)}{\text{second shuffles}}}.$$ Then we prove:
The second shuffles form a coideal in $\widetilde{\mathcal{D}}(G)$. The coproduct on $\widetilde{\mathcal{D}}(G)$ descends to a well-defined coproduct on ${\mathcal{D}}(G)$. \[thm:main\_qdih\_const\]
Theorem \[thm:main\_qdih\_const\] provides us with a Lie coalgebra generated by sequences of elements of $G\cup{\left\{0\right\}}$ that satisfies dihedral symmetry, scaling, and the two shuffle relations.
Let ${\mathcal{C}}^\circ(G)$ the subspace of ${\mathcal{C}}(G)$ generated by elements $C(x_0,\dots,x_n)$ where not all $x_i$ are equal. It is a subcoalgebra, which we call the *restricted cyclic Lie coalgebra*. The image of ${\mathcal{C}}^\circ(G)$ in ${\mathcal{D}}(G)$ is the *restricted quasidihedral Lie coalgebra*, denoted ${\mathcal{D}}^\circ(G)$.
The Hodge correlators satisfy cyclic symmetry, first shuffle, distribution, and scaling relations. Equivalently, the function ${\text{\rm Cor}}_{\mathcal{H}}{^{*}}$ factors through $\widetilde{\mathcal{D}}({\mathbb{C}}{^{*}})$ and a map $$C{^{*}}(w_0|n_0,\dots,w_k|n_k)\mapsto{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_0|n_0,\dots,w_k|n_k).$$ An equivalent form of Theorem \[thm:period\_main\] is that, restricted to the set of arguments where not all $w_i=1$ or not all $n_i=0$, this function factors through the quotient ${\mathcal{D}}^\circ({\mathbb{C}}{^{*}})$.
### Depth filtration
The Lie coalgebra ${\mathcal{D}}(G)$ is filtered by the *depth*, where a generator has depth $d$ if it includes $d+1$ elements of $G$ (not counting 0s). Consider ${\text{\rm gr}}^D{\mathcal{D}}(G)$. In this coalgebra, the second shuffle relations lose their lower-depth terms.
Relations for motivic correlators: Hodge realization
----------------------------------------------------
We present the construction of motivic correlators of [@goncharov-hodge-correlators] and state our main result in this setting.
This section concerns the Hodge realization of motivic correlators. They are objects in the fundamental Lie coalgebra of the category of ${\mathbb{R}}$-mixed Hodge structures, and are Hodge-theoretic upgrades of the Hodge correlator functions.
### Summary
In [@goncharov-hodge-correlators], given any collection of complex numbers $z_0,\dots,z_n$, the Hodge correlators ${\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)$ were upgraded to elements of the Tannakian Lie coalgebra ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ of the category of real mixed Hodge structures: $${\text{\rm Cor}}_{\text{\rm Hod}}(z_0,\dots,z_n)\in{\text{\rm Lie}}_{\text{\rm HT}}^\vee.\label{eqn:elem_cor_hod}$$ Furthermore, if follows easily from the construction of the upgraded Hodge correlators (\[eqn:elem\_cor\_hod\]) that they satisfy the dihedral and first shuffle relations, and that their coproduct in the coalgebra ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ is given precisely by the formula (\[eqn:cyc\_coproduct\]).
One of the main results of this paper is that the elements (\[eqn:elem\_cor\_hod\]) satisfy the second shuffle relations. In other words, they provide a map of Lie coalgebras ${\mathcal{D}}^\circ({\mathbb{C}}{^{*}})\to{\text{\rm Lie}}_{\text{\rm HT}}^\vee$.
### Hodge-theoretic setup
Let ${\mathrm{MHT}}_{\mathbb{R}}$ of be the tensor category of ${\mathbb{R}}$-mixed Hodge-Tate structures and ${\mathrm{HT}}_{\mathbb{R}}$ the category of ${\mathbb{R}}$-pure Hodge-Tate structures. Every object of ${\mathrm{MHT}}_{\mathbb{R}}$ is filtered by weight, and ${\mathrm{MHT}}_{\mathbb{R}}$ is generated by the simple objects ${\mathbb{R}}(n)$, the pure Hodge-Tate structures of weight $-n$. The cohomology of a punctured projective line is a mixed Hodge-Tate structure, nontrivial in weights 0 and 1.
The *Galois Lie algebra* of the category of mixed Hodge-Tate structures, ${\text{\rm Lie}}_{\text{\rm HT}}$, is the algebra of tensor derivations of the functor ${\text{\rm gr}}^W:{\mathrm{MHT}}_{\mathbb{R}}\to{\mathrm{HT}}_{\mathbb{R}}$. It is a graded Lie algebra in the category ${\mathrm{HT}}_{\mathbb{R}}$, and ${\mathrm{MHT}}_{\mathbb{R}}$ is equivalent to the category of graded ${\text{\rm Lie}}_{\text{\rm HT}}$-modules in ${\mathrm{HT}}_{\mathbb{R}}$. Let ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ be its graded dual. A canonical *period map* $$p:{\text{\rm Lie}}_{\text{\rm HT}}^\vee\to{\mathbb{R}}$$ was defined in [@goncharov-hodge-correlators].
Let $X={\mathbb{P}}^1({\mathbb{C}})$, $S\subset X$ a finite set of punctures containing $\infty$, and $v_\infty={\frac{-1}{z^2}}{\frac{{\partial}}{{\partial}z}}$ a distinguished tangent vector at $\infty$. The pronilpotent completion $\pi_1^{\text{\rm nil}}(X{\setminus}(S\cup{\left\{\infty\right\}}),v_\infty)$ of the fundamental group $\pi_1(X{\setminus}S,\infty)$ carries a mixed Hodge-Tate structure, depending on $v_\infty$, and thus there is a map $${\text{\rm Lie}}_{\text{\rm HT}}\to{\text{\rm Der}}{\left({\text{\rm gr}}^W\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)\right)}.$$
### Hodge correlator coalgebra {#sec:hodge_cor_coalg}
The *Hodge correlator coalgebra* is defined by [@goncharov-hodge-correlators] as $${\mathcal{CL}}_{X,S,v_\infty}^\vee:={\frac{T({\mathbb{C}}{\left[S{\setminus}{\left\{\infty\right\}}\right]}^\vee)}{\text{relations}}}\otimes H_2(X).$$ Note that $H_2(X)\cong{\mathbb{R}}(1)$. If $[h]\in H_2(X)$ is the fundamental class, we write $x(1)$ for $x\otimes[h]$.
The relations are the following:
(1) Cyclic symmetry: $x_0\otimes\dots\otimes x_n=x_1\otimes\dots\otimes x_n\otimes x_0$.
(2) (First) shuffle relations: $$\sum_{\sigma\in\Sigma_{p,q}}x_0\otimes x_{\sigma{^{-1}}(1)}\otimes\dots\otimes x_{\sigma{^{-1}}(p+q)}=0.$$
(3) Take the quotient by the weight $-1$ elements $(x_0)$.
There is a Lie coalgebra structure on ${\mathcal{CL}}_{X,S,v_0}^\vee$, defined by the same formula as for the cyclic Lie coalgebra: $$\delta{\left((x_0\otimes\dots\otimes x_n)(1)\right)}=\sum_{\rm cyc}\sum_{i=1}^{n-1}{\left((x_0\otimes x_1\otimes\dots\otimes x_i)(1)\right)}\wedge{\left((x_0\otimes x_{i+1}\otimes\dots\otimes x_n)(1)\right)}.
\label{eqn:clie_coproduct}$$
An action of the graded dual Lie algebra ${\mathcal{CL}}_{X,S,v_\infty}$ by derivations on $L_{X,S,s_0}$ was constructed by [@goncharov-hodge-correlators]. The action $${\mathcal{CL}}_{X,S,v_\infty}\to{\text{\rm Der}}{\left(L_{X,S,v_\infty}\right)}$$ is injective. Its image consists of the *special derivations* ${\text{\rm Der}}^S{\left(L_{X,S,v_\infty}\right)}$, those which act by 0 on the loop around $\infty$ and preserve the conjugacy classes of all the loops $s\in S{\setminus}{\left\{\infty\right\}}$.
Dualizing this map composed with the action of ${\text{\rm Lie}}_{\text{\rm HT}}$, we get the *Hodge correlator map* of Lie coalgebras: $${\text{\rm Cor}}_{\text{\rm Hod}}:{\mathcal{CL}}^\vee_{X,S,v_{\infty}}\to{\text{\rm Lie}}_{\text{\rm HT}}^\vee.$$ We will also write ${\text{\rm Cor}}_{\text{\rm Hod}}(x_0,\dots,x_n)$ for ${\text{\rm Cor}}_{\text{\rm Hod}}{\left((x_0\otimes\dots\otimes x_n)(1)\right)}$, and similarly define $${\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_0|n_0,\dots,w_k|n_k).$$
### Period map and Hodge correlator functions
Recall that the Hodge correlator functions ${\text{\rm Cor}}_{\mathcal{H}}(x_0,\dots,x_n)$ satisfy cyclic symmetry and shuffle relations, so we may also denote by ${\text{\rm Cor}}_{\mathcal{H}}$ the function $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}:{\mathcal{CL}}^\vee_{X,S,v_\infty}&\to{\mathbb{C}},\\
(x_0\otimes\dots\otimes x_n)(1)&\mapsto{\text{\rm Cor}}_{\mathcal{H}}(x_0,\dots,x_n).\end{aligned}$$ The dual to the Hodge correlator ${\text{\rm Cor}}_{\mathcal{H}}:{\mathcal{CL}}_{X,S,v_\infty}^\vee\to{\mathcal{C}}$, an element of ${\mathcal{CL}}_{X,S,v_\infty}$, is called the *Green operator* $\mathbf{G}_{v_\infty}$. It can be viewed as a special derivation of ${\text{\rm gr}}^W\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)\otimes{\mathbb{C}}$, and defines a real mixed Hodge structure on $\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)$. An element $x\in{\mathcal{CL}}^\vee_{X,S,v_\infty}$ provides a framing ${\mathbb{R}}(n)\to{\text{\rm gr}}_{2n}^W\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)$, and ${\text{\rm Cor}}_{\text{\rm Hod}}(x)$ is the element of ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ induced by this framing.
As made precise by a main result of [@goncharov-hodge-correlators], ${\text{\rm Cor}}_{\mathcal{H}}$ factors through the Hodge correlator map to ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ and the period map ${\text{\rm Lie}}_{\text{\rm HT}}^\vee\to{\mathbb{C}}$, and the resulting mixed Hodge structure on $\pi_1^{\text{\rm nil}}$ coincides with the standard one.
(a) Let $x\in{\mathcal{CL}}^\vee_{X,S,v_\infty}$ be homogeneous of weight $n$. Then ${\text{\rm Cor}}_{\mathcal{H}}(x)=(2\pi i)^{-n}p({\text{\rm Cor}}_{\text{\rm Hod}}(x))$, where $p$ is the canonical period map ${\text{\rm Lie}}_{\text{\rm HT}}^\vee\to{\mathbb{R}}$.
(b) The mixed Hodge structure on $\pi_1^{\text{\rm nil}}$ determined by the dual Hodge correlator map coincides with the standard mixed Hodge structure on $\pi_1^{\text{\rm nil}}$.
\[thm:hc\_main\_point\]
### Second shuffle relations
We state the version of the main result for the Hodge correlators, on the level of the map ${\text{\rm Cor}}_{\text{\rm Hod}}$.
(a) Restricted to the subspace of ${\mathcal{CL}}_{X,S,v_\infty}^\vee$ generated by elements $(x_0\otimes\dots\otimes x_n)(1)$ with not all $x_i$ equal, the map ${\text{\rm Cor}}_{\text{\rm Hod}}$ factors through ${\mathcal{D}}^\circ({\mathbb{C}}{^{*}})$.
(b) Suppose that $r,s>1$ and that not all $n_i=0$ or not all $w_i=1$. Then the Hodge correlators satisfy the relation: $$\begin{aligned}
&\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}{\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_{\sigma{^{-1}}(1)}|n_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)}|n_{\sigma{^{-1}}(M_\sigma)},w_0|n_0)\\
&-{\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_1|n_1,\dots,w_r|n_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}|n_{{\left\{r+1,\dots,r+s,0\right\}}})\\
&-{\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_{r+1}|n_{r+1},\dots,w_{r+s}|n_{r+s},w_{{\left\{1,\dots,r,0\right\}}}|n_{{\left\{1,\dots,r,0\right\}}})&=0,
\end{aligned}$$ where $$n_S=\sum_{i\in S}(n_i+1)-1,\quad w_S=\prod_{i\in S}w_i.$$
(c) The Hodge correlators satisfy all specializations of this relation as any subset of the $w_i$ $(1\leq i\leq n)$ approaches 0.
\[thm:hodge\_main\]
While Theorem \[thm:period\_main\] was an equality between functions, Theorem \[thm:hodge\_main\] is a relation in the fundamental Lie coalgebra of mixed Hodge-Tate structures. Theorem \[thm:period\_main\] follows immediately from Theorem \[thm:hodge\_main\] by applying the period map.
Relations for motivic correlators over a number field
-----------------------------------------------------
We now state the most general version of the result by upgrading the constructions of the previous section from mixed Hodge structures to mixed motives over a number field.
### Motivic setup
Let $F$ be a number field and ${\mathcal{MTM}}_F$ the category of mixed Tate motives over $F$. It is generated by objects ${\mathbb{Q}}(n)={\mathbb{Q}}(1)^{\otimes n}$ for $n\in{\mathbb{Z}}$, where ${\mathbb{Q}}(1)$ is the Tate motive, pure of weight $-1$. This induces a canonical weight filtration on objects of ${\mathcal{MTM}}_F$. There is a functor ${\text{\rm gr}}^W:{\mathcal{MTM}}_F\to{\mathcal{PM}}_F$, where ${\mathcal{PM}}_F$ is the category of pure motives over $F$.
The *fundamental (motivic Tate) Lie algebra* ${\text{\rm Lie}}_{{\text{\rm MT}/F}}$ is the algebra of tensor derivations of the functor ${\text{\rm gr}}^W$, a graded Lie algebra in the category ${\mathcal{PM}}_F$, and ${\mathcal{MTM}}_F$ is equivalent to the category of graded ${\text{\rm Lie}}_{{\text{\rm MT}/F}}$-modules.
An embedding $\sigma:F\to{\mathbb{C}}$ induces a *realization functor* $r:{\mathcal{MTM}}_F\to{\mathrm{MHT}}_{\mathbb{R}}$ and a map $r:{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\to{\text{\rm Lie}}_{\text{\rm HT}}^\vee$.
Let $X={\mathbb{P}}^1$, $S\subset X(F)$ a finite set of punctures containing $\infty$, and $v_\infty$ the distinguished tangent vector at $\infty$. Deligne and Goncharov’s *motivic fundamental group* ([@deligne-goncharov]) $\pi_1^{\text{\rm Mot}}(X{\setminus}S,v_\infty)_{\rm un}$ is a prounipotent group scheme in the category ${\mathcal{MTM}}_F$. The Hodge realization of its Lie algebra is $\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)$. As it is an object in ${\mathcal{MTM}}_F$, there is an action ${\text{\rm Lie}}_{{\text{\rm MT}/F}}\to{\text{\rm Der}}{\left({\text{\rm gr}}^W\pi_1^{\text{\rm Mot}}\right)}$.
### Motivic correlator coalgebra
The construction of the Hodge correlator coalgebra ${\mathcal{CL}}^\vee_{X,S,v_\infty}$ can be upgraded to the motivic setting. The definition of the *motivic correlator coalgebra* mimics that of its Hodge realization: $${\left({\mathcal{CL}}^{\text{\rm Mot}}_{X,S,v_\infty}\right)}^\vee:={\frac{T{\left(({\mathbb{Q}}(1)^{S{\setminus}{\left\{\infty\right\}}})^\vee\right)}}{\text{relations}}}\otimes H_2(X),$$ a graded Lie coalgebra in the category of pure motives over $F$, where the relations imposed are the cyclic symmetry, first shuffles, and quotient by weight $0$. Then ${\mathcal{CL}}^{\text{\rm Mot}}_{X,S,v_0}$ is isomorphic to the algebra of special derivations of ${\text{\rm gr}}^W\pi_1^{\text{\rm Mot}}(X-S,v_\infty)$, and there is a map $${\text{\rm Cor}}_{\text{\rm Mot}}:{\left({\mathcal{CL}}^{\text{\rm Mot}}_{X,S,v_\infty}\right)}^\vee\to{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee.$$ We will write ${\text{\rm Cor}}_{\text{\rm Mot}}(x_0,\dots,x_n)$ for ${\text{\rm Cor}}_{\text{\rm Mot}}({\left(x_0\otimes\dots\otimes x_n\right)}(1))$, and likewise ${\text{\rm Cor}}_{\text{\rm Mot}}(w_0|n_0,\dots,w_k|n_k)$.
Let us describe how motivic correlators are related to Hodge correlators. Fix an embeding $r:F\to{\mathbb{C}}$. The Hodge realization provides coalgebra maps ${\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\to{\text{\rm Lie}}_{\text{\rm HT}}^\vee$ and $$r:{\left({\mathcal{CL}}^{\text{\rm Mot}}_{X,S,v_\infty}\right)}^\vee\otimes{\mathbb{C}}\to{\mathcal{CL}}^\vee_{X,S,v_\infty}\otimes{\mathbb{C}},$$ and thus a period map $${\text{\rm Cor}}_{\mathcal{H}}\circ r:{\left({\mathcal{CL}}^{\text{\rm Mot}}_{X,S,v_\infty}\right)}^\vee\otimes{\mathbb{C}}\to{\mathcal{CL}}_{X,S,v_\infty}^\vee\otimes{\mathbb{C}}\to{\mathbb{C}}.$$
By Theorem \[thm:hc\_main\_point\], it coincides with the composition $${\left({\mathcal{CL}}^{\text{\rm Mot}}_{X,S,v_\infty}\right)}^\vee\to{\text{\rm Lie}}_{\text{\rm Mot}}^\vee\to{\text{\rm Lie}}_{\text{\rm HT}}^\vee\to{\mathbb{C}}.$$ We can summarize the objects and maps defined thus far as follows: $$\xymatrix{
{\text{\rm Der}}^S({\text{\rm gr}}^W\pi_1^{\text{\rm Mot}}(X{\setminus}S,v_\infty))^\vee\ar@{-}[r]
&({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}})^\vee\ar[r]^{\quad{\text{\rm Cor}}_{\text{\rm Mot}}}\ar[d]^r
&{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\ar[d]^r
\\
{\text{\rm Der}}^S({\text{\rm gr}}^W\pi_1^{\text{\rm nil}}(X{\setminus}S,v_0))^\vee\ar@{-}[r]
&({\mathcal{CL}}_{X,S,v_\infty}^\vee)\ar[r]^{\quad{\text{\rm Cor}}_{\text{\rm Hod}}}\ar[dr]_{(2\pi i)^w{\text{\rm Cor}}_{\mathcal{H}}}&{\text{\rm Lie}}_{\text{\rm HT}}^\vee\ar[d]^p
\\
&&{\mathbb{R}}.
}$$
Under certain conditions, relations on motivic correlators hold can be proven by showing that they hold in the Hodge realization under any complex embedding. This is a key fact in the proof of the motivic upgrade of our relations on Hodge correlators:
Let $X{\setminus}S$ be a rational curve over $F$. Suppose $x\in{\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^\vee$ has weight $>1$, $\delta{\text{\rm Cor}}_{\text{\rm Mot}}(x)=0$, and ${\text{\rm Cor}}_{\mathcal{H}}(r(x))=0$ for every embedding $r:F\to{\mathbb{C}}$. Then ${\text{\rm Cor}}_{\text{\rm Mot}}(x)=0$. \[lma:d0h0\_rational\]
### Dependence on $S$
If $S\subseteq S'$, there is an induced inclusion $\iota:({\mathcal{CL}}_{X,S,v_0}^{\text{\rm Mot}})^\vee\to({\mathcal{CL}}_{X,S',v_0}^{\text{\rm Mot}})^\vee$.
The following diagram commutes: $$\xymatrix{
&({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}})^\vee\ar[d]^\iota\ar[ddl]_{{\text{\rm Cor}}_{\text{\rm Mot}}}\ar[ddr]^{{\text{\rm Cor}}_{\mathcal{H}}\circ r}\\
&({\mathcal{CL}}_{X,S',v_\infty}^{\text{\rm Mot}})^\vee\ar[dl]^{{\text{\rm Cor}}_{\text{\rm Mot}}}\ar[dr]_{{\text{\rm Cor}}_{\mathcal{H}}\circ r}\\
{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\ar[rr]_{p\circ r}&&{\mathbb{C}}.
}$$ This allows us to write down elements of $({\mathcal{CL}}_{X,S,v_0}^{\text{\rm Mot}})^\vee$ without explicitly specifying $S$.
### Second shuffle relations
We are ready to state the most general version of the main result.
Let $F$ be a number field.
(a) Restricted to the subspace of ${\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^\vee$ generated by elements $(x_0\otimes\dots\otimes x_n)(1)$ with not all $x_i$ equal, the map ${\text{\rm Cor}}_{\text{\rm Mot}}$ factors through ${\mathcal{D}}^\circ(F^\times)$.
(b) Suppose that $r,s>1$ and that not all $n_i=0$ or not all $w_i=1$. Then the motivic correlators satisfy the same relation as in Theorem \[thm:hodge\_main\], with ${\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}$ replaced by ${\text{\rm Cor}}_{\text{\rm Mot}}{^{*}}$.
(c) The motivic correlators satisfy all specializations of this relation as any subset of the $w_i$ $(1\leq i\leq n)$ approaches 0.
\[thm:mot\_main\]
Background: Hodge and motivic correlators
=========================================
Hodge realization of motivic correlators
----------------------------------------
### Mixed Hodge theory
We recall the relevant definitions from [@deligne-mixed-hodge]. A real mixed Hodge structure consists of the following data:
(1) A real vector space $V$;
(2) An increasing weight filtration $W_\bullet$ on $V$;
(3) A decreasing Hodge filtration $F^\bullet$ on its complexification $V_{\mathbb{C}}=V\otimes_{\mathbb{R}}{\mathbb{C}}$, with conjugate ${\overline{F}}^\bullet$,
such that $F^\bullet$ and ${\overline{F}}^\bullet$ induce a pure real Hodge structure of weight $n$ on ${\text{\rm gr}}_n^WV_{\mathbb{C}}$, i.e., $${\text{\rm gr}}^W_nV_{\mathbb{C}}=\bigoplus_{p+q=n}F_{(n)}^p\cap{\overline{F}}_{(n)}^q,\quad
F_{(n)}^p={\frac{F^p\cap(W_n)_{\mathbb{C}}+(W_{n-1})_{\mathbb{C}}}{(W_{n-1})_{\mathbb{C}}}},\quad{\overline{F}}_{(n)}^q=\dots.$$
A mixed Hodge structure is a *mixed Hodge-Tate structure* if $V_{\mathbb{C}}^{p,q}=0$ for $p\neq q$. For the real mixed Hodge structures that are Tate, which are the ones we consider, the associated graded pure Hodge-Tate structures are trivial in odd weight. Therefore, we reindex the filtration by semiweight (so ${\mathbb{R}}(1)$ has weight $-1$, rather than $-2$).
Mixed Hodge-Tate structures are iterated extensions of the one-dimensional pure mixed Hodge-Tate structures of weight $-n$, denoted ${\mathbb{R}}(n)$. Equivalently, in the category ${\mathrm{MH}}_{\mathbb{R}}$ of real mixed Hodge-Tate structures, the subcategory of mixed Hodge-Tate structures ${\mathrm{MHT}}_{\mathbb{R}}$ is the full subcategory generated by the simple objects ${\mathbb{R}}(n)$.
The map ${\text{\rm gr}}^W$ provides a fiber functor from mixed to pure real Hodge-Tate structures: $${\text{\rm gr}}^W:{\mathrm{MHT}}_{\mathbb{R}}\to{\mathrm{HT}}_{\mathbb{R}}.$$ The Tannakian reconstruction theorem implies that there is a graded Lie algebra ${\text{\rm Lie}}_{\text{\rm HT}}$ in the category ${\mathrm{HT}}_{\mathbb{R}}$ such that ${\mathrm{MHT}}_{\mathbb{R}}$ is equivalent to the category of finite-dimensional graded ${\text{\rm Lie}}_{\text{\rm HT}}$-modules in ${\mathrm{HT}}_{\mathbb{R}}$. Specifically, ${\text{\rm Lie}}_{\text{\rm HT}}={\text{\rm Der}}^\otimes{\text{\rm gr}}^W$, the graded Lie algebra in ${\mathrm{HT}}_{\mathbb{R}}$ of tensor derivations of the functor ${\text{\rm gr}}^W$. That is, every mixed Hodge-Tate structure $X$ determines an action $${\text{\rm Lie}}_{\text{\rm HT}}\to{\text{\rm Der}}{\left({\text{\rm gr}}^WX\right)}.$$ Let ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ be the graded dual of ${\text{\rm Lie}}_{\text{\rm HT}}$.
The simple objects of the category ${\mathrm{HS}}_{\mathbb{R}}$ are ${\mathbb{R}}(n)$, and ${\text{\rm Lie}}_{\text{\rm HT}}$ is free on $$\bigoplus_{n<0}\operatorname{Ext}^1_{{\mathrm{MHT}}_{\mathbb{R}}}({\mathbb{R}}(0),{\mathbb{R}}(n))^\vee\otimes_{{\text{\rm End}}{\mathbb{R}}(n)}{\mathbb{R}}(n).$$
A *framing* of a mixed Hodge-Tate structure $V$ of weight $n$ consists of a pair of morphisms ${\mathbb{R}}(0)\to{\text{\rm gr}}_0^WV$, ${\text{\rm gr}}_{-2n}^WV\to{\mathbb{R}}(n)$. The isomorphism classes of framed real mixed Hodge-Tate structures generate a Hopf algebra ${\mathcal{H}}_\bullet$, with the structure defined by [@bgsv], which is canonically isomorphic to the dual to the universal enveloping algebra of ${\text{\rm Lie}}_{\text{\rm HT}}$. An element of ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ of weight $n$ is represented by a framed real mixed Hodge-Tate structure of weight $n$, modulo products in ${\mathcal{H}}$, that is, $${\text{\rm Lie}}_{{\text{\rm HT}}/B}^\vee\cong{\frac{{\mathcal{H}}}{{\mathcal{H}}_{>0}\cdot{\mathcal{H}}_{>0}}}.
\label{eqn:lie_quot_hopf}$$
The $\operatorname{Ext}^1({\mathbb{R}}(0),{\mathbb{R}}(n))$ are trivial for $n\geq0$ and 1-dimensional for $n<0$, in which case $$\operatorname{Ext}^1({\mathbb{R}}(0),{\mathbb{R}}(n))=({\mathbb{R}}(n)\otimes{\mathbb{C}})/{\mathbb{R}}(n)={\mathbb{R}}(n)\otimes_{\mathbb{R}}i{\mathbb{R}}.$$
According to [@goncharov-hodge-correlators], a choice of generators $n_w$ of ${\text{\rm Lie}}_{\text{\rm HT}}\otimes{\mathbb{C}}$ satisfying $n_w=-{\overline{n}}_w$ amounts to a map $${\text{\rm Lie}}_{\text{\rm HT}}^\vee\to\bigoplus_{n<0}\operatorname{Ext}^1_{{\mathrm{MHT}}_{\mathbb{R}}}({\mathbb{R}}(0),{\mathbb{R}}(n))\otimes{\mathbb{R}}(n)^\vee=\bigoplus_{n<0}{\mathbb{R}}(n)\otimes_{\mathbb{R}}i{\mathbb{R}},$$ and thus defines a canonical period map $$p:{\text{\rm Lie}}_{\text{\rm HT}}^\vee\to{\mathbb{R}}.$$ Such generators were originally defined by Deligne for the larger category of real mixed Hodge structures ([@deligne-mixed-hodge]). However, we use the different set of generators proposed by Goncharov ([@goncharov-hodge-correlators]), the *Green’s operators* $G_{w}$. They have the property that, for Hodge structures varying over a base, the Griffiths transversality condition needed to define variations of Hodge structures is expressed by a Maurer-Cartan differential equation on the $G_{w}$, which is essential for the construction of Hodge correlators. Contrary to this, the differential equations for Deligne’s generators are difficult to write.
A *variation* of real mixed Hodge-Tate structures on a complex variety $B$ is a variation of the linear data of real mixed Hodge-Tate structure that satisfies the Griffiths transversality condition. Precisely, it is a real vector bundle with flat connection $(V,\nabla)$ with a weight filtration $W_\bullet$ on $V$ and a Hodge filtration $F^\bullet$ on $V\otimes_{\mathbb{R}}{\mathcal{O}}_B$ such that $F$ and $V$ induce a real mixed Hodge-Tate structure over each point of $X$ and $\nabla^{0,1}F^p\subseteq F^{p-1}\otimes\Omega_B^1$.
A consequence of the transversality condition is that for $n>1$, $\operatorname{Ext}^1({\mathbb{R}}(0),{\mathbb{R}}(n))$ is rigid in the category of variations of mixed Hodge-Tate structures over $B$: if the coproduct of a variation of Hodge-Tate structures of weight $w>1$ is 0, then the variation is isomorphic to a constant one.
### Pronilpotent fundamental group
Let $X={\mathbb{P}}^1({\mathbb{C}})$, $S\subset X$ a finite set of punctures containing $\infty$, and $v_\infty={\frac{-1}{z^2}}{\frac{d}{dz}}$ a distringuished tangent vector at $\infty$. Let $\pi_1=\pi_1(X{\setminus}S,\infty)$ be the classical fundamental group. The group algebra $A={\mathbb{Q}}[\pi_1]$ is a free group generated by loops around the points of $S{\setminus}{\left\{\infty\right\}}$. Let ${\mathcal{I}}=\ker(A\to{\mathbb{Q}})$ be the augmentation ideal. Then form a Hopf algebra $$A^{\text{\rm nil}}(X{\setminus}S,v_\infty):=\lim_{\leftarrow}{\left(\dots\to A/{\mathcal{I}}^{n+1}\to A/{\mathcal{I}}^n\to\dots\to A\right)},$$ with coproduct defined by $g\to g\otimes g$ for $g\in\pi_1$. The subset of primitive elements is denoted $\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)$. It is actually a pronilpotent Lie algebra, the Mal’cev completion of $\pi_1$.
There is a canonical weight filtration on $H_1(X{\setminus}S,{\mathbb{Q}})$, where the loops around punctures lie in weight $-1$. This induces a weight filtration $W$ on $A^{\text{\rm nil}}$, and we have $${\text{\rm gr}}^WA^{\text{\rm nil}}(X{\setminus}S,v_\infty)=T({\text{\rm gr}}^WH_1(X{\setminus}S,{\mathbb{Q}})).$$ Furthermore, let $L_{X,S,v_\infty}$ be the free Lie algebra generated by ${\mathbb{C}}{\left[S{\setminus}{\left\{\infty\right\}}\right]}$. Then there is a canonical isomorphism $$L_{X,S,v_\infty}\cong{\text{\rm gr}}^W\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)\otimes{\mathbb{C}}.$$
There is a real mixed Hodge-Tate structure on $\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)\otimes{\mathbb{R}}$, which depends on the choice of the tangent vector $v_\infty$, and thus an action ${\text{\rm Lie}}_{\text{\rm HT}}\to{\text{\rm Der}}{\left({\text{\rm gr}}^W\pi_1^{\text{\rm nil}}(X{\setminus}S,v_\infty)\right)}$.
### Correlators in families
The construction of the Hodge correlator coalgebra (§\[sec:hodge\_cor\_coalg\]) can be performed over a base. Let $X\to B$ be a smooth family of genus 0 curves. Generalizing from the case of $B$ a point, one simply replaces the punctures $S$ by *nonintersecting* sections $s:B\to X$ and the tangential base point by a nonvanishing section $v_\infty:B\to T^1_{X/B}$ factoring through a distinguished section $s_\infty:B\to X$. This construction yields a family of coalgebras $${\left({\mathcal{CL}}^\vee_{X_t,{\left\{(s_i)_t\right\}},(v_\infty)_t}\right)}_{t\in B}.
\label{eqn:coalg_base}$$ We will denote this coalgebra by ${\mathcal{CL}}^\vee_{X/B,S,v_\infty}$ when the objects $X,S,v_\infty$ vary over $B$.
The Green’s function $(2\pi i){^{-1}}\log{\left|x-y\right|}$, used in the definition of the Hodge correlator, becomes a distribution on $X\times_BX$ with logarithmic singularities along the relative divisors $x=s_\infty$, $y=s_\infty$, and $x=y$. As we explain below, the higher-weight correlators also determine smooth variations over the base. In particular, the period map $Cor_{\mathcal{H}}:{\mathcal{CL}}^\vee_{X,S,v_\infty}\to{\mathbb{C}}$ is upgraded to a map $${\text{\rm Cor}}_{\mathcal{H}}:{\mathcal{CL}}^\vee_{X/B,S,v_\infty}\to{\mathcal{A}}^0_B,$$ and the map ${\text{\rm Cor}}_{\text{\rm Hod}}$ to a map $${\text{\rm Cor}}_{\text{\rm Hod}}:{\mathcal{CL}}^\vee_{X/B,S,v_\infty}\to{\text{\rm Lie}}_{{\text{\rm HT}}/B}^\vee$$ to the fundamental Lie coalgebra of the category of variations of real mixed Hodge-Tate structures.
The case of specialization at intersecting sections, as well as degeneration to nodal curves, is related to the behavior of the Hodge structure on $\pi_1^{\text{\rm nil}}$ at the boundary of the moduli space of Riemann surfaces with $n$ punctures. We will examine this question in §\[sec:nodal\].
As $X,S,v_0$ vary over the moduli space ${\mathcal{M}}_{0,n}'$ of Riemann surfaces of genus $0$ with $n$ distinct marked points and a tangential base point $v_0$, we get a family $\mathbf{V}$ of framed ${\mathbb{R}}$-mixed Hodge structures on $\pi_1^{\text{\rm nil}}(X{\setminus}S,s_0)$. Theorem \[thm:hc\_main\_point\] is generalized to the following.
(a) There is a flat connection on $\mathbf{V}$ making it a variation of mixed Hodge structures over ${\mathcal{M}}_{0,n}'$.
(b) This variation coincides with the standard variation of mixed Hodge structures on $\pi_1^{\text{\rm nil}}$.
\[thm:hc\_main\]
A consequence of Theorem \[thm:hc\_main\] is that the coalgebra structure on ${\mathcal{CL}}_{X,S,s_0}^\vee$ should translate into differential equations on the periods over ${\mathcal{M}}_{0,n}'$. We now describe these equations.
Extend the period map ${\text{\rm Cor}}_{\mathcal{H}}$ to a map $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}:\wedge^2{\mathcal{CL}}_{X/B,S,v_\infty}&\to{\mathcal{A}}^1_B,\\
C_1\wedge C_2&\mapsto{\frac{2w_2-1}{2(w-1)}}{\text{\rm Cor}}_{\mathcal{H}}(C_2)\,d_B^{\mathbb{C}}{\text{\rm Cor}}_{\mathcal{H}}(C_1)\\&\quad-{\frac{2w_1-1}{2(w-1)}}{\text{\rm Cor}}_{\mathcal{H}}(C_1)\,d_B^{\mathbb{C}}{\text{\rm Cor}}_{\mathcal{H}}(C_2),\end{aligned}$$ where $w_i=\operatorname{wt}C_i$ and $w=w_1+w_2$. Then we have a diagram that commutes in weight $>1$: $$\xymatrix{
{\mathcal{CL}}_{X/B,S,v_\infty}\ar[r]^\delta\ar[d]_{{\text{\rm Cor}}_{\mathcal{H}}}&\bigwedge_B^2{\mathcal{CL}}_{X/B,S,v_\infty}\ar[d]^{{\text{\rm Cor}}_{\mathcal{H}}}\\{\mathcal{A}}^0_B\ar[r]^{d_B}&{\mathcal{A}}^1_B.
}
\label{eqn:diff_eq_p1}$$
For the simplest example, consider the Hodge correlator ${\text{\rm Cor}}_{\mathcal{H}}(1,0,z)$ as $z$ varies over ${\mathbb{P}}^1{\setminus}{\left\{0,1,\infty\right\}}$. Noting that ${\text{\rm Cor}}_{\mathcal{H}}(1,0)=0$, we have $$\begin{aligned}
d{\text{\rm Cor}}_{\mathcal{H}}(1,0,z)
&={\text{\rm Cor}}_{\mathcal{H}}{\left(C(1,z)\wedge C(0,z)\right)}\\
&=(2\pi i)^{-2}{\left(\log{\left|z\right|}\,d^{\mathbb{C}}\log{\left|z-1\right|}-\log{\left|z-1\right|}\,d^{\mathbb{C}}\log{\left|z\right|}\right)}&=-{\frac{1}{2}}(2\pi i)^{-2}d{\mathcal{L}}_2(z),\end{aligned}$$ and indeed, by (\[eqn:h\_w2\_l2\]), ${\text{\rm Cor}}_{\mathcal{H}}(1,0,z)=-{\frac{1}{2}}(2\pi i)^{-2}{\mathcal{L}}_2(z)$.
We emphasize that the sections have so far required to be nonintersecting. In §\[sec:nodal\] we will prove a specialization theorem, which allows to pass to the boundary of ${\mathcal{M}}_{0,n}'$. It will imply the statement about periods:
The Hodge correlators ${\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)$ are continuous on ${\mathbb{C}}^{n+1}{\setminus}{\left\{z_0=\dots=z_n\right\}}$. \[thm:corrh\_cts\]
### Distribution relations {#sec:distr_rel}
The formula expressing how the Hodge correlators transform under endomorphisms of $X$ appears in [@goncharov-hodge-correlators], Lemma 12.3. We translate this result to our setting, showing that it gives a relation of the form (\[eqn:distr\_rel\]).
Consider the map $[l]:{\mathbb{P}}^1\to{\mathbb{P}}^1$, $z\mapsto z^l$ ($l\in{\mathbb{Z}}_{>0}$). Let $S'=[l]{^{-1}}(S)$. Then there is an induced map $$\begin{aligned}
[l]{^{*}}:{\mathcal{CL}}_{X,S,v_\infty}&\to{\mathcal{CL}}_{X,S',v_\infty},\\
(z_0\otimes\dots\otimes z_n)(1)&\mapsto{\frac{1}{l}}(z_0'\otimes\dots\otimes z_n')(1),\end{aligned}$$ where $$z_i'=\begin{cases}\sum_{y_i^l=z_i}(y_i)&z_i\neq0\\l\cdot(0)&z_i=0\end{cases}.$$ That is, each point is pulled back to the sum of its preimages, counted with multiplicity. The factor ${\frac{1}{l}}$ comes from the degree of the induced map on $H_2(X)$.
Then the diagram commutes: $$\xymatrix{
{\mathcal{CL}}_{X,S,v_\infty}\ar[r]^{[l]^*}\ar[dr]_{{\text{\rm Cor}}_{\text{\rm Hod}}}&{\mathcal{CL}}_{X,S',v_\infty}\ar[d]^{{\text{\rm Cor}}_{\text{\rm Hod}}}\\&{\text{\rm Lie}}_{\text{\rm HT}}^\vee.
}$$ For example, in weight 1, we have $${\text{\rm Cor}}_{\text{\rm Hod}}(x,y)={\frac{1}{2}}{\left({\text{\rm Cor}}_{\text{\rm Hod}}(\sqrt x,\sqrt y)+{\text{\rm Cor}}_{\text{\rm Hod}}(\sqrt x,-\sqrt y)+{\text{\rm Cor}}_{\text{\rm Hod}}(-\sqrt x,\sqrt y)+{\text{\rm Cor}}_{\text{\rm Hod}}(-\sqrt x,-\sqrt y)\right)},$$ where a branch of the square root has been chosen. On the level of periods, this becomes the equality $$\log{\left|x-y\right|}={\frac{1}{2}}{\left(\log{\left|\sqrt x-\sqrt y\right|}+\log{\left|\sqrt x+\sqrt y\right|}+\log{\left|-\sqrt x-\sqrt y\right|}+\log{\left|-\sqrt x+\sqrt y\right|}\right)}.$$
Motivic correlators over a number field
---------------------------------------
### Mixed motives
Let $F$ be a number field. There is a semisimple abelian category ${\mathcal{PM}}_F$ of Grothendieck pure motives over $F$ and a functor $H:\mathbf{SmProj}_F\to{\mathcal{PM}}_F$ assigning to every smooth projective variety over $F$ the sum of its motivic cohomology objects: $$H(X)=\bigoplus_{i=0}^{2\dim(X)}H^i(X).$$ Every Weil cohomology theory $\mathbf{SmProj}_F\to\mathbf{Vect}$ factors through $H$ and a realization functor ${\mathcal{PM}}_F{\xrightarrow{r}}\mathbf{Vect}$: $$\begin{aligned}
r_{\text{\rm Betti}}H(X)&=\bigoplus_{i=0}^{2\dim(X)}H^i_{\text{\rm Betti}}(X_{\mathbb{C}},{\mathbb{Z}})\otimes{\mathbb{Q}}&\text{(Betti)},\\
r_{\text{\rm Hod}}H(X)&=\bigoplus_{i=0}^{2\dim(X)}\bigoplus_{p+q=i}H^{p,q}_{\text{\rm Hod}}(X_{\mathbb{C}},{\mathbb{R}})&\text{(real de Rham (Hodge))},\\
r_\ell H(X)&=\bigoplus_{i=0}^{2\dim(X)}H^i_{\text{\rm\'et}}(X_{{\overline{F}}},{\mathbb{Z}}_\ell)\otimes{\mathbb{Q}}_\ell&\text{($\ell$-adic \'etale)}. \end{aligned}$$ This category is graded by the *weight*, where the weight of $H^i(X)$ is $i$. There is an invertible Tate object ${\mathbb{Q}}(1)$ of weight $-2$; we write $M(n)$ for the Tate twist $M\otimes{\mathbb{Q}}(1)^{\otimes n}$. The various realization functors respect the weight. For example, for $X$ a variety over $F$ and a fixed embedding $F\to{\mathbb{C}}$, the $r_{\text{\rm Hod}}H^i(X)$ carries a pure Hodge structure of weight $i$. For $X$ with good reduction modulo $p$, the Frobenius automorphism acts on $r_\ell H^i(X)$ with eigenvalues of norm $p^{i/2}$.
There is a conjectural category of *mixed motives* ${\mathcal{MM}}_F$ that should extend this construction to arbitrary varieties over $F$. The desired properties of ${\mathcal{MM}}_F$ were conjectured by Beilinson [@beilinson-height-pairing], see also Deligne [@deligne-motifs]. It is expected to be an abelian tensor category, in which every object has a canonical weight filtration $W_\bullet$. There should be a fiber functor ${\text{\rm gr}}^W:{\mathcal{MM}}_F\to{\mathcal{PM}}_F$ such that ${\text{\rm gr}}^W_iX$ is pure of weight $i$.
The Hodge realization of a mixed motive should be a mixed Hodge structure. Deligne [@deligne-hodge3] showed that for any complex variety $X$, there is a mixed Hodge structure on $\bigoplus H^i_{\text{\rm Hod}}(X,{\mathbb{R}})$. In this way, ${\text{\rm gr}}^W$ is a motivic lift of the associated graded functor from mixed to pure real Hodge structures: ${\text{\rm gr}}^W:{\mathrm{MH}}_{\mathbb{R}}\to{\mathrm{HS}}_{\mathbb{R}}$.
The full tensor subcategory of ${\mathcal{MM}}_F$ generated by ${\mathbb{Q}}(1)$ is the category of *mixed Tate motives* ${\mathcal{MTM}}_F$. Such a category with desirable properties has been constructed by [@deligne-goncharov]. If $X$ is a rational curve, then $H(X)$ is a mixed Tate motive. The simple objects of ${\mathcal{MTM}}_F$ are ${\mathbb{Q}}(n)={\mathbb{Q}}(1)^{\otimes n}$, $n\in{\mathbb{Z}}$, and every object of ${\mathcal{MTM}}_F$ is an iterated extension of these objects. They satisfy $$\begin{aligned}
\operatorname{Hom}({\mathbb{Q}}(m),{\mathbb{Q}}(n))&=0,\quad m<n;\\
\operatorname{Ext}^1({\mathbb{Q}}(0),{\mathbb{Q}}(n))&=\begin{cases}0&n\leq 0\\K_{2n-1}(F)\otimes{\mathbb{Q}}&n>0\end{cases},\\
\operatorname{Ext}^i({\mathbb{Q}}(0),{\mathbb{Q}}(n))&=0,\quad i>1.\end{aligned}$$ The real Hodge realizations of mixed Tate motives are mixed Hodge-Tate structures. The images of the ${\mathbb{Q}}(n)$, the real mixed Hodge-Tate structures ${\mathbb{R}}(n)$ generate the subcategory ${\mathrm{MHT}}_{\mathbb{R}}$ in ${\mathrm{MH}}_{\mathbb{R}}$.
We will consider only the mixed Tate motives. As in the Hodge realization, the associated graded objects of the weight filtration are trivial in odd weight, so we reindex the filtration by semiweight (so ${\mathbb{Q}}(1)$ has weight $-1$, rather than $-2$).
### Fundamental Lie algebra and period map
Assume the mixed motivic formalism above. The Tannakian reconstruction theorem implies that there would be a negatively graded Lie algebra ${\text{\rm Lie}}_{{{\text{\rm MT}/F}}}$ in the category ${\mathcal{PM}}_F$, the *fundamental (motivic Tate) Lie algebra*, such that ${\mathcal{MTM}}_F$ is canonically equivalent to the category of finite-dimensional graded ${\text{\rm Lie}}_{{\text{\rm MT}/F}}$-modules in ${\mathcal{PM}}_F$. That is, for any $X\in{\mathcal{MTM}}_F$, there is an action by derivations ${\text{\rm Lie}}_{{\text{\rm MT}/F}}\to{\text{\rm Der}}({\text{\rm gr}}^WX)$. We prefer to study its graded dual ${\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee$.
This Lie coalgebra breaks into isotypical components over the isomorphism classes of simple Tate objects of ${\mathcal{PM}}_F$: $${\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee=\bigoplus_{[M]\in{\mathcal{PM}}_F}{\left({\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\right)}_M\boxtimes_{{\text{\rm End}}(M)}M{^{*}}.$$ As a consequence, the cohomology of ${\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee$ can be expressed as Ext-groups in the category of mixed motives: $$H^i{\left({\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee{\xrightarrow{\delta}}\wedge^2{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee{\xrightarrow{\delta}}\dots\right)}=\bigoplus_{[M]}\operatorname{Ext}^i_{{{\text{\rm MT}/F}}}({\mathbb{Q}}(0),M)\boxtimes_{{\text{\rm End}}(M)}M{^{*}}.$$ For $F$ a number field, the $\operatorname{Ext}^i$ in ${\mathcal{MTM}}_F$ are trivial for $i>1$; equivalently, ${\text{\rm Lie}}_{\text{\rm Mot}}^\vee$ is free on the generators $\operatorname{Ext}^1_{{\mathcal{MTM}}_F}({\mathbb{Q}}(0),M)$.
Fix an embedding $r:F\to{\mathbb{C}}$. The Hodge realization functor induces a Lie coalgebra morphism $r_{\text{\rm Hod}}:{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\to{\text{\rm Lie}}_{\text{\rm HT}}^\vee$. This means that there is a period map $p\circ r_{\text{\rm Hod}}:{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee\to{\mathbb{R}}$.
For every integer $n>0$, there is the Beilinson regulator map $${\text{\rm reg}}:\operatorname{Ext}^1_{{\mathcal{MTM}}_F}({\mathbb{Q}}(0),{\mathbb{Q}}(n))\to\bigoplus_{F\to{\mathbb{C}}/\sigma}\operatorname{Ext}^1_{{\mathrm{MHT}}_{\mathbb{R}}}({\mathbb{R}}(0),{\mathbb{R}}(n)),$$ where $\sigma$ is complex conjugation. By Beilinson’s theorem ([@beilinson-regulators]) it coincides for $n>1$ with the Borel regulator on $K_{2n-1}(F)$, i.e., the diagram commutes: $$\xymatrix{
\operatorname{Ext}^1_{{\mathcal{MTM}}_F}({\mathbb{Q}}(0),{\mathbb{Q}}(n))\ar[r]^{{\text{\rm reg}}}\ar[d]&\bigoplus_{F\to{\mathbb{C}}/\sigma}\operatorname{Ext}^1_{{\mathrm{MHT}}_{\mathbb{R}}}({\mathbb{R}}(0),{\mathbb{R}}(n))\ar[d]\\K_{2n-1}(F)\otimes{\mathbb{Q}}\ar[r]^{\text{\rm reg}}&{\mathbb{R}}^{d_n(F)}
},$$ $$d_n(F)=\begin{cases}r_1(F)+r_2(F)&\text{$n$ odd},\\r_2(F)&\text{$n$ even}\end{cases}.$$
Borel’s theorem states that this regulator map – the second row in the diagram – is injective [@borel-stable-cohomology]. So there is an injective map on the first cohomology of the fundamental Lie coalgebras $$\ker({\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee{\xrightarrow{\delta}}\wedge^2{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee)
\to
\bigoplus_{F\to{\mathbb{C}}/\sigma}\ker({\text{\rm Lie}}_{{\text{\rm HT}}}^\vee{\xrightarrow{\delta}}\wedge^2{\text{\rm Lie}}_{{\text{\rm HT}}}^\vee).$$ In particular, we get the following basic theorem, which plays a crucial role in this paper:
If $x\in{\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee$ is of weight at least 2 with $\delta(x)=0$ and $p(r_{\text{\rm Hod}}(x))=0$ for every embedding $r:F\to{\mathbb{C}}$, then $x=0$. \[thm:d0h0m0\]
Specifically, we obtain Lemma \[lma:d0h0\_rational\]:
Let $X{\setminus}S$ be a rational curve over $F$. Suppose $x\in{\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^\vee$ has weight $>1$, $\delta{\text{\rm Cor}}_{\text{\rm Mot}}(x)=0$, and ${\text{\rm Cor}}_{\mathcal{H}}(r(x))=0$ for every embedding $r:F\to{\mathbb{C}}$. Then ${\text{\rm Cor}}_{\text{\rm Mot}}(x)=0$.
${\text{\rm Cor}}_{\text{\rm Mot}}(x)$ is an element of ${\text{\rm Lie}}_{{\text{\rm MT}/F}}^\vee$ with coproduct 0. The canonical period of its Hodge realization in ${\text{\rm Lie}}_{{\text{\rm HT}}}^\vee$ coincides with the correlator period ${\text{\rm Cor}}_{\mathcal{H}}(r(x))=0$. By Theorem \[thm:d0h0m0\], it is 0.
This does not hold in weight 1. For example, choose $z$ to be an element of $F$ that is not a root of unity, but has norm 1 under every complex embedding (e.g., $F={\mathbb{Q}}(i)$ and $z={\frac{1}{5}}{\left(3+4i\right)}$). Then $((0)\otimes(z))(1)$ has coproduct 0 and period $\log{\left|\sigma(z)\right|}=0$ under both of the embeddings ${\mathbb{Q}}(i){\xrightarrow{\sigma}}{\mathbb{C}}$. However, the object ${\text{\rm Cor}}_{\text{\rm Mot}}(0,z)$ is not 0 as an element of $\operatorname{Ext}^1_{{\mathcal{MTM}}/F}({\mathbb{Q}}(0),{\mathbb{Q}}(1))\cong F^\times\otimes{\mathbb{Q}}$.
### Distribution relations {#distribution-relations}
Suppose $x_i\in F$ are such that $x^l-x_i$ splits in $F$ for all $i$. Then the distribution relations from §\[sec:distr\_rel\] hold: $${\text{\rm Cor}}_{\text{\rm Mot}}(x_0,\dots,x_n)={\frac{1}{l}}\sum_{y_i^l=x_i}{\text{\rm Cor}}_{\text{\rm Mot}}(y_0,\dots,y_n),$$ where $y_i=0$ is taken with multiplicity $l$ if $x_i=0$.
Construction of the quasidihedral Lie coalgebra
===============================================
Definitions
-----------
For an abelian group $G$, we defined the Lie coalgebra $\widetilde{\mathcal{D}}(G)$ as the quotient of the tensor algebra of ${\mathbb{Q}}[G\cup{\left\{0\right\}}]$ by cyclic symmetry, first shuffle, distribution, and scaling relations.
Recall Theorem \[thm:main\_qdih\_const\]:
The second shuffles form a coideal in $\widetilde{\mathcal{D}}(G)$. The coproduct on $\widetilde{\mathcal{D}}(G)$ descends to a well-defined coproduct on ${\mathcal{D}}(G)$.
The proof of this theorem is the goal of this section.
The extra term in the scaling relation in weight 1, and the presence of terms of lower depth in the coproduct formula (\[eqn:cyc\_coproduct\]), makes the proof more difficult than that in [@goncharov-dihedral]’s construction of the dihedral Lie coalgebra. We find Theorem \[thm:main\_qdih\_const\] to be a small combinatorial miracle. Unfortunately, we do not know a simpler proof.
### Generating functions
The second shuffle relations can be expressed in a compact form in terms of generating functions. This simplifies their proof.
We package the elements of $\widetilde{\mathcal{D}}(G)$ into a generating function as follows: $${\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;t_0,\dots,t_k\big)}:=\sum_{n_i\geq0} C^*(w_0|n_0,\dots,w_k|n_k)\prod_{i=0}^kt_i^{n_i}, \label{eqn:gf_def_simple}$$ where $\prod_{i=0}^kw_i=1$ and the $t_i$ are formal variables.
We allow multisets of variables to appear in place of the $t_i$: if $S_i={\left\{t_{i,1},\dots,t_{i,d_i}\right\}}$, then $$\begin{aligned}
{\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;S_1,\dots,S_k\big)}
&=\sum_{n_i\geq0}\sum_{\substack{n_{i,j}\geq0\\\sum_{j=1}^{d_i}n_{i,j}=n_i-d_i+1}}C{^{*}}(w_0|n_0,\dots,w_k|n_k)\prod_{i=0}^k\prod_{j=1}^{d_i}t_{i,j}^{n_{i,j}}\nonumber\\
&=\sum_{n_{i,j}\geq0}C{^{*}}(x_0|N_0,\dots,x_k|N_k)\prod_{i=0}^k\prod_{j=1}^{d_i}t_{i,j}^{n_{i,j}}, \label{eqn:gf_multiset}\end{aligned}$$ where in the last expression $N_i=n_{i,1}+1+n_{i,2}+1+\dots+1+n_{i,d_j}$. The corresponding operation on the correlator coefficients is combining adjacent segments of 0s, with additional 0s being inserted between them, such as $$\substack{\displaystyle(\text{$n_{i,1}$ 0s indexed by $t_{i,1}$})\\\displaystyle(\text{$n_{i,2}$ 0s indexed by $t_{i,2}$})}\to(\text{$n_{i,1}+1+n_{i,2}$ 0s indexed by ${\left\{t_{i,1},t_{i,2}\right\}}$}).$$
There is a useful identity
$$\begin{aligned}
{\mathbf{\Lambda}^*}{\big(\dots,w,\dots\;|\;\dots,{\left\{t\right\}}\sqcup T,\dots\big)}&-{\mathbf{\Lambda}^*}{\big(\dots,w,\dots\;|\;\dots,{\left\{u\right\}}\sqcup T,\dots\big)}\nonumber\\&=(t-u){\mathbf{\Lambda}^*}{\big(\dots,w,\dots\;|\;\dots,{\left\{t,u\right\}}\sqcup T,\dots\big)}.
\label{eqn:multi_identity}\end{aligned}$$
Clear by comparing the coefficients of $t^ru^s$.
Theorem \[thm:main\_qdih\_const\] can then be expressed in terms of the generating functions:
The subspace of $\widetilde{\mathcal{D}}(G){\ldbrt_1,\dots,t_k{\,\right]\hspace{-0.02in}\right]}}$ generated by elements of the form $$\begin{aligned}
&\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}{\mathbf{\Lambda}^*}{\big(w_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)},w_0\;|\;S_{\sigma{^{-1}}(1)},\dots,S_{\sigma{^{-1}}(M_\sigma)},S_0\big)}\\
&-{\mathbf{\Lambda}^*}{\big(w_1,\dots,w_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}\;|\;S_1,\dots,S_r,S_{{\left\{r+1,\dots,r+s,0\right\}}}\big)}\\
&-{\mathbf{\Lambda}^*}{\big(w_{r+1},\dots,w_{r+s},w_{{\left\{1,\dots,r,0\right\}}}\;|\;S_{r+1},\dots,S_{r+s},S_{{\left\{1,\dots,r,0\right\}}}\big)}&=0,
\end{aligned}$$ where $$S_I=\bigsqcup_{i\in I}S_i,\quad w_I=\prod_{i\in I}w_i$$ forms a coideal.
### Coproduct
Let us write down the formula defining the coproduct (\[eqn:cyc\_coproduct\]) in terms of the elements $C{^{*}}$.
Let $C= C^*(w_0|n_0,\dots,w_k|n_k)$ and suppose $\text{\rm wt}(C)>2$. Then $$\begin{aligned}
\delta C\nonumber
=&\sum_{\rm cyc}\bigg(\sum_{i=0}^{k}\sum_{n_i'+n_i''=n_i} C^*(\underbracket{w_i\dots w_k}|n_i',w_0|n_0,\dots,w_{i-1}|n_{i-1})\wedge\nonumber\\
&\hspace{1.04in}\wedge C^*(w_{i+1}|n_{i+1},\dots,w_k|n_k,\underbracket{w_0w_1\dots w_i}|n_i'')\bigg)\label{eqn:mult_cop_1}\\
+&\sum_{\rm cyc}\bigg(\sum_{i=1}^k\sum_{\substack{n_i'+n_i''=n_i\\n_0'+n_0''=n_0+1}} C^*(w_1|n_1,\dots,w_{i-1}|n_{i-1},\underbracket{w_i\dots w_k w_0}|n_i'+n_0'')\wedge\nonumber\\
&\hspace{1.19in}\wedge C^*(\underbracket{w_0\dots w_i}|n_0'+n_i'',w_{i+1}|n_{i+1},\dots,w_k|n_k)\bigg)\label{eqn:mult_cop_2}\\
+&\sum_{i=0}^kL_i\wedge C(0,w_i),\label{eqn:mult_cop_3}\end{aligned}$$ where $$L_i=
\begin{cases}
C^*(w_0|n_0,\dots,w_i|n_i-1,\dots,w_k|n_k),&n_i>0,\\
\substack{\displaystyle\quad C^*(w_0|n_0,\dots,\underbracket{w_{i-1}w_i}|n_{i-1},w_{i+1}|n_{i+1},\dots,w_k|n_k)\\\displaystyle+ C^*(w_0|n_0,\dots,w_{i-1}|n_{i-1},\underbracket{w_iw_{i+1}}|n_{i+1},\dots,w_k|n_k)},&n_i=0
\end{cases},\label{eqn:log_cop}$$ and the sums are taken over cyclic permutations of the indices $0,\dots,k$.
If $\text{\rm wt}(C)=2$, this formula holds modulo terms of the form $C(0,a)\wedge C(0,b)$. \[lma:mult\_cop\]
Classify the terms $C'\wedge C''$ of $\delta C$ by the common point of the two resulting parts $C'$ and $C''$. Let $x_i=w_1\dots w_i$ be the point counterclockwise from the segment $w_i$. Up to cyclic symmetry, any cut is either:
(a) a cut from $x_0$ to the segment $w_i$ (between $x_{i-1}$ and $x_i$) (Fig. \[fig:cut\](a));
(b) cut from a 0 on the segment $w_0$ (between $x_k$ and $x_0$) to the segment $w_i$ (Fig. \[fig:cut\](b)).
-------------------- --------------------
 
(a) (b)
-------------------- --------------------
[1emFigure \[fig:cut\]. \[fig:cut\]]{}
We first write the terms arising from these cuts modulo elements of form $C(0,x)$.
Case (a) contributes the terms (\[eqn:mult\_cop\_1\]) and case (b) contributes the terms (\[eqn:mult\_cop\_2\]), noting that modulo elements of the form $C(0,a)$ the $C{^{*}}$ have cyclic symmetry.
Now we handle the terms (\[eqn:mult\_cop\_3\]). Let $w=n_0+n_1+\dots+n_k+k$ be the weight. Consider the $(\text{weight $w-1$})\wedge(\text{weight 1})$ terms of the coproduct.
Such elements, of form $C_{w-1}\wedge C$, fall into two cases, depending on which point is present in $C$ but not in $C_{w-1}$.
(1) 0 on the segment $w_i$ (from $x_{i-1}$ to $x_i$).
(2) $x_i$.
If $w>2$, the $C_{w-1}$ are invariant under scaling. If $w=2$, then the cyclic permutation of the arguments $w_0,\dots,w_{i-1}w_i,w_{i+1},\dots,w_k$ and $w_0,\dots,w_{i-1},w_iw_{i+1},\dots,w_k$ in (\[eqn:log\_cop\]) modifies those terms by an element of weight 1, so the expressions in (\[eqn:mult\_cop\_3\]) are determined up to $(\text{weight 1})\wedge(\text{weight 1})$.
In case (1), we have $$C_{w-1}=(w_0|n_0,\dots,x_i|n_i-1,\dots,x_k|n_k).$$ The only nonzero terms that appear are $(C_{w-1}\wedge(-C(0,x_i))$ (cut clockwise of $x_i$) and $C_{w-1}\wedge C(0,x_{i-1})$ (cut counterclockwise of $x_i$).
On the other hand, (\[eqn:mult\_cop\_1\]) produces no terms for these two cuts (they correspond to to $i=1$ and $i=k$). Thus this case contributes the terms $$C_{w-1}\wedge(C(0,x_i)-C(0,x_{i-1}))=C_{w-1}\wedge C(0,w_i),$$ which are the $n_i>0$ terms in (\[eqn:mult\_cop\_3\]).
In case (2), $$C_{w-1}= C^*(x_0|n_0,\dots,x_ix_{i+1}|n_i+n_{i+1},\dots,x_k|n_k).$$ Let $C_1'$ and $C_1''$ be the elements formed by $x_i$ and the point clockwise and counterclockwise from $x_i$, respectively. Then the resulting terms are $-C_{w-1}\wedge C_1'$ and $C_{w-1}\wedge C_1''$.
If $n_i=0$, then $C_1'=C(x_i,x_{i-1})=C{^{*}}(w_i|0,w_i{^{-1}}|0)+C(0,x_i)$, while (\[eqn:mult\_cop\_1\]) contributes $C{^{*}}(w_i{^{-1}}|0,w_i|0)\wedge C_{w-1}$. Thus we get an added term $$-C_{w-1}\wedge(C(0,x_i)-C(0,w_i)).$$
If $n_i\neq0$, then $C_1'=C(0,x_i)$, while (\[eqn:mult\_cop\_2\]) contributes 0. Thus we get a term $-C_{w-1}\wedge C(0,x_i)$.
Similarly, we get terms $C_{w-1}\wedge(C(0,x_i)+C(0,w_{i+1}))$ if $n_{i+1}=0$ and $C_{w-1}\wedge C(0,x_i)$ if $n_{i+1}>0$.
Collecting terms, the total contribution from this case is $C_{w-1}\wedge(M_i'+M_i'')$, where $$M_i'=\begin{cases}C(0,x_i)&n_i=0,\\0&n_i\neq0\end{cases},\quad
M_i''=\begin{cases}C(0,x_{i+1})&n_{i+1}=0,\\0&n_{i+1}\neq0\end{cases}.
\label{eqn:log_cop_mi}$$ Reindexing, we get exactly the $n_i=0$ terms of (\[eqn:mult\_cop\_3\]).
We remark that if a cyclic permutation is applied to the arguments in (\[eqn:mult\_cop\_1\]), so that it is written $$\begin{aligned}
&C^*(w_0|n_0,\dots,w_{i-1}|n_{i-1},\underbracket{w_i\dots w_k}|n_i')\wedge\nonumber\\
&\wedge C^*(\underbracket{w_0w_1\dots w_i}|n_i'',w_{i+1}|n_{i+1},\dots,w_k|n_k)\bigg),\end{aligned}$$ then the $n_i=0$ terms in (\[eqn:mult\_cop\_3\]) disappear.
Then there is the following formula for the coproduct of generating functions:
Suppose $k>2$ and let $X={\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;t_0,\dots,t_k\big)}$. Then $$\begin{aligned}
\delta X=\sum_{\rm cyc}\bigg(\sum_{i=0}^{k}&{\mathbf{\Lambda}^*}{\big(\underbracket{w_i\dots w_k},w_0,\dots,w_{i-1},\;|\;t_i,t_1,\dots,t_{i-1}\big)}\nonumber\\
&\wedge {\mathbf{\Lambda}^*}{\big(w_{i+1},\dots,w_k,\underbracket{w_0\dots w_i}\;|\;t_{i+1},\dots,t_k,t_i\big)}\bigg)\label{eqn:gf_cop_1}\\
+\sum_{\rm cyc}\bigg(\sum_{i=1}^k&t_0{\mathbf{\Lambda}^*}{\big(w_1,\dots,w_{i-1},\underbracket{w_i\dots w_kw_0}\;|\;t_1,\dots,t_{i-1},{\left\{t_i,t_0\right\}}\big)}\nonumber\\
&\wedge {\mathbf{\Lambda}^*}{\big(\underbracket{w_0\dots w_i},w_{i+1},\dots,w_k\;|\;{\left\{t_i,t_0\right\}},t_{i+1},\dots,t_k\big)}\label{eqn:gf_cop_2}\\
+\sum_{i=1}^k&L_i\wedge\log w_i,\label{eqn:gf_cop_3}
\end{aligned}$$ where $$\begin{aligned}
L_i=&t_i{\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;t_0,\dots,t_k\big)}\label{eqn:log_cop_gf_lr}\\
&+{\mathbf{\Lambda}^*}{\big(w_0,\dots,\underbracket{w_{i-1}w_i},w_{i+1},\dots,w_k\;|\;t_1,\dots,t_{i-1},t_{i+1},\dots,t_k\big)}\label{eqn:log_cop_gf_l}\\
&+{\mathbf{\Lambda}^*}{\big(w_0,\dots,w_{i-1},\underbracket{w_iw_{i+1}},\dots,w_k\;|\;t_1,\dots,t_{i-1},t_{i+1},\dots,t_k\big)}.\label{eqn:log_cop_gf_r}
\end{aligned}$$ If $k=2$, this formula holds modulo terms of the form $C(0,a)\wedge C(0,b)$. \[lma:cop\_of\_gf\]
Directly reinterpret Lemma \[lma:mult\_cop\] via the definition (\[eqn:gf\_def\_simple\]) by summing the expressions (\[eqn:mult\_cop\_1\]), (\[eqn:mult\_cop\_2\]), (\[eqn:mult\_cop\_3\]) over choices of ${\left\{n_i\right\}}_{i=0}^k$ taken with a monomial $\prod_it_i^{n_i}$.
The expressions (\[eqn:mult\_cop\_1\]) and (\[eqn:mult\_cop\_2\]) yield (\[eqn:gf\_cop\_1\]) and (\[eqn:gf\_cop\_2\]) in an obvious manner.
The $n_i>0$ cases in (\[eqn:log\_cop\]) give the terms with (\[eqn:log\_cop\_gf\_lr\]), and the $n_i=0$ cases give (\[eqn:log\_cop\_gf\_l\])-(\[eqn:log\_cop\_gf\_r\]).
We also remark that if a cyclic permutation is applied to the arguments in (\[eqn:gf\_cop\_1\]), so that it is written $$\begin{aligned}
&{\mathbf{\Lambda}^*}{\big(w_0,\dots,w_{i-1},\underbracket{w_i\dots w_k}\;|\;t_1,\dots,t_{i-1},t_i\big)}\nonumber\\
&\wedge {\mathbf{\Lambda}^*}{\big(\underbracket{w_0\dots w_i},w_{i+1},\dots,w_k\;|\;t_i,t_{i+1},\dots,t_k\big)}\bigg)\end{aligned}$$ then the terms (\[eqn:log\_cop\_gf\_l\]) and (\[eqn:log\_cop\_gf\_r\]) disappear.
### Dual generating function and homogeneity
For a more complete analogy with the generating functions $L, L{^{*}}$ for multiple polylogarithms (§\[sec:zeta\_shuffle\]), we define a dual generating function ${\mathbf{\Lambda}}$: $$\begin{aligned}
{\mathbf{\Lambda}}{\big(x_0,\dots,x_k\;|\;t_0,\dots,t_k\big)}&:=\sum_{n_i\geq0} C(x_0,\underbrace{0,\dots,0}_{n_0},x_1\dots,x_k,\underbrace{0,\dots,0}_{n_k})\prod_{i=0}^k(t_0+\dots+t_i)^{n_i},\label{eqn:hgf_def_simple}\end{aligned}$$ where the formal variables $t_i$ satisfy the relation $\sum_{i=0}^kt_i=0$. The pair of generating functions ${\mathbf{\Lambda}^*},{\mathbf{\Lambda}}$ resemble those used by [@goncharov-dihedral] in the definition of the dihedral Lie coalgebra.
The duality is made clear by the following statement:
(a) The generating functions are related by $${\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;t_0,\dots,t_k\big)}={\mathbf{\Lambda}}{\big(1,w_0,\dots,w_0\dots w_{k-1}\;|\;t_0,t_1-t_0,\dots,t_k-t_{k-1}\big)}.$$
(b) For $k>1$, the generating functions ${\mathbf{\Lambda}^*}$ are homogeneous in the $t_i$ (invariant under a shift $t_i\mapsto t_i+t$), and the ${\mathbf{\Lambda}}$ are homogeneous in the $x_i$ (invariant under a shift $x_i\mapsto x_i\cdot x$).
(c) Both generating functions are invariant under cyclic permutation of the indices.
Part (a) is clear from the definitions.
For ${\mathbf{\Lambda}^*}$, (c) is clear from the scaling relations imposed in $\widetilde{\mathcal{D}}(G)$. For ${\mathbf{\Lambda}}$, (b) is also immediate. Part (c) for ${\mathbf{\Lambda}}$ would follow easily from (a) and (b,c) for ${\mathbf{\Lambda}^*}$, recalling that $t_1+\dots+t_k=0$.
The nontrivial part is (b) for ${\mathbf{\Lambda}^*}$. We must show $${\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;t_0+t,\dots,t_k+t\big)}
=
{\mathbf{\Lambda}^*}{\big(w_0,\dots,w_k\;|\;t_0,\dots,t_k\big)}.$$ Consider the coefficient of $t^n\cdot\prod_it_i^{n_i}$ on each side. If $k=0$, the coefficients on both sides are equal. If $k>0$, the coefficient on the left side is precisely a first shuffle relation (where the $n$ 0s indexed by the variable $t$ are shuffled with all other points, with the point 1 remainining fixed), while the right side is 0.
The first shuffle relation imposed in $\widetilde{\mathcal{D}}(G)$ can be expressed in terms of the ${\mathbf{\Lambda}}$:
The generating functions ${\mathbf{\Lambda}}$ obey a shuffle relation for $r,s>1$: $$\sum_{\sigma\in\Sigma_{r,s}}{\mathbf{\Lambda}}{\big(x_{\sigma{^{-1}}(1)},\dots,x_{\sigma{^{-1}}(r+s)},x_0\;|\;t_{\sigma{^{-1}}(1)},\dots,t_{\sigma{^{-1}}(r+s)},t_0\big)}=0.
\label{eqn:first_shuffle_corr}$$ \[lma:first\_shuf\_gf\]
Similar to the previous lemma. It follows from the shuffle relation on the coefficients, where we fix $x_0$ and shuffle the $x_1,\dots,x_r$ and the zeros indexed by $t_1,\dots,t_r$ with the other points.
Proof of Theorem \[thm:main\_qdih\_const\]
------------------------------------------
### Summary of the proof
The proof of the Theorem \[thm:main\_qdih\_const\] will be by induction on the depth of the second shuffles.
Define $$\begin{aligned}
{{\rm QSh}}^{r,s}(w_1|S_1,\dots,w_n|S_n,w_0|S_0)=\\
=\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}{\mathbf{\Lambda}^*}{\big(w_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)},w_0\;|\;S_{\sigma{^{-1}}(1)},\dots,S_{\sigma{^{-1}}(M_\sigma)},S_0\big)},\end{aligned}$$ where $w_i\in G$ with $\prod_iw_i=1$, and $$\begin{aligned}
{\overline{{{\rm QSh}}}}^{r,s}(w_1|S_1,\dots,w_n|S_n,w_0|S_0)&=
{{\rm QSh}}^{r,s}(w_1|S_1,\dots,w_n|S_n,w_0|S_0)\label{eqn:def_olqsh_q}\\
&-{\mathbf{\Lambda}^*}{\big(w_1,\dots,w_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}\;|\;S_1,\dots,S_r,S_{{\left\{r+1,\dots,r+s,0\right\}}}\big)}\label{eqn:def_olqsh_ra}\\
&-{\mathbf{\Lambda}^*}{\big(w_{r+1},\dots,w_{r+s},w_{{\left\{1,\dots,r,0\right\}}}\;|\;S_{r+1},\dots,S_{r+s},S_{{\left\{1,\dots,r,0\right\}}}\big)}.\label{eqn:def_olqsh_rb}\end{aligned}$$ We must show that the elements ${\overline{{{\rm QSh}}}}$ form a coideal, i.e., their coproducts vanish modulo other elements of this form.
To make the notation more transparent, when $r$ and $s$ are fixed, we will relabel $$\begin{aligned}
T_1,\dots,T_r&=S_1,\dots,S_r,\\
U_1,\dots,U_s&=S_{r+1},\dots,S_{r+s},\\
V&=S_0,\\
a_1,\dots,a_r&=w_1,\dots,w_r\\
b_1,\dots,b_s&=w_1,\dots,w_{r+s},\\
c&=w_0,\end{aligned}$$ so that we consider elements $${\overline{{{\rm QSh}}}}^{r,s}(a_1|T_1,\dots,a_r|T_r,b_1|U_1,\dots,b_s|U_s,c|V).$$
The main steps will be the following:
1. Fix the $a_i$ and $b_j$. Show that it suffices to assume ${\left|T_i\right|}={\left|U_i\right|}={\left|V\right|}=1$. Denote the three terms (\[eqn:def\_olqsh\_q\]), (\[eqn:def\_olqsh\_ra\]), (\[eqn:def\_olqsh\_rb\]) by $Q$, $R_A$, and $R_B$, respectively.
2. Show that $\delta(Q-R_A-R_B)$ is zero modulo shuffle relations of lower depth and elements of the form $C(0,x)$ (Lemma \[lma:pf\_lwsr\]).
(a) Group the terms of $\delta Q$ according to a combinatorial classification and reduce them using shuffle relations of lower depth (Lemma \[lma:pf\_dq\]).
(b) Group the terms of $\delta(R_A)$ and $\delta(R_B)$ in the same way and show that they coincide with the terms found in (a) (Lemma \[lma:pf\_dr\]).
3. Show that the (weight 1)$\wedge$(weight $\geq 1$) component of $\delta(Q-R_A-R_B)$ is 0, modulo shuffle relations of lower depth (Lemma \[lma:pf\_log\]).
Throughout the proof, in a term ${\mathbf{\Lambda}^*}{\big(w_1,\dots,w_k,w_0\;|\;s_1,\dots,s_k,s_0\big)}$ appearing in the definition of ${\overline{{{\rm QSh}}}}$, call the segment ${\big(w_0\;|\;s_0\big)}$ the *distinguished segment* (i.e., ${\big(c\;|\;v\big)}$ in (\[eqn:def\_olqsh\_q\]) and the collapsed segments in (\[eqn:def\_olqsh\_ra\]) and (\[eqn:def\_olqsh\_rb\])). In the following lemmas, we will always use the following classification of terms of the coproduct of a generating function (see Fig. \[fig:cut\_class\]).
(1) Terms $g\wedge h$ where one of the parts $g$ or $h$ contains the distinguished segment (i.e., the distinguished segment is not cut). In this case, we always write the term in the form $\pm g\wedge h$, where $g$ contains the distinguished segment.
(a) Cut from a point $x_i$ to the segment ${\big(w_j\;|\;s_j\big)}$ ($0\leq i<j\leq k$).
(b) Cut from a point $x_j$ to the segment ${\big(w_{i+1}\;|\;s_{i+1}\big)}$ ($0\leq i<j\leq k$).
(c) Cut from a 0 on the segment ${\big(x_{i+1}\;|\;t_{s+1}\big)}$ to the segment ${\big(w_j\;|\;t_j\big)}$ ($0\leq i<j\leq k$).
(d) Cut from a 0 on the segment ${\big(x_j\;|\;s_j\big)}$ to the segment ${\big(w_{i+1}\;|\;t_{i+1}\big)}$ ($0\leq i<j\leq k$).
(2) Terms $g\wedge h$ where the distinguished segment is cut. In this case, we always write $\pm g\wedge h$, where $g$ contains the point $x_0$ and $h$ the point $x_k$.
(a) Cut from a point $x_i$ to the distinguished segment.
(b) Cut from a 0 on the segment ${\big(w_i\;|\;s_i\big)}$ to the distinguished segment ($0<i<k$).
(c) Cut from a 0 on the distinguished segment to the segment ${\big(s_i\;|\;t_i\big)}$ ($0<i<k$).
--------------------- --------------------- --------------------- ---------------------
   
(1a) (1b) (1c) (1d)
--------------------- --------------------- --------------------- ---------------------
--------------------- --------------------- ---------------------
  
(2a) (2b) (2c)
--------------------- --------------------- ---------------------
[1emFigure \[fig:cut\_class\]. \[fig:cut\_class\]]{}
### Step 0
As stated in Step 0 above, we fix $m>0$ and $n>0$, the $a_i$, $b_j$, $c$, and the $T_i$, $U_j$, $V$, and let $Q,R_A,R_B$ be the three terms of the expression defining ${\overline{{{\rm QSh}}}}$: (\[eqn:def\_olqsh\_q\]), (\[eqn:def\_olqsh\_ra\]), and (\[eqn:def\_olqsh\_rb\]), respectively.
We may assume $T_i={\left\{t_i\right\}}$, $U_j={\left\{u_j\right\}}$, and $V={\left\{v\right\}}$, by the following:
\[lma:pf\_red\_multi\] The shuffle relations for ${\left|T_i\right|}={\left|U_j\right|}={\left|V\right|}=1$ imply the shuffle relations for general index sets.
Obvious by induction using (\[eqn:multi\_identity\]).
\[lma:pf\_lwsr\] Modulo shuffle relations of lower depth and elements $C(0,x)$, $\delta(Q-R_A-R_B)=0$.
Modulo lower-depth shuffle relations and terms $C(0,x)\wedge C(0,y)$, $$\delta(Q-R_A-R_B)={\left[\sum_{i=1}^mC(0,a_i)(t_i-v)+\sum_{j=1}^nC(0,b_j)(u_j-v)\right]}\wedge(Q-R_A-R_B).
\label{eqn:step2}$$ \[lma:pf\_log\]
### Proof of Step 1
\[lma:pf\_dq\] Modulo shuffle relations of lower depth and elements $C(0,x)$, $\delta Q$ is given by the sum of expressions
(\[eqn:dq\_2\_total2\])-(\[eqn:dq\_sum\_ffgg\]) below.
Group all terms of $\delta Q$ by the type of cut as defined in the outline above. Some computational lemmas will simplify the contributions to $\delta Q$ coming from the cuts of each type. The contribution of cuts (1a/b/c/d) is computed in Lemma \[lma:pf\_dq\_type1\], and cuts (2a/b/c) are dealt with in Lemma \[lma:pf\_dq\_type2\].
\[lma:pf\_dq\_type1\] The contribution of cuts of type (1a/b/c/d) to $\delta Q$, modulo shuffle relations of lower depth and elements $C(0,x)$, is given by (\[eqn:dq\_1\_total\]) below.
The cuts of types (1a) and (1b) contribute terms of the form (\[eqn:gf\_cop\_1\]), while cuts of types (1c) and (1d) contribute terms of the form (\[eqn:gf\_cop\_2\]) below.
Consider the upper parts $g$ of terms $\pm g\wedge h$ as shown in Fig. \[fig:cut\_class\]; by cyclic invariance modulo $C(0,x)$ we may write $$g={\mathbf{\Lambda}^*}{\big(w_1,\dots,w_l,c\;|\;S_1,\dots,S_l,V\big)}.$$ Let ${\big(w_p\;|\;S_p\big)}$ be the new segment arising from the cut (that is, the bracketed segment in (\[eqn:gf\_cop\_1\]) or (\[eqn:gf\_cop\_2\])).
We say that $a_i$ appears in $g$ if either the segment ${\big(a_i\;|\;t_i\big)}$ or some ${\big(a_ib_j\;|\;{\left\{t_i,u_j\right\}}\big)}$ is present in $g$ as one of the ${\big(w_l\;|\;S_l\big)}$ ($l\neq i$), and similarly for $b_j$. Then the set of segments that *do not* appear in $g$ (“appear below g”) is determined by the $w_1,\dots,\widehat{w_p},\dots,w_l$ and consists of consecutively indexed elements $a_i$ and $b_j$, i.e., $a_{i_0},\dots,a_{i_1}$ and $b_{j_0},\dots,b_{j_1}$, where by convention $i_0=i_1+1$ if no $a_i$ appear, and likewise for $j_0,j_1$.
Group the terms $g\wedge h$ by the sequence of segments $w_1,\dots,\widehat{w_p},\dots,w_l$. To shorten notation, write
$${\widetilde{g}{\left(S\right)}}={\mathbf{\Lambda}^*}{\big(w_1,\dots,w_p,\dots,w_l,c\;|\;S_1,\dots,S_p=S,\dots,S_l,V\big)}.$$
There are three cases:
(1) $i_1-i_0>0$ and $j_1-j_0>0$: at least two $a_i$ and two $b_j$ appear below $g$ (Lemma-Computation \[lma:dq\_1\_case1\]).
(2) $i_1-i_0=-1$ or $j_1-j_0=-1$: only $a_i$’s or only $b_j$’s appear below $g$ (Lemma-Computation \[lma:dq\_1\_case2\]).
(3) $i_1-i_0=0$ or $j_1-j_0=0$: only one $a_i$ or only one $b_j$ appear below $g$ (Lemma-Computation \[lma:dq\_1\_case3\]).
We compute the contribution of each case in the next three lemmas.
*Case 1* ($i_1-i_0>0$ and $j_1-j_0>0$) contributes 0 to $\delta Q$. \[lma:dq\_1\_case1\]
Consider a term $g\wedge h$ coming from a cut in Case 1.
Let $i_0'\geq i_0$ be minimal such that $a_{i_0'}$ appears in $h$, and $i_1'\leq i_1$ be maximal such that $a_{i_1'}$ appears in $h$. Define $j_0',j_1'$ in the analogous way. For example, for cuts of type (1a), $i_0'=i_0$; for cuts of type (1c), $$i_0'=\begin{cases}
i_0&\text{if ${\big(w\;|\;S\big)}$ is ${\big(b_{i_0}\;|\;u_{i_0}\big)}$}\\
i_0+1&\text{if ${\big(w\;|\;S\big)}$ is ${\big(a_{i_0}\;|\;t_{i_0}\big)}$ or ${\big(a_{i_0}b_{i_0}\;|\;{\left\{t_{i_0},u_{j_0}\right\}}\big)}$}
\end{cases}
,$$ where ${\big(w\;|\;S\big)}$ is the segment that contains the vertex of the cut.
Notice that $i_0'-i_0\leq 1$ and $j_0-j_0'\leq 1$, and $i_1-i_0>0$ implies $i_1'-i_0'\geq-1$.
Group all terms of $\delta Q$ coming from Case 1 by the type of cut and by $i_0',j_0',i_1',j_1'$. These groups can be expressed in terms of $$\begin{aligned}
{\widetilde{g}{\left(S_1\right)}}\wedge{{\rm QSh}}{\big(&a_{i_0'},\dots,a_{i_1'},b_{j_0'},\dots,b_{j_1'},{\left(a_{i_0'}\dots a_{i_1'}\cdot b_{j_0'}\dots b_{j_1'}\right)}{^{-1}}\;|\;\\&t_{i_0'},\dots,t_{i_1'},u_{j_0'},\dots,u_{j_1'},S_2\big)}\end{aligned}$$ for some $S_1,S_2$. Indeed, the arrangements of segments that may occur in the lower part of the cut, given $i_0,j_0$ and $i_1',j_1'$, are precisely the quasishuffles. Applying the lower-weight shuffle relations, this expression becomes $$\begin{aligned}
{\widetilde{g}{\left(S_1\right)}}&\wedge{\left({\mathbf{\Lambda}^*}{\big(a_{i_0'},\dots,a_{i_1'},{\left(a_{i_0'}\dots a_{i_1'}\right)}{^{-1}}\;|\;t_{i_0'},\dots,t_{i_1'},{\left\{u_{j_0'},\dots,u_{j_1'}\right\}}\sqcup S_2\big)}\right)}\nonumber\\
+\,{\widetilde{g}{\left(S_1\right)}}&\wedge{\left({\mathbf{\Lambda}^*}{\big(b_{j_0'},\dots,b_{j_1'},{\left(b_{j_0'}\dots b_{j_1'}\right)}{^{-1}}\;|\;u_{j_0'},\dots,u_{j_1'},{\left\{t_{i_0'},\dots,t_{i_1'}\right\}}\sqcup S_2\big)}\right)}\label{eqn:fafb}.\end{aligned}$$
Fix $i_0',i_1',j_0',j_1'$, and introduce the notation $$\begin{aligned}
{\widetilde{f}}_A(i_0',i_1',S)&={\mathbf{\Lambda}^*}{\big(a_{i_0'},\dots,a_{i_1'},{\left(a_{i_0'}\dots a_{i_1'}\right)}{^{-1}}\;|\;t_{i_0'},\dots,t_{i_1'},{\left\{u_{j_0+1},\dots,u_{j_1-1}\right\}}\sqcup S\big)},\\
{\widetilde{f}}_B(j_0',j_1',S)&={\mathbf{\Lambda}^*}{\big(b_{j_0'},\dots,b_{j_1'},{\left(b_{j_0'}\dots b_{j_1'}\right)}{^{-1}}\;|\;u_{j_0'},\dots,u_{j_1'},{\left\{t_{i_0+1},\dots,t_{i_1-1}\right\}}\sqcup S\big)}.\end{aligned}$$ The expressions in (\[eqn:fafb\]) can be rewritten with ${\widetilde{f}}_A$ and ${\widetilde{f}}_B$.
Now let us collect these terms coming from different cuts and show that they yield 0. By symmetry, it suffices to show this for three kinds of terms ${\widetilde{f}}_A(i_0',i_1',j_0',j_1',S_2)$: where $i_0'=i_0$ and $i_1'=i_1$; where $i_0'=i_0$ and $i_1'=i_1-1$; and where $i_0'=i_0+1$ and $i_1'=i_1-1$.
Look at the terms with $i_0'=i_0$ and $i_1'=i_1$ (all $a_i$ that are not in $g$ are in $f_A$). They arise from cuts (1a) and (1b) where the cut segment is $b_{i_0+1}$ or $b_{i_1-1}$ and from cuts (1c) and (1d) where the cut segment and the segment containing the vertex are $b_{i_0+1}$ and $b_{i_1-1}$, or vice versa. These cases give: $$\begin{aligned}
-{\widetilde{g}{\left(u_{j_1}\right)}}&\wedge{\widetilde{f}}_A(i_0,i_1,{\left\{u_{j_0}\right\}}\sqcup{\left\{u_{j_1}\right\}}),\\
{\widetilde{g}{\left(u_{j_0}\right)}}&\wedge{\widetilde{f}}_A(i_0,i_1,{\left\{u_{j_1}\right\}}\sqcup{\left\{u_{j_0}\right\}}),\\
(u_{j_1}-u_{j_0}){\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}}_A(i_0,i_1,{\left\{u_{j_0},u_{j_1}\right\}}),\end{aligned}$$ the sum of which is 0 by (\[eqn:multi\_identity\]).
The terms with $i_0'=i_0$ and $i_1'=i_1-1$ (all $a_i$ that are not in $g$, except the last, are in $f_A$) come from three sources:
- cuts of type (1a) where the cut segment $x_2$ is either $a_{i_1}$ or $a_{i_1}b_{j_1}$;
- cuts of type (1c) and (1d) where the segment $x_1$ containing the vertex and the segment $x_2$ that is cut are $b_{j_0}$ and $a_{i_1}$, or vice versa;
- cuts of type (1c) and (1d) where the segment $x_1$ containing the vertex and the segment $x_2$ that is cut are $b_{j_0}$ and $a_{i_1}b_{j_1}$, or vice versa.
A similar computation shows their total contribution is 0.
Finally, consider terms with $i_0'=i_0+1$ and $i_1'=i_1-1$ (all $a_i$ not in $g$ except the first and last are in $f_A$). They arise from cuts of type (1c) and (1d), where the segment $x_1$ is either $a_{i_0}$ or $a_{i_0}b_{j_0}$ and the segment $x_2$ is either $a_{i_1}$ or $a_{i_1}b_{j_1}$, yielding four cases: $$(x_1,x_2)=(a_{i_0},a_{j_1}),(a_{i_0}b_{j_0},a_{j_1}),(a_{i_0},a_{j_1}b_{j_1}),(a_{i_0}b_{j_0},a_{i_1}b_{j_1}).$$ The sum of their contributions is also 0.
\[lma:dq\_1\_case2\] The contribution of Case 2 ($i_1-i_0=-1$) to the $\delta Q$ is given by the sum of expressions (\[eqn:dq\_1a\_allb\])-(\[eqn:dq\_1d\_allb\]) below.
Suppose that $i_1-i_0=-1$. The cuts of types (1a), (1b), (1c), and (1d) contribute $$\begin{aligned}
-{\widetilde{g}{\left(u_{j_1}\right)}}&\wedge{\widetilde{f}_B(j_0,j_1-1,u_{j_1})}\label{eqn:dq_1a_allb},\\
{\widetilde{g}{\left(u_{j_0}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1,u_{j_0})}\label{eqn:dq_1b_allb},\\
-u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})}\label{eqn:dq_1c_allb},\\
u_{j_1}{\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})},\label{eqn:dq_1d_allb}\end{aligned}$$
respectively.
By symmetry, analogous expressions will result if $j_1-j_0=-1$.
\[lma:dq\_1\_case3\] The contribution of Case 3 ($i_1-i_0=0$) to the $\delta Q$ is given by the sum of expressions (\[eqn:onea\_contrib3\]) and (\[eqn:onea\_contrib\]) below.
Suppose $i_1-i_0=0$, so only one segment $a_i$ occurs below $g$.
If $j_1-j_0=0$, then it is easy to see that only cuts of type (1a) and (1b) contribute nonzero terms, and that the (1a) terms cancel with the (1b) terms. So assume $j_1-j_0>0$.
The cuts of type (1a) fall into three classes depending on which segment is cut: (i) $a_{i_0}$, (ii) $b_{j_1}$, or (iii) $a_{i_0}b_{j_1}$. The first two contribute $$\begin{aligned}
-{\widetilde{g}{\left(t_{i_0}\right)}}&\wedge{\widetilde{f}_B(j_0,j_1,t_{i_0})}\label{eqn:dq_1a_onea_ca},\\
-{\widetilde{g}{\left(u_{j_1}\right)}}&\wedge{{\rm QSh}}^{1,j_1-j_0}{\big(a_{i_0},b_{j_0},\dots,b_{j_1-1},{\left(a_{i_0}b_{j_0}\dots b_{j_1-1}\right)}{^{-1}}\;|\;\nonumber\\&\quad\quad\quad\quad\quad\quad t_{i_0},u_{j_0},\dots,u_{j_1-1},u_{j_1}\big)}\nonumber\\
\equiv-{\widetilde{g}{\left(u_{j_1}\right)}}&\wedge\biggl({\widetilde{f}_B(j_0,j_1-1,{\left\{t_{i_0},u_{j_1}\right\}})}
\nonumber\\&\quad\quad+{\mathbf{\Lambda}^*}{\big(a_{i_0},a_{i_0}{^{-1}}\;|\;t_{i_0},{\left\{u_{j_0},\dots,u_{j_1}\right\}}\big)}\biggr)\label{eqn:dq_1a_onea_cb}, \end{aligned}$$
respectively, where we have used that the sequences that may occur in the lower part of the cut are precisely the shuffles of $a_i$ and $b_j$ appearing below $g$, except the cut segment $b_{j_1}$. Finally, the third class gives $$\begin{aligned}
{\frac{1}{t_{i_0}-u_{j_1}}}\biggl(
{\widetilde{g}{\left(t_{i_0}\right)}}&\wedge{\widetilde{f}_B(j_0,j_1-1,t_{i_0})}\nonumber\\
-{\widetilde{g}{\left(u_{j_1}\right)}}&\wedge{\widetilde{f}_B(j_0,j_1-1,u_{j_1})}\biggr)\label{eqn:dq_1a_onea_cab},\end{aligned}$$
where we have applied (\[eqn:multi\_identity\]) to break the generating functions with ${\left\{t_{i_0},u_{j_1}\right\}}$ into ones with only $t_{i_0}$ or $u_{j_1}$.
The cuts of type (1c) fall into five classes, depending on the segment where the vertex of the cut lies and the segment that is cut: (i) vertex on $a_{i_0}$ and $b_{j_1}$ is cut, (ii) vertex on $b_{j_0}$ and $b_{j_1}$ is cut, (iii) vertex on $b_{j_0}$ and $a_{i_0}$ is cut, (iv) vertex on $b_{j_0}$ and $a_{i_0}b_{j_1}$ is cut, (v) vertex on $a_{i_0}b_{j_0}$ and $b_{j_1}$ is cut. They contribute the following terms: $$\begin{aligned}
-t_{i_0}{\widetilde{g}{\left({\left\{t_{i_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0,j_1-1,{\left\{t_{i_0},u_{j_1}\right\}})}\label{eqn:dq_1c_onea_zacb},\\
-u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge {{\rm QSh}}^{1,j_1-j_0-1}{\big(a_{i_0},b_{j_0+1},\dots,b_{j_1-1},{\left(a_{i_0}b_{j_0+1}\dots b_{j_1-1}\right)}{^{-1}}\;|\;\nonumber\\&\quad\quad\quad\quad\quad\quad\quad t_{i_0},u_{j_0+1},\dots,u_{j_1-1},{\left\{u_{j_0},u_{j_1}\right\}}\big)}\nonumber\\
\equiv-u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge\biggl({\widetilde{f}_B(j_0+1,j_1-1,{\left\{t_{i_0},u_{j_0},u_{j_1}\right\}})}\nonumber\\&\quad\quad+{\mathbf{\Lambda}^*}{\big(a_{i_0},a_{i_0}{^{-1}}\;|\;t_{i_0},{\left\{u_{j_0},\dots,u_{j_1}\right\}}\big)}\biggr)
\label{eqn:dq_1c_onea_zbcb},\\
-u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},t_{i_0}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1,{\left\{u_{j_0},t_{j_0}\right\}})}\label{eqn:dq_1c_onea_zbca},\\
{\frac{1}{t_{i_0}-u_{j_1}}}\biggl(u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},t_{i_0}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},t_{i_0}\right\}})}\nonumber\\
-u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0};u_{j_1}\right\}})}\label{eqn:dq_1c_onea_zbcab}\biggr),\\
{\frac{1}{t_{i_0}-u_{j_0}}}\biggl(t_{i_0}{\widetilde{g}{\left({\left\{t_{i_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{t_{i_0},u_{j_1}\right\}})}\nonumber\\
-u_{j_0}{\widetilde{g}{\left({\left\{u_{j_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})}\label{eqn:dq_1c_onea_zabcb}\biggr).\end{aligned}$$
The cuts (1b) and (1d) contribute antisymmetric terms, i.e., $u_{j_0}$ and $u_{j_1}$ are exchanged and ${\widetilde{f}_B(j_0+d_0,j_1-d_1,S)}$ becomes $-{\widetilde{f}_B(j_0+d_1,j_1-d_0,S)}$. The entire contribution of case 3 is then the
symmetrization of the sum of expressions (\[eqn:dq\_1a\_onea\_ca\])-(\[eqn:dq\_1c\_onea\_zabcb\]).
The expression (\[eqn:dq\_1a\_onea\_ca\]) with its symmetrization cancels to 0.
The remaining terms form the contribution of Case 3, and are simplified to $$\begin{aligned}
{\widetilde{g}{\left({\left\{t_{i_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0,j_1-1,u_{j_1})}
-{\widetilde{g}{\left({\left\{t_{i_0},u_{j_0}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1,u_{j_0})}
\label{eqn:onea_contrib3}\\
-(u_{j_1}-u_{j_0}){\widetilde{g}{\left({\left\{t,u_{j_0},u_{j_1}\right\}}\right)}}&\wedge{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})}.\label{eqn:onea_contrib}\end{aligned}$$
Analogous expressions result if $j_1-j_0=0$.
Let us collect the terms obtained from cases 2 and 3: (\[eqn:dq\_1a\_allb\])-(\[eqn:dq\_1d\_allb\]), (\[eqn:onea\_contrib3\]), and (\[eqn:onea\_contrib\]).
Consider first the expressions of the form ${\widetilde{f}_B(j_0,j_1-1,u_{j_1})}$, arising from (\[eqn:dq\_1a\_allb\]) and (\[eqn:onea\_contrib3\]). (The notation $\thhh$, which by definition depends on $i_0$ and $i_1$, is unambiguous here since no $a_i$ appear in the expression for $\thhh$ when $i_1-i_0\leq0$.) We claim that for fixed $j_0$ and $j_1$, the sum of these terms over all $g$ is precisely $$\begin{aligned}
-{{\rm QSh}}^{m,n-(j_1-j_0)}{\big(&a_1,\dots,a_r,b_1,\dots,\underbracket{b_{j_0}\dots b_{j_1}},\dots,b_s,c\;|\;\nonumber\\&t_1,\dots,t_r,u_1,\dots,u_{j_1},\dots,u_s,v\big)}
\quad\quad\quad\wedge\,{\widetilde{f}_B(j_0,j_1-1,u_{j_1})}.
\label{eqn:qsh_red_1a}\end{aligned}$$ Indeed, the term that appears on the left side for a fixed $g$ is $-{\widetilde{g}{\left(u_{j_1}\right)}}$ if $i_1-i_0=-1$ and ${\widetilde{g}{\left({\left\{t_{i_0},u_{j_1}\right\}}\right)}}$ if $i_1-i_0=0$
. The quasishuffles for which the underlined segment collides with no $a_i$ provides the terms with $i_1-i_0=-1$, while the quasishuffles for which the underlined segment collides with some $a_i$ provide the terms with $i_0=i_1=i$.
In a similar way, the expressions with $u_{j_0}{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})}$, coming from (\[eqn:onea\_contrib\]) and (\[eqn:dq\_1c\_allb\]), yield $$\begin{aligned}
-u_0{{\rm QSh}}^{m,n-(j_1-j_0)}{\big(&a_1,\dots,a_r,b_1,\dots,\underbracket{b_{j_0}\dots b_{j_1}},\dots,b_s,c\;|\;\nonumber\\&t_1,\dots,t_r,u_1,\dots,{\left\{u_{j_0},u_{j_1}\right\}},\dots,u_s,v\big)}
\quad\quad\quad\wedge\,{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})}.
\label{eqn:qsh_red_1c}\end{aligned}$$ The expressions with ${\widetilde{f}_B(j_0+1,j_1,u_{j_1})}$ and $u_{j_1}{\widetilde{f}_B(j_0+1,j_1-1,{\left\{u_{j_0},u_{j_1}\right\}})}$ give the antisymmetric terms.
Applying the shuffle relations of lower depth to (\[eqn:qsh\_red\_1a\]) and (\[eqn:qsh\_red\_1c\]), we get the total contribution of cases 2 and 3 for fixed $j_0$ and $j_1$: $$\begin{aligned}
&-\biggl({\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,b_1\dots b_s\cdot c\;|\;t_1,\dots,t_r,{\left\{u_1,\dots,u_{j_0-1},u_{j_1},u_{j_1+1},\dots,u_s,v\right\}}\big)}\nonumber\\
&\quad\quad+{\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_0}\dots b_{j_1},\dots,b_s,a_1\dots a_r\cdot c\;|\;u_1,\dots,u_{j_1},\dots,u_s,{\left\{t_1,\dots,t_r,v\right\}}\big)}\biggr)\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0},b_{j_0+1},\dots,b_{j_1-1},{\left(b_{j_0}\dots b_{j_1-1}\right)}{^{-1}}\;|\;u_{j_0},u_{j_0+1},\dots,u_{j_1-1},u_{j_1}\big)}\label{eqn:qsh_red_a}\\
&+\biggl({\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,b_1\dots b_s\cdot c\;|\;t_1,\dots,t_r,{\left\{u_1,\dots,u_{j_0-1},u_{j_0},u_{j_1+1},\dots,u_s,v\right\}}\big)}\nonumber\\
&\quad\quad+{\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_0}\dots b_{j_1},\dots,b_s,a_1\dots a_r\cdot c\;|\;u_1,\dots,u_{j_1},\dots,u_s,{\left\{t_1,\dots,t_r,v\right\}}\big)}\biggr)\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0+1},\dots,b_{j_1-1},b_{j_1},{\left(b_{j_0+1}\dots b_{j_1}\right)}{^{-1}}\;|\;u_{j_0+1},\dots,u_{j_1-1},u_{j_1},u_{j_0}\big)}\label{eqn:qsh_red_b}\\
&+(u_{j_1}-u_{j_0})\biggl({\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,b_1\dots b_s\cdot c\;|\;t_1,\dots,t_r,{\left\{u_1,\dots,u_{j_0-1},u_{j_0},u_{j_1},u_{j_1+1},\dots,u_s,v\right\}}\big)}\nonumber\\
&\quad\quad+{\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_0}\dots b_{j_1},\dots,b_s,a_1\dots a_r\cdot c\;|\;u_1,\dots,{\left\{u_{j_0},u_{j_1}\right\}},\dots,u_s,{\left\{t_1,\dots,t_r,v\right\}}\big)}\biggr)\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0+1},\dots,b_{j_1-1},{\left(b_{j_1}\dots b_{j_1-1}\right)}{^{-1}}\;|\;u_{j_0+1},\dots,u_{j_1-1},{\left\{u_{j_0},u_{j_1}\right\}}\big)}\label{eqn:qsh_red_cd}.\end{aligned}$$ Notice that this expression does not depend on $i_0,i_1$, and all but one of the segments in each generating function $f$ depends only on the $a_i$ or only on the $b_j$.
Reindexing leads to cancelation of all terms $f(a_1,\dots,a_r,\dots)$ except the term in (\[eqn:qsh\_red\_a\]) where $j_0=1$ and the term in (\[eqn:qsh\_red\_b\]) where $j_1=n$. That is, if $j_0\neq 1$ and $j_1\neq n$, then this expression becomes $$\begin{aligned}
F(j_0,j_1):=
&-\biggl({\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_0}\dots b_{j_1},\dots,b_s,a_1\dots a_r\cdot c\;|\;u_1,\dots,u_{j_1},\dots,u_s,{\left\{t_1,\dots,t_r,v\right\}}\big)}\biggr)\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0},b_{j_0+1},\dots,b_{j_1-1},{\left(b_{j_0}\dots b_{j_1-1}\right)}{^{-1}}\;|\;u_{j_0},u_{j_0+1},\dots,u_{j_1-1},u_{j_1}\big)}\label{eqn:qsh_red_a_noa}\\
&+\biggl({\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_0}\dots b_{j_1},\dots,b_s,a_1\dots a_r\cdot c\;|\;u_1,\dots,u_{j_0},\dots,u_s,{\left\{t_1,\dots,t_r,v\right\}}\big)}\biggr)\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0+1},\dots,b_{j_1-1},b_{j_1},{\left(b_{j_0+1}\dots b_{j_1}\right)}{^{-1}}\;|\;u_{j_0+1},\dots,u_{j_1-1},u_{j_1},u_{j_0}\big)}\label{eqn:qsh_red_b_soa}\\
&+(u_{j_1}-u_{j_0})\biggl({\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_0}\dots b_{j_1},\dots,b_s,a_1\dots a_r\cdot c\;|\;u_1,\dots,{\left\{u_{j_0},u_{j_1}\right\}},\dots,u_s,{\left\{t_1,\dots,t_r,v\right\}}\big)}\biggr)\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0+1},\dots,b_{j_1-1},{\left(b_{j_1}\dots b_{j_1-1}\right)}{^{-1}}\;|\;u_{j_0+1},\dots,u_{j_1-1},{\left\{u_{j_0},u_{j_1}\right\}}\big)}\label{eqn:qsh_red_cd_noa}.\end{aligned}$$ If $j_0=1$ or $j_1=s$, the following terms remain, respectively: $$\begin{aligned}
F_L(j_1):=
&-{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,b_1\dots b_s\cdot c\;|\;t_1,\dots,t_r,{\left\{u_{j_1},u_{j_1+1},\dots,u_s,v\right\}}\big)}\nonumber\\
&\quad\wedge\,{\mathbf{\Lambda}^*}{\big(b_1,\dots,b_{j_1-1},{\left(b_1\dots b_{j_1-1}\right)}{^{-1}}\;|\;u_1,\dots,u_{j_1-1},u_{j_1}\big)}\label{eqn:qsh_red_a_l},\\
F_R(j_0):=\,
&{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,b_1\dots b_s\cdot c\;|\;t_1,\dots,t_r,{\left\{u_1,\dots,u_{j_0},v\right\}}\big)}\nonumber\\
&\wedge\,{\mathbf{\Lambda}^*}{\big(b_{j_0+1},\dots,b_s,{\left(b_{j_0+1}\dots b_s\right)}{^{-1}}\;|\;u_{j_0+1},\dots,u_s,u_{j_0}\big)}.\label{eqn:qsh_red_b_r}\end{aligned}$$
Identical terms $G(i_0,i_1)$, $G_L(i_1)$, $G_R(i_0)$ with the ${\big(a_i\;|\;t_i\big)}$ and ${\big(b_j\;|\;u_j\big)}$ exchanged appear in the cases $j_1-j_0=\text{$0$ or $-1$}$.
So the total contribution of cuts of type 1 is $$\begin{aligned}
\sum_{\substack{1\leq j_0,j_1\leq s\\j_1-j_0\geq-1}}F(j_0,j_1)&+\sum_{\substack{1\leq i_0,i_1\leq r\\i_1-i_0\geq-1}}G(i_0,i_1)\nonumber\\
+\sum_{1\leq j_1\leq s}F_L(j_1)+\sum_{1\leq j_0\leq s}F_R(j_0)
&+\sum_{1\leq i_1\leq r}G_L(i_1)+\sum_{1\leq i_0\leq r}G_R(i_0)
\label{eqn:dq_1_total}\end{aligned}$$ finishing the computation.
\[lma:pf\_dq\_type2\] Computation of cuts of type (2).
A cut of type (2a/b/c) divides the circle into a left part $g$ and a right part $h$ (see Fig. \[fig:cut\_class\]). Let $i_0$ be maximal such that $a_{i_0}$ appears in $g$ and $i_1$ minimal such that $a_{i_1}$ appears in $h$, with $i_0=-1$ or $i_1=m+1$ if the corresponding segments do not appear. Define $j_0,j_1$ in the same manner, for the $b_j$.
Let $$\begin{aligned}
f_L(i_0,j_0,S)&={\mathbf{\Lambda}^*}{\big(a_1,\dots,a_{i_0},{\left(a_1\dots a_{i_0}\right)}{^{-1}}\;|\;t_1,\dots,t_{i_0},{\left\{u_1,\dots,u_{j_0}\right\}}\sqcup S\big)},\\
f_R(i_1,j_1,S)&={\mathbf{\Lambda}^*}{\big(a_{i_1},\dots,a_r,{\left(a_{i_1}\dots a_r\right)}{^{-1}}\;|\;t_{i_1},\dots,t_r,{\left\{u_{j_1},\dots,u_m\right\}}\sqcup S\big)},\end{aligned}$$ and define $g_L(i_0,j_0,S)$ and $g_R(i_1,j_1,S)$ in a similar way for the ${\big(b_j\;|\;u_j\big)}$. (As usual, one interprets these expressions as 0 if the index set is empty.) Also let $$\begin{aligned}
q_L(i_0,j_0,S)&={{\rm QSh}}^{i_0,j_0}{\big(a_1,\dots,a_{i_0},b_1,\dots,b_{j_0},{\left(a_1\dots a_{i_0}\cdot b_1\dots b_{j_0}\right)}{^{-1}}\;|\;t_1,\dots,t_{i_0},u_1,\dots,u_{j_0},S\big)},\\
&=f_L(i_0,j_0,S)+g_L(i_0,j_0,S)\\
q_R(i_1,j_1,S)&={{\rm QSh}}^{r-i_1+1,s-j_1+1}{\big(a_{i_1},\dots,a_r,b_{j_1},\dots,b_s,{\left(a_{i_1}\dots a_r\cdot b_{j_1}\dots b_s\right)}{^{-1}}\;|\;t_{i_1},\dots,t_r,u_{j_1},\dots,u_s,S\big)}\\
&=f_R(i_1,j_1,S)+g_R(i_1,j_1,S).\end{aligned}$$
Consider cuts (2a) for fixed $i_0,i_1,j_0,j_1$. For such cuts, $$i_1-i_0=j_1-j_0=1,
\quad-1\leq i_0\leq r,\quad-1\leq j_0\leq s.$$ The $g$ that occur in the resulting terms are exactly the quasishuffles of ${\left\{a_i:i\leq i_0\right\}}$ and ${\left\{b_j:j\leq j_0\right\}}$. The analogous statement holds for $h$. The contribution of cuts (2a) is $$\begin{aligned}
-q_L(i_0,j_0,v)\wedge q_R(i_0+1,j_0+1,v).\label{eqn:dq_2a}\end{aligned}$$
Now look at cuts (2b) and (2c). The non-distinguished segment containing the vertex or the cut is either $a_{i_0+1}$ ($i_0<r$), $b_{j_0+1}$ ($j_0<s$), or $a_{i_0+1}b_{j_0+1}$ ($i_0<r,j_0<s$). The terms coming from the sum of (2b) and (2c) are, for these three cases respectively, $$\begin{aligned}
(v-t_{i_0+1})q_L(i_0,j_0,{\left\{t_{i_0+1},v\right\}})&\wedge q_R(i_0+2,j_0+1,{\left\{t_{i_0+1},v\right\}}),\label{eqn:dq_2bc_a}\\
(v-u_{j_0+1})q_L(i_0,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge q_R(i_0+1,j_0+2,{\left\{u_{j_0+1},v\right\}}),\label{eqn:dq_2bc_b}\\
{\frac{-1}{t_{i_0+1}-u_{j_0+1}}}\biggl(
(v-t_{i_0+1})q_L(i_0,j_0,{\left\{t_{i_0+1},v\right\}})&\wedge q_R(i_0+2,j_0+2,{\left\{t_{i_0+1},v\right\}})\nonumber\\
-(v-u_{i_0+1})q_L(i_0,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge q_R(i_0+2,j_0+2,{\left\{u_{j_0+1},v\right\}})\biggr).\label{eqn:dq_2bc_ab}\end{aligned}$$
Let us assemble the terms of the form $f_L\wedge g_R$ and $g_L\wedge g_R$ coming from application of the shuffle relations to the $q_L$ and $q_R$. (The terms $g_L\wedge f_R$ and $f_L\wedge f_R$ are symmetrical.)
The terms $f_L\wedge g_R$, for $-1\leq i_0<r$ and $-1\leq j_0<s$, are: $$\begin{aligned}
-f_L(i_0,j_0,v)&\wedge g_R(i_0+1,j_0+1,v)\nonumber\\
+(v-t_{i_0+1})f_L(i_0,j_0,{\left\{t_{i_0+1},v\right\}})&\wedge g_R(i_0+2,j_0+1,{\left\{t_{i_0+1},v\right\}}),\nonumber\\
+(v-u_{j_0+1})f_L(i_0,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(i_0+1,j_0+2,{\left\{u_{j_0+1},v\right\}}),\nonumber\\
-{\frac{1}{t_{i_0+1}-u_{j_0+1}}}\biggl(
(v-t_{i_0+1})f_L(i_0,j_0,{\left\{t_{i_0+1},v\right\}})&\wedge g_R(i_0+2,j_0+2,{\left\{t_{i_0+1},v\right\}})\nonumber\\
-(v-u_{i_0+1})f_L(i_0,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(i_0+2,j_0+2,{\left\{u_{j_0+1},v\right\}})\biggr)\nonumber\\
=f_L(i_0,j_0,t_{i_0+1})&\wedge g_R(i_0+1,j_0+1,v)\nonumber\\
-f_L(i_0,j_0+1,t_{i_0+1})&\wedge g_R(i_0+1,j_0+2,v)\nonumber.\end{aligned}$$ Summing this over $j_0$ leaves $$\begin{aligned}
f_L(i_0,0,t_{i_0+1})\wedge g_R(i_0+1,1,v)&-f_L(i_0,s,{\left\{u_s,v\right\}})\wedge g_R(i_0+1,s+1,v)\\
&=f_L(i_0,0,t_{i_0+1})\wedge g_R(i_0+1,1,v)&=-G_L(i_0+1).
\label{eqn:dq_2}\end{aligned}$$ If $i_0=r,j_0<s$, then from (\[eqn:dq\_2a\]) and (\[eqn:dq\_2bc\_b\]) we also have the terms $$\begin{aligned}
-f_L(r,j_0,v)&\wedge g_R(r+1,j_0+1,v),\label{eqn:dq_2_t1}\\
(v-u_{j_0+1})f_L(r,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(r+1,j_0+2,{\left\{u_{j_0+1},v\right\}})\nonumber\\
=f_L(r,j_0+1,v)&\wedge{\left(g_R(r+1,j_0+2,v)-g_R(r+1,j_0+2,u_{j_0+1})\right)}\label{eqn:dq_2_t2}.\end{aligned}$$ The last term $f_L(r,j_0+1,v)\wedge g_R(r+1,j_0+2,u_{j_0+1})$ is $F_R(j_0+1)$. The remaining term and (\[eqn:dq\_2\_t1\]) mostly cancel when summed over $j_0$, leaving only $$\begin{aligned}
Z:=-f_L(r,0,v)\wedge g_R(r+1,1,v)+f_L(r,s,v)\wedge g_R(r+1,s+1,v)\nonumber\\
=-{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,\prod_jb_j\cdot c\;|\;t_1,\dots,t_r,v\big)}\wedge {\mathbf{\Lambda}^*}{\big(b_1,\dots,b_s,\prod_ia_i\cdot c\;|\;u_1,\dots,u_s,v\big)}\label{eqn:dq_2_nice}.\end{aligned}$$
If $j_0=s,i_0<r$, there are terms $$\begin{aligned}
-f_L(i_0,s,v)&\wedge g_R(i_0+1,s+1,v)&=0,\nonumber\\
(v-t_{i_0+1})f_L(i_0,s,{\left\{t_{i_0+1},v\right\}})&\wedge g_R(i_0+2,s+1,{\left\{t_{i_0+1},v\right\}})&=0.\end{aligned}$$ Finally, $i_0=r,j_0=s$ also produces 0.
Thus the sum of terms $f_L\wedge g_R$ is $$Z-\sum_{i_0=0}^{r-1}G_L(i_0+1)-\sum_{j_0=0}^{s-1}F_R(j_0+1)\label{eqn:dq_all_2_fg}.$$
Similarly, terms of the form $g_L\wedge f_R$ give $$-Z-\sum_{j_0=0}^{s-1}F_L(j_0+1)-\sum_{i_0=0}^{r-1}G_R(i_0+1)\label{eqn:dq_all_2_gf}.$$
The terms $g_L\wedge g_R$ where $i_0<r$, $j_0<s$ are, similarly: $$\begin{aligned}
-g_L(i_0,j_0,t_{i_0+1})&\wedge g_R(i_0+2,j_0+1,{\left\{t_{i_0+1},v\right\}})\nonumber\\
+g_L(i_0,j_0,{\left\{t_{i_0+1},u_{j_0+1}\right\}})&\wedge g_R(i_0+2,j_0+2,{\left\{t_{i_0+1},v\right\}}).\label{eqn:dq_all_2_gg}\end{aligned}$$ If $i_0=r,j_0<s$, we get the terms $$\begin{aligned}
-g_L(r,j_0,v)&\wedge g_R(r+1,j_0+1,v),\nonumber\\
(v-u_{j_0+1})g_L(r,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(r+1,j_0+2,{\left\{u_{j_0+1},v\right\}}).
\label{eqn:dq_all_2_ggm}\end{aligned}$$ If $j_0=s,i_0<r$, there are terms $$\begin{aligned}
-g_L(i_0,s,v)&\wedge g_R(i_0+1,s+1,v)&=0\nonumber,\\
(v-t_{i_0+1})g_L(i_0,s,{\left\{t_{i_0+1},v\right\}})&\wedge g_R(i_0+2,s+1,{\left\{t_{i_0+1},v\right\}})&=0.\end{aligned}$$ The case $i_0=r$, $j_0=s$ again contributes 0.
The terms $f_L\wedge f_R$ are symmetrical.
Assembling (\[eqn:dq\_all\_2\_gf\])-(\[eqn:dq\_all\_2\_ggm\]), the total contribution of cuts (2a/b/c) is $$\begin{aligned}
-\sum_{i=1}^rG_L(i)&-\sum_{j=1}^sF_R(j)\label{eqn:dq_2_total1}\\
+\sum_{i_0=0}^{r-1}\sum_{j_0=0}^{s-1}\biggl(
-g_L(i_0,j_0,t_{i_0+1})&\wedge g_R(i_0+2,j_0+1,{\left\{t_{i_0+1},v\right\}})\label{eqn:dq_2_total2}\\
+g_L(i_0,j_0,{\left\{t_{i_0+1},u_{j_0+1}\right\}})&\wedge g_R(i_0+2,j_0+2,{\left\{t_{i_0+1},v\right\}})\biggr)\label{eqn:dq_2_total2v}\\
+\sum_{j_0=0}^{s-1}\biggl(
-g_L(r,j_0,v)&\wedge g_R(r+1,j_0+1,v)\label{eqn:dq_2_total3v}\\
+(v-u_{j_0+1})g_L(r,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(r+1,j_0+2,{\left\{u_{j_0+1},v\right\}})\biggr),\label{eqn:dq_2_total3}\end{aligned}$$ plus symmetrical terms.
Cancellation of (\[eqn:dq\_1\_total\]) with (\[eqn:dq\_2\_total1\]) leaves $$\sum_{\substack{1\leq j_0,j_1\leq s\\j_1-j_0\geq-1}}F(j_0,j_1)
\label{eqn:dq_sum_ffgg}$$ plus the symmetrical term.
Thus $\delta Q$ is the symmetrized sum of expressions (\[eqn:dq\_2\_total2\])-(\[eqn:dq\_sum\_ffgg\]).
\[lma:pf\_dr\] Modulo elements $C(0,x)$, $\delta R_B$ and $\delta R_A$ are given by expression (\[eqn:drb\_all\]) below and its symmetric expression, respectively.
We compute $\delta R_B$.
Recall that the distinguished segment of $R_B$ is $\prod_ja_i\cdot c$. We use the above classification of cuts of type (1a/b/c/d) and (2a/b/c).
Consider first the terms $f\wedge g$ coming from cuts of type (1). For each such term, let $j_0,j_1$ be the minimal and maximal indices of $b_j$ that do not appear in $g$. For fixed $j_0,j_1$, the cuts of type (1a), (1b), and (1c/d) produce precisely the expressions (\[eqn:qsh\_red\_a\_noa\]), (\[eqn:qsh\_red\_b\_soa\]), and (\[eqn:qsh\_red\_cd\_noa\]) above. Thus the contribution of cuts of type (1) is $F(j_0,j_1)$, and the total contribution is $$\sum_{\substack{0\leq j_0,j_1\leq s\\j_1-j_0\geq-1}}F(j_0,j_1).
\label{eqn:dr_1}$$
Next, we look at cuts of type (2). We will need a simplified formula for terms where either the vertex or the cut are on a segment indexed with $S={\left\{s_1,\dots,s_k\right\}}$. If $k=1$ and the vertex is at a nonzero point, we get terms of the form $${\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,s_1\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;s_1,\dots\big)}.$$ Applying (\[eqn:multi\_identity\]), it is easy to show by induction that, for general $k$, the resulting terms are $$\sum_{i=1}^k{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,{\left\{s_1,\dots,s_i\right\}}\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;{\left\{s_i,\dots,s_k\right\}},\dots\big)}.\label{eqn:multi_cop_id_1}$$ For example, if $k=2$, this becomes $$\begin{aligned}
{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,s_1\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;{\left\{s_1,s_2\right\}},\dots\big)}&+{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,{\left\{s_1,s_2\right\}}\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;s_2,\dots\big)}\nonumber
\\={\frac{1}{s_1-s_2}}\biggl({\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,s_1\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;s_1,\dots\big)}&+{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,s_2\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;s_2,\dots\big)}\biggr)\nonumber,\end{aligned}$$ agreeing with the formula following directly from (\[eqn:multi\_identity\]) that has been used in the previous computations.
Similarly, if the vertex is on some segment $s'$, the term for $k=1$, $$s'{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,{\left\{s_1,s'\right\}}\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;{\left\{s_1,s'\right\}},\dots\big)},$$ expands into $$s'\sum_{i=1}^k{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,{\left\{s_1,\dots,s_i,s'\right\}}\big)}\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;{\left\{s_i,\dots,s_k,s'\right\}},\dots\big)}.\label{eqn:multi_cop_id_2}$$
Finally, if the vertex is at a 0 on the segment $s_i$ and the cut is on the segment $s'$, we get terms $$\begin{aligned}
\sum_{i=1}^ks_i{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,{\left\{s_1,\dots,s_i,s'\right\}}\big)}&\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;{\left\{s_i,\dots,s_k,s'\right\}},\dots\big)}\nonumber\\
+\sum_{i=1}^{k-1}{\mathbf{\Lambda}^*}{\big(\dots\;|\;\dots,{\left\{s_1,\dots,s_i,s'\right\}}\big)}&\wedge {\mathbf{\Lambda}^*}{\big(\dots\;|\;{\left\{s_{i+1},\dots,s_k,s'\right\}},\dots\big)}.
\label{eqn:multi_cop_id_3}\end{aligned}$$
These identities can also be shown combinatorially, by interpreting the definition of the multiple generating functions in terms of collapsing segments.
For a term $f\wedge g$ coming from a cut of type (2), let $j_0$ be the maximal index of $b_j$ appearing in $f$ and $j_1$ the minimal index in $g$, so $j_1-j_0=1$ for cuts (2a) and $j_1-j_0=2$ for cuts (2b/c). By (\[eqn:multi\_cop\_id\_1\]), for fixed $j_0$, the cuts of type (2a) contribute $$\begin{aligned}
-\sum_{i=1}^mg_L(i,j_0,\emptyset)&\wedge g_R(i,j_0+1,v)\label{eqn:dr_2a}
\\+\,g_L(r,j_0,v)&\wedge g_R(r+1,j_0+1,v). \label{eqn:dr_2av}\end{aligned}$$ By (\[eqn:multi\_cop\_id\_2\]), the cuts of type (2b) contribute $$\begin{aligned}
-\sum_{i=1}^mu_{j_0+1}g_L(i,j_0,u_{j_0+1})&\wedge g_R(i,j_0+2,{\left\{u_{j_0+1},v\right\}})\label{eqn:dr_2b}\\
-\,u_{j_0+1}g_L(r,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(r+1,j_0+2,{\left\{u_{j_0+1},v\right\}}).\label{eqn:dr_2bv}\end{aligned}$$ By (\[eqn:multi\_cop\_id\_3\]), the cuts of type (2c) contribute $$\begin{aligned}
\sum_{i=1}^mt_ig_L(i,j_0,u_{j_0+1})&\wedge g_R(i,j_0+2,{\left\{u_{j_0+1},v\right\}})\label{eqn:dr_2c1}\\
+\,vg_L(r,j_0,{\left\{u_{j_0+1},v\right\}})&\wedge g_R(r+1,j_0+2,{\left\{u_{j_0+1},v\right\}})\label{eqn:dr_2cv}\\
+\sum_{i=1}^mg_L(i,j_0,u_{j_0+1})&\wedge g_R(i+1,j_0+2,{\left\{u_{j_0+1},v\right\}}).
\label{eqn:dr_2c2}\end{aligned}$$ The sum of expressions (\[eqn:dr\_2b\]), (\[eqn:dr\_2c1\]), and (\[eqn:dr\_2c2\]) simplifies to $$g_L(i,j_0,u_{j_0+1})\wedge g_R(i,j_0+2,v).
\label{eqn:dr_2bc}$$ Then, letting $H_B(j_0)$ be the sum of expressions (\[eqn:dr\_2a\]), (\[eqn:dr\_2av\]), (\[eqn:dr\_2bv\]), (\[eqn:dr\_2cv\]), and (\[eqn:dr\_2bc\]), the coproduct of $R_B$ is $$\sum_{\substack{0\leq j_0,j_1\leq s\\j_1-j_0\geq-1}}F(j_0,j_1)
+\sum_{j=0}^{s-1}H_B(j).
\label{eqn:drb_all}$$
The coproduct of $R_A$ is the symmetric expression.
We now compare the results of the computations in Lemmas \[lma:pf\_dq\] and \[lma:pf\_dr\].
We have computed that $\delta Q$ is the symmetrization of $$(\ref{eqn:dq_2_total2})
+(\ref{eqn:dq_2_total2v})
+(\ref{eqn:dq_2_total3v})
+(\ref{eqn:dq_2_total3})
+(\ref{eqn:dq_sum_ffgg})$$ and $\delta R_B+\delta R_A$ is the symmerization of $$(\ref{eqn:dr_1})
+\sum_{j=0}^{n-1}\bigl[
(\ref{eqn:dr_2a})
+(\ref{eqn:dr_2av})
+(\ref{eqn:dr_2bv})
+(\ref{eqn:dr_2cv})
+(\ref{eqn:dr_2bc})\bigr].$$
Obviously $(\ref{eqn:dr_1})=(\ref{eqn:dq_sum_ffgg})$. Now $$(\ref{eqn:dq_2_total2})=\sum(\ref{eqn:dr_2a}),\quad
(\ref{eqn:dq_2_total2v})=\sum(\ref{eqn:dr_2bc}),\quad
(\ref{eqn:dq_2_total3v})=\sum(\ref{eqn:dr_2av}),\quad
(\ref{eqn:dq_2_total3})=\sum{\left[(\ref{eqn:dr_2bv})+(\ref{eqn:dr_2cv})\right]},$$ which finishes the proof.
### Proof of Step 2
Here we show the terms of weight $(1)\wedge(w-1)$ coming from $\delta Q-\delta R_A-\delta R_B$ are 0.
We first examine the relevant terms of $\delta Q$. Let us compute the coefficient $L_i^A$ occurring with $C(0,a_i)$. These come from shuffles containing segment ${\big(a_i\;|\;t_i\big)}$, a segment ${\big(a_ib_j\;|\;{\left\{t_i,u_j\right\}}\big)}$, and the segment ${\big(c\;|\;v\big)}$ (where we write $c=\prod_ia_i{^{-1}}\prod_jb_j{^{-1}}$).
Inspect the generating functions of depth 1 ${\mathbf{\Lambda}^*}{\big(w,w{^{-1}}\;|\;s_1,s_2\big)}$ that appeared in the proof of Lemma \[lma:pf\_lwsr\]. All generating functions in the lower half of cuts (1a/b/c/d) were written in a form where ${\big(w\;|\;s_1\big)}$ is the first segment counterclockwise of the distinguished segment, rather than with the segment counterclockwise of the vertex of the cut as in (\[eqn:gf\_cop\_1\])).
So, by the remark following Lemma \[lma:cop\_of\_gf\],
the terms (\[eqn:log\_cop\_gf\_r\]) vanish in the coproduct, so the terms arising from these cuts are canceled by the lower-depth shuffle relations in Lemma \[lma:pf\_lwsr\]. Similarly, for cuts of type (2), we only have terms (\[eqn:log\_cop\_gf\_lr\]) contributing the coefficient of $C(0,a_i{^{-1}})$.
For quasishuffles in which ${\big(a_i\;|\;t_i\big)}$ appears, the terms (\[eqn:log\_cop\_gf\_l\]) where some $b_j$ appears immediately clockwise of $a_i$ gives terms $$\begin{aligned}
{{\rm QSh}}^{r,s}_{(ij)}{\big(a_1,\dots,a_i,\dots,a_r,b_1,\dots,b_j,c\;|\;&t_1,\dots,\emptyset,\dots,t_r,u_1,\dots,u_j,\dots,u_s,v\big)},\label{eqn:dq_log_ab}\end{aligned}$$ where ${{\rm QSh}}^{r,s}_{(ij)}$ denotes the sum over only those quasishuffles where $a_i$ collapses with $b_j$.
The terms (\[eqn:log\_cop\_gf\_l\]) where either $a_{i-1}$ or some $a_{i-1}b_j$ appears immediately clockwise of $a_i$ sum to $$\begin{aligned}
{{\rm QSh}}^{r-1,s}{\big(a_1,\dots,a_{i-1}a_i,a_{i+1},\dots,a_r,b_1,\dots,b_s,c\;|\;t_1,\dots,t_{i-1},t_{i+1},\dots,t_r,u_1,\dots,u_s,v\big)}.\label{eqn:dq_log_al}\end{aligned}$$
Finally, the terms (\[eqn:log\_cop\_gf\_lr\]) contribute to $L^A_i$ the terms $$\begin{aligned}
t_i{{\rm QSh}}^{r,s}_{(i)}{\big(a_1,\dots,a_r,b_1,\dots,b_s,c\;|\;t_1,\dots,t_r,u_1,\dots,u_s,v\big)}
\label{eqn:dq_log_alr}\end{aligned}$$ where ${{\rm QSh}}_{(i)}$ denotes the the quasishuffles in which $a_i$ does not collapse with any $b_j$.
For quasishuffles in which some ${\big(a_ib_j\;|\;{\left\{t_i,u_j\right\}}\big)}$ appears, the terms (\[eqn:log\_cop\_gf\_l\]) contribute 0, since they arise from cuts of segments containing no 0s. The terms (\[eqn:log\_cop\_gf\_lr\]) give $$\begin{aligned}
{\frac{-1}{t_i-u_j}}\biggl(
t_i{{\rm QSh}}^{r,s}_{(ij)}{\big(a_1,\dots,a_i,\dots,a_r,b_1,\dots,b_j,\dots,b_s,c\;|\;&t_1,\dots,t_i,\dots,t_r,u_1,\dots,\emptyset,\dots,u_s,v\big)}\nonumber\\
-u_j{{\rm QSh}}^{r,s}_{(ij)}{\big(a_1,\dots,a_i,\dots,a_r,u_1,\dots,u_j,\dots,u_s,v\;|\;&t_1,\dots,\emptyset,\dots,t_r,u_1,\dots,u_j,\dots,u_s,v\big)}\biggr)\nonumber\\
=-t_i{{\rm QSh}}^{r,s}_{(ij)}{\big(a_1,\dots,a_i,\dots,a_r,b_1,\dots,b_s,c\;|\;&t_1,\dots,t_i,\dots,t_r,u_1,\dots,u_j,\dots,u_s,v\big)}\label{eqn:dq_log_ablr1}\\
-{{\rm QSh}}^{r,s}_{(ij)}{\big(a_1,\dots,a_i,\dots,a_r,b_1,\dots,b_s,c\;|\;&t_1,\dots,\emptyset,\dots,t_r,u_1,\dots,u_j,\dots,u_s,v\big)}.
\label{eqn:dq_log_ablr2}\end{aligned}$$
For the segment ${\big(c\;|\;v\big)}$, which includes a factor of $a_i{^{-1}}$, we get a contribution of $$\begin{aligned}
-v{{\rm QSh}}^{r,s}{\big(a_1,\dots,a_r,b_1,\dots,b_s,c\;|\;t_1,\dots,t_r,u_1,\dots,u_s,v\big)}=-vQ.\label{eqn:dq_log_vq}\end{aligned}$$ from (\[eqn:log\_cop\_gf\_lr\]) and $$\begin{aligned}
-{{\rm QSh}}^{r-1,s}{\big(a_1,\dots,a_{m-1},b_1,\dots,b_s,a_rc\;|\;&t_1,\dots,t_{r-1},u_1,\dots,u_s,t_r\big)}\\
-{{\rm QSh}}^{r,s-1}{\big(a_1,\dots,a_r,b_1,\dots,b_{n-1},b_sc\;|\;&t_1,\dots,t_r,u_1,\dots,u_{s-1},u_s\big)}\\
+{{\rm QSh}}^{r-1,s-1}{\big(a_1,\dots,a_{m-1},b_1,\dots,b_{n-1},a_rb_sc\;|\;&t_1,\dots,t_{r-1},u_1,\dots,u_{s-1},{\left\{t_r,u_s\right\}}\big)},\end{aligned}$$ from (\[eqn:log\_cop\_gf\_l\]), with three terms, depending on which segment ($a_r$, $b_s$, or $a_rb_s$) appears clockwise of $c$. By the lower-depth shuffle relations, this simplifies to $$\begin{aligned}
-{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,\prod_jb_j\cdot c\;|\;&t_1,\dots,t_r,{\left\{u_1,\dots,u_s\right\}}\big)}\label{eqn:dq_log_ca}\\
-{\mathbf{\Lambda}^*}{\big(b_1,\dots,b_s,\prod_ia_i\cdot c\;|\;&u_1,\dots,u_s,{\left\{t_1,\dots,t_r\right\}}\big)}.\label{eqn:dq_cb}\end{aligned}$$
The terms (\[eqn:dq\_log\_ab\]) cancel with (\[eqn:dq\_log\_ablr2\]). Summing (\[eqn:dq\_log\_ablr1\]) over $j$ and adding to (\[eqn:dq\_log\_alr\]) results in $$t_i{{\rm QSh}}^{r,s}{\big(a_1,\dots,a_r,b_1,\dots,b_s,c\;|\;t_1,\dots,t_r,u_1,\dots,u_s,v\big)}=t_iQ.
\label{eqn:dq_log_x}$$ Thus $L_i^A$ is the sum of (\[eqn:dq\_log\_al\]), (\[eqn:dq\_log\_vq\]), (\[eqn:dq\_log\_ca\]), (\[eqn:dq\_cb\]) and (\[eqn:dq\_log\_x\]). Applying lower-depth shuffle relations and (\[eqn:multi\_identity\]), this sum simplifies to $$\begin{aligned}
L_i^A={\mathbf{\Lambda}^*}{\big(a_1,\dots,a_{i-1}a_i,a_{i+1},\dots,a_r,\prod_jb_j\cdot c\;|\;&t_1,\dots,t_{i-1},t_{i+1},\dots,t_r,{\left\{u_1,\dots,u_s\right\}}\sqcup{\left\{v\right\}}\big)}\label{eqn:dq_log_qa}\\
-{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,\prod_jb_j\cdot c\;|\;&t_1,\dots,t_r,{\left\{u_1,\dots,u_s\right\}}\big)}\label{eqn:dq_log_qab}\\
+(t_i-v)(Q-R_B).\label{eqn:dq_log_qb}\end{aligned}$$
Now let us compute the coefficient $M_i^A$ occuring with $C(0,a_i)$ in $\delta(R_A)$. For the segment ${\big(a_i\;|\;t_i\big)}$ in $R_A$, (\[eqn:log\_cop\_gf\_lr\]) and (\[eqn:log\_cop\_gf\_l\]) contribute the terms $$\begin{aligned}
t_i{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_i,\dots,a_r,\prod_jb_j\cdot c\;|\;&t_1,\dots,t_i,\dots,t_r,{\left\{u_1,\dots,u_s\right\}}\sqcup{\left\{v\right\}}\big)}&=t_iR_A,\label{eqn:dr_log_x}\\
{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_{i-1}a_i,a_{i+1},\dots,a_r,\prod_jb_j\cdot c\;|\;&t_1,\dots,t_{i-1},t_{i+1},\dots,t_r,{\left\{u_1,\dots,u_s\right\}}\sqcup{\left\{v\right\}}\big)},\label{eqn:dr_log_al}\end{aligned}$$ where the second term appears only if $i>1$.
The distinguished segment ${\big(\prod_ia_i{^{-1}}\;|\;{\left\{u_j\right\}}\sqcup{\left\{v\right\}}\big)}$ contributes only a term (\[eqn:log\_cop\_gf\_lr\]). By an argument similar to that in Lemma \[lma:pf\_dr\], this term can be written $$\begin{aligned}
-vR_A-{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,\prod_jb_j\cdot c\;|\;t_1,\dots,t_r,{\left\{u_1,\dots,u_s\right\}}\big)}.
\label{eqn:dr_log_v}\end{aligned}$$
Combining (\[eqn:dq\_log\_qa\])-(\[eqn:dr\_log\_v\]), we find that $$(L_i^A-M_i^A)=(t_i-v)(Q-R_A-R_B).$$ Therefore, adding the symmetric terms for the $C(0,b_j)$, $$\delta(Q-R_A-R_B)={\left[\sum_{i=1}^rC(0,a_i)(t_i-v)+\sum_{j=1}^sC(0,b_j)(u_j-v)\right]}\wedge(Q-R_A-R_B)$$ modulo lower-depth shuffle relations and elements $(\text{weight 1})\wedge(\text{weight 1})$.
### Conclusion
We are ready to use the coproduct we have computed to reduce the proof of the relations to a simple base case.
We induct on the depth $r+s$. When $r=0$ or $s=0$, ${\overline{{{\rm QSh}}}}^{r,s}$ is identically 0.
If $r,s>0$, taking coproduct on both sides of (\[eqn:step2\]) and using that $\delta^2=0$, one deduces that $\delta(Q-R_A-R_B)=0$ modulo shuffle relations of depth $<r+s$ and terms $\text{(weight 1)}\wedge\text{(weight 1)}$.
When no terms $C(0,x)\wedge C(0,y)$ are present in the coproduct, Lemma \[lma:pf\_red\_multi\] and Lemma \[lma:pf\_log\] imply that $\delta(Q-R_A-R_B)$ lies in the ideal generated by lower-depth relations.
These terms appear only in a base case: the constant term of the shuffle relation for $r=s=1$. Showing the coproduct of this term is 0 amounts to proving the identity $$\delta{\left({\left[ C^*(a|0,b|0,c|0)+ C^*(b|0,a|0,c|0)- C^*(ab|1,c|0)\right]}- C^*(a|0,bc|1)- C^*(b|0,ac|1)\right)}=0.\label{eqn:dilog_identity}$$ We compute directly that the left side of (\[eqn:dilog\_identity\]) is $$\begin{aligned}
\quad C(1,a)\wedge C(1,ab)+C(1,ab)\wedge (C(1,b)+C(0,a))+(C(1,b)+C(0,a))&\wedge C(1,a)\\
+\,C(1,b)\wedge C(1,ab)+C(1,ab)\wedge (C(1,a)+C(0,b))+(C(1,a)+C(0,b))&\wedge C(1,b)\\
-\,C(1,ab)\wedge C(0,ab)+C(1,a)\wedge C(0,a)+C(1,b)&\wedge C(0,b)\\
=\quad C(1,ab)\wedge C(0,a)+C(0,a)&\wedge C(1,a)\\+\,C(1,ab)\wedge C(0,b)+C(0,b)&\wedge C(1,b)\\-\,C(1,ab)\wedge C(0,ab)+C(1,a)\wedge C(0,a)+C(1,b)&\wedge C(0,b)&=0.\end{aligned}$$
The theorem is proved.
Specialization theorem for Hodge correlators {#sec:nodal}
============================================
We now study how the Hodge correlators over a base $B$ behave when the sections collide. This will require extending the theory of Hodge correlators to nodal curves.
### The correlator Lie coalgebra for nodal curves
Recall the moduli space ${\mathcal{M}}_{0,n}'$ of $n$ distinct points and a distinguished tangent vector on ${\mathbb{P}}^1$. Its Deligne-Mumford compactification ${\overline{{\mathcal{M}}}}_{0,n}'$ consists of the nodal curves of genus 0, i.e., those whose dual graph is a tree and in which every component is a punctured projective line. with $n$ marked points and a distinguished tangent vector $v_\infty$.
Let $X=\bigcup_iX_i$ be a genus 0 nodal curve with a set of punctures $S$. Let $T$ be the dual tree of $X$, with vertices indexed by $i$ corresponding to $X_i$, rooted at the component $0$ with the base point $s_0\in X_0$, oriented away from the root (write $i\to j$ if $(i,j)$ is an edge). Choose a coordinate $z_i$ on each $X_i$ such that the point joining the component to its parent $X_j$ is $(z_i=\infty,z_j=v_{ji})$, and the base point on $X_0$ is at $z_0=\infty$ with tangent vector $v_\infty$. Let $S_i$ be the set of punctures on $X_i$. Let $N_i={\left\{v_{ij}:i\to j\right\}}$.
We define the correlator Lie coalgebra for the nodal curve $X$ by $${\mathcal{CL}}^\vee_{X,S,v_\infty} =
\bigoplus_{i}{\mathcal{CL}}^\vee_{X_i,S_i\cup N_i,v_i}
,\label{eqn:def_nodal_clie}$$ where $v_i$ is the tangent vector ${\frac{-1}{z_i^2}}{\frac{{\partial}}{{\partial}z_i}}$ at $z_i=\infty$.
It coincides with the usual definition if $X$ is smooth, justifying the notation. If $X$ is not smooth, it is different from $\widetilde{{\mathcal{CL}}}^\vee_{X,S,s_0}$, the coalgebra naively defined as the tensor algebra of $S$ modulo cyclic symmetry and shuffle relations with a $H_2(X)$ coefficient. They are related in the following way. For each $i$, there is a surjective coalgebra morphism to the component of the direct sum corresponding to $X_i$: $$\widetilde{{\mathcal{CL}}}^\vee_{X,S,s_0}{\xrightarrow{\pi_i}}{\mathcal{CL}}^\vee_{X_i,S_i\cup N_i,v_i}.$$ To define it on a generator $(x_1\otimes\dots\otimes x_n)\otimes[X_i]$, let $p$ be the common parent of the components containing the $x_j$. If $p\neq i$, the $i$-th component of the map is 0. Otherwise, set $$\begin{aligned}
\pi_i(x)=\begin{cases}x,&x\in X_i\\z_i=v_{ij}\in N_i,&\text{$x\in X_k$, where $\exists$ path $i\to j\to\dots\to k$}\end{cases},\end{aligned}$$ extended to preserve the tensor product. That is, points in $X_i$ remain, while points in components below $X_i$ collapse to the nearest node on $X_i$. Evidently this map preserves the coproduct and defining relations. Taking the direct sum of the maps $\pi_i$, we have produced a coalgebra morphism: $$\pi:\widetilde{{\mathcal{CL}}}^\vee_{X,S,v_\infty}\to{\mathcal{CL}}^\vee_{X,S,v_\infty}.$$ It preserves the decomposition of the domain by $H_2(X)=\bigoplus_iH_2(X_i)$.
In particular, if $(X,S,v_0)$ vary over a base $B\to{\mathcal{M}}_{0,n}'$, and the variation extends to ${\overline{B}}\to{\overline{{\mathcal{M}}}}_{0,n}'$, with $D={\overline{B}}{\setminus}B$, then we have a *degeneration map* $$\pi_D:{\mathcal{CL}}^\vee_{X/B,S,v_\infty}\to\widetilde{{\mathcal{CL}}}^\vee_{X/D,S,v_\infty}\to{\mathcal{CL}}^\vee_{X/D,S,v_\infty},$$ where the first map simply applies the induced map on $H_2$ and the second map is the quotient defined above. The composition forgets the way in which the sections in $S$ collided at boundary of $B$.
### Specialization theorem
Recall that an element of ${\mathcal{CL}}^\vee_{X/B,S,v_\infty}$ over a base $B\to{\mathcal{M}}_{0,n}'$ determines, by the map ${\text{\rm Cor}}_{\text{\rm Hod}}$, a variation of Hodge structures over $B$, and, by the period map $p$, a smooth function on $B$. The maps ${\text{\rm Cor}}_{\text{\rm Hod}}$ and $p$ also exist for $X$ a nodal curve, extended by linearity from the definition (\[eqn:def\_nodal\_clie\]).
Suppose $B\to{\mathcal{M}}_{0,n}'$ is a family of curves $(X,S,v_\infty)$ extending to ${\overline{B}}\to{\overline{{\mathcal{M}}}}_{0,n}'$, with $D={\overline{B}}{\setminus}B$ a normal crossings divisor, and suppose $x\in{\mathcal{CL}}^\vee_{X/B,S,s_0}$ of weight $n>1$.
(a) The Deligne’s canonical extension to $D$ of the variation of framed mixed Hodge structures determined by ${\text{\rm Cor}}_{\text{\rm Hod}}(x)$ is independent of the normal vector to $D$. Thus there is a specialized map $\operatorname{Spec}_D{\text{\rm Cor}}_{\text{\rm Hod}}:{\left({\mathcal{CL}}_{X/B,S,v_\infty}\to{\text{\rm Lie}}_{{\text{\rm HT}}/D}^\vee\right)}_{w>1}$.
(b) This specialized map coincides with the Hodge correlator of the degeneration map: $$\xymatrix{
{\left({\mathcal{CL}}^\vee_{X/B,S,v_\infty}\right)}_{w>1}\ar[r]^{\pi_D}\ar[d]_{{\text{\rm Cor}}_{\text{\rm Hod}}}
&{\left({\mathcal{CL}}^\vee_{X/D,S,v_\infty}\right)}_{w>1}\ar[d]^{{\text{\rm Cor}}_{\text{\rm Hod}}}\\({\text{\rm Lie}}_{{\text{\rm HT}}/B}^\vee)\ar[r]^{\operatorname{Spec}_D}&({\text{\rm Lie}}_{{\text{\rm HT}}/D}^\vee).
}
.$$
(c) Let $t=0$ be a local equation for $D$. Then $$\lim_{t\to0}p({\text{\rm Cor}}_{\text{\rm Hod}}(x_t))=p(Cor_Hod(x_{t=0})).$$
\[thm:degen\_nodal\]
Let $x\in{\mathcal{CL}}^\vee_{X/B,S,v_\infty}$ be a generator of weight $w>1$. For any $v$ a normal vector to $D$, we get the specialized framed mixed Hodge-Tate structure $\operatorname{Spec}_D^v{\text{\rm Cor}}_{\text{\rm Hod}}(x)$.
We must show that:
(1) The periods of ${\text{\rm Cor}}_{\text{\rm Hod}}(x)$ extend continuously to $D$.
(2) The coproduct of $\operatorname{Spec}_D^v{\text{\rm Cor}}_{\text{\rm Hod}}(x)$ does not depend on the direction of specialization $v$ at any smooth point of $D$.
(3) The periods of the specializations (i.e., the limits of the periods at $D$) coincide with the periods of the degeneration to $D$.
We will prove (1-3) by induction on the weight. First, let us see how they imply the result.
Assuming (2), the coproduct of $\operatorname{Spec}_D^v{\text{\rm Cor}}_{\text{\rm Hod}}(x)$ is independent of $v$. Because the coproduct commutes with $\operatorname{Spec}_D^v$, this element is independent of $v$ up to $\operatorname{Ext}^1({\mathbb{R}}(0),{\mathbb{R}}(n))$, which is 1-dimensional and controlled by the period. By (1), the period is independent of the direction of specialization, which gives (a). By (3), it coincides with the period of the degeneration, which gives (b). Then (c) follows by the definitions from (b).
To show (1), we let ${\left\{{\varepsilon}_i=0\right\}}$ be a set of smooth local equations for $D$ and prove that $p({\text{\rm Cor}}_{\text{\rm Hod}}(x))$ can be represented locally as a polynomial in the $\log{\varepsilon}_i$ such that the terms with $\log{\varepsilon}_i$ appearing in positive degree have coefficients vanishing along ${\left\{{\varepsilon}_i=0\right\}}$ (*tame logarithmic singularities*). This will follow from the differential equations on the periods. Note that in weight 1, the period of $C(x,y)$ has a (not tame) logarithmic singularity along $x=y$. In weight $>1$, we proceed by induction.
Consider a simple element $x=x_0\otimes\dots\otimes x_n\in{\mathcal{CL}}^\vee_{X/B,S,v_\infty}$ ($n>1$). Suppose that not all $x_i$ collide on $D$, so we must only consider the summand of the nodal ${\mathcal{CL}}^\vee_{X/D,S,s_0}$ corresponding to the component containing the base point. The terms of $\delta(x)$ can be grouped into those of two forms:
(i) $x'\wedge x''$, where not all sections in $x'$ and in $x''$ collide to the same section on $D$;
(ii) $x'\wedge(x''_1-x''_2)$, where not all sections in $x'$ collapse on $D$, but $x''_1$ and $x''_2$ coincide on $D$.
By the inductive hypothesis, the specialization of $\delta(x)$ does not depend on the direction of specialization: for terms (i), $x'$ and $x''$ satisfy (2), while in terms (ii) the $x''_1-x''_2$ vanish under specialization to $D$. This gives (2).
For (1), from the differential equations on the periods (\[eqn:diff\_eq\_p1\]), we see that $d_Bp({\text{\rm Cor}}_{\text{\rm Hod}}(x))$ is a sum of terms that are smooth over $B$ with logarithmic singularities along $D$ (from type (i)) and terms that vanish along $D$ by the inductive hypothesis (from type (ii)). We conclude that $p^{v_\infty}(x)$ has tame logarithmic singularities along $D$.
If all $x_i$ collide on $D$, we simply pass to their common parent component and apply the same argument.
We conclude with (3). We have shown that the specializations of ${\text{\rm Cor}}_{\text{\rm Hod}}^{v_0}$ and its coproduct to $D$ exist at every point and their periods are independent of $v$, and thus the specialized period map $p\circ{\text{\rm Cor}}_{\text{\rm Hod}}$ is equal to the period of the degeneration up to adding a constant for each smooth component of the smooth locus of $D$. We must show the constant 0.
It is enough to show this for $D$ a lowest-codimension boundary stratum in ${\overline{{\mathcal{M}}}}_{g,n}'$. We are done by the next lemma.
Let $I$ be a proper subset of ${\left\{0,1,\dots,n\right\}}$ ($n>1$) and $x_0,\dots,x_n\in{\mathbb{C}}^*$ with $x_i\neq x_j$ if $i\neq j$ and either $i,j\in I$ or $i,j\notin I$. Let $$x_i(t)=\begin{cases}tx_i&i\in I,\\x_i&i\notin I\end{cases}.$$ Then $${\text{\rm Cor}}_{\mathcal{H}}(x_0(t),x_1(t),\dots,x_n(t))$$ is continuous at $t=0$.
\[lma:circ\_spec\]
For $n=2$, this amounts to continuity of ${\mathcal{L}}_2$ at 1.
In the proof of Theorem \[thm:degen\_nodal\] it was established that $$\lim_{t\to0}{\text{\rm Cor}}_{\mathcal{H}}(x_0(t),x_1(t),\dots,x_n(t))-{\text{\rm Cor}}_{\mathcal{H}}(x_0(0),x_1(0),\dots,x_n(0))$$ is independent of the $x_i$, for generic $x_i$. Let us integrate this difference over $(x_0,\dots,x_n)\in(S^1)^{n+1}$, with respect to the standard measures $\mu(x_i)$ of volume 1 on $S^1={\left\{{\left|z\right|}=1\right\}}\subset{\mathbb{C}}$.
The limit is uniform in the directions $x_i$ ($i\in I$), and so $$\int\lim_{t\to0}{\text{\rm Cor}}_{\mathcal{H}}(x_0(t),x_1(t),\dots,x_n(t))\,\prod d\mu(x_i)=\lim_{t\to0}\int{\text{\rm Cor}}_{\mathcal{H}}(x_0(t),x_1(t),\dots,x_n(t))\,\prod d\mu(x_i).$$ To conclude, it suffices to show that $$\int{\text{\rm Cor}}_{\mathcal{H}}(x_0(t),x_1(t),\dots,x_n(t))\,\prod d\mu(x_i)=0.\label{eqn:circ_norm_int}$$ for all $t$.
For any tree $T$ entering into the Feynman integral expression for (\[eqn:circ\_norm\_int\]), choose a pair of boundary vertices (without loss of generality, labeled $x_0$ and $x_1$) incident to a common internal vertex $v$ with corresponding variable $x_v$, and let $x_w$ be variable corresponding to the third vertex incident to $v$. Then the integral over the $x_i$ contains the term $$\int_{x_0,x_1}{\left(\int{\mathcal{L}}_2{\left({\frac{x_w-x_0(t)}{x_w-x_1(t)}}\right)}\wedge(\text{terms independent of $x_0(t),x_1(t)$})\right)}\,d\mu(x_0)\,d\mu(x_1).$$ Exchanging the two integrals and noting that ${\mathcal{L}}_2({\frac{z-a}{z-b}})$ changes sign under the involution $$a\mapsto{\overline{a}}{\frac{z^2}{{\left|z\right|}^2}},\quad b\mapsto{\overline{b}}{\frac{z^2}{{\left|z\right|}^2}},$$ we conclude that this expression is 0.
The specialization theorem states is that when the punctures labeling an element of ${\mathcal{CL}}^\vee$ collide, only the nearest possible to the base point component of the resulting nodal curve determines the limit Hodge correlator. We obtain as a corollary Theorem \[thm:corrh\_cts\]:
The Hodge correlators ${\text{\rm Cor}}_{\mathcal{H}}(z_0,\dots,z_n)$ are continuous on ${\mathbb{C}}^{n+1}{\setminus}{\left\{z_0=\dots=z_n\right\}}$.
For example, ${\mathcal{L}}_2$ is continuous with a tame logarithmic singularity at 1, but ${\mathcal{L}}_2{\left({\frac{a-c}{b-c}}\right)}$ has no limit as $a,b,c\to0$.
The second shuffle relations
============================
Proofs of Theorems \[thm:hodge\_main\] and \[thm:mot\_main\]
------------------------------------------------------------
In this section we will prove the second shuffle relations for Hodge and motivic correlators.
### Proof for Hodge correlators
Recall Theorem \[thm:hodge\_main\]:
(a) Restricted to the subspace of ${\mathcal{CL}}_{X,S,v_\infty}^\vee$ generated by elements $(x_0\otimes\dots\otimes x_n)(1)$ with not all $x_i$ equal, the map ${\text{\rm Cor}}_{\text{\rm Hod}}$ factors through ${\mathcal{D}}^\circ({\mathbb{C}}{^{*}})$.
(b) Suppose that $r,s>1$ and that not all $n_i=0$ or not all $w_i=1$. Then the Hodge correlators satisfy the relation: $$\begin{aligned}
&\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}{\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_{\sigma{^{-1}}(1)}|n_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)}|n_{\sigma{^{-1}}(M_\sigma)},w_0|n_0)\\
&-{\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_1|n_1,\dots,w_r|n_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}|n_{{\left\{r+1,\dots,r+s,0\right\}}})\\
&-{\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}(w_{r+1}|n_{r+1},\dots,w_{r+s}|n_{r+s},w_{{\left\{1,\dots,r,0\right\}}}|n_{{\left\{1,\dots,r,0\right\}}})&=0,
\end{aligned}$$ where $$n_S=\sum_{i\in S}(n_i+1)-1,\quad w_S=\prod_{i\in S}w_i.$$
(c) The Hodge correlators satisfy all specializations of this relation as any subset of the $w_i$ $(1\leq i\leq n)$ approaches 0.
For fixed $r$, $s$, and $n_i$, consider the $(r,s)$-second shuffle relation in (b). It is a family of framed mixed Hodge-Tate structures over $$S={\left\{(w_0,\dots,w_n)\in({\mathbb{C}}{^{*}})^{n+1}:w_0\dots w_n=1\right\}}.$$ To show (b), it suffices to show the family is trivial as an element of ${\text{\rm Lie}}_{\text{\rm HT}}^\vee$ over every point of $S$, except at $(1,\dots,1)$ if all $n_i=0$. This is equivalent to (a) by the definitions, as the Hodge correlators are already known to satisfy the defining relations in $\widetilde{\mathcal{D}}^\circ({\mathbb{C}}{^{*}})$.
Each term of this relation is an element $${\text{\rm Cor}}_{\text{\rm Hod}}(1,z_1,\dots,z_n),$$ where each $z_k$ is either 0 or monomial in the $w_i$. By Theorem \[thm:hc\_main\], it is a variation $\mathbf V$ of framed mixed Hodge-Tate structures over $$T={\left\{(z_1,\dots,z_n)\in({\mathbb{C}}^*)^n\right\}}{\setminus}{\left(\text{\rm diagonals}\right)}.$$ We first show by induction on the weight $n$ that all such variations is trivial.
In the base case $n=1$, there are no second shuffle relations.
For the induction hypothesis, suppose $n>1$ and (b) holds in weights $1<w<n$. Fix $r$, $s$, and $n_i$ and let $\mathbf V$ be the variation defined above. By the induction hypothesis, $\delta{\text{\rm Cor}}_{\text{\rm Hod}}(\mathbf{V})$ vanishes, and thus, by rigidity, $\mathbf V$ is a constant variation, determined pointwise as an element of $\operatorname{Ext}^1({\mathbb{R}}(0),{\mathbb{R}}(n))$ by the period. We show the period is 0.
The specialization theorem (§\[sec:nodal\]) implies that the period of $\mathbf{V}$ is continuous away from the main diagonal in ${\mathbb{C}}^{n+1}$. Unless all $n_i=0$ or all $w_i=1$, in no term of the relation (b) do all points collide to the main diagonal. By Corollary \[thm:corrh\_cts\], the specialization of the period at $w_1,\dots,w_n=0$ is equal to the substitution $w_i=0$. Under this substitution, the period of each term of the relation becomes $${\text{\rm Cor}}_{\mathcal{H}}(1,0,\dots,0)=0.$$ Therefore, $\mathbf V$ is trivial over $T$.
Because $T$ is dense in ${\mathbb{C}}^n$, the relation at all points – except $w_1=\dots=w_n=1$ if all $n_i=0$ – follows by the specialization theorem. This completes the proof of (b) and (c).
Applying the period map, we immediately obtain Theorem \[thm:period\_main\]:
(a) Suppose that $r,s>1$ and that not all $n_i=0$ or not all $w_i=1$. Then the Hodge correlators satisfy the relation: $$\begin{aligned}
&\sum_{\sigma\in{\overline{\Sigma}}_{r,s}}(-1)^{r+s-M_\sigma}{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_{\sigma{^{-1}}(1)}|n_{\sigma{^{-1}}(1)},\dots,w_{\sigma{^{-1}}(M_\sigma)}|n_{\sigma{^{-1}}(M_\sigma)},w_0|n_0)\\
&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_1|n_1,\dots,w_r|n_r,w_{{\left\{r+1,\dots,r+s,0\right\}}}|n_{{\left\{r+1,\dots,r+s,0\right\}}})\\
&-{\text{\rm Cor}}_{\mathcal{H}}{^{*}}(w_{r+1}|n_{r+1},\dots,w_{r+s}|n_{r+s},w_{{\left\{1,\dots,r,0\right\}}}|n_{{\left\{1,\dots,r,0\right\}}})&=0,
\end{aligned}$$ where $$n_S=\sum_{i\in S}(n_i+1)-1,\quad w_S=\prod_{i\in S}w_i.$$
(b) The Hodge correlators satisfy all specializations of this relation as any subset of the $w_i$ $(1\leq i\leq n)$ approaches 0.
### Proof for motivic correlators
Recall Theorem \[thm:mot\_main\]:
Let $F$ be a number field.
(a) Restricted to the subspace of ${\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^\vee$ generated by elements $(x_0\otimes\dots\otimes x_n)(1)$ with not all $x_i$ equal, the map ${\text{\rm Cor}}_{\text{\rm Mot}}$ factors through ${\mathcal{D}}^\circ(F^\times)$.
(b) Suppose that $r,s>1$ and that not all $n_i=0$ or not all $w_i=1$. Then the motivic correlators satisfy the same relation as in Theorem \[thm:hodge\_main\], with ${\text{\rm Cor}}_{\text{\rm Hod}}{^{*}}$ replaced by ${\text{\rm Cor}}_{\text{\rm Mot}}{^{*}}$.
(c) The motivic correlators satisfy all specializations of this relation as any subset of the $w_i$ $(1\leq i\leq n)$ approaches 0.
Fix an embedding $F{\xrightarrow{r}}{\mathbb{C}}$. It induces a map ${\mathcal{D}}^\circ(F^\times)\to{\mathcal{D}}^\circ(F^\times)$, which we also denote by $r$.
Denoting by ${\mathcal{CL}}^{\vee\circ}$ the subalgebras generated by elements $(x_1\otimes\dots\otimes x_n)(1)$ where not all $x_i$ are equal, we have the diagram $$\xymatrix{
{\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^{\vee\circ}
\ar[rr]^{{\text{\rm Cor}}_{\text{\rm Mot}}}\ar[dd]^r\ar[dr]
&&{\text{\rm Lie}}_{{{\text{\rm MT}/F}}}^\vee\ar[dd]^r
\\&{\mathcal{D}}^\circ(F^\times)\ar@{-->}[ur]\ar[dd]^{r}\\
{\mathcal{CL}}_{X,S,v_\infty}^{\vee\circ}\ar[dr]
\ar[rr]^{{\text{\rm Cor}}_{\text{\rm Hod}}\quad\quad}
&&{\text{\rm Lie}}_{{\text{\rm HT}}}^\vee\ar[r]^p&{\mathbb{R}}\\&{\mathcal{D}}^\circ({\mathbb{C}}{^{*}})\ar[ur]\\
}
,$$ where the lower half commutes by Theorem \[thm:hodge\_main\] and the vertical maps are induced by $r$.
It is necessary to show the dashed arrow is well-defined, i.e., that ${\text{\rm Cor}}_{\text{\rm Mot}}$ vanishes on the kernel of the map ${\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^{\vee\circ}\to{\mathcal{D}}^\circ(F^\times)$.
Commutativity of the diagram for every embedding $r$ implies the result. Precisely, we argue by induction.
In weight 1, then there are no first or second shuffles, and the shuffle relations are mapped to 0 by ${\text{\rm Cor}}_{\text{\rm Mot}}$. Indeed, we have ${\text{\rm Cor}}_{\text{\rm Mot}}(0,0)=0$ and ${\text{\rm Cor}}_{\text{\rm Mot}}(ab,ac)={\text{\rm Cor}}_{\text{\rm Mot}}(0,a)+{\text{\rm Cor}}_{\text{\rm Mot}}(b,c)$, since $${\text{\rm Cor}}_{\text{\rm Mot}}(a,b)=(a-b)\in({\text{\rm Lie}}_{{\text{\rm MT}/F}})^\vee_{w=1}\cong F^\times\otimes{\mathbb{Q}}.$$
For the inductive step, if $x\in{\left({\mathcal{CL}}_{X,S,v_\infty}^{\text{\rm Mot}}\right)}^{\vee\circ}$, homogeneous of weight $>1$, vanishes in ${\mathcal{D}}^\circ(F^\times)$, then ${\text{\rm Cor}}_{\text{\rm Hod}}(r(x))=0\in{\text{\rm Lie}}_{\text{\rm HT}}^\vee$ under every embedding $r$, and ${\partial}{\text{\rm Cor}}_{\text{\rm Mot}}(x)=0$ by the inductive hypothesis. By Lemma \[lma:d0h0\_rational\], ${\text{\rm Cor}}_{\text{\rm Mot}}(x)=0$.
Applications
------------
### Additive shuffle relation
Specializing all $w_i$ to 1 in the second shuffle relation, where all $n_i=0$, we extract an *additive* second shuffle relation, which does not have lower-depth terms:
Let $m,n>0$. The additive shuffle $$\sum_{\sigma\in\Sigma_{m,n}}{\text{\rm Cor}}_{\mathcal{H}}({\varepsilon}_{\sigma{^{-1}}(1)},{\varepsilon}_{\sigma{^{-1}}(1)}+{\varepsilon}_{\sigma{^{-1}}(2)},\dots,{\varepsilon}_{\sigma{^{-1}}(1)}+\dots+{\varepsilon}_{\sigma{^{-1}}(m+n)},0).$$ is a constant independent of ${\varepsilon}_1,\dots,{\varepsilon}_n\in{\mathbb{C}}^n{\setminus}0$.
It is easy to see that this constant is 0 if $m+n$ is even. If $m+n$ is odd, it is equal, in particular, to a sum of Hodge correlators at roots of unity.
### Proofs of Corollaries \[cly:gr28\] and \[cly:gr2729\]
Recall Corollary \[cly:gr28\]
For $n>2$, every Hodge correlator of weight $n$ is a linear combination of Hodge correlators of weight $n$ and depth at most $n-2$.
Precisely, for $z_1,\dots,z_n\in{\mathbb{C}}{^{*}}$, we have $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(z_1,\dots,z_n,0)
&=\sum_{i=1}^n{\text{\rm Cor}}_{\mathcal{H}}{\left(z_1,\dots,z_{i-1},z_i,z_i{\frac{z_1}{z_n}},\dots,z_{n-1}{\frac{z_1}{z_n}},z_n{\frac{z_1}{z_n}}\right)}\nonumber\\
&\quad-\sum_{i=2}^n{\text{\rm Cor}}_{\mathcal{H}}{\left(z_1,\dots,z_{i-1},0,z_i{\frac{z_1}{z_n}},\dots,z_{n-1}{\frac{z_1}{z_n}},z_n{\frac{z_1}{z_n}}\right)}\nonumber\\
&\quad-{\text{\rm Cor}}_{\mathcal{H}}{\left(z_1,z_1\cdot{\frac{z_1}{z_n}},0,\dots,0\right)}.\label{eqn:lower_depth_reduction_app}
\end{aligned}$$
By multiplicative invariance, we may assume $z_1=1$. Then this is precisely the $(n-1,1)$-second shuffle relation applied to the segments
$${\left({z_2/z_1}\mid 0\right)},{\left({z_3/z_2}\mid 0\right)},\dots,{\left({z_n/z_{n-1}}\mid 0\right)}$$ and $${\left({z_1/z_n}\mid 0\right)},$$ where the segment $(1|0)$ is left fixed. Indeed, the two summations come from the $n$ shuffles and the $n-1$ additional quasishuffles, with the remaining terms giving the left side and the last summand.
All terms on the right side have at least two coinciding arguments. After an additive shift, they have at least two arguments equal to 0, so they are equal to those of depth at most $n-2$.
Recall Corollary \[cly:gr2729\]:
The Hodge correlators in weight 3 satisfy the relations: $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,x)&+{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-x)+{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-x{^{-1}})={\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1),\label{eqn:corr_27_app}\\
{\text{\rm Cor}}_{\mathcal{H}}(0,x,1,y)&=
-{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-x{^{-1}})
-{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1-y{^{-1}})
-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{y}{x}}\right)}\nonumber\\\label{eqn:corr_29_app}
&\quad-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-y}{1-x}}\right)}
+{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-y{^{-1}}}{1-x{^{-1}}}}\right)}
+{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1).\end{aligned}$$
Apply the $(1,1)$-second shuffle relation to the segments ${\left(x\mid 0\right)}$ and ${\left(x{^{-1}}\mid 1\right)}$, keeping the segment ${\left(1\mid 1\right)}$ fixed: $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(1,x,0,0)&+{\text{\rm Cor}}_{\mathcal{H}}(1,0,x{^{-1}},1)-{\text{\rm Cor}}_{\mathcal{H}}(1,0,0,1)\\&-{\text{\rm Cor}}_{\mathcal{H}}(1,x,0,0)-{\text{\rm Cor}}_{\mathcal{H}}(1,0,x{^{-1}},0)&=0.\end{aligned}$$ Multiplicative invariance and the first shuffle relation imply $$-{\text{\rm Cor}}_{\mathcal{H}}(1,x,0,0)-{\text{\rm Cor}}_{\mathcal{H}}(1,0,x{^{-1}},0)={\text{\rm Cor}}(1,0,0,x).$$ Rearranging terms and applying additive invariance gives (\[eqn:corr\_27\_app\]).
Now apply (\[eqn:lower\_depth\_reduction\_app\]) to ${\text{\rm Cor}}_{\mathcal{H}}(x,1,y,0)$ and apply the dihedral symmetry and additive invariance to change all terms to the form ${\text{\rm Cor}}_{\mathcal{H}}(1,0,0,z)$: $$\begin{aligned}
{\text{\rm Cor}}_{\mathcal{H}}(0,x,1,y)&=
{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-y}{x-y}}\right)}
+{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-x{^{-1}}}{1-y{^{-1}}}}\right)}
+{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{x-1}{x-y}}\right)}\\
&\quad-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,1-y{^{-1}}\right)}
-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,x{^{-1}}\right)}
-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{x}{y}}\right)}.\end{aligned}$$ Finally, by (\[eqn:corr\_27\_app\]), $${\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-y}{x-y}}\right)}+{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{x-1}{x-y}}\right)}={\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,1\right)}-{\text{\rm Cor}}_{\mathcal{H}}{\left(1,0,0,{\frac{1-x}{1-y}}\right)},$$ which gives the result.
Stuff
=====
### A direct proof for generic parameters
Here we give a direct proof that the periods of the shuffle relations vanish for parameters lying in a generic subset of $({\mathbb{C}}^*)^{m+n}$ that does not depend on specialization to nodal curves (but resembles the proof of Lemma \[lma:circ\_spec\]).
We prove that it is enough to show they are constant by means of an analytic ingredient:
If the canonical real period of $\mathbf{V}^{m,n}_{r_1,\dots,r_{m+n+1}}$ is constant on $P$, then it is 0. \[lma:period\_const\_zero\]
The canonical real period is computed by the Feynman integral construction.
Assume that is it equal to a constant for $p\in P$. Observe that the complement of $P$ is closed in $(S^1)^{m+n}\subset( C^*)^{m+n}$ (where $S^1={\left\{{\left|z\right|}=1\right\}}\subset{\mathbb{C}}^*$). We conclude that the normalization $$\int_{(S^1)^{m+n}}p^{v_\infty}{\overline{{{\rm QSh}}}}^{m,n}_{n_1,\dots,n_{m+n+1}}(a_1,\dots,a_r,b_1,\dots,b_s,c)\,\prod_id\mu(a_i)\prod_jd\mu(b_j)$$ is constant, where $\mu$ is the rotation-invariant measure of volume 1 on $S^1$.
We will show that, in fact, the normalization of each *term* in the quasishuffle is 0. As above, write such a term as $C(z_1(p),\dots,z_{m+n+1}(p))$. Each $z_i$ is either 0 or a product of distinct parameters (coordinates of $p$). Let $I$ be set of those $i$ for which $z_i$ is not 0. Rotation invariance allows to change coordinates: $$\begin{aligned}
\int_{(S^1)^{m+n}}p^{v_\infty}C(z_1(p),\dots,z_{m+n+1}(p))\,\prod_id\mu(a_i)\prod_jd\mu(b_j)
&=\int_{(z_i)_{i\in I}\in(S^1)^{{\left|I\right|}}}p^{v_\infty}C(1,z_1,\dots,z_{m+n+1})\,\prod_{i\in I}d\mu(z_i),\end{aligned}$$ where $z_{m+n+1}=(z_1\dots z_{m+n}){^{-1}}$. The measure is again a product of the rotation-invariant measures, now in the $z_i$. We claim this integral is 0.
Fix a trivalent plane tree $T$ contributing to the period map of the integrand. Because there are at least three boundary vertices, we may choose two of them, labeled by $z_i,z_{i+1}$, distinct from the vertex labeled 1, incident to a common internal vertex $v$. Let $w$ be the third vertex incident to $v$. Then the integral defining the period map is a constant multiple of $$\int_{\text{\rm other internal vertices}}\int_{z_v}\log{\left|z_v-z_i\right|}\,d\log{\left|z_v-z_{i+1}\right|}\wedge\,d\log{\left|z_v-z_w\right|}\wedge{\left(\text{\rm terms from other edges}\right)}.
\label{eqn:period_integral_normalization}$$ Consider the inner integral $\int_{z_v}\log{\left|z_v-z_i\right|}\,d\log{\left|z_v-z_{i+1}\right|}\wedge\,d\log{\left|z_v-z_w\right|}$. It is equal to ${\mathcal{L}}_2{\left({\frac{z_i-z_w}{z_{i+1}-z_w}}\right)}$. When this is integrated over the parameters, we have three cases, all of which give 0:
(1) $i,i+1\notin I$. The inner integral is clearly 0.
(2) $i\in I$, $i+1\notin I$ or $i\notin I$, $i+1\in I$. In the former case, we have $$\int_{z_i\in S^1}{\mathcal{L}}_2{\left({\frac{z_i}{z_w}}\right)}\,d\mu(z_i)=0$$ because ${\mathcal{L}}_2$ is antisymmetric across the real axis. The latter case is symmetric.
(3) $i,i+1\in I$, so we integrate over both $z_i$ and $z_{i+1}$. This is 0 because of the identity ${\mathcal{L}}_2{\left({\frac{z_i-z_w}{z_{i+1}-z_w}}\right)}+{\mathcal{L}}_2{\left({\frac{z_{i+1}-z_w}{z_i-z_w}}\right)}=0$.
Taking the sum over all $T$, this shows the integral \[eqn:period\_integral\_normalization\] is 0 for all terms in the quasishuffle relation. Therefore, the relation’s period is the constant 0.
We remark again that this proof only applies for generic values of the parameters. For example, consider the period of the variation $\textbf{V}^{2,1}_{0,0,0}$:
$$\begin{aligned}
p(C(1,x,xy,xyz)+C(1,x,xz,xyz)+C(1,z,xz,xyz)\\-C(1,0,xz,xyz)-C(1,x,0,xyz)\\-C(1,x,xy,0)-C(1,z,0,0)).\end{aligned}$$
For $x,y,z\neq1$, $xy,xz,yz\neq1$, $xyz\neq1$, the period is 0. However, at $x=y=z=1$, it becomes $$\begin{aligned}
p{\left(3C(1,1,1,1)-3C(1,1,1,0)-C(1,1,0,0)\right)}&=\\
=(2\pi i)^{-2}q\zeta(3)&\quad(q\in{\mathbb{Q}}^*),\end{aligned}$$ which is not 0. However, the content of Theorem \[thm:main\](b) is that the relations hold even when some of the points in the terms of the shuffle relations collide, unless all parameters collapse to 1.
Appendix: Multiple polylogarithms
=================================
We review the properties of multiple polylogarithms ([@goncharov-polylogs-modular]).
It is well known that these functions obey a family of double shuffle relations similar to our relations for the Hodge correlators. However, they do not enjoy some of their other properties. They are multi-valued and do not satisfy dihedral symmetry relations. The shuffle relations between multiple polylogarithms involve products, while for Hodge correlators they are linear.
\[sec:zeta\_shuffle\]
### Multiple polylogarithms
The multiple polylogarithms are defined by $${\text{\rm Li}}_{n_1,\dots,n_r}(z_1,\dots,z_r)=\sum_{0<k_1<\dots<k_r}{\frac{z_1^{k_1}\dots z_r^{k_r}}{k_1^{n_1}\dots k_r^{n_r}}},\quad{n_1,\dots,n_r>0}.
\label{eqn:def_mult_plog}$$ (The *depth* of this formal expression is $r$ and the *weight* is $w:=n_1+\dots+n_r$.) These series converge for ${\left|z_i\right|}<1$ and have analytic continuations to multivalued functions with singularities on ${\mathbb{C}}^r$. The multivalued structure is encoded by a smooth variation of mixed Hodge-Tate structures of weight $w$ over a dense open subset of ${\mathbb{C}}^r$.
When $r=1$, the multiple polylogarithms are the classical polylogarithms ${\text{\rm Li}}_n(z)$. Their monodromy and associated mixed Hodge-Tate structures are well understood ([@hain-polylogs]).
We can form an algebra $L$ generated over ${\mathbb{Q}}$ by the multiple polylogarithms, filtered by the weight and the depth. The expression (\[eqn:def\_mult\_plog\]) yields expansions for products of polylogarithms, which shows that $L$ has a well-defined multiplication. For example, $$\begin{aligned}
{\text{\rm Li}}_{n_1}(z_1){\text{\rm Li}}_{n_2}(z_2)
&={\left(\sum_{0<k_1}{\frac{z_1^{k_1}}{k_1^{n_1}}}\right)}{\left(\sum_{0<k_2}{\frac{z_2^{k_2}}{k_2^{n_2}}}\right)}
={\left[\sum_{0<k_1<k_2}+\sum_{0<k_2<k_1}+\sum_{0<k_1=k_2}\right]}{\frac{z_1^{k_1}z_1^{k_2}}{k_1^{n_1}k_2^{n_2}}}\\
\vphantom{\sum}
&={\text{\rm Li}}_{n_1,n_2}(z_1,z_2)+{\text{\rm Li}}_{n_2,n_1}(z_2,z_1)+{\text{\rm Li}}_{n_1+n_2}(z_1z_2).\end{aligned}$$ Notice that the left side and all terms on the right side have weight $n_1+n_2$; however, the left side and the first two terms on the right side have depth 2, while ${\text{\rm Li}}_{n_1+n_2}(z_1z_2)$ has depth 1.
The general relation is: $$\begin{aligned}
&{\text{\rm Li}}_{n_1,\dots,n_r}(z_1,\dots,z_r){\text{\rm Li}}_{n_{r+1},\dots,n_{r+s}}(z_{r+1},\dots,z_{r+s})\nonumber\\
&=
\sum_{\sigma\in\Sigma_{r,s}}{\text{\rm Li}}_{n_{\sigma{^{-1}}(1)},\dots,n_{\sigma{^{-1}}(r+s)}}(z_{\sigma{^{-1}}(1)},\dots,z_{\sigma{^{-1}}(r+s)})+\text{lower-depth terms},
\label{eqn:first_shuffle_polylog}\end{aligned}$$ Expressions (\[eqn:first\_shuffle\_polylog\]) are called *first shuffle relations* for multiple polylogarithms. It is convenient to express them with generating functions. Let $$L{\big(z_1,\dots,z_r\;|\;t_1:\dots:t_r\big)}\sum_{n_r>0}{\text{\rm Li}}_{n_1,\dots,n_r}(z_1,\dots,z_r)\prod_it_i^{n_i-1};$$ then $$\begin{aligned}
&L{\big(z_1,\dots,z_r\;|\;t_1:\dots:t_r\big)}L{\big(z_{r+1},\dots,z_{r+s}\;|\;t_{r+1},\dots,t_{r+s}\big)}
=\nonumber\\&=
\sum_{\sigma\in\Sigma_{r,s}}L{\big(z_{\sigma{^{-1}}(1)},\dots,z_{\sigma{^{-1}}(r+s)}\;|\;t_{\sigma{^{-1}}(1)}:\dots:t_{\sigma{^{-1}}(r+s)}\big)}+\text{lower-depth terms}.\label{eqn:l_shuf_inhom}\end{aligned}$$
To describe the lower-depth terms in the right side of (\[eqn:first\_shuffle\_polylog\]), one needs to work with the set of *quasishuffles* $\widetilde\Sigma_{r,s}$. Then $$\begin{aligned}
{\text{\rm Li}}_{n_1,\dots,n_r}(z_1,\dots,z_r){\text{\rm Li}}_{n_{r+1},\dots,n_{r+s}}(z_{r+1},\dots,z_{r+s})
=\nonumber\\=
\sum_{\sigma\in\widetilde\Sigma_{r,s}}{\text{\rm Li}}_{\tilde n_{\sigma{^{-1}}(1)},\dots,\tilde n_{\sigma{^{-1}}(M_\sigma)}}(\tilde z_{\sigma{^{-1}}(1)},\dots,\tilde z_{\sigma{^{-1}}(M_\sigma)}),
\label{eqn:polylog_quasishuf}\end{aligned}$$ where $$\tilde n_{\sigma{^{-1}}(i)}=\sum_{\sigma(j)=i}n_j,\quad\tilde z_{\sigma{^{-1}}(i)}=\prod_{\sigma(j)=i}z_j.$$ Such relations are easily proved by interpreting the terms as the simplicial decomposition of the product of an $r$-simplex and an $s$-simplex.
### Iterated integrals
The analytic continuation of the multiple polylogarithms has a presentation in terms of iterated integrals. Let $$I_{n_1,\dots,n_r}(z_1:z_2:\dots:z_{r+1})=\int_{\gamma}\underbrace{{\frac{dt}{z_1-t}}\circ{\frac{dt}{t}}\circ\dots\circ{\frac{dt}{t}}}_{n_1}\circ\dots\circ\underbrace{{\frac{dt}{z_r-t}}\circ{\frac{dt}{t}}\circ\dots\circ{\frac{dt}{t}}}_{n_r},$$ where $\gamma:{\left[0,1\right]}\to{\mathbb{C}}$ is a path from 0 to $z_{r+1}$. Here, for 1-forms $\omega_1,\dots,\omega_r$, $$\int_\gamma\omega_1\circ\dots\circ\omega_r:=\int_{0\leq t_1\leq\dots\leq t_r\leq 1}\bigwedge_{i=1}^m\gamma^*\omega_i(t_i)$$ is Chen’s iterated path integral ([@chen-iterated-integrals]). Then ([@goncharov-polylogs-modular], Theorem 2.1) $${\text{\rm Li}}_{n_1,\dots,n_r}(z_1,\dots,z_r)=I_{n_1,\dots,n_r}(1:z_1:z_1z_2:\dots:z_1\dots z_r).
\label{eqn:li_iterint}$$ Iterated path integrals also satisfy a shuffle product formula, whose terms correspond to the top-dimensional cells of a decomposition of the product of two simplices: $$\int_\gamma\omega_1\circ\dots\circ\omega_r\int_\gamma\omega_{m+1}\circ\dots\circ\omega_{m+n}=\sum_{\sigma\in\Sigma_{m,n}}\omega_{\sigma{^{-1}}(1)}\circ\dots\circ\omega_{\sigma{^{-1}}(m+n)}.$$ This gives a different kind of shuffle relations (*second shuffle relations*) on the iterated integrals $I_{n_1,\dots,n_r}$, which can also be expressed in terms of generating functions. Let $$\begin{aligned}
&L'{\big(z_1:\dots:z_{r+1}\;|\;t_1,\dots,t_r\big)}=\nonumber\\=&\sum_{n_i>0}I_{n_1,\dots,n_r}{\left(z_1:\dots:z_{r+1}\right)}t_1^{n_1-1}(t_1+t_2)^{n_2-1}\dots(t_1+\dots+t_r)^{n_r-1},
\label{eqn:def_hom_polylog}\end{aligned}$$ so $$L{\big(z_1,\dots,z_r\;|\;t_1:\dots:t_r\big)}=L'{\big(1:z_1:\dots:z_1\dots z_r\;|\;t_1,t_2-t_1,\dots,t_r-t_{r-1}\big)}.
\label{eqn:gf_duality_polylog}$$ Then $$\begin{aligned}
L'{\big(z_1:\dots:z_r:1\;|\;t_1,\dots,t_r\big)}L'{\big(z_{r+1}:\dots:z_{r+s}:1\;|\;t_{r+1},\dots,t_{r+s}\big)}
=\nonumber\\=
\sum_{\sigma\in\Sigma_{r,s}}L'{\big(z_{\sigma{^{-1}}(1)}:\dots:z_{\sigma{^{-1}}(r+s)}:1\;|\;t_{\sigma{^{-1}}(1)},\dots,t_{\sigma{^{-1}}(r+s)}\big)}.\label{eqn:l_shuf_hom}\end{aligned}$$
### Double shuffle relations
Note the similarity between (\[eqn:l\_shuf\_inhom\]) and (\[eqn:l\_shuf\_hom\]). There is a duality between the relations with homogeneous and inhomogeneous $z_i$ and $t_i$ arguments. Together, they form systems of *double shuffle relations*.
The combinatorics of such relations are studied by [@goncharov-polylogs-modular; @goncharov-polylogs-tate], allowing them to describe a connection between an algebra of values of the multiple polylogarithms at roots of unity and the geometry of some locally symmetric spaces for ${\mathrm{GL}}_n({\mathbb{Z}})$ ($n=2,3$; and recently for $n=4$ in [@goncharov-motivic-modular]).
### Relation to Hodge correlators {#sec:rel_hodge_polylog}
In depth 1, the Hodge correlators are related to the multiple polylogarithms. We have seen this in weights 1 and 2. In higher weight, define a single-valued version of the polylogarithm by $${\mathcal{L}}_n(z)=\begin{cases}\Re&\text{$n$ odd}\\\Im&\text{$n$ even}\end{cases}{\left(\sum_{k=0}^{n-1}\beta_k\log^k{\left|z\right|}\cdot{\text{\rm Li}}_{n-k}(z)\right)}\quad(n\geq2),$$ where $\beta_k$, close relatives of the Bernoulli numbers, are the coefficients of the Taylor expansion ${\frac{2x}{e^{2x}-1}}=\sum\beta_kx^k$. Then $${\text{\rm Cor}}_{\mathcal{H}}(1,\underbrace{0,\dots,0}_{n-1},z)=-(2\pi i)^{-n}\binom{2n-2}{n-1}{^{-1}}\sum_{\stackrel{0\leq k\leq n-2}{\text{$k$ even}}}\binom{2n-k-3}{n-1}{\frac{2^{k+1}}{(k+1)!}}{\mathcal{L}}_{n-k}(z)\log^k{\left|z\right|}.
\label{eqn:corr_polylog}$$ The precise relationship between the multiple polylogarithms and Hodge correlators in depth $>1$ is unknown.
### The relations modulo depth
There is a simpler relation modulo the depth filtration:
Fix a set $S\subset{\mathbb{P}}^1({\mathbb{C}})$ such that all terms below are defined. Let $D$ be the depth filtration induced by ${\left\{0\right\}}\subset S$, and let $z_i\in{\mathbb{C}}^*$. The elements $$\begin{aligned}
{{\rm Sh}}^{m,n}_{n_1,\dots,n_{m+n+1}}(z_1,\dots,z_{m+n+1}):=\nonumber\\=\sum_{\sigma\in\Sigma_{m,n}} C^*_{n_{\sigma{^{-1}}(1)},\dots,n_{\sigma{^{-1}}(m+n)},n_{m+n+1}}(z_{\sigma{^{-1}}(1)},\dots,z_{\sigma{^{-1}}(m+n)},z_{m+n+1})
\label{eqn:def_sh}\end{aligned}$$ vanish in ${\text{\rm gr}}^D{\text{\rm Cor}}_{\text{\rm Hod}}^{v_\infty}{\left({\mathcal{CL}}^\vee_{{\mathbb{P}}^1,S,v_\infty}\right)}$. The motivic versions of these elements vanish in ${\text{\rm gr}}^D{\text{\rm Cor}}_{\text{\rm Mot}}^{v_\infty}{\left({\mathcal{CL}}^\vee_{{\mathbb{P}}^1,S,v_\infty}\right)}$. \[cly:main\]
Compare (\[eqn:first\_shuffle\_polylog\]).
Suppose not all $n_i=0$ and not all $z_i=1$. Modulo the depth filtration, the terms of ${\overline{{{\rm QSh}}}}$ coming from quasishuffles that are not proper shuffles (where some $\sigma(i)=\sigma(j)$) and the two correction terms vanish, and the statement follows from Theorem \[thm:main\](b,c).
Suppose now that all $n_i=1$ and all $z_i=1$. Consider the specialization of ${\overline{{{\rm QSh}}}}$ as $z_i$ to $1$ along $z_i=1+{\varepsilon}_it$. In the terms of ${\overline{{{\rm QSh}}}}$ corresponding to the improper quasishuffles and correction terms, not all points collide at $t=0$. The terms from proper shuffles are all equal to $C(1,1,\dots,1)=0$. This means the desired relation holds up to the specialization at 0 of the proper shuffle (\[eqn:def\_sh\]).
We must show this specialization has lower depth. Supposing not all ${\varepsilon}_i=0$, the proper shuffle terms land in components of the limiting nodal curve $X$ distinct from the main component $X_0$, with their common parent $X_1$ attached to the $X_0$ at $z=1$. Monomials $\prod z_i$ degenerate to $z=\sum{\varepsilon}_i$ on $X_1$, and $1$ degenerates to $z=0$, where $z$ is the coordinate on $X_1$. By Theorem \[thm:degen\_nodal\], every term in the specialization has the form $C(0,{\overline{z}}_1,\dots,{\overline{z}}_{m+n})$, where each ${\overline{z}}_i$ is a sum of the ${\varepsilon}_i$ – an element of lower depth. (The specialization of the shuffle relation is an “additive shuffle”.) Choosing any ${\varepsilon}_i\in S$, not all 0, such that all appearing terms are in $S$ (for example, ${\varepsilon}_i=0$ unless $i=1$) we get the result.
The relations (\[eqn:def\_olqsh\]) of Theorem \[thm:main\] can be seen as a deformation of the relations (\[eqn:def\_sh\]) of Corollary \[cly:main\]. The elements form a *nontrivial* coideal modulo depth: they are not 0; in particular, their periods are not 0.
[@goncharov-hodge-correlators] points out that modulo the depth filtration, the coproduct concides with the dihedral Lie algebra of [@goncharov-polylogs-modular]. However, this is true only under multiplicative invariance, which does not hold in weight 1 if the $C(0,a_i)$, $C(0,b_j)$ do not vanish.
### Expression in generating functions
We can express the second shuffle relations using generating functions, just as we did for the first shuffle relations (\[eqn:first\_shuffle\_corr\]). Define $$\begin{aligned}
{{\rm QSh}}^{m,n}{\big(z_1,\dots,z_{m+n+1})\;|\;T_1,\dots,T_{m+n+1}\big)}:=\\
=\sum_{\sigma\in\widetilde\Sigma_{m,n}}(-1)^{\operatorname{sgn}\sigma}{\mathbf{\Lambda}^*}{\left(\tilde z_{\sigma{^{-1}}(1)},\dots,\tilde z_{\sigma{^{-1}}(M_\sigma)},z_{m+n+1}\right)}{\tilde T_1,\dots,\tilde T_{M_\sigma},T_{m+n+1}},\end{aligned}$$ where $$\tilde T_i=\bigsqcup_{j\in\sigma{^{-1}}(i)}T_i,\quad\tilde z_{\sigma{^{-1}}(i)}=\prod_{\sigma(j)=i}z_j,$$ That is, we rearrange the ${\big(z_i\;|\;T_i\big)}$ in the order specified by $\sigma$, and if ${\big(z_i\;|\;T_i\big)}$ and ${\big(z_j\;|\;T_j\big)}$ collide, we write ${\big(z_iz_j\;|\;T_i\sqcup T_j\big)}$ and change the sign.
Also define the corrected quasishuffle $$\begin{aligned}
{\overline{{{\rm QSh}}}}^{m,n}{\big(a_1,\dots,a_r,b_1,\dots,b_s,c)\;|\;T_1,\dots,T_m,U_1,\dots,U_n,V\big)}:=\nonumber\\
={{\rm QSh}}^{m,n}{\big(a_1,\dots,a_r,b_1,\dots,b_s,c)\;|\;T_1,\dots,T_m,U_1,\dots,U_n,V\big)}\nonumber\\
-{\mathbf{\Lambda}^*}{\big(a_1,\dots,a_r,\prod_jb_j\cdot c\;|\;T_1,\dots,T_m,\bigsqcup_jU_j\sqcup V\big)}\nonumber\\
-{\mathbf{\Lambda}^*}{\big(b_1,\dots,b_s,\prod_ia_i\cdot c\;|\;U_1,\dots,U_n,\bigsqcup_iT_i\sqcup V\big)}.\label{eqn:def_olqsh}\end{aligned}$$
For example: $$\begin{aligned}
{\overline{{{\rm QSh}}}}^{2,1}{\big(a_1,a_2,b,c\;|\;T_1,T_2,U,V\big)}
&=[{\mathbf{\Lambda}^*}{\big(a_1,a_2,b,c\;|\;T_1,T_2,U,V\big)}\\
&\quad+{\mathbf{\Lambda}^*}{\big(a_1,b,a_2,c\;|\;T_1,U,T_2,V\big)}\\
&\quad+{\mathbf{\Lambda}^*}{\big(b,a_1,a_2,c\;|\;U,T_1,T_2,V\big)}\\
&\quad-{\mathbf{\Lambda}^*}{\big(a_1b,a_2,c\;|\;T_1\sqcup U,T_2,V\big)}\\
&\quad-{\mathbf{\Lambda}^*}{\big(a_1,a_2b,c\;|\;T_1,T_2\sqcup U,V\big)}]\\
&\quad-{\mathbf{\Lambda}^*}{\big(a_1,a_2,bc\;|\;T_1,T_2,U\sqcup V\big)}\\
&\quad-{\mathbf{\Lambda}^*}{\big(b,a_1a_2c\;|\;T_1\sqcup T_2\sqcup V\big)}&=0.\end{aligned}$$
[^1]: In this paper, “Hodge correlators” will refer only to Hodge correlators associated to the curve ${\mathbb{P}}^1$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Any graph can be represented pictorially as a figure. Moreover, it can be represented as two or more figures that can be have different properties to each other. For the purpose of HCP, we represent a graph by two such figures. In each of them, there is an exterior part called the [*contour*]{}, and an interior part. These two figures differ from each other by the constitution of the edges in the interior part. That is, any edges in the interior part for one figure are not in the interior for the other figure. We call these two figures [*basic objects*]{}.
We develop rules and algorithms that allow us to represent any graph of degree $d \leq 3$ by two basic objects. Individually, neither of these representations possess the features to easily determine the Hamiltonicity of the graph. However, the combination of these two figures, once certain weights are assigned to their edges, allows us to determine the Hamiltonicity with a polynomial-time check.
The rules for the assignment of weights are:
1. The weight of any edge of the interior part is 0, for both objects.
2. In both figures any common edge of the contour has the same weight.
The weights of the edges allow us to extend the number of parameters of the objects, that is sufficient to determine the Hamiltonicity of the graph. Then, if the graph is Hamiltonian, then both figures possess the same set of parameters. If the sets of parameters for two figures are different, then the graph is not Hamiltonian. The parameters that determine the Hamiltonicity of the graph are the sums of weights of edges and [*windows*]{} of contours in the figure. The algorithms of their construction do not contain a combinatorial number of elements and have polynomial complexity. We also supply an estimate of the complexity of each algorithm.
author:
- 'Ivan I. Goray[^1]'
title: On polynomial solvability of the Hamiltonian cycle problem for graphs of degree less than or equal to 3
---
[P vs NP, HCP, Hamiltonicity of graphs, graphs of degree less than or equal to 3, equivalent graphs, graph representation by basic objects, contour and interior parts of graphs, assignment weights to edges and nodes, Hamiltonian cycle length]{}
68Q25, 68R10, 03D15
Introduction {#sec1}
============
To date, a solution of the general problem P $\neq$ NP has not be found. If this problem can be solved then it will be most likely on the basis of some common features of the classes P and NP. Looking for a solution of any particular P vs NP problem seems to lack perspective. If it were possible to obtain, for a particular NP problem, that we cannot solve it in polynomial time, it would be necessary to supply the complete set of algorithms used in the solution of this problem, which is most likely impossible. Thinking deeper about the NP phenomenon and about the possibility of solving such problems leads to some questions: Which algorithms, in principle, could be constructed to solve particular NP problems ? Can we avoid combinatorial enumeration in such algorithms ? Can we overcome NP-completeness with the help of such algorithms ? Garey and Johnson [@1] say that polynomial algorithms can be constructed only when it is possible to penetrate deep into the essence of the problem in question. The success in solving particular NP problems using optimization methods is very high (e.g. [@2; @3; @4; @5]). Most likely, here is the key of the difficulty of finding the precise solution, because one can not substitute the essence of the problem by advances in modern optimization technologies. It is commonly accepted [@6; @7; @8] that the branch and bound method in all NP problems is the most universal and effective. It is possible that branch and bound methods mostly captures the difficulty of the problem as it is based in investigation and transformation of the network.
The solution of the HCP that we offer here is not based on optimization. It does not contain a progression to a better solution. This solution is based on the analysis of structural differences between two objects constructed on the basis of the same non-oriented graph with degree less than equal to 3.
Objects {#sec2}
=======
Basic object
------------
### Properties of basic object
![An example of a first basic object and its properties. Nodes forming windows are 2, 6, 10, 18, 14, 25. Windows are $w(2-6)$, $w(10-18)$, $w(14-25)$. Free edges are $e(22-24)$, $e(9-12)$, $e(8-15)$, $e(17-23)$. Segments are $S_{2,25}$, $S_{14,18}$, $S_{6,10}$.\[fig2-1\]](Fig2-1.eps){width="30.00000%"}
Given a graph $\Gamma$ of size $N$, we define object $G_N$ as a figure representing $\Gamma$ that, by some property or properties, differs from other figures representing the same graph $\Gamma$. We can solve the HCP on a set of objects, and construct a parameter for each object such that the parameters should not coincide unless the graph is Hamiltonian. While, individually, these objects do not solve HCP, the solution of the problem is obtained on a finite set of the objects, by considering their special properties. In fact, for any two such figures, we will find a parameter that will distinguish between them.
For certain special cases, it is only possible to construct a single, unique figure. In this case, the solution of the HCP is given trivially by construction of this figure. However, in general, for the solution of HCP, one has to construct two objects that differ in the constitution of edges in different parts of the objects, and in the method of construction of those parts. As an example, in Figure 2.1, we display the first (basic) object constructed from a 25-node graph.
For convenience, we will distinguish between the exterior part of the figure, that we call the contour, and the interior part. The contour contains the edges of the graph and so-called [*windows*]{}. A window is an edge that does not belong to the graph. In the example in Figure 2.1, there are three windows: $w(2-6)$, $w(10-18)$, and $w(14-25)$, that are formed by the nodes 2 and 6, 10 and 15, and 14 and 25 of the graph $\Gamma$.
The edges of the interior part of the object, that are not connected at either end to the nodes that form the windows (for convenience, we refer to these nodes as [*window nodes*]{}), are called [*free*]{} edges. In the example given in Figure 2.1, the free edges are $e(22-24)$, $e(17-23)$, $e(8-15)$, $e(9-12)$.
As follows from the example, the basic object has the following properties.
1. Window nodes are not directly connected by the edges of the interior part. We refer to such edges as [*links*]{}.
2. The basic object contains a [*maximal*]{} number of windows (three in this example). That is, no further windows can be inserted without violating the previous property.
The contour is partitioned by windows into several parts, and can be determined and enumerated. The parts of the object separated by windows will be called [*segments*]{}. In the example in Figure 2.1, there are three segments, between nodes 2 and 25, 14 and 18, and 6 and 10, which we denote by $(S_{2,25}$, $S_{14,18}$, $S_{6,10})$. A segment is called [*degenerate*]{} if it consists of a single node such that has a window on both sides in the contour.
The basic objects that are constructed for the HCP should satisfy the following properties.
1. The number of windows in the object should be maximal. As we show later, we will use a simple algorithm to increase the number of windows, and we will determine the criterion that signals the end of the algorithm.
2. Window nodes should not be connected to other window nodes by the edges of the interior part.
3. The object should not contain degenerate segments if the degree of the graph is $d = 3$. Degenerate segments can only exist for nodes of degree $d = 2$.
4. For a given graph, the set of edges in the interior part should be different for any pair of basic objects.
[**Lemma 1.**]{} For a degenerate segment-free graph $\Gamma$ of size $N$, and degree $d \leq 3$, the number of segments in all basic objects is less than or equal to $\nu = [N/6]$.
First, we observe that, by property 1, the nodes that form the windows are not interconnected. Therefore, every node that forms a window is connected by the edges of the interior part with another two nodes that do not form windows. From the fragment of the graph shown in Figure 2.2, we see that window $w(r-j)$, together with nodes $a, b, c$ and $e$ constitutes a set of six nodes in the graph. Since in the absence of degenerate segments this is true for every window, then there can not be more than $\nu = [N/6]$ segments.
![The nodes $a$, $b$, $c$, $e$, $r$ and $j$ that correspond to a window $w(r-j)$, and its incident nodes. None of these nodes can be incident to any other window.\[fig2-2\]](Fig2-2.eps){width="25.00000%"}
1. The number of windows can be larger if there are degenerate segments.
2. If there are free edges, then the number of free edges is less than $\nu$.
3. If the degree of the graph was not bounded by 3 then the number of windows in the basic object constructed for the graph would be strictly less than $\nu$. This is because the nodes that form the windows for $d > 3$ will be connected to a larger number of nodes in the graph that, in view of property 2, can not form windows themselves.
### Construction of the basic object
![A transformation that eliminates a link $e(b-c)$ between windows $w(a-b)$ and $w(c-d)$ (in Figure 2.3a). Edge $e(b-c)$ is moved to the contour, and windows $w(a-b)$ and $w(c-d)$ are substituted by a single window, $w(a-d)$ (in Figure 2.3b).\[fig3ab\]](Fig2-3.eps){width="50.00000%"}
Consider the problem of constructing a basic object that satisfies properties (1)–(4) above. This can be performed in two steps. In the first step, we can use elementary procedures that are from known optimization algorithms [@9], that are described below.
[*Step 1: Construction of a graph whose windows are not connected by the edges of the interior part.*]{}
It is known that optimization algorithms ensure an optimal solution is found by gradually eliminating links between the windows by means of elementary network transformations. For this reason, they can be used to construct a basic object such that nodes that form the windows are not interconnected. It is useful to note that for the first step it is sufficient to use the procedure of checking and substitution of no more than two edges on the contour.
![A transformation that eliminates a link $e(b-d)$ between windows $w(a-b)$ and $w(c-d)$ (in Figure 2.4a). Edge $e(b-d)$ is moved to the contour, and windows $w(a-b)$, $w(c-d)$, $w(e-f)$ are substituted by two windows, $w(a-e)$ and $w(c-f)$ (in Figure 2.4b).\[fig4ab\]](Fig2-4.eps){width="50.00000%"}
There are four possible cases of elimination of links:
[**Case 1.1**]{}: We use the transformation that substitutes two windows $w(a-b)$, $w(c-d)$ (see Figure 2.3a), with a single window $w(a-d)$ and edge $e(b-c)$ (see Figure 2.3b). In this case, the interior edge between nodes $b$ and $c$ is moved to the contour.
[**Case 1.2**]{}: We use the transformation that substitutes three windows $w(a-b)$, $w(c-d)$, $w(e-f)$ (see Figure 2.4a), with two windows $w(a-e)$, $w(c-f)$, and the interior edge $e(b-d)$ is moved to the contour (see Figure 2.4b).
[**Case 1.3**]{}: Suppose the contour contains a segment $S_{b,f}$, and its endpoints are connected with an interior edge $e(b-f)$. Suppose that $S_{b,f}$ contains a node $l$ that is a neighbour to node $m$ on the contour, and that is linked by an interior edge $e(l-c)$ with another node that forms a window (see Figure 2.5a). In this case, we use the transformation that substitutes three windows $w(a-b)$, $w(c-d)$, $w(e-f)$ with two windows $w(a-e)$, $w(m-d)$, and moves two interior edges $e(b-f)$ and $e(l-c)$ to the contour (see Figure 2.5b). The edge $e(l-m)$ is moved to the interior.
![A transformation that eliminates a link $e(b-f)$ between windows $w(a-b)$ and $w(e-f)$ (in Figure 2.5a). Edges $e(b-f)$ and $e(c-l)$ are moved to the contour, and windows $w(a-b)$, $w(c-d)$ and $w(e-f)$ are substituted by two windows, $w(a-e)$ and $w(d-m)$ (in Figure 2-5b).\[fig2-5\]](Fig2-5.eps){width="50.00000%"}
[**Case 1.4**]{}: This case differs from case 3 in that $S_{b,f}$ does not contain any nodes that are connected with a node that forms a window. In this case, we select a node $l$ on $S_{b,f}$ for which an interior arc $e(l-s)$ goes to a different segment that contains a node $t$ that is a neighbour to node $s$ on the contour (see Figure 2.6a). Then, we use the transformation that moves edges $e(b-f)$, $e(s-l)$ to the contour, and substitutes windows $w(a-b)$ and $w(e-f)$ with two windows $w(a-t)$, $w(e-n)$ (see Figure 2.6b).
![A transformation that eliminates a link $e(b-f)$ between windows $w(a-b)$ and $w(e-f)$ (in Figure 2.6a). Edges $e(b-f)$ and $e(l-s)$ are moved to the contour, and windows $w(a-b)$ and $w(e-f)$ are substituted by two windows, $w(a-t)$ and $w(e-m)$ (in Figure 2.6b).\[fig2-6\]](Fig2-6.eps){width="50.00000%"}
In all of these cases, an interior edge that joins two window nodes is eliminated from the set of interior edges. The object that we construct after Step 1 can contain an arbitrary number of windows, but cannot contain degenerate segments formed by nodes of degree $d = 3$. We can achieve this by elimination of degenerate segments from the object as described as follows.
[**Case 1.5**]{}: Suppose that a degenerate segment contains node $a$ that form two windows $w(a-b)$, $w(a-c)$ (see Figure 2.7a). Suppose that node $a$ has degree $d = 3$, and is connected to nodes $k$, $l$ and $n$, none of which form windows. To eliminate $S_{a,a}$, we use the transformation that moves two interior edges $e(a-l)$ and $e(a-m)$ to the contour, and substitutes windows $w(a-b)$ and $w(a-c)$ with two windows $w(s-b)$, $w(t-c)$ (see Figure 2.7b).
![A transformation that eliminates a degenerate segment containing a node with degree $d = 3$ (in Figure 2.7a). Edges $e(a-l)$ and $e(a-m)$ are moved to the contour, and windows $w(a-b)$ and $w(a-c)$ are substituted by two windows, $w(b-s)$, $w(c-t)$(in Figure 2.7b).\[fig2-7\]](Fig2-7.eps){width="50.00000%"}
[**Case 1.6**]{}: If at least one of nodes $k$, $l$ and $m$ forms a window, then the degenerate segment $S_{a,a}$ is eliminated from the object by either of Cases 1.1 and 1.2 above.
We note that the above cases are sufficient to eliminate all interior edges between windows, and all degenerate segments, which completes Step 1.
[*Step 2: Introduction of additional windows*]{}
We can introduce additional windows if there are special structures involving the free edges in the object formed in Step 1. Recall that the free edges are those that are not incident to any window nodes. We note that introduction of additional windows should not introduce any links, or else property 1 will be violated.
The windows are added by means of elementary transformations of one free edge to the contour, and two edges of the contour to the interior. There are two cases when it is possible to introduce an additional window to the object formed in Step 1:
![A transformation that creates a windows $w(k-m)$ (in Figure 2.8b), using free edges $e(s-t)$, $e(k-l)$ and $e(m-n)$ (in Figure 2.8a).\[fig2-8\]](Fig2-8.eps){width="50.00000%"}
[**Case 2.1**]{}: Suppose that an object formed in Step 1 contains free edges $e(k-l)$, $e(m-n)$ and $e(s-t)$ such that $m$ is a neighbour of $t$ on the contour, $k$ is a neighbour of $s$ on the contour, and $k$ and $m$ are on different sides of edge $e(s-t)$. Nodes $l$ and $n$ can be anywhere on the contour (see Figure 2.8a). An additional window $w(k-m)$ is introduced by the transformation that moves interior edge $e(s-t)$ to the contour, and moves edges $e(k-s)$ and $e(t-m)$ to the interior (see Figure 2.8b). This transformation is also valid if one or both of the edges $e(k-l)$ and $e(m-n)$ are not present. In this case, one or both of nodes $k$ and $m$ will have degree 2.
[**Case 2.2**]{}: Suppose than on object formed in Step 1 contains free edges $e(k-l)$, $e(m-n)$ and $e(s-t)$, such that $m$ is a neighbour of $t$ on the contour, and $k$ is a neighbour of $s$ on the contour, and nodes $k$, $m$ and $n$ are all on the same side of edge $e(s-t)$, and a window $w(a-b)$ exists (see Figure 2.9a). An additional window can be formed by a transformation of the part of the contour bounded by nodes $m$ and $k$. In Figure 2.9a, this is the part of the contour to the left of edge $e(s-t)$) to become a segment between nodes $a$ and $b$ (see Figure 2.9b). As a result of this transformation, we substitute window $w(a-b)$ with windows $w(a-k)$ and $w(b-m)$. This transformation is also valid if one or both of the edges $e(k-l)$ and $e(m-n)$ are not present. In this case, one or both of nodes $k$ and $m$ will have degree 2.
![A transformation that creates an additional window by moving a part of the contour bounded by $m$ and $k$ between $a$ and $b$ that form window $w(a-b)$ (in Figure 2.9a). This transformation substitutes window $w(a-b)$ by two windows, $w(a-k)$ and $w(b-m)$ (in Figure 2.9b).\[fig2-9\]](Fig2-9.eps){width="50.00000%"}
Therefore, in both cases, if we have a special configuration of free edges, we can construct additional windows so that the edges $e(k-l)$, $e(m-n)$, $e(s-t)$ are no longer free. If the object contains configurations analogous to those considered above (see Figures 2.8a, 2.9a), then additional edges should be moved to its contour. We note that none of these transformations should be performed if they result in degenerate segments that contain a node whose degree $d = 3$.
A basic object is completed if it is does not contain free edges that could lead to the introduction of additional windows using the transformations described above. In such a case, we say that the number of windows in the object is maximal. Note that depending on the construction of the graph, it may be possible to construct a different object containing more windows, but it is not necessary to do so. A basic object may contain one or more free edges, if the configuration of those free edges does not coincide with those described above. This is the only criteria required to complete the second step of the construction of a basic object.
Second basic object
-------------------
The second basic object should possess the same properties (see Subsection 2.1.1) as the first basic object. The second basic object is constructed from the first basic object. The two basic objects will have no common interior edges. This requires us to introduce additional restrictions in the algorithm for constructing the second basic object, that is otherwise analogous to the algorithm that constructs the first basic object. These restrictions concern the structure of the contour, and specify the nodes which cannot not become window nodes in the second basic object. We satisfy these new requirements in the first step of the construction of the second object.
### Preliminary formation of the second basic object
Suppose that the first basic object is constructed, and satisfies all of the required properties. Then, we identify:
1. The window nodes, and the interior edges incident to these nodes. The number of such edges will be either two (if the node has degree $d = 3$), or one (if the node has degree $d = 2$).
2. The free edges.
3. Degenerate segments that contain a single node of degree $d = 2$.
4. The set of nodes with degree $d = 2$ that do not form windows.
Note that the first basic object cannot contain any interior edges or nodes of degree $d = 2$ that can be different from the cases described above. This phenomenon is determined by the properties of the construction of the first basic object: the original graph has degree $d \leq 3$; the windows of the first basic object are not linked by interior edges; and the first basic object does not contain degenerate segments of degree $d = 3$.
The first basic object is constructed without restrictions on the contour edges, interior edges, or on the configuration of nodes with degree $d = 2$. The construction of the second basic object depends on the particular structure of the first basic object.
[*Algorithm of the preliminary construction of the second basic object:*]{}
1. Identify in the first basic object:
1. [*Chains*]{} that consist of two interior edges that are both incident to a window node of degree $d = 3$.
2. Interior edges that are incident to the window nodes of degree $d = 2$. These nodes will form windows in the second basic object as well.
3. Free edges.
4. Chains that consist of two interior edges incident to a node of degree $d = 2$ that form a degenerate segment.
5. Nodes of degree $d = 2$ that do not form windows. These nodes will form degenerate segments in the second basic object.
2. The identified chains consisting of two edges, free interior edges, and the degenerate segments of degree $d = 2$ all move to the contour, and are all separated by windows.
3. Connect the nodes of the contour in the second basic object by interior edges that belong to the contour of the first basic object.
4. Now, we eliminate some interior edges some interior edges from the second basic object that cannot be moved to the contour. These are edges that are:
1. Incident to window nodes of degree $d = 2$.
2. Incident to window nodes in the first basic object, and belong to its contour.
5. Mark those nodes that cannot form windows. By means of the first two algorithms of eliminating links between windows (see Subsection 2.1.2, Cases 1.1 and 1.2) move, to the contour, interior edges incident to these marked nodes.
6. Mark edges of the contour that cannot be moved to the interior.
![A transformation that allows us to delete node $c$ from the list of candidates for window nodes by elimination of the link between windows $w(c-m)$, $w(d-n)$ (see Figures 2.10a and 2.10b).\[fig2-10\]](Fig2-10.eps){width="50.00000%"}
![A transformation that allows us to delete nodes $b$ and $c$ from the list of candidates for window nodes by elimination of the link between windows $w(b-m)$, $w(l-n)$ (edge $e(b-l)$), and between windows $w(c-e)$, $w(k-f)$ (edge $e(k-c)$) (see Figures 2.11a and 2.11b).\[fig2-11\]](Fig2-11.eps){width="80.00000%"}
1. Consider, in more detail, the possible situations of the appearance of nodes in the second basic object that cannot form windows (corresponding to step (5) above), and the elimination of those nodes. Suppose that in the first basic object, there are window nodes of degree $d = 2$. Then, if edge $e(a-c)$ that is incident to node $a$ belongs to the contour of the first basic object, then the second basic object will contain on the contour an edge $e(a-b)$ incident to the same node $a$, which again forms a window in the second basic object (see Figure 2.10a). In the second basic object the node $c$ should not form a window. If it does, then the second basic object violates the rules outlined in Section 2. In this case, by removing a link between windows $w(c-m)$ and $w(d-n)$, we can delete node $c$ from the candidates to form windows in the second basic object. We eliminate this link (interior edge $e(c-d)$) in one of two ways described in Section 2. In Figure 2.10a, the positions of windows $w(c-m)$ and $w(d-n)$ imply that the first method should be used. In another configuration of those windows, the second method would be used. The situation when nodes $a$ and $c$ have degree $d = 2$ (edge $e(c-d)$ is not present) cannot happen, because two adjacent nodes of degree $d = 2$ are substituted by a single node of degree $d = 2$.
2. Suppose that the first basic object contains node $a$ of degree $d = 2$ that does not form a window, and edges $e(a-b)$ and $e(a-c)$ belong to the contour. Then, node $a$ should form a degenerate segment in the second basic object, and nodes $b$ and $c$ should not form windows (see Figure 2.11a). If this condition was violated, then the second basic object would have windows linked by interior edges, which would violate the rules outlined in Subsection 2.1.1. As in item (A) above, nodes $b$ and $c$ are eliminated from the set of nodes that can form windows in the second basic object by temporarily deleting edges $e(b-l)$, $e(c-k)$, between pairs of windows $w(b-m)$ and $w(l-n)$, and $w(c-e)$ and $w(f-k)$ respectively (see Figure 2.11b). The situation when nodes $a$ and $c$, and $a$ and $b$ have degree $d =
2$ (edges $e(b-l)$ and $e(c-k)$ are absent) is not possible, as two or more adjacent nodes of degree $d = 2$ can be substituted by a single node of degree $d = 2$.
We note that elimination of links between windows by means of the algorithms described in Subsection 2.1.2, Cases 1.1 and 1.2, are performed so as to substitute one of the two windows of the contour by a single interior edge. By doing this, neither edge of the contour is moved to the interior.
1. One of the properties of the basic objects is that they have a disjoint set of interior edges. In the preliminary construction of the second basic object, the interior edges of the first basic object are all moved to the contour of the second basic object. During the subsequent steps, their movement to the interior will be prohibited.
2. This restriction on moving these particular contour edges to the interior is achieved by temporarily deleting interior edges that cannot belong to the contour (item (4) above). Then, the subproblems of the construction of the second basic object whose windows are not linked by interior edges, and of the introduction of additional windows into the second basic object, are performed as outlined in Subsection 2.1.2. This completes the preliminary construction of the second basic object.
### Completion of the second basic object
![Configuration of individual edges, chains of two edges, and windows needed to form the contour of the second object.\[fig2-12\]](Fig2-12.eps){width="40.00000%"}
![The second basic object, in which the missing interior edges belonging to the contour of the first basic object are added.\[fig2-13\]](Fig2-13.eps){width="40.00000%"}
![The second basic object, with deleted interior edges that can not belong to its contour. Node 24 that can not form a window is marked.\[fig2-14\]](Fig2-14.eps){width="40.00000%"}
Construction of the second basic object is performed using the same algorithm as the construction of the first basic object, that is, the construction of an object whose windows are not linked by interior edges, and the subsequent introduction of additional windows.
Consider the following example.
In Figure 2.1, the first basic object for a certain 25-node graph is displayed. The construction of the second basic object proceeds as follows:
1. In the first basic object, we select:
1. Chains of the interior that consist of two edges incident to the window nodes: $e(16-14)$, $e(14-13)$; $e(4-10)$, $e(10-11)$; $e(1-6)$, $e(6-5)$; $e(7-2)$, $e(2-3)$; $e(21-18)$, $e(18-19)$.
2. Interior edges that are incident to window nodes of degree $d = 2$: $e(20-25)$.
3. Free edges: $e(8-15)$, $e(22-24)$, $e(17-23)$, $e(9-12)$.
2. The above selected edges separated by windows, form the preliminary contour (see Figure 2.12): (16-14-13 W 4-10-11 W 1-6-5 W 7-2-3 W 21-18-19 W 20-25 W 8-15 W 22-24 W 17-23 W 9-12).
3. Connect nodes (see Figure 2.13) by interior edges that were not selected in (1), and correspond to contour edges in the first basic object: $e(1-2)$, $e(1-3)$, $e(3-4)$, $e(4-5)$, $e(5-11)$, $e(11-12)$, $e(12-13)$, $e(13-19)$, $e(19-20)$, $e(20-21)$, $e(21-22)$, $e(22-23)$, $e(23-24)$, $e(24-25)$, $e(14-15)$, $e(15-16)$, $e(16-17)$, $e(17-18)$, $e(9-10)$, $e(8-9)$, $e(7-8)$, $e(6-7)$.
4. From the set of interior edges, we temporarily remove those edges that cannot be moved to the contour of the second basic object: $e(17-18)$, $e(14-15)$, $e(6-7)$, $e(1-2)$, $e(9-10)$, $e(24-25)$ (see Figure 2.14). Once these edges are temporarily removed, node 25 becomes an end node (that is, node 25 acquires degree $d = 1$), and nodes 24, 9, 15, 1, 7, 17 acquire degree $d = 2$.
5. Mark the nodes that cannot form windows. In this example, node $24$ can not form a window. This node is circled (see Figure 2.14). By the algorithm of eliminating links between window nodes (see Subsection 2.1.2, Case 1.2), we move the edge $e(23-24)$, incident to node 24, to the contour. This completes the preliminary construction of the second basic object (see Figure 2.15).
![An object in which node 24 is deleted from the list of nodes that can form windows, by means of eliminating the link between windows $w(9-23)$, $w(17-24)$.\[fig2-15\]](Fig2-15.eps){width="40.00000%"}
![A segment of the second object where the links between windows are eliminated by means of interior edges. Free edges $e(19-20)$, $e(5-11)$, $e(3-4)$, $e(12-13)$ are identified that determine a possibility of forming additional windows.\[fig2-16\]](Fig2-16.eps){width="40.00000%"}
We note that the second object at this stage does not contain window nodes of degree $d = 2$. For this reason, this second object does not contain degenerate segments. The edges of the contour of the first basic object that should also belong to the contour of the second basic object, and therefore cannot be moved to its interior, are displayed in bold (see Figure 2.15).
An object whose windows are not linked by interior edges (see Figure 2.16) is constructed using the algorithm described in [@9] (which uses the steps outlined in Subsection 2.1.2, Cases 2.1–2.4). At this stage, construction of the object whose windows are not linked by interior edges is complete.
The next step is to introduce additional windows. We use the same algorithm (in Subsection 2.1.2) that was used in the construction of the first basic object. If we encounter an instance where an additional window can be introduced by means of free edges, then we introduce this window and continue to search until all instances are exhausted. In our example (see Figure 2.16), there are at most four possibilities of introduction of an additional windows. The object contains four free edges: $e(12-13)$, $e(3-4)$, $e(9-11)$, $e(19-20)$, that determine the number of possibilities. Consider all possible ways to introduce an additional window by using any of the listed four free edges.
![A segment of the second object where an additional window is formed by means of moving, between window nodes 22 and 25, the segment $S_{1,5}$.\[fig2-17\]](Fig2-17.eps){width="40.00000%"}
![A segment of the second object where an additional window $w(11-19)$ is formed by moving edge $e(12-13)$ to the contour, and edges $e(11-12)$, $e(19-13)$ to the interior.\[fig2-18\]](Fig2-18.eps){width="40.00000%"}
![A segment of the second object, where additional window $w(4-12)$ is formed by moving edge $e(5-11)$ to the contour, and edges $e(4-5)$, $e(11-12)$ to the interior.\[fig2-19\]](Fig2-19.eps){width="40.00000%"}
![The second basic object where temporarily deleted interior edges are reinstated. Common contour edges are marked with dashed lines.\[fig2-20\]](Fig2-20.eps){width="40.00000%"}
1. There exists a free edge $e(5-11)$ and node 1 of degree $d = 2$ that are adjacent to the free edge $e(3-4)$ on one side (see Figure 2.16). An additional window is introduced by moving part of the contour bounded by nodes $1$ and $5$ between any two nodes that form a window. In this example, window $w(22-25)$ or $w(9-17)$. We move this part of the contour between nodes 22 and 25 (see Figure 2.17).
2. Free edges $e(5-11)$ and $e(19-20)$ are adjacent to free edge $e(12-13)$ on two different sides (see Figure 2.16). An additional window $w(11-19)$ is introduced by moving edge $e(12-13)$ to the contour, and edges $e(11-12)$ and $e(19-13)$ to the interior (see Figure 2.18).
3. Free edges $e(3-4)$ and $e(12-13)$ are adjacent to free edge $e(5-11)$ on two different sides (see Figure 2.16). An additional window $w(4-12)$ is introduced by moving edge $e(5-11)$ to the contour, and edges $e(4-5)$ and $e(11-12)$ to the interior (see Figure 2.19).
4. Free edge $e(19-20)$ (see Figure 2.16) can not be used to introduce an additional window because node 20 does not have adjacent nodes that satisfy the possibility of introduction of additional windows (edge $e(21-22)$ is not a free edge, and node 25 forms a window).
From all of these possibilities, only one can be implemented. Two or more possibilities can not be implemented in this example because, otherwise, the condition of absence of links between windows will be violated. The second object with the temporarily deleted edges being reintroduced is displayed (see Figure 2.20). Comparing the first (see Figure 2.1) and the second basic objects (see Figure 2.20) we observe the following:
1. The sets of interior edges for each basic object are disjoint.
2. The contours for each basic object contain a number of common edges: $e(3-4)$, $e(11-12)$, $e(7-8)$, $e(15-16)$, $e(13-19)$, $e(21-20)$, $e(23-24)$, that appeared during the construction of the second basic object. These edges are marked by dotted lines (see Figure 2.20).
3. No links exist in either basic objects.
Selection of edges and windows {#sec3}
==============================
Suppose that we have constructed both basic objects for a graph of degree $d \leq 3$. Suppose that such a graph contains one or more Hamiltonian cycles. Edges of the graph can be naturally separated in three disjoint groups:
1. Edges that belong to all Hamiltonian cycles.
2. Edges that belong to no Hamiltonian cycles.
3. Edges that belong to some but not all Hamiltonian cycles.
Some edges from both basic objects can be separated into two disjoint groups:
1. Interior edges of both basic objects (these sets of disjoint).
2. Common edges of the contours of both basic objects.
The procedure of selection of edges includes not only separation into group, but also the assigning of weights to some edges that determines the weights of all other edges, and of windows in both basic objects. The assigning of weights can be performed for an additional edge or window of the contour, or to a group of edges and windows.
Assigning of weights to the edges of both basic objects
-------------------------------------------------------
### Selection of contour edges and windows
To any edge of the contour, a weight can be assigned. We assign a particular weight (-1, 0, or +1) to an edge or window of the contour of a basic object, that determines the weights of all other edges and windows of the basic object. Initially, the weights of all edges and windows are 0. Suppose that we select a contour edge $e(m-n)$. If $N$ is odd, we assign weight -1 to this edge, and then we perform the following operations:
1. Assign weight -0.5 to nodes $m$ and $n$.
2. Moving along the nodes of the contour, starting with node $s$ adjacent to node $m$ we assign to each node alternating weights +0.5, -0.5.
3. The weight of every edge and every window is assigned by adding the weights of their adjacent nodes. With this approach we obtain:
1. The weight of one particular edge (edge $e(m-n)$) becomes -1.
2. The weight of all contour edges and windows other than $e(m-n)$ remain 0.
3. Any interior edge attains the weight attains a value that can be -1, 0 or 1.
We could also assign the initial weight of -1 to a window $w(k-l)$, and perform the same operations as above.
If $N$ is even, we assign weight 0 to edge $e(m-n)$, and then we perform the following operations:
1. Assign weights +0.5 to node $m$, and -0.5 to node $n$.
2. Moving along the nodes on the contour, starting with node $t$ adjacent to node $m$ we assign to each node alternative weights -0.5, +0.5.
3. The weight of every edge and every window is assigned by adding the weights of their adjacent nodes. With this approach we obtain:
1. The weight of all contour edges and windows (including $e(m-n)$) remain 0.
2. Any interior edge attains the weight attains a value that can be -1, 0 or 1.
Consider the result of the selection of edges of a basic object, assuming that the contour of this basic object is a fictitious Hamiltonian cycle. A fictitious Hamiltonian cycle is a contour of a basic object whose windows, initially, have the same 0 weight as its edge. Suppose that a graph contains one or more Hamiltonian cycles. Suppose that one of the contour edges is selected and assigned a weight of -1. Independently of whether or not this edge belongs to a Hamiltonian cycle, the combined length of any Hamiltonian cycle will be equal to -1. If the selected edge is given a weight of +1, then the length of every Hamiltonian cycle will also be +1. This is because, on the contour, by our assumption, a fictitious Hamiltonian cycle, where the windows become fictitious edges with 0 weight, and the selection of any edge or window implies the same change in the length of any real Hamiltonian cycles as the same of any fictitious Hamiltonian cycle. Consider the following cases:
1. Suppose that a selected edge belongs to all Hamiltonian cycles. Then, by the construction of one such Hamiltonian cycle, part of the contour edges should be substituted by interior edges. Since the combined weight of the substituted contour edges equals 0, and the total length of the Hamiltonian cycle is determined by the weight of the selected edge, then the combined weight of the interior edges that substitute the contour edges should also be 0.
2. Suppose that the selected edge of weight -1 belongs to a some but not all Hamiltonian cycles. Regardless of which Hamiltonian cycle is constructed (with or without this edge), the combined weight of the interior edges that substitute contour edges in the Hamiltonian cycle will be equal to the combined weights of those contour edges. If the selected edge belongs to the constructed Hamiltonian cycle, then this combined weight will be 0. If the selected edge does not belong to the constructed Hamiltonian cycle, then the combined weight will be -1.
3. Suppose that the selected edge of weight -1 does not belong to any Hamiltonian cycle. Then, in the construction of a Hamiltonian cycle, this edge should be substituted, and the combined weight of the interior edges that are used in the Hamiltonian cycle, and the combined weights of the contour edges that are not used in the Hamiltonian cycle, will be -1.
4. Suppose that we select a window of weight -1. It can be considered as a fictitious edge that does not belong to any Hamiltonian cycle, and so should be substituted by an interior edge. In this case, the previous argument is valid: in the construction of a Hamiltonian cycle that does not contain the selected window, the combined weights of the interior edges that are used in the Hamiltonian cycle, and the combined weights of the contour edges that are not used in the Hamiltonian cycle, will be -1.
5. The claims (1)–(4) are valid in the case of selection of any contour edge or window in either basic object. However, one can select a set of contour edges and/or windows. Since in the construction of a Hamiltonian cycle, relative to any selected contour edge or window, the combined weight of the substituting interior edges is equal to the combined weight of the contour edges that are being substituted, this property will be valid for any number of simultaneously selected contour edges and windows.
Hence, the common argument is the following:
1. In the construction of a Hamiltonian cycle, the combined weight of substituted contour edges and windows is equal to the combined weight of the substituting interior edges, and this effect does not depend on whether selected edges belong to the Hamiltonian cycle.
2. If we simultaneously select a set of contour edges and windows then the combined weights of the substituting interior edges, and this effect does not depend on whether selected edges belong to the Hamiltonian cycle.
We note that the assigning of weights in turn assigns weights to the edges of the basic objects. However, it is also convenient to assign weights to the nodes of the basic objects. If a node is assigned -0.5 or +0.5, then it is means that edges incident to this node gain the weight -0.5 or +0.5.
### Selection of a group of edges
We can select not just a single edge, but a group of edges. Suppose that two basic objects of the same graph are constructed. The contours of both basic objects contain common contour edges. Each edge of this group could or could not belong to a particular Hamiltonian cycle, if any exist. Now we solve the problem of selection of common contour edges. By means of this subproblem, we will define some characteristics of the basic objects that will determine the Hamiltonicity of the graph.
The procedure of selection of common contour edges is as follows. The task is to select common contour edges of basic objects under the following conditions:
1. The weights of the nodes incident to common contour edges in each basic object should be identical.
2. The weights of the common contour edges could be -1, 0 or 1.
3. The interior edges of both basic objects should have 0 weights. If this is impossible, the correction procedure is performed (see Subsection 3.2.3) and then the combined weight of interior edges should be 0 for each object.
We can formulate a dual to the subproblem of selecting common contour edges of the basic objects. It differs from the above problem by the condition that the contour edges that are not common should all have zero weight.
We can request, in both primal and dual subproblems, that weights of the nodes incident to common contour edges should be identical, not only in both basic objects, but in both subproblems as well. In this case, the difference in the solution of this subproblem will be in assigning weights to other nodes that are not incident to common contour edges. In the primal subproblem, weights of those nodes should be assigned in such a way that all interior edges have zero weights. In the dual subproblem, the contour edges that are not common for both basic objects should have 0 weight. These two can be simultaneously achieved. Furthermore, we show that the solution of each of these subproblems can be used to solve HCP.
Assigning of weights to edges of the objects
--------------------------------------------
### Selection of common edges of contours of the objects
![Two basic objects, where the weights of the nodes are displayed next to the node numbers. The nonzero weights of the edges is displayed next to the edges.\[fig3-1\]](Fig3-1.eps){width="85.00000%"}
![Selected subgraph that determines the weight of the nodes incident to common contour edges.\[fig3-2\]](Fig3-2.eps){width="40.00000%"}
Assigning of weights to edges of the objects is a part of the solution to HCP. First, we will consider examples of assigning weights to contour edges and interior edges that satisfy the required conditions. Also, we will determine the sequence of such assignment algorithm.
1. Determine the set of contour edges for each object. The set of common contour edges is determined by comparing the two sets of contour edges together.
2. One of the conditions of assigning weights to the edges implies that the nodes of the common contour edges should have identical weights in both objects. For this reason, these nodes, and the contour edges between them, should be selected as a separate subgraph common to both objects.
On the basis of two objects, select a subgraph that contains:
1. Common contour edges.
2. Edges of each object that link two common contour edges. In both objects, they are the same.
3. Interior edges that are incident to window nodes. These sets will be disjoint for each object.
Consider the following example of the selection of a subgraph for the two objects displayed in Figure 3.1.
The subgraph (see Figure 3.2) consists of the following edges:
1. Edges $e(3-4)$, $e(11-12)$, $e(13-19)$, $e(20-21)$, $e(23-24)$, $e(15-16)$, $e(7-8)$ are common contour edges.
2. Interior edges $e(12-13)$ and $e(19-20$ that link common contour edges.
3. Pairs of edges that in each object incident are incident to the window nodes. In the first object they are $e(18-19)$, $e(18-21)$; $e(10-11)$, $e(10-4)$; $e(2-3)$, $e(2-7)$; $e(14-13)$, $e(14-16)$. In the second object they are $e(20-21)$, $e(20-23)$; $e(5-11)$, $e(5-4)$. For the node $25$ having $d=2$ edges $e(25-20)$, $e(25-24)$.
<!-- -->
1. For the nodes of the selected subgraph, we need to assign the weights. This algorithm has to satisfy a number of conditions, as follows.
1. The nodes that are incident to common contour edges should have equal weight in both objects. Because the subgraph is common in both objects, then the weights assigned to its edges will be the same in both objects.
2. The weight of the common edges can be -1, 0 or +1.
3. The edges of the subgraph that link nodes incident to common contour edges should have 0 weight.
4. The pairs of edges that are adjacent to a window node of degree 3, that in one object are interior edges, should have 0 weights. In the other object, this pair of edges will be contour edges and the weight of these edges is not yet determined, and could be -1, 0 or +1 depending on other weight assignments.
![Selected islands of the subgraph that determine the weights of the nodes.\[fig3-3\]](Fig3-3.eps){width="20.00000%"}
![The set of chains that determines the possibilities of assigning weights to the nodes on the islands.\[fig3-4\]](Fig3-4.eps){width="25.00000%"}
Consider an example of weight assignment for the subgraph shown in Figure 3.2.
1. To all window nodes, we assign the weight +0.5. In Figure 3.2 (and other figures) we represent this assignment only by displaying the sign of the weight. These nodes are nodes 2, 10, 18, 14, 22, 5 and 25. In general, we can assign to these nodes the weight +0.5 or -0.5. If there is a node in the subgraph linked to two window nodes, then the two window nodes should be assigned the same weight, either +0.5 or -0.5 simultaneously. In this example, the weights for nodes 2, 14 and 25 can be assigned independently. The weights of nodes 18 and 22, however, must be the same as each other. This is because edges $e(18-21)$ and $e(22-21)$ are incident to node 21, that is incident to the common contour edge $e(20-21)$. This is determined by the fact that two edges $e(18-21)$ and $e(22-21)$ are interior edges (one edge is interior in one object, and the other edge is interior in the other object), and should therefore have 0 weight. An analogous argument is valid for nodes 10 and 5, for which the weight should therefore be equal.
2. Because the weight of the edges incident to window nodes 2, 10, 18, 14, 22, 5 and 25 should be equal to 0, then nodes 23, 21, 19, 13, 11, 4, 3, 7, 16, 20, 24 have weight -0.5. Since edge $e(20-19)$ is interior in one of the basic objects, and has weight -1, then we need to reassign the weight of one of the windows nodes from +0.5 to -0.5. Nodes 19 and 20 that are incident to this node in both object have interior edges that link them to window nodes (18 and 25). The weight of one of these nodes should be changed from +0.5 to -0.5. We assign the weight -0.5 to node 25 having $d=2$. Then, nodes 20 and 24 should be assigned the weight +0.5.
<!-- -->
1. In the subgraph, we select islands that are parts of the subgraph such that their extreme (endpoint) nodes have weights -0.5, or the weight is unassigned. An island can not contain pairs of edges that are incident to window nodes. For this example, the islands are displayed in Figure 3.3. In assigning weights to nodes of an island, one can use patterns displayed in the example of assigning weights to the nodes of chains, including different number of common edges. If the subgraph contains chains with common contour edges (see Figure 3.4), then the extreme nodes $a$ and $b$ of the chains can have any weights associated with the choice of the weight of common edges, in particular, -0.5. In Figure 3.4, we display examples of assigning weights to the nodes of chains containing one, two or three common edges. By variation of the choice of weight for a particular edge, with any number of common edges in the chain, the extreme nodes can be assigned equal or opposite weights. We note that the situation when the extreme nodes are assigned opposite weights remains unchanged for any number of common edges in the chain. Furthermore, we need to assign weights to nodes of the islands that can have only one value. Such is node 12 whose weight can only be +0.5. We assign weights to chain $7-8-15-16$ (see Figure 3.3). They are assigned according to the rule displayed in Figure 3.4. If, to node 8, we assign the weight +0.5, then to node 15 we should assign the weight -0.5, and vice versa.
This completes the algorithm of selection of common contour edges, and assigning weights to the corresponding nodes. As a result of this algorithm, the nodes that are incident to common contour edges are assigned identical weights in both objects. Also, the edges of the subgraph that belong to three selected sets are assigned admissible weights. In Figures 3.1, 3.2 and 3.3, the signs of the weights are displayed next to the nodes.
### Assignment of weights to the nodes of the objects
As a result of the selection of common contour edges, some but not all nodes of the objects are assigned weights. The above algorithm allows us to assign weights to the nodes incident to common contour edges, and also to the window nodes. The next problem is to assign weights to the nodes that have not yet been assigned a weight. The missing weights are assigned according to the condition that interior edges must have 0 weights, and are therefore determined by the weights assigned already to other nodes in the above algorithm. In Figure 3.1, the nodes whose weights are assigned at this step are circled. We note that objects may contain nodes whose weights can not be assigned using the above conditions. For example, in Figure 3.1, nodes 1, 6 and 5 are not assigned weights. This is because nodes 1 and 5 are not incident to common contour edges. The choice of weight of a window node (in this example, node 6) is made on the basis of comparison of the sums of the weights of both objects, and is considered later. The weight of any edge is defined as the sum of the weights of the nodes incident to this edge. In Figure 3.1, the zero weights are not displayed.
### Algorithm for assignment of weights to the edges of the objects
1. Determine the contour edges for each object.
2. On the basis of two objects, select a subgraph that contains:
1. Common contour edges.
2. Edges of each object that link two common contour edges. In both objects, they are the same.
3. The sets of interior edges that are incident to window nodes. These sets of disjoint for each object.
3. Assign weights to the nodes of this subgraph using the algorithm outlined in Subsection 3.2.1.
4. Identify the islands in the above subgraph and assign weights to the nodes contained in the islands using the algorithm also outlined in Subsection 3.2.1.
5. Assign weights to remaining nodes using the algorithm outlined in Subsection 3.2.2.
6. If in one or both objects there remains nodes without weights assigned, then we assign to those nodes weights according to the rules will be described later.
7. Determine the weight of the edges as the sum of the weights of the nodes incident to this edge.
8. If, among the interior edges of one or both objects, there appears a nonzero weight, then we apply the following correction algorithm for this object.
[*Correction algorithm*]{}
![Two basic objects, where the weights of the nodes are displayed next to the node numbers. The nonzero weights of the edges is displayed next to the edges.\[fig3-5\]](Fig3-5.eps){width="85.00000%"}
1. Determine the sum of the weights of interior edges for both objects. Suppose that the sum is $-\gamma$ ($+\gamma$). The number $\gamma$ can only be integer (-1, 0 or +1) because the weight of each edge is determined by the sum of the weights of its adjacent nodes.
2. Change the weight of the edges incident to $\gamma$ nodes by +1 (-1) in such a way that the sum of interior edges becomes 0. These edges cannot be
1. Window nodes.
2. Nodes incident to a common contour edges.
3. Nodes with degree $d =
2$.
![Selected subgraph that determines the weight of the nodes incident to common contour edges.\[fig3-6\]](Fig3-6.eps){width="40.00000%"}
![Selected islands of the subgraph that determine the weights of the nodes.\[fig3-7\]](Fig3-7.eps){width="20.00000%"}
At this stage, the general algorithm of assigning weights to the edges of the objects is complete. Next, we consider examples of assigning weights to edges of objects.
In Figure 3.5, we display two basic objects that satisfy all required conditions. Now we explain how we assign weights to the edges of these objects.
1. Determine the common contour edges. In Figure 3.5, the common contour edges are displayed in bold, and they are $e(2-3)$, $e(4-8)$, $e(12-15)$, $e(6-11)$, $e(7-9)$.
2. On the basis of these two objects, select a subgraph that contains:
1. Common contour edges.
2. Edges $e(3-6)$, $e(4-7)$, $e(6-12)$, $e(7-11)$ that link nodes incident to common contour edges.
3. Interior edges incident to the window nodes. These are $e(10-11)$ and $e(10-8)$ from the first object, and $e(1-3)$, $e(1-4)$; $e(13-15)$, $e(13-18)$ from the second object.
3. Assign weights to the nodes of the selected subgraph.
1. To window nodes 10, 1, 13, assign weight +0.5. If there is a node in the subgraph linked to two window nodes, then these window nodes must be assigned the same weight +0.5 (nodes 10 and 13).
2. To nodes 11, 3, 4, 8, 15, that are linked with window nodes by a single edge, assign the weight -0.5.
4. In the subgraph, select the islands. The pairs of edges incident to a window node can not be part of any island. Assign weights to the nodes in these islands. First, determine the nodes whose weights are predetermined by previous weight assignments. In this example, nodes 7 and 6 can only have weight +0.5. This choice is determined by the condition that edges $e(7-11)$ and $e(3-6)$ must have 0 weight. Then, the weight of node 12 can only be -0.5. The choice of weights for nodes 9 and 2 can be taken arbitrarily. In our example, both edges are assigned the weight -0.5. In Figures 3.5, 3.6 and 3.7, the weights are displayed next to the nodes.
5. Assign weights to the nodes that have not been assigned a weight in the previous steps. They are assigned in accordance with the condition that the weights of interior edges should be 0. The final assignment of weights is displayed in Figure 3.5, where the weights of edges are also displayed. The nodes for which the weights are determined during this stage are circled.
![Two basic objects, where the weights of the nodes are displayed next to the node numbers. The nonzero weights of the edges is displayed next to the edges.\[fig3-8\]](Fig3-8.eps){width="85.00000%"}
![Selected subgraph that determines the weight of the nodes incident to common contour edges.\[fig3-9\]](Fig3-9.eps){width="40.00000%"}
In Figure 3.8, two basic objects satisfying the required conditions are displayed. We assign weights to the edges of these objects.
1. Determine the common contour edges. In Figure 3.8, the common contour edges are displayed in bold, and they are $e(3-4)$, $e(6-7)$, $e(9-10)$, $e(12-13)$, $e(14-15)$, $e(20-24)$, $e(18-21)$.
2. On the basis of these two objects, select a subgraph (see Figure 3.9) that contains:
1. Common contour edges.
2. Edges $e(9-12)$, $e(20-21)$, $e(4-10)$ that link nodes incident to common contour edges.
3. Interior edges incident to the window nodes. These are $e(8-9)$ and $e(8-15)$; $e(1-3)$, $e(1-6)$; $e(19-13)$, $e(19-20)$; $e(22-21)$, $e(22-24)$ from the first object, and $e(25-13)$, $e(25-14)$; $e(5-4)$, $e(5-6)$ from the second object.
3. Assign weights to the nodes of the selected subgraph.
1. To window nodes 8, 1, 19, 5, 22, 25, assign weight +0.5.
2. To nodes 13, 9 , 4, 6, 3, 21, 20, 24, 14, 15, that are linked with window nodes by a single edge, assign the weight -0.5. Since edge $e(20-21)$ is interior in one of the basic objects, and has weight -1, then we need to reassign the weight of one of the windows nodes from +0.5 to -0.5. Nodes 20 and 21 that are incident to this node in both object have interior edges that link them to window nodes (19 and 22). The weight of one of these nodes should be changed from +0.5 to -0.5. We assign the weight -0.5 to node 22. Then, nodes 21 and 24 should be assigned the weight +0.5.
4. In the subgraph, select the islands. The pairs of edges incident to a window node can not be part of any island. Assign weights to the nodes in these islands. First, determine the nodes whose weights are predetermined by previous weight assignments. In this example, nodes 10 and 12 can only have weight +0.5. This choice is determined by the condition that edges $e(9-12)$ and $e(4-10)$ must have 0 weight. Then, the weight of nodes 7 and 18 can be taken arbitrarily. In our example, both edges are assigned the weight -0.5. In Figures 3.8, 3.9 and 3.10, the weights are displayed next to the nodes.
5. Assign weights to the nodes that have not been assigned a weight in the previous steps. They are assigned in accordance with the condition that the weights of interior edges should be 0. The final assignment of weights is displayed in Figure 3.8, where the weights of edges are also displayed. The nodes for which the weights are determined during this stage are circled.
![Two basic objects, where the weights of the nodes are displayed next to the node numbers. The nonzero weights of the edges is displayed next to the edges.\[fig3-10\]](Fig3-10.eps){width="20.00000%"}
![Selected subgraph that determines the weight of the nodes incident to common contour edges.\[fig3-11\]](Fig3-11.eps){width="85.00000%"}
In Figure 3.11, two basic objects satisfying the required conditions are displayed. We assign weights to the edges of these objects.
1. Determine the common contour edges. In Figure 3.11, the common contour edges are displayed in bold, and they are $e(1-2)$, $e(5-6)$, $e(7-9)$, $e(11-12)$.
2. On the basis of these two objects, select a subgraph (see Figure 3.12) that contains:
1. Common contour edges.
2. Edges $e(5-7)$, $e(9-11)$, $e(9-12)$ that link nodes incident to common contour edges.
3. Interior edges incident to the window nodes. These are $e(3-5)$ and $e(3-2)$; $e(8-6)$, $e(8-7)$; from the first object, and $e(4-2)$, $e(4-6)$ from the second object.
3. Assign weights to the nodes of the selected subgraph.
1. In the subgraph there are nodes 2 and 6, that are linked to two window nodes. Node 6 is linked to window nodes 4 and 8, and node 2 is linked to window nodes 3 and 4. In this case, all window nodes 3, 4, 8 are assigned a weight +0.5.
2. Nodes 2, 6, 5, 7 that are linked to the window nodes, are assigned a weight -0.5.
4. In the subgraph, select the islands (see Figure 3.13). The pairs of edges incident to a window node can not be part of any island. Assign weights to the nodes in these islands. First, determine the nodes whose weights are predetermined by previous weight assignments. Because one of the islands contains a link between nodes 9 and 12, then these nodes should be assigned opposite weights. Assign weight +0.5 to node 9, and weight -0.5 to node 12. Then, the weight of node 11 must be -0.5. In Figures 3.11, 3.12 and 3.13, the weights are displayed next to the nodes.
5. Assign weights to the nodes that have not been assigned a weight in the previous steps. They are assigned in accordance with the condition that the weights of interior edges should be 0. The final assignment of weights is displayed in Figure 3.11, where the weights of edges are also displayed. The nodes for which the weights are determined during this stage are circled.
![Selected subgraph that determines the weight of the nodes incident to common contour edges.\[fig3-12\]](Fig3-12.eps){width="40.00000%"}
![Selected islands of the subgraph that determine the weights of the nodes.\[fig3-13\]](Fig3-13.eps){width="25.00000%"}
As follows from Figure 3.11, edge $e(5-7)$ that links nodes incident to common contour edges $e(6-5)$ and $e(7-9)$ is assigned a weight -1. In the second object, edge $e(5-7)$ is an interior edge. This weight assignment contradicts the condition that interior edges must have zero weight. After performing the correction algorithm, which we now outline, we will still be able to perform the test for Hamiltonicity of the graph. The underlying principal of the correction algorithm will be that in the basic objects, the sum of weights of interior edges must be 0. We need to increase or decrease the weight of the edges incident to a node of an object that contains an interior edge of nonzero weight. This node can not be a window node (in this case, node 4 or 10), nor nodes that are incident to common contour edges (in this case, nodes 1, 2, 5, 6, 7, 9, 11, 12), as the weights of these node should be the same in each object. Therefore, in this example, we can only use nodes 3 and 8. We select node 3, and increase change its weight from -0.5 to +0.5. This increases the weight of all edges incident to node 3 by +1. This will ensure that the sum of weights of the interior edges in the second object is zero. The second object, with the corrected edge weights, is displayed in Figure 3.14.
![The second object with the changed weights of edges after the correction algorithm has been performed.\[fig3-14\]](Fig3-14.eps){width="40.00000%"}
Therefore, we can now determine the sum of weights of the contour edges, and the sum of weights of the windows. These parameters are necessary to determine the Hamiltonicity of the original graph.
Main theorem {#sec5}
============
If both basic objects, after undergoing the algorithms outlined in Section 3, have equal sums of weights of contour edges ($L_1 = L_2$), and equal sums of weights of their windows ($S_1 =
S_2$), then the original graph is Hamiltonian. If either of these equalities are violated, the graph is not Hamiltonian.
1. On transformed objects:
Consider basic objects that we subject to testing. Suppose that we construct two basic objects. They do not intersect on interior edges, or on uncommon contour edges. Suppose that the weight assignment algorithm is completed.
We next take the union of both basic objects. The interpretation of the union of one basic object with a second basic object is that we receive the first basic object, where the weight for each edge is the sum of weights of that edge in each basic object. Thus, by finding the union of the first basic object with the second, and also finding the union of the second basic object with the first, we form two new basic objects that have the structure of the two original basic objects, and equal weights. The interior edges of the first union object will consist of the uncommon contour edges of the second basic object, and vice versa. If, in the first and second basic objects, the sum of contour weights was the same ($L_1 = L_2$) and the sum of window weights was the same ($S_1 = S_2$), then in the union objects, the weight of the contour will increase by the sum of weights of the common contour edges, which is equal in both objects. The weight of the interior edges for each union object will be equal to the sum of weights of uncommon contour edges in the alternate basic object. If $L_1 =
L_2$, this implies that the sum of weights of interior edges in both union objects will be the same. Thus, if in the first two basic objects, $L_1 = L_2$, and $S_1 = S_2$, then likewise the sum of weights of contour edges, and windows, will be the same in the union objects as well. Using analogous arguments, it can be seen that if $L_1 \neq L_2$, or $S_1 \neq S_2$, these inequalities are preserved in the union objects as well.
We note that if we were required to perform the correction algorithm, that led to $L_1 = L_2$ and $S_1 = S_2$, then these equalities will be preserved for the union objects as well. This is because the sum of the weights of interior edges is equal to 0, and therefore will not change the sum of weights of the contour edges in the union objects. If we now perform [*equivalent transformations*]{}, which are described below, then the weight of interior edges in the union objects will become 0 by means of the transformations of the first kind listed below, that do not lead to the violation of the equalities. If $L_1 \neq L_2$, or $S_1
\neq S_2$, then this inequalities will also be preserved in the union objects after the correction algorithm is performed.
Suppose that in the first object the length of the Hamiltonian cycle was $H_1$, and in the second it was $H_2$. Then in the union objects, it will be equal to $H_1 + H_2$. The interior edges of the union objects will have weights -1, 0 or +1.
2. On equivalent transformations.
Equivalent transformations are performed with the aim to confirm the equality of the length of Hamiltonian cycles in the constructed basic objects. These lengths should be such that if interior edges have zero weights, then the length of a Hamiltonian cycle in each basic object should be equal to the length of its contour. Since the interior edges in the constructed basic objects have weights -1, 0 and +1, an equivalent transformation would consist of consecutive annihilation of weights of -1 and +1 of interior edges. These equivalent transformations can be divided into two groups:
1. The first group consists of the transformations that, while annihilating the weight of an edge, does not change the length of the contour, and therefore does not change the length of a Hamiltonian cycle.
2. The second group consists of the transformations that do change the length of the contour while annihilating the weight of an edge, and consequently change the length of a Hamiltonian cycle by the same value simultaneously in two basic objects.
Consider an equivalent transformation of the first type. Suppose we can select $k$ interior edges of weight +1 and $k$ interior edges of weight -1. We change the weight of $k$ interior edges of weight +1 by using weight -1 that is assigned to nodes that are incident to these interior edges, excluding window nodes. Analogously, we annihilate the weight of $k$ interior edges of weight -1 using weight +1 that is assigned to nodes that are incident to these interior edges, excluding window nodes. This allows us to annihilate the weight of $2k$ interior edges.
Consider a transformation of the second type. This transformation is equivalent if, in both basic objects, there appears a weight of the same sign. Suppose that $H_1 = H_2$ ($L_1 = L_2$, $S_1 =
S_2$). Then, the entire annihilation of the weights of interior edges requires adding the same number of weights of the same sign in each basic object that leads to equal, simultaneous, change in the length of a Hamiltonian cycle, by assigning zero weights to interior edges of both basic objects. As in the previous case, the weights are added to nodes that do not form windows. The number of added weights depends on the number of interior edges of one basic object with positive weight ($r+$), and the number of interior edges in the same basic object with negative weight ($r-$).
It is possible to identify other equivalent transformations of the second type, using window nodes, however all of them change the length of contour, and therefore a Hamiltonian cycle by the same value.
3. On Hamiltonian graphs.
Suppose that in two union objects, the sum of contour weights are equal ($L_1 = L_2$) and the sum of window weights are equal ($S_1
= S_2$), that was also true in the original basic objects. It is known that if the weights of the interior edges are all zero, then:
1. In the construction of cycles, we will omit some of the contour edges, and all of the window edges.
2. We substitute these edges with some interior edges.
3. The sum of the weights of contour edges and windows that are omitted is equal to the sum of the weights of the interior edges they are substituted with.
Since the sum of the weights of interior edges that we use for the substitution equals zero (because all interior edges have zero weight), then the length of any possible Hamiltonian cycle is equal to the length of the contour ($H_1 = L_1 + S_1$, $H_2 = L_2
+ S_2$).
The remainder of the proof will show, by means of equivalent transformations, that two union objects will take a form that will allow us to determine the length of any Hamiltonian cycle in these forms. This will be done by ensuring that the interior edges in both union objects have zero weight.
We perform equivalent transformations in two union objects. Because the sum of weights of the interior edges in both union objects are equal to each other, then the composition of the set of interior edges in these objects is the following:
1. Without loss of generality, assume that the weight of the interior is nonnegative. In the first union object, there are $\mu_1$ edges with negative weight, and at least $\mu_1$ edges with positive weight. After equivalent transformations of type 1, we denote the number of remaining interior edges of positive weight as $m \geq 0 $.
2. Using analogous arguments to in (1), and the fact that the interior edges in both union objects have equal weight, then after equivalent transformations of type 1, the number of remaining interior edges of positive weight is $m$ in the second union object as well.
We can alter the weight of these $m$ edges in both union objects using equivalent transformations of the second type, to reduce their weight to zero. These equivalent transformations will change the weight of the contour in both union objects by $-2m$, and therefore the weights of the contours will remain equal to each other. Thus, is in the basic objects $L_1 = L_2$ and $S_1 = S_2$, then the equivalent transformations to the form where $H_1 = L_1 +
S_1$ and $H_2 = L_2 + S_2$, and therefore $H_1 = H_2$, which confirms the existence of a Hamiltonian cycle.
4. On non-Hamiltonian graphs.
Suppose that in two union objects, $L_1 \neq L_2$ and $S_1 \neq
S_2$. Without loss of generality, assume that the sum of weights of the interior edges in the first union object is nonnegative. Then, after we perform equivalent transformations of type 1, there will remain $m \geq 0$ interior edges of positive weight in the first union object. Then, in the second union object, there will remain $k$ interior edges with nonzero weight, and the sum of weights of the interior edges in each union object after these equivalent transformations of type 1 is different.
At this stage, it is not possible to reduce the weights of all interior edges in both union objects to zero by means of equivalent transformations of type 2. Since the existence of a Hamiltonian cycle implies that the length of any Hamiltonian cycle should be equal in both union objects, this proves that graph is non-Hamiltonian. This concludes the proof.
Finally, we check the Hamiltonicity for the graphs displayed in Figures 3.1, 3.5, 3.8 and 3.11.
In Figure 3.1, the two basic objects with the completed algorithm of assignment of weights are displayed (see Example 4). It was suggested that node $6$ in the first basic object (see Figure 3.1) could be assigned weight -0.5 (or +0.5). The nodes 1 and 5 are assigned weights +0.5 (or -0.5). Since we are trying to solve HCP, which is determined by equating $H_1$ and $H_2$, the choice of weights +0.5 or -0.5 is entirely determined by the condition $H_1
= H_2$. If node $6$ is assigned the weight +0.5 in the first basic object, it implies that $H_1 = -5$ ($L_1 = -7$, $S_1 = 2$), and for the second basic object $H_2 = -5$ ($L_2 = -7$, $S_2 = 2$). Since $H_1 = H_2$, the graph, according to the main theorem, is Hamiltonian.
In Figure 3.5, the two basic objects with the completed algorithm of assignment of weights are displayed (see Example 5). As a result of this algorithm, all edges and windows of the objects are assigned weights -1, 0 or +1. We determine the sum of weights of the contours of each object, which is supposed to be the length of a Hamiltonian cycle. We obtain $H_1 = -2$ ($L_1 = -3$, $S_1 = 1$), and $H_2 = -4$ ($L_2 = -6$, $S_2 = 2$). Because $H_1 \neq H_2$, $(L_1 \neq L_2$, $S_1 \neq S_2$), then the graph, according to the main theorem, is non-Hamiltonian.
In Figure 3.8, the two basic objects with the completed procedure of assignment of weights are displayed (see Example 6). As a result of this procedure, all edges and windows of the objects are assigned weights -1, 0 or +1. We determine the sum of weights of the contours of each object, which is supposed to be the length of a Hamiltonian cycle. We obtain $H_1 = -1$ ($L_1 = -2$, $S_1 = 1$), and $H_2 = -3$ ($L_2 = -5$, $S_2 = 2$). Because $H_1 \neq H_2$ ($L_1 \neq L_2$, $S_1 \neq S_2$), then the graph, according to the main theorem, is non-Hamiltonian.
In Figure 3.11, the two basic objects with the completed procedure of assignment of weights, after the correction procedure is also applied, are displayed (see Example 7). As a result of these algoritms, all edges and windows of the objects are assigned weights -1, 0 or +1. We determine the sum of weights of the contours of each object, which is supposed to be the length of a Hamiltonian cycle. We obtain $H_1 = -2$ ($L_1 = -3$, $S_1 = 1$), and $H_2 = -2$ ($L_2 = -3$, $S_2 = 1$). Because $H_1 = H_2$ ($L_1
= L_2$, $S_1 = S_2$), then the graph, according to the main theorem, is Hamiltonian.
The general solution of the HCP with complexity estimate
========================================================
The solution of the HCP consists of several consecutive subproblems:
1. Construction of a basic object:
1. Construction of an object whose windows are not linked by interior edges.
2. Introduction of additional windows.
2. Construction of the second object:
1. Preliminary formation of the second object.
2. Construction of an object whose windows are not linked by interior edges.
3. Introduction of additional windows.
3. Selection of common edges of the contours of both objects, including assignment of weights to the nodes of both objects (and therefore to the edges of both object).
4. Determination of parameters of both objects, and comparison of these parameters to determine the existence of a Hamiltonian cycle in the graph.
We will now estimate the complexity of each of the above subproblems that are equivalent to the HCP.
1. Construction of a basic object:
1. Construction of an object whose windows are not linked by interior edges - this subproblem consists of separate individual algorithms that eliminate links between windows. The order in which these individual algorithms are performed is determined by the configuration of the windows and the presence of degenerate segments containing nodes whose degree $d = 3$. Each of these procedures moves an interior edge linking window nodes to the contour, and moves one of the contour edges to the interior. The initial number of windows (some of which may violate the required properties) is less than or equal to $N$. Each procedure can be performed by a single search of all nodes, and no corrections (or doubling back) are required. This proves that the complexity of this subproblem is polynomial, because the number of algorithms can not be more than $N$.
2. Introduction of additional windows – this subproblem consists of separate individual algorithms that introduce additional windows. The order in which these individual algorithms are performed is determined by the configuration of windows, and by the presence of free edges. Each of these algorithms moves two of the contour edges to the interior, and one free edge to the contour, creating an additional window. The initial number of windows after the previous step is less than or equal to $[N/6]$. During this algorithm, two contour edges are substituted with one free edge, and with an additional window. This substitution can be performed by a single search of all free edges of the object, and no corrections (or doubling back) is required. This proves that the complexity of this subproblem is polynomial, because the number of algorithms can not be more $[N/6]$.
2. Construction of the second object:
1. Preliminary formation of the second object – this subproblem consists of separate individual algorithms that construct the initial formation of the second object. These algorithms are the construction of the initial contour, the determination of nodes that should not form windows, and temporary elimination of these nodes. The order in which these individual algorithms are performed is determined by the configuration of the first basic object, i.e., the number its windows, the position of the windows on the contour, the presence of nodes of the degree $d
= 2$, presence of degenerate segments, and the presence of free edges. Once these algorithms have been completed, the final algorithm is the temporary elimination of interior edges that cannot belong to the contour of the second object, and for which no further algorithms apply.
1. Construction of the initial contour – this algorithm consists of a sequence of rearrangements of the first basic object by moving the interior edges to the contour, and rearranging some nodes of degree $d = 2$. The number of such nodes is less than $N$.
2. Determination and elimination of nodes that should not form windows – this algorithm identifies nodes that are incident to edges that are linked to window nodes of degree $d = 2$. The algorithm of eliminating these nodes also involves the elimination of links between windows. The number of nodes that can not form windows of the second object is determined by the number of nodes of degree $d = 2$, which is less than $N$.
3. Elimination of interior edges that cannot belong to the contour – this algorithm identifies interior edges and temporarily deletes them from the object. The number of interior edges that are deleted is dependent on the number of windows in the first basic object, and the number of nodes with degree $d = 2$ in the second object.
Hence, since each of the individual algorithms that construct the preliminary form of the second object is performed less than $N$ times, each performed by a single search on the nodes of the contour, and no corrections (or doubling back) are required, the complexity of this subproblem is polynomial.
Elimination of interior edges that cannot belong to the contour prohibits the movement of certain contour edges to the interior. The subproblems 2(b) and 2(c) can be performed as described as in 1(a) and 1(b).
3. Selection of common edges and assignment of weights to the edges of the object. This subproblem consists of individual algorithms that:
1. Identify contour edges that are common for both basic objects.
2. Assigns equal weights, in both basic objects, to nodes incident to common edges.
3. Assigns zero weights to all interior edges in either basic object. If this is impossible, apply the correction algorithm to ensure that the sum of weights of the interior edges is zero.
The algorithm of identifying common contour edges consists of comparing contour edges of both objects, and is performed by a single search of the contour edges of both objects.
The weights of the nodes that are incident to the common edges should be the same. This is achieved through the assigning of weights to the subgraph that is common to both objects. This will ensure that the weight of every common contour edge will be the same in both objects. The assignment of weights to the edges that link nodes incident to common contour edges is achieved by means of islands that are identified in the subgraph. In those cases when the weight of an interior edge is nonzero, then we apply the correction algorithm that makes the sum of the weight of interior edges equal to zero. As a result of the weight assignment, the sum of weights of the interior edges is zero. This is achieved by assigning the weights +0.5 (or -0.5) to a node $a$, and if there is an interior edge $e(a-b)$, then the weight -0.5 (or +0.5) should be assigned to node $b$.
Therefore, since all algorithms of this subproblem are performed no more than $N$ times, the complexity of this subproblem is polynomial.
4. Determination of parameters of both objects – the parameters that determine the Hamiltonicity of the graph are sums of weights of edges, and the sums of weights of windows of both basic objects. If for both basic objects, these parameters are equal, the graph is Hamiltonian. These parameters are determined by adding the weights of less than $N$ contour edges, and the weights of no more than $[N/6]$ windows, in each basic object. Therefore, the complexity of computing these parameters is polynomial.
Therefore, the complexity of performing all of the above subproblems is polynomial.
Conclusion
==========
The described solution of the HCP based on the construction of two basic objects for a given graph can be considered as surprising. The HCP for graphs of degree $d \leq 3$ is NP-complete, but it is just a particular case of HCP for general non-oriented graphs. However, its solution provides a hope for the solution of the general $P = NP$. The question of construction of basic objects for other NP-complete problems remains open, and further hard work is required to prove the existence of such objects.
It is worth mentioning that these results on the existence of Hamiltonian cycles identify an algorithm to construct the cycle, if it exists. A separate article will be devoted to the construction of this algorithm.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author is grateful to V. Ejov (School of Mathematics and Statistics, UniSA), and M. Haythorpe (School of Mathematics and Statistics, UniSA) for useful discussions and questions.
[9]{} M. R. Garey and D. S. Johnson. [*Computers and Intractability: A Guide to the Theory of NP-Completeness*]{} (W. H. Freeman & Co., New York, 1979). I. I. Goray and L. I. Goray, [*Computer system configured in supporter of solving NP-complete problems at high speed*]{} (United States Patent No. 5, 535,840 B1, International Intellectual Group, Inc., 2003). G. Gutin and A. P. Punnen (eds), [*The Traveling Salesman Problem and Its Variations*]{} (Kluwer Academic Publishers, Boston, 2002). D. S. Hochbaum (ed), [*Approximation algorithms for NP-hard problems*]{} (PWS Publ. Co., Boston, 1997). S. Arora and M. Sudan, “Improved Low-Degree Testing and its Applications”, [*ACM STOC*]{} (1997) http://www.cs.princeton.edu/$\sim$arora/pubs/ldtest.ps. E. Balas and N. Christofides, “A restricted lagrangean approach to the traveling salesman problem”, [*Math. Progr.*]{} [**. 21**]{} (1) (1981) 19–46. M. Held and R. M. Karp, “Traveling Salesman Problem and Minimum Spanning Trees: Part II”, [*Math. Progr.*]{}, [**1**]{} (1971) 6–25. E. L. Lawler, D. E. Wood, “Branch-and-Bound Methods: A Survey”, [*OR*]{} [**14**]{} (1966) 699–719. Website at: http://www.pcgrate.com/download.
[^1]: I.I.G., Inc., Discrete Mathematics Division, 8 Obruchevyh str. Ap. 190, Saint Petersburg, 194064, Russian Federation ([[email protected]]{}).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Yu Zhang, Xiaoguang Di, Bin Zhang, and Chunhui Wang [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'mybibfile.bib'
title: 'Self-supervised Image Enhancement Network: Training with Low Light Images Only'
---
=1
[UTF8]{}[gbsn]{}
[Yu : Self-supervised Image Enhancement Network: Training with Low Light Images Only]{}
Introduction {#sec:introduction}
============
, various algorithms based on deep learning have achieved surprising results in some image processing and computer vision tasks, such as object detection [@liu2016ssd], [@ren2015faster], [@redmon2018yolov3],[@he2017mask], image segmentation[@he2017mask],[@long2015fully],[@badrinarayanan2017segnet], etc. One important reason for the rapid development of deep learning in these tasks is that we can obtain a large number of data sets with clear and unambiguous labels. In these tasks, although the construction of the data set requires some cost, it is still acceptable, also on the Internet, a large number of open source data sets can be found for these tasks to support the training of the network. However, in low-level image processing tasks such as low light image enhancement, image dehazing, and image restoration, etc., it is difficult to obtain a large number of true input/label image pairs.
As for low light image enhancement task, in the previous work, some solutions such as synthesizing low light images [@Lore2015LLNet], using different exposure time images to obtain data [@chen2018learning], and so on, have achieved good visual effects. However, there are still two problems with those methods. One is how to ensure that the pre-trained network can be used for images collected from different devices, different scenes, and different lighting conditions rather than building new training data set. The other is how to determine whether the normal light image used for supervision is the best, there can be lots of normal light images for a low light image. Usually, the builder of the data set gets the normal light images by experience or artificial adjustment, which will cost lots of time and energy and we cannot make sure that the enhanced image can show the information contained in the low light image to the greatest extent with those normal light images.
For those two questions, this paper proposes a self-supervised low light image enhancement network based on information entropy theory and Retinex model, and achieves the state-of-the-art in terms of enhancement quality and efficiency. In this paper, the only data we need are the low light images, without any paired or unpaired normal light images. To our knowledge, this is the first fully self-supervised image enhancement method based on deep learning. The proposed method does not rely on a well-designed complex network structure, only with a simple fully convolutional neural network (CNN) as shown in Fig.2 and minute-level training, we can complete low-light image enhancement tasks.
There are some image enhancement networks based on Retinex model [@wei2018deep],[@park2018dual], but they all require paired data, and then use the assumptions that images captured in different light conditions should have the same reflectance and the illumination map should be smooth to decompose low light images into corresponding reflectance and illumination map. Similar to these works, we also use a network to decompose low light image into reflectance and illumination, but unlike those previous works, we use self-supervised methods to train the network. Only low light images are required (even a single low light image) for training, then we can get the reflectance with good visual effects, and it can be treated as an enhanced image.
We think that low light image enhancement task is to display the information contained in low light images in a more intuitive way, rather than creating new information. At the same time, according to the entropy theory, images whose histogram are uniform distribution have the maximum entropy and contain the most information. Based on the above analysis, we propose an assumption that the histogram distribution of the maximum channel of the enhanced image should conform to the histogram distribution of the maximum channel of the low light image after histogram equalization. With this assumption, the loss function can be designed without normal light images, and it can not only retain the authenticity of the enhanced image, but also ensure that the enhanced image has sufficient information. The proposed method does not have any dependence on the way of acquiring low light images, and the training process is completely self-supervised, so the method proposed in this paper has good generalization ability, even if the pre-trained network is not well enough in new environment, retraining or fine-tuning without building paired/unpaired normal light images data set is possible for the network. Our contributions include:
- We propose a new maximum entropy based Retinex model, and give its theoretical source.
- Combined with deep learning, we propose a self-supervised low light image enhancement network, which can complete the training with even one single low light image.
- The proposed method only requires minute-level training and has a good real-time performance. We verify the enhancement effect and stability of the algorithm through some experiments and objective indexes.
Related Works
=============
Our method mainly comes from histogram equalization, model-based methods, and deep learning based image enhancement methods.\
**Histogram Equalization**
In low light image enhancement tasks, Histogram Equalization(HE) is the most simply and wildly used method. It can let the histogram of the enhancement image have a uniform distribution to get the maximum entropy. However, HE cannot avoid the problems of details disappearance (over enhancement, under enhancement), poor color restoration, noise amplification and so on.
To solve those problems, various improved algorithms are proposed, such as Adaptive Histogram Equalization(AHE)[@pizer1987adaptive] and Contrast-limited Adaptive Histogram Equalization(CLAHE)[@pisano1998contrast] for details, Hue-preserving color image enhancement [@naik2003hue] for hue preserving, Brightness Bi-Histogram Equalization Method (BBHE) [@kim1997contrast], Dualistic Sub-Image Histogram Equalization Method (DSIHE) [@wang1999image] for brightness preserving, etc. In [@celik2011contextual] and [@lee2013contrast], the method considering the relationship between adjacent pixels and large gray-level difference is proposed. Although many improved methods have been proposed, there are still many problems in applying histogram equalization directly to image enhancement.
![The enhancement results by proposed method. (a) Input. (b) The network was trained 200 epochs with low-light images without (a). (c) The network was trained 10000 times with image (a) only. (d) Reference[]{data-label="fig1"}](fig1){width="3.5in"}
\
**Model Based Image Enhancement Method**
Among the model-based low-light image enhancement methods, there are mainly based on the dehazing model [@dong2011fast] and Retinex model[@land1977retinex]. The method based on the dehazing model is mainly based on the discovery that the low light image is similar to the haze image after inversion. Dong et al. proposed an enhancement method that performs the dehazing operation after inverting the low light image and then inverts the image back [@dong2011fast]. Some studies have extended these works [@li2015low], although these methods have achieved some good effect, they lack corresponding physical model, which limits the application of the method in various scenes.
According to Retinex theory, the collected images can be decomposed into illumination and reflectance, but this is a highly ill-posed problem to obtain them from low light images only. Therefore, other constraints must be introduced: the early researches like single-scale Retinex [@jobson1997properties] model and the multi-scale Retinex [@jobson1997multiscale] model, only use the constraint that the illumination map is smooth to solve the problem. The captured image is smoothed by using Gaussian filters with one or more scales to obtain illumination, however, the enhanced image often have unreal phenomena such as over-enhancement and whitening. [@wang2013naturalness] proposes a Bright-Pass Filter that preserves natural characteristics. [@fu2016fusion] performs global illumination adjustment and local contrast enhancement on the initially estimated illumination map. Although they have obtained some good effects, without considering the structural characteristics, information loss is prone to occur in areas with rich details. LIME [@guo2016lime] introduces a filter considering the structure characteristic to smooth the illumination map and uses BM3D to denoise the enhanced image. Although those methods are proposed to maintain image details and naturalness, neither before nor after enhancement tasks, the denoising process will still cause blur or loss of details.
It is difficult to add some additional priors to those methods based on Retinex model only, which leads to the problems of noise, halo, detail preservation and so on, so in recent years, many algorithms based on the variational Retinex model are proposed. Kimmel et al. [@kimmel2003variational] first proposes a variational Retinex model, and uses $L_{2}$ regularization to obtain a smooth illumination map. Fu et al. [@fu2013variational] introduces the bright channel prior to the variational Retinex model to suppress the halo effect. Park et al. [@park2017low] proposes a weighted $L_{2}$ regularization to constraint reflectance image, which has a slight noise suppression effect. Fu et al. [@fu2019hybrid] proposes a $L_{2}$-$L_{p}$ norm to constrain the illumination map and keep more details. Although these variation based methods have achieved good results, it is very time-consuming to process images due to the need of multiple iterations to solve the variational equation. Even with Fast Fourier Transform(FFT), it is difficult to ensure the real-time.
In addition, in those model-based low light image enhancement algorithms, spatial smoothing prior [@lagendijk2012iterative] and its improvement are mostly used to constrain the illumination map, and there is no constraint on the contrast information of the reflectance. In this paper, we use the maximum information entropy to constrain the reflectance image, so as to further improve its contrast information.\
**Learning based methods**
Learning based methods have achieved good results in some low-level image processing tasks, such as image denoising [@zhang2017beyond],[@guo2019toward], super-resolution reconstruction [@dong2015image],[@li2019feedback],restoration[@nah2017deep],[@kupyn2018deblurgan], etc. However, most of the current algorithms based on deep learning are supervised, and it is difficult to obtain both degraded and normal images in those low-level image processing tasks. It is proposed to synthesize low light image data with normal light image for training in some researches. For example, LLNET [@lore2017llnet] is the first work to use deep learning to solve image enhancement problem, it proposes to train the networks with synthetically noisy and dark images separately, but it does not consider the natural images characteristic. In [@li2018lightennet] and [@yang2016enhancement], gamma transformation is applied to natural image patches to generate low light image patches for training, but they do not consider other degradation of the real collected low light images like noise, color changing, etc. In MSR-net [@shen2017msr], high quality (HQ) data are obtained by artificial selection and Photoshop, and low light images are obtained by processing HQ data with random brightness and contrast reduction and gamma transformation. The data obtained by those methods seems to look like low light images, however, it is difficult to truly reflect the characteristics of low light images, such as noise, overexposed and underexposed areas existed in the same image, etc.
In order to solve this problem ,some methods propose to use real low light images for training. In [@cai2018learning], a large multi-exposure image database is established, and the reference images are obtained by combing different exposure images and subjective selection. Retinex-net [@wei2018deep] tries to obtain the low/normal light image pairs through adjusting the exposure time, and achieves good enhancement effects, but the exposure time is still artificially determined, and it is difficult to choose the best exposure time to get a reference image. In [@chen2018learning], it introduces a parameter to link two images with different exposure time, and with end-to-end training, it can well deal with the noise problem, but it can only be used for raw images. In [@zhang2019kindling], a light adjustment network is introduced to link paired images with any different exposure time, which solves the problem in acquisition of normal light images. However, in practical applications, if we want to get better images, we may need to choose a hyper-parameter for each low light image.
Although these deep learning based methods have achieved good visual effects in low light image enhancement, they are all based on paired images, and the cost of building training data is so high, and they do not solve the two problems we mentioned before, i.e. how to obtain an optimal reference image and how to ensure the adaptability of the method to new environments or new equipments.
Method
======
Maximum Entropy Based Retinex model
------------------------------------
![Structure of the Self-supervised image enhancement network[]{data-label="fig2"}](fig2){width="3.5in"}
Base on Retinex model, an image can be decomposed into reflectance and illumination map as follows:
$$\label{eqn1}
S=R\circ I$$
where $S$ represents the captured image, $R$ represents the reflectance and $I$ represents the illumination map. This is a highly ill-posed problem, its solution needs additional prior. According to Bayesian formula, the problem can be expressed as follows:
$$\label{eqn2}
p(R,I\mid S)\propto p(S\mid R,I)p(R)p(I)$$
Where, $p(R,I\mid S)$ is posterior probability, $p(S\mid R,I)$ is the class conditional probability, and $p(R)$ and $p(I)$ are prior probabilities of reflectance and illumination. Existing methods generally add the prior probabilities $p(R)$ and $p(I)$ to find the maximum posterior probability, and estimate the reflectance and illumination.
By calculating the negative logarithm of equation (\[eqn2\]), the problem of image enhancement can be transformed into the form of three distance terms, as can be seen in formula (\[eqn3\]):
$$\label{eqn3}
\underset{R,S}{min} l_{rcon}+\lambda_{1}l_{R}+\lambda_{2}l_{I}$$
Where, $l_{rcon}$ represents reconstruction loss, $l_{R}$ represents reflectance loss, and $l_{I}$ represents illumination loss. $\lambda_{1}$ and $\lambda_{2}$ are weight parameters.
In this paper, we use the $L_{1}$ norm to constrain all the losses, we do not compare the impact of $L_{1}$, $L_{2}$, SSIM and other loss functions on low level image processing tasks, there are some related studies such as[@zhao2016loss]. The reconstruction loss $l_{rcon}$ can be expressed as:
$$\label{eqn4}
l_{rcon}=\left \| S-R\circ I \right \|_{1}$$
As for the reflectance loss, different from the existing methods only using $ \left \|\bigtriangleup R \right \|_{1}$ [@park2017low],[@fu2015probabilistic], we propose a new distance measurement method for reflectance loss based on the following reasons:
- For image enhancement task, the processed image should have enough information
- The processed image shall conform to the original image information
- Histogram equalization can greatly improve the information entropy of image
Based on the above considerations, we propose equation (\[eqn5\]) as the loss of reflectance image, which also uses $L_{1}$ loss:
$$\label{eqn5}
l_{R}=\left \| \underset{c\in{R,G,B}}{max} R^{c} - F(\underset{c\in{R,G,B}}{max}S^{c}) \right \|_{1} + \lambda \left \|\bigtriangleup R \right \|_{1}$$
Where, $F(X)$ means the histogram equalization operator to image X. $\lambda$ is weight parameters. This loss function means that maximum channel of the reflectance should conform to the maximum channel of the low light image and has the maximum entropy. There are three main reasons why we choose the maximum channel to constrain. Firstly, for a low light image, the maximum channel has the greatest impact on its visual effect. Secondly, if other channels are selected, there is no doubt that saturation will occur according to the prior that the maximum channel must be greater than the other two channels. Thirdly, if we choose one of the color channel, such as R, G or B channel, it is obviously not in line with the natural image.
For the illumination loss, we adopt the structure-aware smoothness loss proposed in [@wei2018deep]:
$$\label{eqn6}
l_{I}=\left \| \bigtriangleup I\circ exp\left ( -\lambda_{3}\bigtriangleup R \right ) \right \|_{1}$$
It is proposed that equation (\[eqn6\]) can make the illumination loss aware of the image structure in [@wei2018deep]. And the loss means that the original TV function $\left \| \bigtriangleup I \right \|_{1}$ is weighted with the gradient of reflectance.
From the equation (\[eqn3\]) to equation (\[eqn6\]), we get the maximum entropy based Retinex model, as can be seen in equation (\[eqn7\]):
$$\begin{aligned}
\label{eqn7}
Z&{}={}&\left \| S-R\circ I \right \|_{1}+\lambda_{1} \left \| \underset{c\in{R,G,B}}{max} R^{c} - F(\underset{c\in{R,G,B}}{max}S^{c}) \right \|_{1} \nonumber\\
&&{+}\:\lambda_{2} \left \| \bigtriangleup I\circ \lambda exp\left ( -\lambda_{3}\bigtriangleup R \right ) \right \|_{1} \nonumber\\
&&{+}\:\lambda_{4}\left \|\bigtriangleup R \right \|_{1}\end{aligned}$$
Variational methods or FFT are generally used to solve equation (\[eqn7\]) with $L_{2}$ loss, however, they both need multiple iterations which will bring time consumption problems, and with more constraints, the solution will be more complicated. In order to enhance the image in real time, we propose a solution based on deep learning. The network uses equation (\[eqn7\]) as the loss function. We can find that in equation (\[eqn7\]), there is only low light images, so the network can be trained through a self-supervised way.
The values of $\lambda_{1}$,$\lambda_{2}$,$\lambda_{3}$,$\lambda_{4}$ are $0.1$, $0.1$, $10$ and $0.01$ in this paper. The influence of values of $\lambda_{1}$ and $\lambda_{2}$ is not so obvious in visual effect that we just choose 0.1, the value of $\lambda_{3}$ comes from [@wei2018deep]. As for $\lambda_{4}$, in our experiments, we found that it can be used to control the noise. When its value increases, the noise decreases, and at the same time, the image will be more blurry. Through some experiments, we choose 0.01 for $\lambda_{4}$ and if $\lambda_{4}=0.1$, the enhanced image will appear obvious blur.
Self-supervised Network Based Solution
--------------------------------------
If we use the variational methods or FFT to solve the model proposed from equation (\[eqn3\]) to equation (\[eqn7\]), then it means that we need to carry out the same iterative processing for each low light image, which will not only bring time-consuming problems, but also the iteration times for each low light image may be uncertain, which is almost a disaster in many real applications. At the same time, this kind of solution can not take advantage of big data, the previous data processing can do nothing helpful to the new data processing.
In the previous deep learning based researches, due to the lack of models that can support self-supervised training, only the paired or unpaired low/normal light images collected in advance can be used to complete the network training. However, the data collected in advance can not contain all the real low light situations, such as different environments, devices, or degradation problems, etc., which also limits the application scope of the pre-trained network. After all, it is impossible to build the data set when we are using them.
However, based on the model proposed from equation (\[eqn3\]) to equation (\[eqn7\]), we can achieve the self-supervised training, which means that we can build the data set online and avoid the problem of applicability. And compared with the supervised learning whose supervisor is selected by artificial method, the model based on maximum entropy can ensure that the enhanced image has enough information entropy.
We only need a very simple CNN structure to achieve the decomposition of the illumination and reflectance. The specific structure of the CNN we finally adopted is shown in Fig.2. The input of the network is low light image and its maximum channel, after some convolution and concat layers, reflectance and illumination can be gotten with a sigmod layer. Table \[tab:structure\] is the specific information of each layer of the network.
In fact, we have experimented with different network structures, and the stacking of convolutional layers and a sigmod layer can also produce acceptable results. However, if we add some concat layers, the enhancement results will become clearer. It can be seen that we use down-sampling and up-sampling in the network, its prime function is to reduce the noise. In some experiments, we find that adding the down-sampling layer will make the image blur, however, it will reduce the noise too.
\[tab:structure\]
Experiment
==========
We use the LOL database [@wei2018deep] which contains 500 low/normal light image pairs, 485 of which are used for training and images size are $400*600$. Note that during the training process, we only use natural low light images and do not use synthetic data and normal light images. During the training process, our batch size is set to 16 and the patch size is set to 48 \* 48. We use Adam stochastic optimization [@kingma2014adam] to train the network and the update rate is set to 0.001. The training and testing of the network are completed on a Nvidia GTX 2080Ti GPU and Inter Core i9-9900K CPU, and the code is based on the tensorflow framework.
In section 4.1, we introduce some objective evaluation indexes. In section 4.2, we measure the influence of the training times on loss and evaluation indexes. In section 4.3, we measure the stability of the algorithm through repeated experiments. In section 4.4, we compare our algorithm with some existing methods. In section 4.5, we give some enhancement results when the network is trained with one single low light image.
Evaluation Indexes
------------------
There are many indexes with or without reference that can be used to evaluate the quality of the enhanced image. However, the constrain we use in this paper does not conform to the natural image characteristics, so it is difficult for us to evaluate the enhanced image accurately with those existing evaluation indexes. In this paper, we use gray entropy (GE), color entropy (CE, color entropy is the sum of entropy of [R,G,B]{} channels), gray mean illumination (GMI), gray mean gradient (GMG), LOE[@wang2013naturalness], NIQE[@mittal2012making], PSNR, SSIM to evaluate the enhanced image. It should be noted that these indexes can only reflect the image quality in some aspects, which are not completely consistent with the evaluation results given by the human visual system. The LOE$_{low}$ and LOE$_{high}$ are calculated with low and high light images respectively.
The Influence of Training Times
-------------------------------
We use 485 low light images in the LOL dataset for training, and 15 for testing. Considering that our method is self-supervised, it lacks an absolute reference, and some parameters and constraints in our loss function come from the individual experience. We cannot determine whether our training has reached the best through the change of the loss. So we train the network for 1,000 epochs, and process the testing data every 20 training epochs and use those indexes to evaluate the training results of the network.
Fig.3 and Fig.4 show the change of loss and indexes with the increase of training times. It can be seen that the loss falling fast at the beginning. On our GPU, it takes less than 0.65s to train one epoch. Fig.5 shows enhancement results of low light images in the testing data with different training times. We only selected the results of the first 200 epochs to display. It can be seen that as the training goes, some indexes which can reflect the image clarity such as entropy and gradient increase, however, the gap between the enhanced images and reference images is also growing. That is caused by noise, although the image becomes more and more clear as the training goes, at the same time, the noise keep increasing too. In order to keep balance between clarity and noise, we just stop training after 200 epochs. And in our experiment, if the training epochs keep increasing more than about 1000 epochs, there will be artifacts in some testing images like [@chen2018learning]. Early stopping is a reasonable method to avoid the noise and artifacts.
![The training loss with 1000 epochs[]{data-label="fig3"}](fig3){width="3.5in"}
{width="\linewidth"}
{width="\linewidth"}
Repeated Learning Stability
---------------------------
Due to the characteristics of learning based methods, in most of cases, we cannot reproduce the optimal results. So we repeat the experiments many times to evaluate the repeatability of the method. In every experiment, we train the network with 200 epochs, and evaluate the network on testing data through indexes mentioned in section 4.1. Fig.6 and Fig.7 show the evaluation indexes and some enhancement results in different experiments respectively. It can be seen that there are large fluctuations in some indexes, like LOE、GMG and NIQE, however, the changings of enhancement results are not so obvious in most experiments. In the fifth experiment, the color of enhanced images are lighter than others. We think that the differences of enhancement results may come from the $L_{1}$ loss functions and the difference among the training data in every experiment.
In every experiment, the only difference is training patch, which seems to have an impact on training results. Those training patches are randomly selected and cropped, and considering the training times and the large size difference between images and patches, they are only a small part of training images. At the same time, we use the $L_{1}$ loss for training, compared with $L_{2}$ loss, $L_{1}$ loss may have multiple solutions, and its solutions will be highly affected by training data. When training data changes, the results may also change greatly. However, from visual effects, the method proposed in this paper is relatively stable.
{width="\linewidth"}
{width="\linewidth"}
Metrics GE GMI CE GMG LOE$_{low}$ LOE$_{high}$ NIQE PSNR SSIM Time
--------------- ----------- ------- ----------- ----------- ------------- -------------- ----------- ----------- ------------ ------------
HE **7.785** 105.3 **23.21** 17.12 **505.3** **898.6** 5.201 15.81 0.5607 0.0158
MSR 6.947 134.7 17.61 16.93 540.3 950.1 8.008 16.69 0.5262 0.0911
NPE 7.070 96.67 20.787 20.38 1607.7 1867.8 9.135 16.97 0.5894 4.71
MF 6.937 85.99 20.16 17.47 840.9 1197.7 9.713 16.97 0.6049 0.128
SRIE 6.296 50.33 18.62 9.380 952.0 1291.4 7.535 11.86 0.4978 3.45
LIME 7.564 114.9 21.72 **23.46** 1303.5 1543.7 9.127 16.76 0.5644 0.211
Gladnet 7.116 119.0 21.52 9.918 902.5 1205.5 6.797 **19.72** 0.7035 0.0212
Retinex-Net 6.835 110.2 21.13 24.00 1990.1 1988.8 9.730 16.77 0.5594 0.0207
$L_{2}-L_{p}$ 6.419 51.92 19.16 8.704 933.4 1250.4 6.234 12.15 0.5103 5.58
Ours 7.180 108.7 21.65 7.653 1210.8 1384.1 4.793 19.15 **0.7108** **0.0145**
Ours-single 7.030 88.08 20.94 10.27 529.0 1028.2 **4.422** 14.18 0.5169 **0.0145**
Reference 7.040 115.5 21.31 6.910 921.9 - 4.253 - 1 -
{width="\linewidth"}
{width="\linewidth"}
{width="\linewidth"}
Comparisons with Existing Algorithms
------------------------------------
We have also compared our algorithm with some existing classic and state-of-the-art methods, including HE, MSR[@jobson1997multiscale], LIME[@guo2016lime], MF[@fu2016fusion], NPE[@wang2013naturalness],SRIE[@fu2016weighted], Gladnet[@wang2018gladnet], Retinex-Net [@wei2018deep], $L_{2}$-$L_{p}$[@fu2019hybrid].
The 15 images from the LOL dataset are used for testing to get the objective indexes, and Fig.8 and Fig.9 show some enhancement results by different methods, and Table 2 displays the objective indexes and time consumption of those methods. The low light image of Fig.10 is from the LIME [@guo2016lime] dataset. All the results of our method come from a randomly selected experiment.
SSIM is generally used to measure the structural similarity between two images. NIQE is a non reference image quality evaluation method. These two indexes can show that our method has good structural similarity and image quality after processing.
In CE, GE and GMG, we can see that our method is lower than some methods. Although the larger these indexes are, the more abundant information these images have and the clearer these images are, we also need to consider that these indexes will be highly affected by noise. It can be seen that compared with most of the methods, our method is closer to the reference image in these indexes.
In LOE$_{low}$ and LOE$_{high}$, our method does not perform well. This is probably because that the overexposure area and underexposure area may exist in low light images or the reference images at the same time, and the model proposed in this paper can avoid this problem to a certain extent, which leads to our poor performance in these two indexes. As shown in the lower left corner of Fig.7, there are overexposure areas in the reference image. If we use such reference image for training, we can not promise that the training results will not be over exposed. It can also explain that it is difficult to obtain the optimal reference image by adjusting the exposure time.
In PSNR, although our method is lower than Gladnet[@wang2018gladnet], we need to pay attention to that our method does not consider the reference image in training, so it is difficult to ensure the similarity of the enhanced image and the reference image in brightness, which has a certain impact on PSNR.
Although our method can not achieve the best results in all indexes, in terms of visual effect and some important indexes, our method achieves the state-of-the-art. Compared with HE method, our method has a slight denoising effect and can better maintain the structure information and color information. As shown in Table 2, the PSNR and SSIM of our method is higher than the HE method. HE method is not suitable for heavy noise environment and it can be seen in Fig.8 and Fig.9, there are still dark areas in the image after directly using HE method, which is caused by the theory of histogram equalization itself. Compared with the model-based methods, our method cost less running time. It can be seen in Table 2, our method is more than 6x faster than MSR[@jobson1997multiscale] which has the least time among model-based method. Compared with the learning based methods, our method does not need to build data sets carefully, which can save a lot of time and energy and has better applicability for new environment and equipment.
Training with Single Low Light Image
------------------------------------
At the same time, in order to further evaluate the performance of the method proposed in this paper, we do an experiment that train the network with single low light image. The training image is one of the test data of the LOL dataset and the test data is all the 15 test data of LOL dataset. In Fig.12, we use image 12-(a) only to train the network, and 12-(b) to 12-(k) are the enhancement results of image 12-(a) by different training epochs. Fig.11 displays the evaluation indexes on the test data by different training epochs. Fig.13 shows the enhancement results on some LOL test data and other low light images. Table 2 shows the indexes on the test data by 10000 training epochs. Through Fig.11 to Fig.13, we can see that our method can be applied to new environments quickly, even if we only have one image of the new environment.
It can be seen that in terms of visual effect and some indexes, the result of single image training is worse than that of multiple images training. However in our experiment, with the increase of training times, single image training does not produce artifacts. That can prove that the artifacts are not produced by the model proposed in this paper. As we train the image with single low light image, there is no need for the network to fit the histogram equalization stretch, that is an main reason why there are no artifacts in single low light training. It is perhaps difficult to fit the histogram equalization stretch when the network is trained without considering the whole image information in multiple images training. We think that if we want to avoid the artifacts in multiple images training, we have to make the network deeper or considering the whole image information, however, that will also increase time consumption.
{width="\linewidth"}
{width="\linewidth"}
{width="\linewidth"}
Conclusion
==========
In this paper, we propose a maximum entropy based Retinex model and a self-supervised image enhancement network. The network can be trained with low light images only and can slightly reduce the noise during enhancement. By testing on real low light images, it shows that, with short time training, the network can produce a well visual effect and has a good real-time performance. It should be noted that our method is self-supervised, so it can adapt to new environments and devices, also, the enhanced image may be different from the real data and look more like the night one in color. The future work will focus on the color restoration, noise and artifact suppression, better detail keeping and so on, we think those can be achieved through Generative Adversarial Networks or new constrains.
[Yu Zhang]{} was born in Hebei, China. He received the B.E. degree and M.S.degree in Control Science and Engineering from Harbin Institute of Technology, China, in 2016 and 2018, respectively. He is currently studying for a Ph.D degree in Electronic science and technology in National Key laboratory of Tunable Laser Technology, Harbin Institute of Technology. His research interests include image restoration and enhancement, SLAM.
[Xiaoguang Di]{} was born in Heilongjiang,China. He received the M.S. and Ph.D degree in navigation, guidance and control from Northwestern Polytechnical University, China, in 1999 and 2004, respectively. He is currently an associate professor with the Control and Simulation Center, Harbin Institute of Technology, Where he is in charge of courses in digital image processing and computer vision. His current research interests include real-time image restoration and enhancement, 3D object detection and recognition, SLAM. Prof. Di is a member of China Simulation Federation and Chinese Society of Astronautics.
[Bin Zhang]{} was born in Gansu, China. He received the B.Sc. degree in apply physics and the M.Sc. degree in physics from China University of Petroleum, in 2010 and 2013, respectively. He is currently studying for a PhD degree in National Key Laboratory of Tunable Laser Technology, Harbin Institute of Technology. His research interests include laser imaging and laser transmission.
[Chunhui Wang]{} Chunhui Wang received the BS, MS, and PhD degrees from Harbin Institute of Technology in 1987, 1991, and 2005，respectively. He is currently a processor and Deputy Director, Institute of Optoelectronic Technology, Harbin Institute of Technology. His current major research interests are laser remote sensing, lidar, laser detection and recognition.
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The classical Waring problem deals with expressing every natural number as a sum of $g(k)$ $k^{\mathrm {th}}$ powers. Similar problems were recently studied in group theory, where we aim to present group elements as short products of values of a given word $w \ne 1$. In this paper we study this problem for Lie groups and Chevalley groups over infinite fields.
We show that for a fixed word $w \ne 1$ and for a classical connected real compact Lie group $G$ of sufficiently large rank we have $w(G)^2=G$, namely every element of $G$ is a product of $2$ values of $w$.
We prove a similar result for non-compact Lie groups of arbitrary rank, arising from Chevalley groups over ${\mathbb{R}}$ or over a $p$-adic field. We also study this problem for Chevalley groups over arbitrary infinite fields, and show in particular that every element in such a group is a product of two squares.
address:
- 'Chun Yin Hui, Einstein Institute of Mathematics, Hebrew University, Givat Ram, Jerusalem 91904, Israel'
- 'Michael Larsen, Department of Mathematics, Indiana University, Bloomington, IN 47405, U.S.A.'
- 'Aner Shalev, Einstein Institute of Mathematics, Hebrew University, Givat Ram, Jerusalem 91904, Israel'
author:
- 'Chun Yin Hui, Michael Larsen, Aner Shalev'
title: The Waring problem for Lie groups and Chevalley groups
---
[^1]
Introduction
============
Let $F_d$ be the free group on $x_1, \ldots , x_d$ and let $w = w(x_1, \ldots , x_d) \in F_d$ be a word. For every group $G$ there is a word map $w = w_G: G^d \rightarrow G$ obtained by substitution. The image of this map is denoted by $w(G)$. The theory of word maps has developed significantly in the past decade; see [@La; @Sh1; @S; @LaSh1; @LaSh2; @LOST; @LST; @AGKSh; @Sh2] and the references therein.
A major goal in these investigations is to prove theorems of “Waring type”, i.e., to find small $k$ such that, for every word $w \ne 1$ and for various groups $G$ we have $w(G)^k = G$, namely every element of $G$ is a product of $k$ values of $w$.
A theorem of Borel [@Bo1] states that if $w$ is a non-trivial word then the word map it induces on simple algebraic groups $G$ is dominant. Thus $w(G)$ contains a dense open subset, which easily implies $w(G(F))^2 = G(F)$ where $F$ is an algebraically closed field (see Corollary 2.2 in [@Sh2]). However, much more effort is required in order to prove similar results for fields $F$ (finite or infinite) which are not algebraically closed.
In [@Sh1] it is shown that, fixing $w \ne 1$, we have $w(G)^3 = G$ for all sufficiently large (nonabelian) finite simple groups $G$. This is improved in [@LaSh1; @LaSh2; @LST] to $w(G)^2 = G$. Results of type $w(G)^3 = G$ were recently obtained in [@AGKSh] for $p$-adic groups $G({\mathbb{Z}}_p)$.
The purpose of this paper is to study similar problems for Lie groups and for infinite Chevalley groups. Our main results are as follows.
\[T11\] For every two non-trivial words $w_1, w_2$ there exists $N = N(w_1, w_2)$ such that if $G$ is a classical connected real compact Lie group of rank at least $N$ then $$w_1(G)w_2(G) = G.$$
In particular, for any $w \ne 1$ there is $N = N(w)$ such that $w(G)^2 = G$ for all classical connected real compact Lie groups of rank at least $N$.
We note that the assumption that the rank of $G$ is large is necessary. By a theorem of E. Lindenstrauss (private communication) and A. Thom [@T Cor. 1.2], for any $n \ge 2$ and $\epsilon > 0$ there exists a word $1 \ne w \in F_2$ such that all elements of $w({\mathrm{U}}(n))$ have distance $\le \epsilon$ from the identity. Embedding a given $G$ in ${\mathrm{U}}(n)$, we can arrange that $w(G)^2 \neq G$ or even $w(G)^k \neq G$ for any fixed $k$. We can also find a sequence $\{ w_i \}$ of non-trivial words in two variables such that, for every compact group $G$, $w_i(G)$ converges to $1$.
We also establish a width $2$ result for non-compact Lie groups which arise from Chevalley groups over ${\mathbb{R}}$ or over a $p$-adic field. Here a Chevalley group over a field $F$ means a group generated by the root groups $X_\alpha(F)$ associated to a faithful representation of a complex semisimple Lie algebra (see [@St $\mathsection3$]), or equivalently, the commutator subgroup of $G_F(F)$, where $G_F$ is a split semisimple algebraic group over $F$. In this case there is no large rank assumption.
\[T12\] Let $F$ be a field that contains either ${\mathbb{R}}$ or ${\mathbb{Q}}_p$ for some prime number $p$. Let $w_1$,$w_2$ be non-trivial words and $G$ a Chevalley group over $F$ associated to a complex simple Lie algebra. Then $$G\setminus Z(G)\subset w_1(G)w_2(G).$$ In particular, if $Z(G) = \{ 1 \}$, then $w_1(G)w_2(G) = G$.
Without assumptions on the center of $G$ this result easily implies $w_1(G)w_2(G)w_3(G) = G$ for any non-trivial words $w_1, w_2, w_3$.
Our last results deal with Chevalley groups over an arbitrary infinite field $F$. Here we have a general width $4$ result, and width 3 and 2 in special cases.
Let $w_1$,$w_2$,$w_3$,$w_4$ be non-trivial words and let $F$ be an infinite field.
1. If $G$ is a Chevalley group over $F$ associated to a complex simple Lie algebra, then $$G\setminus Z(G)\subseteq w_1(G)w_2(G)w_3(G)w_4(G).$$ In particular, if $Z(G) = \{ 1 \}$, then $w_1(G)w_2(G)w_3(G)w_4(G) = G$.
2. If $G={\mathrm{SL}}_n(F)$ and $n>2$, then $$G\setminus Z(G)\subseteq w_1(G)w_2(G)w_3(G).$$ Hence, $w_1({\mathrm{PSL}}(n,F))w_2({\mathrm{PSL}}(n,F))w_3({\mathrm{PSL}}(n,F)) = {\mathrm{PSL}}(n,F)$.
For some specific words we obtain stronger results.
Let $w_1 = x^m$ and $w_2 = y^n$ where $m, n$ are positive integers. Let $G$ be a Chevalley group over an infinite field $F$ associated to a complex semisimple Lie algebra $\mathfrak{g}$.
1. If $\mathfrak{g}$ is simple, then $$G \setminus Z(G)\subseteq w_1(G)w_2(G).$$ In particular, if $Z(G) = 1$ then $w_1(G)w_2(G) = G$.
2. If $m=n=2$, then $$G = w_1(G)w_2(G).$$
We also give an example showing that a non-trivial central element is not in the image of the word map $x^4y^4$ (Proposition 4.1). See also [@LaSh3] for the probabilistic behavior of word maps induced by $x^my^n$ on finite simple groups.
The fact that every element of $G$ above is a product of two squares can be regarded as a non-commutative analogue of Lagrange’s four squares theorem. A similar result for finite quasisimple groups can be found in [@LST2].
This paper is organized as follows. In Section 2 we deal with compact Lie groups and prove Theorem 1.1. Section 3 is devoted to the proof of Theorem 1.2, and Theorems 1.3 and 1.4 are proved in Section 4.
Compact Lie groups
==================
In this section we provide solutions for Waring type problems with length two for classical connected real compact Lie groups $G$ with large rank, thus proving Theorem \[T11\]. It suffices to work with simply connected groups $G$, i.e. with ${\mathrm{SU}}(n)$, ${\mathrm{Sp}}(n)$, and ${\mathrm{Spin}}(n)$. Let us start with Gotô’s theorem.
[@Go] Let $G$ be a connected compact semisimple Lie group and $T$ a maximal torus of $G$. Then there exists $x\in N_G(T)$ such that ${\operatorname{Ad}}(x)-1$ is non-singular on $\mathrm{Lie}(T)$. Hence, every element $g$ of $G$ is conjugate to $[x,t]:=xtx^{-1}t^{-1}$ for some $t\in T$.
Let $w_1$ and $w_2$ be non-trivial words. Every element of $G$ can be conjugated into $T$ so the width two result for $G$ follows if we can prove $T\subset w_1(G) w_2(G)$. By Gotô’s theorem, it suffices to show that $x\in w_1(G)$ and $x^{-1}\in w_2(G)$. This will be achieved by using the principal homomorphism [@Se]. Identify $S^1$, the subgroup of unit circle of ${\mathbb{C}}^*$ as a maximal torus of ${\mathrm{SU}}(2)$.
The primitive $n$th roots of unity $\zeta_n^{\pm 1}:=e^{\pm \frac{2\pi i}{n}}$ both belong to $w_i({\mathrm{SU}}(2))\cap S^1$ for $i=1,2$ if $n$ is sufficiently large.
Since $w_1$ and $w_2$ are non-trivial, $w_i({\mathrm{SU}}(2))$ contains a non-empty open subset of ${\mathrm{SU}}(2)$ for $i=1,2$ [@La Cor. 5]. As ${\mathrm{SU}}(2)$ is compact and connected and $x$ and $x^{-1}$ are conjugate for any $x\in{\mathrm{SU}}(2)$, it follows that $w_i({\mathrm{SU}}(2))\cap S^1$ is a closed arc and also a symmetric neighborhood of $1$ in $S^1$ for $i=1,2$. Hence, the primitive $n$-th roots of unity $\zeta_n,\zeta_n^{-1}\in S^1$ belong to $w_i({\mathrm{SU}}(2))$ for $i=1,2$ if $n$ is sufficiently large.
We make the following definitions.
1. Let $I_n$ be the identity complex $n\times n$ matrix.
2. Let $0_n$ be the zero complex $n\times n$ matrix.
3. Let $E_n^i$ be the diagonal complex $n\times n$ matrix whose $(i,i)$-entry is $1$ and all other entries $0$.
4. Let $L_n^i$ be the linear functional on diagonal complex $n\times n$ matrices such that $L_n^i(E_n^j)=\delta_{ij}$ (Kronecker delta) for all $1\leq j\leq n$.
5. Let $s_n\in{\mathrm{U}}(n)$ be the $n$-cycle $$\begin{pmatrix}
0 &0 &...&0 & 1\\
1 &0&... &0 &0\\
0 &1&...&0& 0\\
\vdots&\vdots &\ddots &\vdots &\vdots\\
0 &0 &... &1 &0
\end{pmatrix}.$$
\[T25\] For any non-trivial words $w_1,w_2$ and a sufficiently large $n$ we have $${\mathrm{SU}}(n)=w_1({\mathrm{SU}}(n))w_2({\mathrm{SU}}(n)).$$
Consider the commutative diagram $$\xymatrix{
{\mathrm{SL}}(2,{\mathbb{C}}) \ar[rd]^{p} \ar[r]^{\widetilde{p}}
&{\mathrm{SL}}(n,{\mathbb{C}}) \ar@{->>}[d]^{\pi}\\
&{\mathrm{PSL}}(n,{\mathbb{C}}) }$$ where $p$ is the principal homomorphism associated to simple roots [@FH] $$\Delta:=\{L^{1}_n-L_n^{2},L_n^2-L_n^3,...,L_n^{n-1}-L_n^{n}\},$$ $\pi$ the adjoint quotient, and $\widetilde{p}$ a lifting of $p$. Let $H=\begin{pmatrix}
1 & 0\\
0 &-1
\end{pmatrix}\in\mathfrak{sl}(2,{\mathbb{C}})$. Homomorphism $\widetilde{p}$ is isomorphic to $\mathrm{Sym}^{n-1}$ since $\alpha(d\widetilde{p}(H))=2$ for all $\alpha\in\Delta$ [@Se $\mathsection2.3$]. By restricting to suitable maximal compact subgroups, we obtain $\widetilde{p}:{\mathrm{SU}}(2)\rightarrow {\mathrm{SU}}(n)$. Identify $S^1$ as a maximal torus of ${\mathrm{SU}}(2)$. The set of eigenvalues of $\widetilde{p}(\zeta_m)$ is $$\{\zeta_m^{n-1},\zeta_m^{n-3},...,\zeta_m^{3-n},\zeta_m^{1-n}\}.$$ Then $\widetilde{p}(\zeta_{2n})$ is conjugate to $x_n:=s_n$ if $n$ is odd and $\widetilde{p}(\zeta_{2n})$ is conjugate to $x_n:=s_n\cdot\zeta_{2n}I_{n}$ if $n$ is even by comparing eigenvalues. Let $T$ be the diagonal maximal torus of ${\mathrm{SU}}(n)$. Since ${\operatorname{Ad}}(x_n)-1$ is non-singular on $\mathrm{Lie}(T)$ for all $n$, $x_n\in w_1(\widetilde{p}({\mathrm{SU}}(2)))$ and $x_n^{-1}\in w_2(\widetilde{p}({\mathrm{SU}}(2)))$ for $n\gg0$ by Lemma 2.2, we are done by Gotô’s theorem.
Next, we work with the real compact symplectic group $${\mathrm{Sp}}(n):={\mathrm{U}}(2n)\cap {\mathrm{Sp}}(2n,{\mathbb{C}}),$$ where ${\mathrm{Sp}}(2n,{\mathbb{C}})$ is the subgroup of ${\mathrm{GL}}(2n,{\mathbb{C}})$ that preserves the form $$\begin{pmatrix}
0_n &I_n\\
-I_n &0_n
\end{pmatrix}.$$
\[T27\] For any non-trivial words $w_1,w_2$ and a sufficiently large $n$ we have $${\mathrm{Sp}}(n)=w_1({\mathrm{Sp}}(n))w_2({\mathrm{Sp}}(n)).$$
Let $T$ be the maximal torus of ${\mathrm{Sp}}(n)$ consisting of diagonal matrices with complex entries. Let $x_n\in N_{{\mathrm{Sp}}(n)}(T)$ be the element $$\begin{pmatrix}
s_n &0_n\\
0_n &s_n
\end{pmatrix}\cdot
\begin{pmatrix}
I_n-E_n^1 &E_n^1\\
-E_n^1 & I_n-E_n^1
\end{pmatrix}.$$ Then it is easy to see that ${\operatorname{Ad}}(x_n)-1$ is non-singular on $\mathrm{Lie}(T)$. By Gotô’s Theorem, it suffices to show that $x_n\in w_1({\mathrm{Sp}}(n))$ and $x_n^{-1}\in w_2({\mathrm{Sp}}(n))$ for all sufficiently large $n$.
Consider the commutative diagram $$\xymatrix{
{\mathrm{SL}}(2,{\mathbb{C}}) \ar[rd]^{p} \ar[r]^{\widetilde{p}}
&{\mathrm{Sp}}(2n,{\mathbb{C}}) \ar@{->>}[d]^{\pi}\\
&{\mathrm{PSp}}(2n,{\mathbb{C}}) }$$ where $p$ is the principal homomorphism associated to simple roots [@FH] $$\Delta:=\{L^{1}_{2n}-L_{2n}^{2},L_{2n}^2-L_{2n}^3,...,L_{2n}^{n-1}-L_{2n}^{n},2L_{2n}^n\},$$ $\pi$ the adjoint quotient, and $\widetilde{p}$ a lifting of $p$. Let $H=\begin{pmatrix}
1 & 0\\
0 &-1
\end{pmatrix}\in\mathfrak{sl}(2,{\mathbb{C}})$. Since $\alpha(d\widetilde{p}(H))=2$ for all $\alpha\in\Delta$ [@Se $\mathsection2.3$], the set of weights of $\widetilde{p}$, viewed as a $2n$-dimensional representation, is $$\{2n-1,2n-3,...,1,-1,...,3-2n,1-2n\}.$$ By restricting to suitable maximal compact subgroups, we obtain $\widetilde{p}:{\mathrm{SU}}(2)\rightarrow {\mathrm{Sp}}(n)$.
Identify $S^1$ as a maximal torus of ${\mathrm{SU}}(2)$. The set of eigenvalues of $\widetilde{p}(\zeta_m)$ is $$\{\zeta_m^{2n-1},\zeta_m^{2n-3},...,\zeta_m^{3-2n},\zeta_m^{1-2n}\}.$$ Define $e_1:=(1,0,...,0)\in {\mathbb{C}}^{2n}$. Since $x_n$ satisfies $x_n^{2n}+1=0$ and the set of vectors $$\{x_ne_1,x_n^2e_1,...,x_n^{2n}e_1\}$$ is linearly independent, the characteristic polynomial of $x_n$ is $t^{2n}+1$. In ${\mathrm{Sp}}(n)$, the characteristic polynomial determines the conjugacy class. (Indeed, the diagonal matrices with entries $\lambda_1,\ldots,\lambda_n, \lambda_1^{-1},\ldots,\lambda_n^{-1}$ form a single orbit under the action of the Weyl group.) Since $\widetilde{p}(\zeta_{4n})$ and $x_n$ have the same characteristic polynomial, it follows that they are conjugate in ${\mathrm{Sp}}(n)$. Hence, $x_n$ and $x_n^{-1}$ respectively belong to $w_1(\widetilde{p}({\mathrm{SU}}(2)))$ and $w_2(\widetilde{p}({\mathrm{SU}}(2)))$ when $n$ is sufficiently large by Lemma 2.2. We are done.
We then consider compact special orthogonal group ${\mathrm{SO}}(n)$ and its simply connected cover ${\mathrm{Spin}}(n)$ for $n\geq3$.
For any non-trivial words $w_1,w_2$ and a sufficiently large $n$ we have $${\mathrm{SO}}(n)=w_1({\mathrm{SO}}(n))w_2({\mathrm{SO}}(n)).$$
Since we have a morphism ${\mathrm{SO}}(2n)\rightarrow {\mathrm{SO}}(2n+1)$ such that the image of a maximal torus of ${\mathrm{SO}}(2n)$ is a maximal torus of ${\mathrm{SO}}(2n+1)$, it suffices to deal with ${\mathrm{SO}}(2n)$. This is a maximal compact subgroup of ${\mathrm{SO}}(2n,{\mathbb{C}})$. Let $K(2n,{\mathbb{C}})$ be the subgroup of ${\mathrm{SL}}(2n,{\mathbb{C}})$ preserving the form $$\begin{pmatrix}
0_n &I_n\\
I_n &0_n
\end{pmatrix}.$$ Since $K(2n,{\mathbb{C}})$ is isomorphic to ${\mathrm{SO}}(2n,{\mathbb{C}})$ [@FH] and has a diagonal maximal torus, we use $K(2n,{\mathbb{C}})$ and $K(2n):={\mathrm{U}}(2n)\cap K(2n,{\mathbb{C}})$ (a maximal compact of $K(2n,{\mathbb{C}})$) instead of ${\mathrm{SO}}(2n,{\mathbb{C}})$ and ${\mathrm{SO}}(2n)$. One checks that the diagonal maximal torus $T$ of $K(2n)$ is equal to the diagonal maximal torus of ${\mathrm{Sp}}(n)$. Let $s_{n-1}'$ be the $n\times n$ matrix $$\begin{pmatrix}
1 &0 &...&0 & 0\\
0 &0&... &0 &1\\
0 &1&...&0& 0\\
\vdots&\vdots &\ddots &\vdots &\vdots\\
0 &0 &... &1 &0
\end{pmatrix}$$ which fixes $e_1:=(1,0,...,0)\in{\mathbb{C}}^n$ and is an $(n-1)$-cycle on the natural complement of $e_1$ in ${\mathbb{C}}^n$. Let $x_n\in N_{K(2n)}(T)$ be the element $$\begin{pmatrix}
s_{n-1}' &0_n\\
0_n &s_{n-1}'
\end{pmatrix}\cdot
\begin{pmatrix}
I_n-E_n^1-E_n^2 & E_n^1+E_n^2\\
E_n^1+E_n^2 & I_n-E_n^1-E_n^2
\end{pmatrix}.$$ By choosing the basis $$\{E_{2n}^1-E_{2n}^{n+1},E_{2n}^2-E_{2n}^{n+2},...,E_{2n}^n-E_{2n}^{2n}\}$$ for $\mathrm{Lie}(T)$, the ${\operatorname{Ad}}(x_n)$-action on $\mathrm{Lie}(T)$ is given by the $n\times n$ matrix $$\begin{pmatrix}
-1 &0 & 0 &...&0 & 0\\
0 &0& 0&... &0 &1\\
0 &-1&0&...&0& 0\\
0 &0&1&...&0& 0\\
\vdots&\vdots&\vdots &\ddots &\vdots &\vdots\\
0 &0&0 &... &1 &0
\end{pmatrix}.$$ One sees that ${\operatorname{Ad}}(x_n)-1$ is non-singular on $\mathrm{Lie}(T)$. By Gotô’s Theorem, it suffices to show that $x_n\in w_1(K(2n))$ and $x_n^{-1}\in w_2(K(2n))$ for all sufficiently large $n$. Since $x_n$ is conjugate in ${\mathrm{GL}}(2n,{\mathbb{C}})$ to the permutation matrix $$\begin{pmatrix}
0 &1 & 0 &...&0 & 0\\
1 &0& 0&... &0 &0\\
0 &0&0&...&0& 1\\
0 &0&1&...&0& 0\\
\vdots&\vdots&\vdots &\ddots &\vdots &\vdots\\
0 &0&0 &... &1 &0
\end{pmatrix},$$ the characteristic polynomial of $x_n$ is $(t^2-1)(t^{2n-2}-1)$.
Consider the commutative diagram $$\xymatrix{
{\mathrm{SL}}(2,{\mathbb{C}}) \ar[rd]^{p} \ar[r]^{\widetilde{p}}
&K(2n,{\mathbb{C}}) \ar@{->>}[d]^{\pi}\\
&K(2n,{\mathbb{C}})/\{\pm I_{2n}\} }$$ where $p$ is the principal homomorphism associated to simple roots [@FH] $$\Delta:=\{L^{1}_{2n}-L_{2n}^{2},L_{2n}^2-L_{2n}^3,...,L_{2n}^{n-1}-L_{2n}^{n},L_{2n}^{n-1}+L_{2n}^{n}\},$$ $\pi$ the adjoint quotient, and $\widetilde{p}$ a lifting of $p$. Let $H=\begin{pmatrix}
1 & 0\\
0 &-1
\end{pmatrix}\in\mathfrak{sl}(2,{\mathbb{C}})$. Since $\alpha(d\widetilde{p}(H))=2$ for all $\alpha\in\Delta$ [@Se $\mathsection2.3$], the multiset of weights of the $2n$-dimensional representation $\widetilde{p}$ is $$\{2n-2,2n-4,...,2,0,0,-2,...,4-2n,2-2n\}.$$ By restricting to suitable maximal compact subgroups, we obtain $\widetilde{p}:{\mathrm{SU}}(2)\rightarrow K(2n)$. Identify $S^1$ as a maximal torus of ${\mathrm{SU}}(2)$. The set of eigenvalues of $\widetilde{p}(\zeta_m)$ is $$\{\zeta_m^{2n-2},\zeta_m^{2n-4},...,\zeta_m^{2},1,1,\zeta_m^{-2},...,\zeta_m^{4-2n},\zeta_m^{2-2n}\}.$$
It is not in general true that two diagonal orthogonal matrices are conjugate in ${\mathrm{SO}}(2n)$ if and only if they have the same characteristic polynomial, because the Weyl group of ${\mathrm{SO}}(2n)$ is only an index $2$ subgroup of $S_n\ltimes ({\mathbb{Z}}/2{\mathbb{Z}})^n$. However, it is true in the case of matrices for which $1$ is an eigenvalue. Since $\widetilde{p}(\zeta_{4n-4})$ and $x_n$ both have characteristic polynomial $$(t^2-1)(t^{2n-2}-1)$$ and have eigenvalue $1$, it follows that they are conjugate in $K(2n)$. Hence, $x_n$ and $x_n^{-1}$ respectively belong to $w_1(\widetilde{p}({\mathrm{SU}}(2)))$ and $w_2(\widetilde{p}({\mathrm{SU}}(2)))$ when $n$ is sufficiently large by Lemma 2.2. We are done.
For any non-trivial words $w_1,w_2$ and a sufficiently large $n$ we have $${\mathrm{Spin}}(n)=w_1({\mathrm{Spin}}(n))w_2({\mathrm{Spin}}(n)).$$
Since we have a morphism ${\mathrm{Spin}}(2n)\rightarrow {\mathrm{Spin}}(2n+1)$ such that the image of a maximal torus of ${\mathrm{Spin}}(2n)$ is a maximal torus of ${\mathrm{Spin}}(2n+1)$, it suffices to deal with ${\mathrm{Spin}}(2n)$. Consider commutative diagram $$\xymatrix{
{\mathrm{SU}}(2) \ar[rd]^{\widetilde{p}} \ar[r]^{\hat{p}}
&{\mathrm{Spin}}(2n) \ar@{->>}[d]^{\widetilde{\pi}}\\
&K(2n) }$$ where $\widetilde{\pi}$ is the natural projection and $\hat{p}$ is a lift of $\widetilde{p}$. Recall maximal torus $T$ and element $x_n\in N_{K(2n)}(T)$ constructed in the proof of Theorem 2.5. Since $x_n$ and $x_n^{-1}$ respectively belong to $w_1(\widetilde{p}({\mathrm{SU}}(2)))$ and $w_2(\widetilde{p}({\mathrm{SU}}(2)))$ when $n$ is sufficiently large, $w_1({\mathrm{Spin}}(2n))$ and $w_2({\mathrm{Spin}}(2n))$ respectively contain $\hat{x}_n$ and $\hat{x}_n^{-1}$ such that $\widetilde{\pi}(\hat{x}_n)=x_n$ by the diagram. Let $\hat{T}$ be the maximal torus of ${\mathrm{Spin}}(2n)$ such that $\widetilde{\pi}(\hat{T})=T$. We also have $\hat{x}_n\in N_{{\mathrm{Spin}}(2n)}(\hat{T})$. Consider commutative diagram
$$\xymatrixcolsep{5pc}\xymatrix{
\mathrm{Lie}(\hat{T}) \ar[d]^{d\widetilde{\pi}} \ar[r]^{{\operatorname{Ad}}(\hat{x}_n)-1} &\mathrm{Lie}(\hat{T})\ar[d]^{d\widetilde{\pi}}\\
\mathrm{Lie}(T) \ar[r]^{{\operatorname{Ad}}(x_n)-1} & \mathrm{Lie}(T)}$$ Since $d\widetilde{\pi}$ and ${\operatorname{Ad}}(x_n)-1$ are non-singular, ${\operatorname{Ad}}(\hat{x}_n)-1$ is also non-singular. We are done by Gotô’s Theorem.
We end this section with a width two result for any compact semisimple real Lie group $G$.
Let $G$ be a compact semisimple real Lie group and $w_1,w_2$ non-trivial words.
1. If $\zeta_4\in S^1\cap w_1({\mathrm{SU}}(2)) \cap w_2({\mathrm{SU}}(2)$, then $w_1(G)w_2(G)=G$.
2. If $w_1^2({\mathrm{SU}}(2))={\mathrm{SU}}(2)$, then $w_1^2(G)=G$.
\(i) Since $S^1\cap w_i({\mathrm{SU}}(2))$ ($i=1,2$) is a connected closed arc, symmetric about the $x$-axis of the complex plane (Lemma 2.2), we obtain $$\zeta_{2n}^{\pm}\in S^1\cap w_1({\mathrm{SU}}(2)) \cap w_2({\mathrm{SU}}(2)$$ for all $n\geq 2$ by the assumption. Hence, $w_1({\mathrm{SU}}(n))w_2({\mathrm{SU}}(n))={\mathrm{SU}}(n)$ for all $n\geq 2$ by the proof of Theorem 2.3. Since every element of $G$ is conjugate to some element in a maximal torus and $G$ contains an equal rank semisimple subgroup $H$ with type A simple factors, we are done.
\(ii) There exists $x_1,x_2\in w_1({\mathrm{SU}}(2))$ such that $x_1x_2=-1$. We may assume $x_1,x_2\in S^1$. Then one sees easily that $\zeta_4\in S^1\cap w_1({\mathrm{SU}}(2))$ by Lemma 2.2. We obtain $w_1^2(G)=G$ by (i)
Non-compact groups
==================
In this section we study Waring type problems for split semisimple Lie groups $G$ over a local field of characteristic $0$ (${\mathbb{R}}$, ${\mathbb{C}}$, or finite extension of ${\mathbb{Q}}_p$) and prove Theorem \[T12\]. A key result we need (related to the Thompson conjecture) was proved by Ellers and Gordeev [@EG1; @EG2; @EG3]:
\[T31\] Let $G$ be a Chevalley group over a field $F$ associated to complex simple Lie algebra. Let $g_1$ and $g_2$ be two regular semisimple elements in $G$ from a maximal split torus and let $C_1$ and $C_2$ be the conjugacy classes of $g_1$ and $g_2$, respectively. Then $$G\setminus Z(G)\subseteq C_1C_2.$$
In order to prove Theorem \[T12\] we also need the following.
\[L32\] Let $F$ be a field that contains either ${\mathbb{R}}$ or ${\mathbb{Q}}_p$ for some $p$ and $w$ a non-trivial word of $d$ letters. Denote by $H$ the group ${\mathrm{SL}}_2(F)$ and by $\mathrm{Tr}$. the trace map on $H$. Define the discriminant on $H^d$ as $$\Delta:=(\mathrm{Tr}\circ w)^2-4.$$ Then $\Delta(H^d)\cap (F^\times)^2$ is an infinite set.
If $A\in {\mathrm{SL}}_2({\mathbb{Q}})$ and $N\in M_2({\mathbb{Q}})$ has trace $0$, then $$f_{A,N}\colon t\mapsto (I_2+tN)A$$ defines a non-constant morphism ${\mathbb{A}}^1\to {\mathrm{SL}}_2$, and the tangent spaces to the curve $f_{A,N}({\mathbb{A}}^1)$ at $A$, as $N$ ranges over trace $0$ matrices, span the tangent space to ${\mathrm{SL}}_2$ at $A$. Thus, for every $(A_i)\in {\mathrm{SL}}_2({\mathbb{Q}})^d$ and every proper subvariety $V$ of ${\mathrm{SL}}_2^d$ which contains $(A_i)$ as a non-singular point, there exists a non-constant morphism $f\colon {\mathbb{A}}^1\to {\mathrm{SL}}_2^d$ defined over ${\mathbb{Q}}$ such that $f(0) = (A_i)$, and $f({\mathbb{A}}^1)$ is not contained in $V$. Any non-constant morphism $\pi\colon {\mathrm{SL}}_2^d\to {\mathbb{A}}^1$ defined over ${\mathbb{Q}}$ is generically smooth, and ${\mathrm{SL}}_2({\mathbb{Q}})^d$ is Zariski-dense in ${\mathrm{SL}}_2$, so there always exists $(A_i)\in {\mathrm{SL}}_2({\mathbb{Q}})^d$ which is a non-singular point of the fiber of $\pi$ to which is belongs. Thus, there exists $f\colon {\mathbb{A}}^1\to {\mathrm{SL}}_2^d$ defined over ${\mathbb{Q}}$ so that the composition $\pi\circ f$ is a non-constant morphism ${\mathbb{A}}^1\to {\mathbb{A}}^1$.
We apply this general observation to the case $\pi := \mathrm{Tr}\circ w$ to obtain a morphism $f$ and a polynomial $P(x)\in {\mathbb{Q}}[x]$, not necessarily zero, such that $\mathrm{Tr}\circ w(f(x)) \equiv P(x)$. To prove the lemma, it suffices to show that $y^2 = P(x)^2-4$ has infinitely many solutions in the field $F$. We write $P(x) = c_0 + \cdots + c_k x^k$, where $c_k\neq 0$ and $k\ge 1$. Consider the curve $X$ over ${\mathbb{Q}}$ given in projective coordinates by $$c_k^2 u^{2k} - \Bigl(\sum_{i=0}^k c_i v^i w^{k-i}\Bigr)^2 + 4w^{2k}.$$ As $P := (1:1:0)$ is a non-singular point, by the (real or $p$-adic) implicit function theorem, there is an infinite (real or $p$-adic) neighborhood of $P$ in $X(F)$. Letting $y := c_k u^k/w^k$ and $x := v/w$, this implies that $y^2 = P(x)^2-4$ has infinitely many solutions in $F$.
***Proof of Theorem \[T12\]*:**
In light of Theorem \[T31\], it suffices to prove the theorem for $F={\mathbb{R}}$ or ${\mathbb{Q}}_p$. Suppose $w$ is a non-trivial word of $d$ letters. Let $D$ be the group of diagonal matrices in ${\mathrm{SL}}_2(F)$. Any Chevalley group $G$ over $F$ is the commutator subgroup of the group of $F$-rational points of a corresponding simple algebraic group $G_F$, and we have the following commutative diagram of algebraic groups over $F$:
$$\xymatrix{
{\mathrm{SL}}_{2,F} \ar[d]^{\pi_1} \ar[r]^{\tilde{p}} &G_{F}\ar[d]^{\pi_2}\\
{\mathrm{PGL}}_{2,F} \ar[r]^p & G_{F}^{{\operatorname{ad}}}}$$
where $\pi_1$ and $\pi_2$ are adjoint quotient maps, $G_{F}^{{\operatorname{ad}}}$ is the adjoint group of $G_{F}$, $p$ is the principal homomorphism associated to a system of simple roots [@Se $\mathsection2$], and $\tilde{p}$ is a lifting of $p$. Since $w({\mathrm{SL}}_2(F))$ contains infinitely many elements in $D$ by Lemma \[L32\] and the image of a generic element of $\pi_1(D)\subset{\mathrm{PGL}}_2(F)$ under $p$ is regular [@Se $\mathsection2.3$], $w(\tilde{p}({\mathrm{SL}}_2(F)))$ contains a regular split semisimple element. This semisimple element belongs to $w(G)$ since ${\mathrm{SL}}_2(F)$ is equal to its commutator subgroup [@Th]. Therefore, we obtain $$G\setminus Z(G)\subseteq w_1(G)w_2(G)$$ for non-trivial words $w_1$ and $w_2$ by Theorem \[T31\].
We now state some easy consequences of Theorem \[T12\].
\[C33\]Let $w_1$,$w_2$,$w_3$ be non-trivial words and $G$ a Chevalley group over a local field of characteristic $0$ associated to complex semisimple Lie algebra $\mathfrak{g}$. Then $$G=w_1(G)w_2(G)w_3(G).$$
Let $\widetilde{G}$ be the universal group [@St p.45] associated to $\mathfrak{g}$. We have central extension $\pi:\widetilde{G}\rightarrow G$ [@St $\mathsection7$] and $\widetilde{G}$ is the direct product of universal groups $\widetilde{G_i}$ associated to simple factors of $\mathfrak{g}$. It suffices to prove the corollary for $\widetilde{G_i}$. Since the center $Z(\widetilde{G_i})$ is finite and $w_3(\widetilde{G_i})$ has non-empty interior (same proof as [@La Cor. 5]), Corollary \[C33\] follows easily from Theorem \[T12\].
Let $w_1$,$w_2$ be non-trivial words and $G^{{\operatorname{ad}}}:=G/Z(G)$ where $G$ is a Chevalley group over $F$ associated to complex semisimple Lie algebra $\mathfrak{g}$. Then $$G^{{\operatorname{ad}}}=w_1(G^{{\operatorname{ad}}})w_2(G^{{\operatorname{ad}}}).$$
Let $\widetilde{G}$ be the universal group associated to $\mathfrak{g}$. We have central extension $\pi:\widetilde{G}\rightarrow G$ and $\widetilde{G}$ is the direct product of universal groups $\widetilde{G_i}$ associated to simple factors of $\mathfrak{g}$. Since $\widetilde{G}/Z(\widetilde{G})=G/Z(G)$, it suffices to prove the corollary for $\widetilde{G_i}$ which is done by Theorem \[T12\].
Chevalley groups
================
The method we used in $\mathsection3$ leads to the proof of Theorem 1.3, given below.
\(i) Applying Theorem 3.1 it suffices to show that $w_1(G)w_2(G)$ and $w_3(G)w_4(G)$ contain split regular semisimple elements. By the principal homomorphism and diagram (5), it suffices to show that $$w_1({\mathrm{SL}}_2(F))w_2({\mathrm{SL}}_2(F))$$ contains an infinite set of split semisimple elements of ${\mathrm{SL}}_2(F)$ of different traces. Since $F$ is infinite and words are non-trivial, $w_1({\mathrm{SL}}_2(F))$ and $w_2({\mathrm{SL}}_2(F))$ both contain infinitely many semisimple elements of different traces. If either $w_1({\mathrm{SL}}_2(F))$ or $w_2({\mathrm{SL}}_2(F))$ contains infinitely many split semisimple elements of different traces, then we are done. Otherwise, let $C_1$ and $C_2$ be conjugacy classes respectively of non-split semisimple elements of $w_1({\mathrm{SL}}_2(F))$ and $w_2({\mathrm{SL}}_2(F))$. Then the diagonal matrix $\mathrm{diag}(\lambda,\lambda^{-1})\in C_1C_2$ if and only if $-\lambda\in\chi(C_1)\chi(C_2)$ [@VW Lemma 6.2], where $\chi(C_i)$ is the set of $(2,1)$ entries of $C_i$ (*corner invariant*) for $i=1,2$ [@VW $\mathsection3$]. Since $F$ is infinite, the corner invariants $\chi(C_1),\chi(C_2)$ are infinite and we are done.\
(ii) An $n$ by $n$ matrix $M$ is said to be *cyclic* if every Jordan block of $M$ is of multiplicity one. Let $G$ be ${\mathrm{SL}}_n(F)$ with $n>2$. A conjugacy class $C$ of $G$ is cyclic if every element of $C$ is cyclic. If $C_1,C_2,C_3$ are cyclic conjugacy classes of $G$, then any non-scalar element of $G$ belongs to the product $C_1C_2C_3$ [@Lev Theorem 3]. Therefore, it suffices to show that $w_1(G)$ contains a cyclic element. Since $F$ is infinite and $w_1$ is non-trivial, $w_1(G)$ contains a regular semisimple (and therefore cyclic) element $g_1$ by the principal homomorphism and diagram (5).
Let us now prove Theorem 1.4.
\(i) Since $G$ is a Chevalley group over an infinite field, it has a maximal split torus $T$. Recall that $w_1=x^{m}$ and $w_2=y^{n}.$ It is easy to see that $w_1(T)$ and $w_2(T)$ both contain split regular semisimple elements. Thus, for $i=1,2$, $w_i(G)$ contains a conjugacy class of $C_i$ of a split regular semisimple element. By Theorem 3.1, we obtain $G \setminus Z(G) \subseteq w_1(G)w_2(G)$, proving the result.\
(ii) Let $\widetilde{G}$ be the universal group associated to $\mathfrak{g}$. We have central extension $\pi:\widetilde{G}\rightarrow G$ and $\widetilde{G}$ is the direct product of universal groups $\widetilde{G_i}$ associated to simple factors of $\mathfrak{g}$. We just need to deal with the case that $G$ is universal and $\mathfrak{g}$ is simple. Since $G \setminus Z(G)\subset w_1(G)w_2(G)$ by (i), it suffices to show $Z(G)\subset w_1(G)w_2(G)$. Let $\Lambda$ and $R$ be respectively the weight lattice and root lattice of $\mathfrak{g}$. We have [@St p.45] $$Z(G)=\mathrm{Hom}(\Lambda/R, F^*)$$ and
Type of $\mathfrak{g}$ $\Lambda/R$
------------------------------------------- --------------------------------------------------------------
$A_n=\mathfrak{sl}(n+1)$ ($n\geq 1$) ${\mathbb{Z}}/(n+1){\mathbb{Z}}$
$B_n=\mathfrak{so}(2n+1)$ ($n\geq 2$) ${\mathbb{Z}}/2{\mathbb{Z}}$
$C_n=\mathfrak{sp}(2n)$ ($n\geq 3$) ${\mathbb{Z}}/2{\mathbb{Z}}$
$D_n=\mathfrak{so}(2n)$ ($n\geq 5$, odd) ${\mathbb{Z}}/4{\mathbb{Z}}$
$D_n=\mathfrak{so}(2n)$ ($n\geq 4$, even) ${\mathbb{Z}}/2{\mathbb{Z}}\times{\mathbb{Z}}/2{\mathbb{Z}}$
$E_6$ ${\mathbb{Z}}/3{\mathbb{Z}}$
$E_7$ ${\mathbb{Z}}/2{\mathbb{Z}}$
$E_8$ $\{1\}$
$F_4$ $\{1\}$
$G_2$ $\{1\}$
Since $|\Lambda/R|$ is odd if $\mathfrak{g}=\mathfrak{sl}(2n+1),E_6,E_8,F_4,G_2$ ($n\geq1)$, every element of $Z(G)$ is a square in these cases. If $\mathrm{char}(F)=2$, then $Z(G)$ is trivial for the remaining $\mathfrak{g}$ and (ii) is true. Assume $\mathrm{char}(F)\neq 2$.\
Case $\mathfrak{g}=\mathfrak{sl}(2n)$ ($n\geq1$):\
We have $G={\mathrm{SL}}_{2n}(F)$ and $Z(G)=\{rI_{2n}:~ r\in F,~r^{2n}=1\}$. Define $J_r:=\begin{pmatrix}
0 &1\\
r &0
\end{pmatrix}$ whenever $r\in F$ is a $2n$-th (not necessary primitive) root of unity. When $n=1$, the non-trivial center of $G$ is a square since $$\begin{pmatrix}
-1 &0\\
0 & -1
\end{pmatrix}=J_{-1}^2.$$ When $n>1$, every center element of ${\mathrm{SL}}_{2n}(F)$ is a product of $2$ squares since $$rI_{2n}=
\begin{pmatrix}
J_r &... &0 &0\\
\vdots &\ddots &\vdots &\vdots\\
0 &... &J_r &0\\
0 &...&0 &J_{r}
\end{pmatrix}^2$$ if $(-r)^n=1$ and $$rI_{2n}=\begin{pmatrix}
I_2 &... &0 &0\\
\vdots &\ddots&\vdots &\vdots\\
0 &... &I_2 &0\\
0 &... &0 & J_{-1}
\end{pmatrix}^2
\begin{pmatrix}
J_r &... &0 &0\\
\vdots &\ddots &\vdots &\vdots\\
0 &... &J_r &0\\
0 &...&0 &J_{-r}
\end{pmatrix}^2$$ if $(-r)^n=-1$.\
Case $\mathfrak{g}=\mathfrak{so}(2n+1)$ ($n\geq2),\mathfrak{sp}(2n)$ ($n\geq3),\mathfrak{so}(4n)$ ($n\geq2),E_7$:\
By using the fact that
- $\mathfrak{sp}(2)=\mathfrak{sl}(2)=\mathfrak{so}(3)$,
- $\mathfrak{sl}(2)\times \mathfrak{sl}(2)=\mathfrak{so}(4)\subset\mathfrak{so}(5)=\mathfrak{sp}(4)$,
- $\prod_1^7\mathfrak{sl}(2)\subset E_7$,
there exists a semisimple subalgebra $\mathfrak{h}$ of $\mathfrak{g}$ such that $\mathfrak{h}$ and $\mathfrak{g}$ have same rank and every simple factor of $\mathfrak{h}$ is $\mathfrak{sl}(2)$. Since $G$ is the commutator subgroup of $G_F(F)$, where $G_F$ is a split, simple algebraic group of type $\mathfrak{g}$, there exists a split, semisimple, algebraic subgroup $H_F\subset G_F$ of type $\mathfrak{h}$ (the Zariski closure in $G_F$ of the group generated by $X_\alpha(F)$ for all $\alpha$ belonging to the root subsystem of $\mathfrak{h}$ in the root system of $\mathfrak{g}$) such that $Z(G)\subset Z(H_F)$. Let $m$ be the rank of $G_F$ and $\pi:\widetilde{H}_F\cong\prod_1^m{\mathrm{SL}}_2\rightarrow H_F$ the universal cover. Since $Z(\widetilde{H}_F)$ surjects on $Z(H_F)$, $Z(\widetilde{H}_F)=\prod_1^mZ({\mathrm{SL}}_2)=\prod_1^mZ({\mathrm{SL}}_2(F))$, and $-I_2$ is a square, every element of $Z(G)$ is a square in $\pi(\prod_1^m{\mathrm{SL}}_2(F))=\pi(\prod_1^m[{\mathrm{SL}}_2(F),{\mathrm{SL}}_2(F)])\subset [H_F(F),H_F(F)]\subset [G_F(F),G_F(F)]=G$. We are done.\
Case $\mathfrak{g}=\mathfrak{so}(4n+2)$ ($n\geq2)$:\
Since $\mathfrak{h}=\mathfrak{so}(6)\times\prod_1^{n-1}\mathfrak{so}(4)$ is a maximal rank semisimple subalgebra of $\mathfrak{g}=\mathfrak{so}(4n+2)$, we find a split, semisimple algebraic subgroup $H_F\subset G_F$ of type $\mathfrak{h}$ such that $Z(G)\subset Z(H_F)$. Let $\pi:\widetilde{H}_F\cong {\mathrm{SL}}_4\times\prod_1^{n-1}{\mathrm{SL}}_3\rightarrow H_F$ be the universal cover. Since $Z(G_F(\bar{F}))={\mathbb{Z}}/4{\mathbb{Z}}\subset Z(H_F(\bar{F}))$ and $|Z({\mathrm{SL}}_3(\bar{F}))|$ is odd, $\pi$ is injective on ${\mathrm{SL}}_4$. Since $Z(G)$ is a subgroup of $H_F(F)$ of order a power of $2$, $Z(G)\subset \pi({\mathrm{SL}}_4)\cap H_F(F)\cong {\mathrm{SL}}_4(F)$ by injectivity. Since every element of ${\mathrm{SL}}_4(F)$ is a product of two squares from above and ${\mathrm{SL}}_4(F)=[{\mathrm{SL}}_4(F),{\mathrm{SL}}_4(F)]\subset [H_F(F),H_F(F)]\subset [G_F(F),G_F(F)]=G$, we are done.
Let $G$ be a Chevalley group of the form ${\mathrm{SL}}_2(F)$ for $F$ an infinite field. If either $m$ or $n$ is congruent to $1$, $2$, or $3$ (mod $4$), then every element of $G$ is of the form $x^m y^n$ since $$\begin{pmatrix} 0&1\\ -1&0\end{pmatrix}^k = \begin{pmatrix} -1&0\\ 0&-1\end{pmatrix}$$ if $k\equiv 2$ (mod $4$) (and of course odd powers preserve elements of $Z(G)$). However, this is not true in general.
If $G={\mathrm{SL}}_2(F)$ where $F$ is of characteristic zero and $[F(\zeta_8):F] = 4$ (e.g., $F={\mathbb{Q}}$), then $x^4 y^4$ does *not* represent $-I_2$.
If $A^4 = -B^4$ for elements $A,B\in {\mathrm{SL}}_2(F)$ and $\lambda^{\pm 1}$ and $\mu^{\pm 1}$ are respectively the eigenvalues of $A$ and $B$, then without loss of generality we may assume $\lambda/\mu = \zeta_8$, and ${\operatorname{Gal}}(F(\zeta_8)/F)\cong {\mathbb{Z}}/2{\mathbb{Z}}\times {\mathbb{Z}}/2{\mathbb{Z}}$ acts on $\{\lambda^{\pm 1}\}$ and $\{\mu^{\pm 1}\}$. One of the automorphisms $\zeta_8\mapsto\zeta_8^3$ or $\zeta_8\mapsto\zeta_8^7$ fixes exactly one of $\lambda$ and $\mu$, so either $\lambda^2$ or $\mu^2$ lies in $\{\pm i\}$. However, $\lambda$ and $\mu$ lie in quadratic extensions of $F$ and by hypothesis, every primitive $8$-th root of unity generates a degree $4$ extension of $F$.
[AGKSh]{}
Avni, Nir; Gelander, Tsachik; Kassabov, Martin; Shalev, Aner: Word values in $p$-adic and adelic groups, arXiv: 1303.1161v1.
Borel, Armand: On free subgroups of semisimple groups, *Enseign. Math.* (2) **29** (1983), no. 1-2, 151–164.
Ellers, E. W.; Gordeev, N. L.: Gauss decomposition with prescribed semisimple part in classical Chevalley groups, *Comm. Algebra* **22** (1994, no.14), 5935-5950.
Ellers, E. W.; Gordeev, N. L.: Gauss decomposition with prescribed semisimple part in classical Chevalley groups II: Exceptional cases, *Comm. Algebra* **23** (1995, no.8), 3085-3098.
Ellers, E. W.; Gordeev, N. L.: Gauss decomposition with prescribed semisimple part in classical Chevalley groups III: Finite twisted groups, *Comm. Algebra* **24** (1996, no.14), 4447-4475.
Elkasapy, Abdelrhman; Thom, Andreas: About Gotô’s method showing surjectivity of word maps, arXiv: 1207.5596.
Fulton, William; Harris, Joe: Representation Theory, Graduate Texts in Mathematics 129 (1st ed.), Springer-Verlag 1991.
Gotô, Morikuni: A theorem on compact semi-simple groups. *J. Math. Soc. Japan* **1**, (1949), 270–272.
Larsen, Michael: Word maps have large image, *Israel J. Math.* **139**, (2004), 149–156.
Larsen, Michael; Shalev, Aner: Word maps and Waring type problems, *J. Amer. Math. Soc.* **22** (2009), 437–466.
Larsen, Michael; Shalev, Aner: Characters of symmetric groups: sharp bounds and applications, *Invent. Math.* **174** (2008), no. 3, 645–687.
Larsen, Michael; Shalev, Aner: On the distribution of values of certain word maps, [*Trans. Amer. Math. Soc.*]{}, to appear.
Lev, Arieh: Products of cyclic conjugacy classes in the groups ${\mathrm{PSL}}(n,F)$, *Linear Algebra Appl.* **179** (1993), 59–83.
Larsen, Michael; Shalev, Aner; Tiep, Pham: The Waring problem for finite simple groups, *Annals of Math.* **174** (2011), 1885–1950.
M. Larsen, A. Shalev and Ph. Tiep, Waring problem for finite quasisimple groups, [*IMRN*]{} rns109 (2012), 26 pages.
Liebeck, Martin W.; O’Brien, E. A.; Shalev, Aner; Tiep, Pham Huu: The Ore Conjecture, [*J. Europ. Math. Soc.*]{} [**12**]{} (2010), 939–1008.
Segal, Dan: [*Words: notes on verbal width in groups*]{}, London Math. Soc. Lecture Note Series [**361**]{}, Cambridge University Press, Cambridge, 2009.
Serre, Jean-Pierre: Exemples de plongements des groupes ${\mathrm{PSL}}_2({\mathbb{F}}_p)$ dans des groupes de Lie simples, *Invent. Math.* **124**, (1996), 525–562.
Shalev, Aner: Word maps, conjugacy classes, and a non-commutative Waring-type theorem, [*Annals of Math.*]{} [**170**]{} (2009), 1383–1416.
Shalev, Aner: Some problems and results in the theory of word maps, [*Erdős Centennial*]{}, eds: Lovász et al., Bolyai Soc. Math. Studies [**25**]{} (2013), 611–649.
Steinberg, Robert: Lectures on Chevalley Groups, Yale University, 1967.
Thom, Andreas: Convergent sequences in discrete groups, *Canad. Math. Bull.* **56** (2013), no. 2, 424–433.
Thompson, Robert C.: Commutators in the special and general linear groups, *Trans. Amer. Math. Soc.* **101** (1961), 16–33.
Vaserstein, Leonid N.; Wheland, Ethel: Products of conjugacy classes of two by two matrices, *Linear Algebra Appl.* **230** (1995), 165–188.
[^1]: The first author was supported by ERC Advanced Grant no. 247034. The second author was partially supported by the Simons Foundation, the MSRI, the NSF, and BSF Grant no. 2008194. The third author was partially supported by ERC Advanced Grant no. 247034, ISF grant no. 1117/13, BSF Grant no. 2008194 and the Vinik Chair of Mathematics which he holds
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Two entangled electron spins, or qubits, are analyzed in terms of ordinary three-dimensional space geometric properties, as are the angles between their angular momenta. This formulation allows concurrence, a measure of quantum entanglement, to be expressed as expectation values of trigonometric functions of the azimuthal angle between the two angular momenta.'
address:
- 'Faculty of Mathematics and Physics, University of Ljubljana, Ljubljana, Slovenia'
- 'J. Stefan Institute, Ljubljana, Slovenia'
author:
- 'A. Ramšak'
title: Geometric analysis of entangled qubit pairs
---
As quantum entanglement does not have a direct classical analogue and as it is not an observable according to the usual rules of quantum mechanics, it is not obvious how it might be visualized beyond the standard formalism. Nevertheless, several approaches have so far been developed in order to intuitively analyse this problem in terms of some geometrical quantities. For example, the entanglement of two spin-1/2 particles can be studied in terms of the distances between states in high-dimensional manifolds [@kus], one can define entanglement geometrically as the distance between a given state and the nearest separable state [@wei03; @uyanik], relate entanglement to the geometrical structure of the quaternionic Hopf fibration [@levay], or analyse it by the approach of the operator trigonometry [@gustafson]. A review of separability criteria and entanglement measures is discussed geometrically in Ref. [@krenaknjiga]. These geometric approaches rely predominantly on abstract higher dimensional spaces not on usual $\Bbb{R}^3$.
Indeed, in general, quantities relating to spin-1/2 systems cannot be described in terms of classical variables, in particular not in terms of usual geometrical coordinates. Still, to some extent, geometric imagery is possible for a spin-1/2 state of a single electron which is customarily parametrized in terms of the Bloch or Poincar' e sphere, where polar and azimuthal angles can be interpreted as the Euler angles of a unity vector pointing along the spin direction. One possible generalisation of the Bloch sphere is the sphere model, which gives a geometrical view of entanglement in terms of constraint functions describing the behaviour of the state of one of spins if measurements are made on the other [@aerts].
The aim of the present paper is firstly to visualize the quantum state of a pair of spin-1/2 particles geometrically in ordinary three-dimensional space and, secondly, to express the quantum entanglement measure concurrence [@wootters] intuitively in terms of angles made by their angular momenta.
The azimuthal angle $\phi$ and the orbital angular momentum $L_z=-i \partial /\partial \phi$ are commonly regarded as conjugate variables connected by the commutator $[\phi,L_z]=i$. However, the use of this angle variable requires avareness because the position in space is a periodic function of $\phi$, which itself is not periodic [@carruthers63]. A similar problem appears in the treatment of the phase variable in a quantum mechanical harmonic oscillator. In particular, it was shown that Dirac’s [@dirac27] assumption of a Hermitian operator $\hat\phi$ conjugate to the number operator leads to contradictions [@susskind64]. A proper description of an angle or a phase variable requires that periodicity be taken into ccount. It turns out that a simple way of locating the azimuthal position is to give $\cos \phi$ and $\sin \phi$ instead of $\phi$ [@louisell63]. These operators are related by commutators $[\sin \phi,L_z]=i\cos\phi$ and $[\cos \phi,L_z]=-i\sin\phi$ [@carruthers63].
Here we investigate the role of azimuthal coordinates of two entangled electrons, or qubits. We apply this picture to the case of a qubit pair described by a mixed state and then quantify the quantum entanglement of the pair [@trije]. Qubits are not restricted to real spin of electrons and can represent any two state quantum system, for example, entangled photons pairs [@fotoni], flux qubits in superconducting rings [@makhlin], charge qubits in double quantum dots [@mravlje], flying qubits in quantum point contacts [@rejec], or two-qubit systems with different types of particles [@delfot].
First we consider two qubits in a state with a vanishing total spin projection, $$| \Psi\rangle=\cos { \vartheta \over 2} \left|\uparrow\downarrow\right\rangle+e^{i \varphi}\sin {\vartheta \over 2} \left|\downarrow\uparrow\right\rangle.
\label{psi2}$$ For convenience we use the notation where $\left|\uparrow\downarrow\right\rangle$, for example, represents the state where the first particle (qubit) is in the “up” state, [*i.e.*]{}, in the direction of the $z$-axis, and the second qubit is in the state “down”. For the Schmidt angle $\vartheta/2$ we assume $\vartheta\in[0,\pi]$. Spin operators are given by $\mathbf{S}_{1(2)}={1\over2}(\sigma_{1(2)x},\sigma_{1(2)y},\sigma_{1(2)z})$ for the first and the second qubit, respectively, and $\sigma$ are the Pauli matrices.
Geometric view in ordinary three-dimensional space of such an entangled qubit pair is gained by the analysis of relative angles between the angular momenta. There are two main alternatives. The first is the angle $\Phi$ made of the angular momenta as shown in Fig. \[Fig1\]. As the expectation value of the corresponding cosine we take $$\left\langle \cos\Phi \right\rangle=\langle\Psi| {\mathbf{S}_{1}\cdot \mathbf{S}_{2} \over \sqrt{\mathbf{S}_{1}^2\mathbf{S}_{2}^2}}|\Psi \rangle=
{1 \over 3}(2\sin\vartheta\cos\varphi-1),
\label{ffi}$$ with the variance $\Delta \cos\Phi=\frac{2}{3}\sqrt{1-\sin^2\vartheta\cos^2\varphi}$ vanishing for the singlet or the triplet state. This is a well known result, where the angular momenta can be visualized as antiparallel for the case of the singlet and not quite parallel for the triplet qubit pair state where $\Phi\approx 71^\circ$.
![Visualization of the angle $\Phi$ between the angular momenta of the first and the second qubit and the difference of azimuthal angles, made by the angular momenta projections onto the $xy$-plane, $\phi=\phi_1-\phi_2$.[]{data-label="Fig1"}](Fig1.eps){width="55mm"}
In this paper we concentrate on the properties of another geometric element, the azimuthal angle difference $\phi=\phi_1-\phi_2$ made of angular momenta projections onto the $xy$-plane, Fig. \[Fig1\]. Here $\phi_{1(2)}$ can be visualized as the azimuthal angle of the first or the second qubit, respectively. We define two [*operators*]{} corresponding to $\cos\phi$ and $\sin\phi$, $$\begin{aligned}
\widehat{\cos\phi}&\equiv&{{S_{1x} S_{2x}+S_{1y} S_{2y}} \over
\sqrt{(S_{1x}^2 +S_{1y}^2)(S_{2x}^2 +S_{2y}^2)} }, \\
\widehat{\sin\phi}&\equiv&{{S_{1y} S_{2x}-S_{1x} S_{2y}} \over
\sqrt{(S_{1x}^2 +S_{1y}^2)(S_{2x}^2 +S_{2y}^2)} }.
\label{cossin}\end{aligned}$$ These operators obey commutator relations analogous to the orbital momentum case [@louisell63; @carruthers63], $$\begin{aligned}
\lbrack{\widehat{\sin\phi},\delta S_z}\rbrack&=&i \widehat{\cos\phi}, \label{comm1}\\
\lbrack\widehat{\cos\phi},\delta S_z\rbrack&=&-i \widehat{\sin\phi},
\label{comm2}\end{aligned}$$ where the relative angular momentum $\delta S_z=S_{1z}-S_{2z}$ plays the role of $L_z$ in the case of orbital motion, which supports the view that $\phi$ corresponds to the relative azimuthal angle.
There are two pairs of eigenstates and eigenvalues for these operators, $$\begin{aligned}
\widehat{\cos\phi}\;( \left|\uparrow\downarrow\right\rangle\pm\left|\downarrow\uparrow\right\rangle)&=&\pm (\left|\uparrow\downarrow\right\rangle\pm\left|\downarrow\uparrow\right\rangle), \label{eigen1}\\
\widehat{\sin\phi}\;(\left|\uparrow\downarrow\right\rangle\pm i \left|\downarrow\uparrow\right\rangle)&=&\pm (\left|\uparrow\downarrow\right\rangle\pm i \left|\downarrow\uparrow\right\rangle).
\label{eigen2}\end{aligned}$$ For the triplet and singlet state the projections of angular momenta make angles $0$ and $\pi$, respectively, and the momenta for eigenstates of $\widehat{\sin\phi}$ are perpendicular.
Note that these results imply that there are only four possible values for angle $\phi$, contrary to the naive intuition that $\varphi$ in equation (\[psi2\]) corresponds to $\phi$, at least for the perfectly entangled qubit pairs, i.e., states with $\vartheta=\frac{\pi}{2}$. Such an interpretation follows from a heuristic argument: consider a triplet state $\left|\uparrow\downarrow\right\rangle+\left|\downarrow\uparrow\right\rangle$, where angle $\phi$ is zero. Then by the unitary transformation $\exp(-i \varphi S_{1z})$ rotate the first spin around the $z$-axis for angle $\varphi$. The outcome is the state proportional to $|\Psi\rangle$, therefore one would expect that the angular momenta then make angle $\phi=\varphi$. Below will be given alternative arguments that this view indeed can be considered correct.
As discussed for the case of orbital motion, a proper definition of angle operator $\hat\phi$ is impossible due to the problem of periodicity [@dirac27; @carruthers63; @note1]. However, one can define a pair of operators as in ordinary trigonometry, $$\begin{aligned}
\hat \phi_c&\equiv&\arccos \widehat{\cos\phi}=\frac{\pi}{2}(1-\widehat{\cos\phi}), \\
\hat \phi_s&\equiv&\arcsin\widehat{\sin\phi}=\frac{\pi}{2}\widehat{\sin\phi}.
\label{koti}\end{aligned}$$ The operator $\hat \phi_c$ is actually the operator for the “absolute value of the angle”, $\hat \phi_c\equiv\widehat{|\phi |}$, due to the periodicity properties of trigonometric functions defined on the interval $[0,\pi]$. $\hat \phi_c$ does not represent any new information as it can be by the Taylor expansion and the known properties of the Pauli matrices reexpressed in terms of $\widehat{\cos\phi}$. The operator $\hat \phi_s$ is expressed by $\widehat{\sin\phi}$ and also can not be considered as a proper angle operator, because the inverse sine maps onto $[-\frac{\pi}{2},\frac{\pi}{2}]$. In total there are four possible eigenvalues for the angle operators: $\phi=0$, $\pi$ and $\pm\frac{\pi}{2}$ for $\hat \phi_c$ and $\hat \phi_s$, respectively, as expected for the four eigenvectors in equations (\[eigen1\]) and (\[eigen2\]).
The interpretation of the expectation values for the angle operators would require, due to the periodicity property, the resolving of a double mapping, which can not be done in a unique way. It is sufficient to study the cosine and sine operators and a direct evaluation reveals $$\begin{aligned}
\langle\Psi|\widehat{\cos\phi}|\Psi \rangle& =&C \cos \varphi, \label{c}\\
\langle\Psi|\widehat{\sin\phi}|\Psi \rangle &=&C \sin \varphi,\end{aligned}$$ where $C=\sin\vartheta$. For perfectly entangled qubit pairs these two equations simplify to intuit forms, $\left\langle\widehat{\cos\phi}\right\rangle=\cos \varphi$ and $\left\langle\widehat{\sin\phi}\right\rangle=\sin \varphi$, which in this formulation supports the anticipated view that $\phi=\varphi$ for fully entangled states. Identical results were found also in a recent analysis of quantum entanglement in the de Broglie-Bohm interpretation of quantum mechanics [@geometrical].
Quantum entanglement can be quantified by various measures [@1995], one of them being the entanglement of formation [@bennett] which is the asymptotic conversion rate to maximally entangled states from an ensemble of copies of a non-maximally entangled state [@vedral]. By the Wootters formula [@wootters] is the entanglement of formation simply related to an associated quantity, the concurrence $C$, introduced in equation (\[c\]).
Entanglement of two qubits can therefore be visualized in terms of functions of a simple geometric quantity – the difference of azimuthal angles of the angular momenta – by a simple relation $$C=\sqrt{
\left\langle\widehat{\cos\phi} \right\rangle
^2 +\left\langle\widehat{\sin\phi }
\right\rangle^2 }.
\label{cc}$$ This expression offers a clear interpretation that the concurrence is small if the angles are spread randomly, leading to small cosine and sine averages. The concurrence is close to unity if angles are packed at some common angle difference $\varphi$ such that the sum of squared averages of cosine and sine adds close to unity. As a measure of angle spreading can serve the variances for cosine and sine operators, to $C$ related by the expression $$(\Delta \widehat{\cos\phi})^2+(\Delta \widehat{\sin\phi})^2=2-C^2.
\label{dde}$$ Larger concurrence thus signals lower fluctuations of cosine and sine. However, even for perfectly entangled states with $C=1$ the sum of variances (\[dde\]) is non-zero, because $|\Psi\rangle$ is never an eigenstate of cosine and sine operators simultaneously [@cossin3].
The state $|\Psi\rangle$ is related to the state $$|\widetilde{ \Psi}\rangle=\cos { \tilde\vartheta \over 2}\left |\uparrow\uparrow\right\rangle+e^{i \tilde\varphi}\sin {\tilde\vartheta \over 2} \left|\downarrow\downarrow\right\rangle
\label{tpsi2}$$ where the spin of the second qubit is reversed by the unitary transformation $\exp(-i \pi S_{2y})$, and $\vartheta \to \tilde\vartheta$, $\varphi+\pi \to \tilde\varphi$. The formalism based on cosine and sine operators can be applied in a similar manner as for the case of $\left |\Psi\right\rangle$. The main distinction is, that for states (\[tpsi2\]) the sum of azimuthal angles, $\tilde \phi=\phi_1+\phi_2$ is important, not the difference $\phi$. The corresponding trigonometric operators $$\begin{aligned}
\widehat{\cos\tilde\phi}&\equiv&{{S_{1x} S_{2x}-S_{1y} S_{2y}} \over
\sqrt{(S_{1x}^2 +S_{1y}^2)(S_{2x}^2 +S_{2y}^2)} }, \\
\widehat{\sin\tilde\phi}&\equiv&{{S_{1y} S_{2x}+S_{1x} S_{2y}} \over
\sqrt{(S_{1x}^2 +S_{1y}^2)(S_{2x}^2 +S_{2y}^2)} }
\label{cossin2}\end{aligned}$$ together with $S_z=S_{1z}+S_{2z}$ exhibit similar commutation relations to Eqs. (\[comm1\]), (\[comm2\]). All results valid for $\phi$ can be mutatis mutandis applied to $\tilde\phi$ and are not shown here. We present here only the concurrence $\widetilde C=\sin\tilde\vartheta$, $$\widetilde{C}=\sqrt{
\left\langle\widehat{\cos\tilde \phi}\right \rangle
^2 +\left\langle\widehat{\sin\tilde \phi }
\right\rangle^2 }.
\label{tcc}$$ It should be noted that $C=0$ for $|\tilde\Psi\rangle$ and $\widetilde C=0$ for $|\Psi\rangle$.
Finally, we can express the concurrence $C_\rho$ for a system of two qubits in a mixed state given by the density matrix $$\rho=\sum_i p_i |\Psi_i\rangle\langle\Psi_i|+\sum_j\tilde p_j |\widetilde\Psi_j\rangle\langle\widetilde\Psi_j|,$$ which represents systems with a conserved square of total spin projection, since $[\rho,S_z^2]=0$. For this class of systems the concurrence is given by [@amico08; @ramsak1] $$C_\rho =
\max\left(0,C-2\sqrt{\langle
P_{1}^{\uparrow}P_{2}^{\uparrow}\rangle\langle
P_{1}^{\downarrow}P_{2}^{\downarrow}\rangle}, \widetilde C-2\sqrt{\langle
P_{1}^{\uparrow}P_{2}^{\downarrow}\rangle\langle
P_{1}^{\downarrow}P_{2}^{\uparrow}\rangle}\right),\label{ccc}$$ where $P^m_{1(2)}=\frac{1}{2}\pm S_{1(2)z}$ is the projector onto the spin state $m=\uparrow, \downarrow$, respectively. The expectation values are evaluated as is customary for mixed states, for example, $\left\langle \widehat{\cos\phi}\right\rangle=\mathrm{Tr} \rho \widehat{\cos\phi}$ or $\langle P_{1}^{\downarrow}P_{2}^{\downarrow}\rangle=\sum_j \tilde p_j \sin^2 \frac{\tilde\vartheta_j}{2}$.
To summarize, we have analyzed a pair of entangled qubits in terms of observables expressed by trigonometric functions of relative angles made of spin vectors of two qubits. Following the approach developed for the analysis of angles and angular momenta of orbital motion we applied analogous operators corresponding to cosine and sine of azimuthal angles difference $\phi=\phi_1-\phi_2$ for the case of spin-1/2. In terms of such operators we constructed appropriate azimuthal angle operators, which reveals that for states with vanishing total spin projection there are four eigenstates of angle operators, with eigenvalues $\phi=0, \pm\frac{\pi}{2}, \pi$. The analysis of the expectation values of trigonometric operators gives support to an argument that the phase factor $\varphi$ in the wave function (\[psi2\]) corresponds to $\phi$.
Average cosine and sine of this angle are plainly related to the concurrence, a measure of the degree of quantum entanglement. In particular, the concurrence is exactly given by $C=\left\langle\widehat{\cos\phi}\right\rangle$, for $\varphi=0$, and is also related to the corresponding variance $\Delta \widehat{\cos\phi}=\sqrt{1-C^2}$. A higher degree of entanglement can thus be visualized as a highly correlated distribution of angular momenta making azimuthal angles close to $\varphi$ and with suppressed fluctuations with progressively increasing entanglement. An analogous analysis was performed for the space spanned by the basis vectors $\left|\uparrow\uparrow\right\rangle$ and $\left|\downarrow\downarrow\right\rangle$. Here the relevant geometric quantity is the sum of azimuthal angles, $\phi_1+\phi_2$, and the concurrence is given in terms of the average cosine of this angle.
The analysis of qubits in pure states was generalized to systems described by mixed states and as the final result the concurrence is expressed in terms of trigonometric operators for a rather general class of systems. Although in any explicit quantification of the entanglement the expressions reduce to the manipulation of ordinary Pauli spin operators, we believe that the present approach of the application of geometrical quantities offers a new insight into the phenomenon of quantum entanglement, which is due to the lack of a direct classical analogue not recognized in standard formal approaches.
The author thanks T. Rejec, J. H. Jefferson, I. Sega, and T. Huljev [Č]{}ade[ž]{} for discussions and he acknowledges the support from the Slovenian Research Agency under Contracts No. J1-0747 and P1-0044.
[99]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
For a fixed bipartite graph $H$ and given $\alpha\in
(0,1)$ we determine the threshold $T_H(\alpha)$ which guarantees that any $n$-vertex graph with at least $T_H(\alpha){n \choose 2}$ edges contains $(1-o(1))\frac\alpha{{v(H)}}n$ vertex-disjoint copies of $H$. [In the proof we use a variant of a technique developed by Komlós [\[Combinatorica 20 (2000), 203–218]{}\]]{}.
address:
- 'Faculty of Automatic Control and Computers, Politehnica University of Bucharest, Splaiul Independenei 313, sector 6, Cod 060042, Bucharest, Romania'
- 'Department of Applied Mathematics, Faculty of Mathematics and Physics, Charles University, Malostranské náměstí 25, 118 00, Prague, Czech Republic DIMAP, University of Warwick, Coventry, CV4 7AL, United Kingdom.'
author:
- Codru Grosu
- Jan Hladký
bibliography:
- 'bibl.bib'
title: The extremal function for partial bipartite tilings
---
Introduction
============
The Turán Theorem [@Tur], [one of the most important results]{} in Extremal Graph Theory, gives a sharp threshold, denoted $\ex(n,K_r)$, for the maximum number of edges of an $n$-vertex graph with no copy of $K_r$. Even though the Turán Theorem applies to any pair of values $n$ and $r$, the interesting instances are rather those when $n$ is large compared to $r$. Erdős and Stone [@ErdosStone1946] extended the result by determining the asymptotic behaviour of the function $\ex(n,H)$ for a fixed non-bipartite graph $H$. The same problem in the case that $H$ is a fixed bipartite graph is — despite considerable effort — wide open for most graphs $H$. [This is known as the *Zarankiewicz problem*.]{} Let us recall that when $H$ has colour classes of sizes $s$ and $t$, $s\le t$, then the Kövari-Sós-Turán Theorem [@KovariST] asserts that $$\label{eq:KTS}
\ex(n,H)\le O(n^{2-1/s})=o(n^2)\;.$$ On the other hand, a standard random graph argument gives that $\ex(n,K_{s,t})\ge
\Omega(n^{2-(s+t-2)/(st-1)})$.
It is natural to extend the above *existential questions* to *tiling questions*. In such a setting one asks for the maximum number of edges of an $n$-vertex graph which does not contain $\ell$ vertex-disjoint copies of a graph $H$. This quantity [is denoted by]{} $\ex(n,\ell\times H)$. Erdős and Gallai [@ErdGall59] gave a complete solution to the problem in the case when $H=K_2$.
\[thm:ErdGall\] Suppose that $\ell\le n/2$. Then $$\ex(n,\ell\times
K_2)=\max\left\{(\ell-1)(n-\ell+1)+{\ell-1\choose
2},{2\ell-1\choose 2}\right\}\;.$$
Given $n,x\in\mathbb N$, $x\le n$, we define [two graphs $M_{n,x}$ and $L_{n,x}$ as follows. The]{} graph $M_{n,x}$ [is]{} an $n$-vertex graph whose vertex set is split into sets $A$ and $B$, $|A|=x, |B|=n-x$, $A$ induces a clique, $B$ induces an independent set, and $M_{n,x}[A,B]\simeq K_{x,n-x}$. The graph $L_{n,x}$ is [the complement of $M_{n,n-x}$,]{} [i.e., it is]{} an $n$-vertex graph whose edges induce a clique of order $x$. Obviously, $e(M_{n,\ell-1})=(\ell-1)(n-\ell+1)+{\ell-1\choose 2}$, and $e(L_{n,2\ell-1})={2\ell-1\choose 2}$. Moreover, it is easy to check that there are [no $\ell$ vertex-]{}disjoint edges in either of the graphs $M_{n,\ell-1}$, $L_{n,2\ell-1}$. Therefore, when $\ell<\frac25 n{ + }O(1)$, the graph $M_{n,\ell-1}$ is (the unique) graph showing that $\ex(n,\ell\times
K_2)\ge (\ell-1)(n-\ell+1)+{\ell-1\choose
2}$. The graph $L_{n,2\ell-1}$ is the unique extremal graph for the problem otherwise.
Moon [@Moon68Tiling] started the investigation of $\ex(n,\ell\times K_r)$. Allen, Böttcher, Hladký, and Piguet [@AllBottHlaPig] only recently determined the behaviour of $\ex(n,\ell\times K_r)$ for the whole range of $\ell$ in the case $r=3$, and they made a substantial progress for larger values of $r$. Simonovits [@Simonovits99Turan] determined the value $\ex(n,\ell
\times H)$ for a non-bipartite graph $H$, fixed value of $\ell$ and large $n$.
[An equally important]{} density parameter which can be considered in the context of tiling questions is the minimum degree of the host graph. That is, we ask what is the largest possible minimum degree of an $n$-vertex graph which does not contain $\ell$ vertex-disjoint copies of $H$. In the case $H=K_r$[,]{} the precise answer is given by the Hajnal-Szemerédi Theorem[^1] [@hajnal70:_proof_p]. An asymptotic threshold for a general fixed graph $H$ was determined by Komlós [@komlosTT]. [In this case, the threshold depends]{} [ on a parameter which Komlós calls the *critical chromatic number*. The critical chromatic number of $H$ is a real between $\chi(H)-1$ and $\chi(H)$. Roughly speaking, graphs $H$ which possess a coloring with $\chi(H)$ colors with one of the color classes small, have the critical chromatic number close to $\chi(H)-1$. On the other hand, graphs $H$ which have only approximately balanced $\chi(H)$-colorings have the critical chromatic number close to $\chi(H)$. There is a natural way how to state our main result, Theorem \[main\_result\], using the critical chromatic number. However, we chose not to as in the bipartite setting of Theorem \[main\_result\] it is possible to give a self-contained formula for the problem. Let us also note that Komlós’ result [@komlosTT] gives an asymptotic min-degree threshold even in the case when $H$ is bipartite. In this case the near-extremal graphs for the problem are complete bipartite graphs.]{}
In the present paper we [use a variation of the technique developed by Komlós to]{} determine [the]{} asymptotic behaviour of the function $\ex(n,\ell\times H)$ for a fixed bipartite graph [$H$]{}. Let $H$ be an arbitrary bipartite graph. Suppose that $b:V(H)\rightarrow [2]$ is a proper coloring of $H$ which minimizes $|b^{-1}(1)|$. We define quantities $s(H):=|b^{-1}(1)|$, $t(H):=|b^{-1}(2)|$. Obviously, $s(H)\le t(H)$, and $s(H)+t(H)={v(H)}$. Furthermore, we define $V_1(H):=b^{-1}(1)$ and $V_2(H):=b^{-1}(2)$. The sets $V_1(H)$ and $V_2(H)$ are uniquely defined provided that $H$ does not contain a balanced bipartite graph as [one of its components; in this other case we fix a coloring $b$ satisfying the above conditions and use it to define uniquely $V_1(H)$ and $V_2(H)$.]{}
Given $s,t \in \mathbb N$, we define a function $T_{s,t}:(0,1)\rightarrow (0,1)$ by setting $$\label{thresh_func}
T_{s,t}(\alpha) :=
\max\left\{
\frac{2s\alpha}{s+t}\left(1 -
\frac{s\alpha}{2(s+t)}\right),
\alpha^2 \right\}\;,$$ for $\alpha\in(0,1)$. Note that $T_{s',t'}=T_{s,t}$ when $s'=ks$ and $t'=kt$. Also, note that $$\label{eq:TssVersusErdosGallai}
T_{s,s}(\alpha){n\choose
2}=\ex\left(n,\frac{\alpha n}2\times
K_2\right)+o(n^2)\;,$$ [and, in general, [for $s\le t$, the number]{} $T_{s,t}(\alpha)\binom{n}{2}$ is asymptotically the maximum between the number of edges of $M_{n, \frac{\alpha s}{s+t}n}$ and $L_{n, \alpha n}$.]{}
Our main result is the following.
\[main\_result\] Suppose that $H$ is a bipartite graph with no isolated vertices, $s:=s(H), t:=t(H)$. Let $\alpha \in
(0,1)$ and $\eps > 0$. Then there exists an $n_0 = n_0(s,t,\alpha,\eps)$ such that for any $n \geq n_0$, any graph $G$ with $n$ vertices and at least $T_{s,t}(\alpha){n \choose 2}$ edges contains more than $(1- \eps)\tfrac{\alpha}{s+t}n$ vertex-disjoint copies of the graph $H$.
[Let $H, s$ and $t$ be as in the hypothesis of the theorem, $\epsilon' > 0$ and $\beta \in (0, 1)$. Then we may find an $\alpha > \beta(s+t)$ and an $\epsilon < \epsilon'$ sufficiently small, such that for $n$ large enough, by Theorem \[main\_result\], any graph $G$ with $n$ vertices and at least $T_{s, t}(\alpha)\binom{n}{2} < (T_{s,t}(\beta(s+t))+\epsilon')\binom{n}{2}$ edges contains at least $\beta n$ vertex-disjoint copies of $H$. Hence $\ex(n,\beta n\times H)\le T_{s,t}(\beta(s+t)){n\choose 2}+\eps' n^2$.]{} This asymptotically matches the lower bound which comes — as in Theorem \[thm:ErdGall\] — from graphs $M_{n,\beta
{s}n-1}$ and $L_{n,\beta{(s+t)}n-1}$. [Indeed, neither of these graphs contains $\beta n$ vertex-disjoint copies of $H$, as any such copy would require at least $s$ vertices in the clique subgraph of $M_{n, \beta sn -1}$, and at least $s+t=v(H)$ non-isolated vertices in $L_{n, \beta(s+t)n-1}$[, respectively]{}.]{} Note however that for most values of [$H$,]{} the graphs $M_{n,\beta
{s}n-1}$ and $L_{n,\beta{(s+t)}n-1}$ are not extremal for the problem. For example, we can replace the independent set in the graph $L_{n,\beta{(s+t)}n-1}$ by any $H$-free graph. [This links us to the Zarankiewicz problem]{}, and suggests that an [exact]{} result is not within the reach of current techniques.
The assumption on $H$ to contain no isolated vertices in Theorem \[main\_result\] is made just for the sake of compactness of the statement. Indeed, let $H'$ be obtained from $H$ by removing all the isolated vertices. Then there is a simple relation [between]{} the sizes of optimal coverings by vertex disjoint copies of $H$ and $H'$ in an $n$-vertex graph $G$. Let $x$ and $x'$ be the number of vertices covered by a maximum family of vertex-disjoint copies of $H$ and $H'$ in $G$, respectively. We have that $$x=\min\left\{{v(H)}\left\lfloor\frac{n}{{v(
H)}}\right\rfloor,\frac{x'{v(H)}}{{v(H')}}\right\}\;.$$
[One can attempt to obtain an analogue of Theorem \[main\_result\] for graphs with higher chromatic number. This however appears to be substantially more difficult. To indicate the difficulty, let us recall that there are two types ($M_{n,x}$ and $L_{n,x}$) of extremal graphs for the $H$-tiling [problem]{} for bipartite $H$. The graphs $M_{n,x}$ and $L_{n,x}$ have a *block structure*, i.e., [their vertex set can be partitioned into *blocks* (two, in this case), such that any two vertices from the same block have almost [the same neighborhoods]{}. These two graphs]{} appear even in the simplest case [of]{} $H=K_2$ (cf. Theorem \[thm:ErdGall\]). However, when $H$ is not balanced, [if we let $\alpha$ go from $0$ to $1$, the transition between the two extremal structures which determine the threshold function occurs]{} at a different time in the evolution. On the other hand, there are five types of extremal graphs for the problem of determining $\ex(n,\ell\times K_3)$ as shown in [@AllBottHlaPig]. All the five types have a block structure. It is plausible that when $H$ is a general $3$-colorable graph, [the same five types of extremal graphs determine the threshold function for $H$-tilings.]{} However, the transitions between them occur at different times and the block [sizes]{} depend on various structural properties of $H$. In particular, we have indications that the critical chromatic number alone does not determine $\ex(n,\alpha n\times H)$ in this situation. ]{}
If $\mathcal{F}$ is a family of graphs, and $G$ is a graph, an *$\mathcal{F}$-tiling* in $G$ is a set of vertex-disjoint subgraphs of $G$, each of them isomorphic to a graph in $\mathcal{F}$. If $\mathcal{F}=\{H\}$ then we simply say $H$-tiling. $V(F)$ denotes the vertices of $G$ covered by an $\mathcal{F}$-tiling $F$, and $|F| = |V(F)|$ is the *size* of the tiling $F$. [If $F$ is a collection of bipartite graphs, we let $V_1(F) = \bigcup_{H \in F}V_1(H)$ and $V_2(F)=\bigcup_{H \in F}V_2(H)$.]{} For $n \in \mathbb N$, we write $[n]$ to denote the set $\{1,2,\ldots,n\}$.
Tools for the proof of the main result
======================================
Our main tool is Szemerédi’s [R]{}egularity [L]{}emma (see [@KS96; @KuhnOsthusSurv] for surveys). To state it we need some more notation.
Let $G = (V,E)$ be an $n$-vertex graph. If $A,B$ are disjoint nonempty subsets of $V(G)$, the *density* of the pair $(A,B)$ is $d(A,B) = e(A,B)/(|A||B|)$. We say that $(A,B)$ is an *$\eps$-regular pair* if $
|d(X,Y) - d(A,B)| < \eps
$ for every $X
\subset A, |X| > \eps|A|$ and $Y \subset B, |Y| >
\eps|B|$.
The following statement asserts that large subgraphs of regular pairs are also regular.
\[lem:slicing\] Let $(A,B)$ be an $\eps$-regular pair with density $d$, and let $A' \subset A, |A'| \geq
\alpha|A|, B' \subset B, |B'| \geq \alpha|B|$, $\alpha\ge \eps$. Then $(A',B')$ is an $\eps'$-regular pair with $\eps' = \max\{\eps/\alpha, 2\eps\}$, and for its density $d'$ we have $|d'-d| < \eps$.
Let $\eps > 0$ and $d \in [0,1]$. An $(\eps,
d)$-*regular partition* of $G$ with *reduced graph* $R = (V',E')$ is a partition $V_0
\dot\cup V_1 \dot\cup \ldots \dot\cup V_k$ of $V$ with $|V_0| \leq \eps n$, $|V_i| = |V_j|$ for any $1 \leq i <
j \leq k$, $V(R) = \{V_1, V_2, \ldots, V_k\}$, such that $(V_i,V_j)$ is an $\eps$-regular pair in $G$ of density greater than $d$ whenever $V_iV_j \in E(R)$, and the subgraph $G' \subset G$ induced by the $\eps$-regular pairs corresponding to the edges of $R$ has more than $e(G) - (d+3\eps)n^2/2$ edges. In this case, we also say that $G$ has an $(\eps, d)$-reduced graph $R$, and call the sets $V_i, 1 \leq i \leq k$, the *clusters* of $G$.
The following lemma is a consequence of the so-called degree version of the Regularity Lemma [@KS96 Theorem 1.10].
\[lem:reg\] For every $\eps > 0$ and $m \in \mathbb N$ there is an $M=M(\eps, m)$ such that, if $G$ is any graph with more than $M$ vertices and $d \in [0,1]$ is any real number, then $G$ has an $(\eps, d)$-reduced graph $R$ on $k$ vertices, with $ m \leq k \leq M$.
Given four positive numbers $a,b,x,y$ we say that the pair $a,b$ *dominates* the pair $x,y$, if $\max\{x,y\}/\min\{x,y\}\ge \max\{a,b\}/\min\{a,b\}$. The following easy lemma states that $K_{a,b}$ has an almost perfect $K_{s,t}$-tiling provided that $a,b$ dominates $s,t$.
\[lem:til\] For any $s,t\in\mathbb N$ there exists a constant $C$ such that the following holds. Suppose that the pair $a,b\in\mathbb N$ dominates $s,t$. Then the graph $K_{a,b}$ contains a $K_{s,t}$-tiling containing all but at most $C$ vertices of $K_{a,b}$.
If $s=t$ then necessarily $a=b$. There obviously exists a $K_{s,t}$-tiling containing all but at most $C:=2(s-1)$ vertices of $K_{a,b}$.
With no loss of generality, we may suppose that $a\le b$ and $s< t$. Then $as\le bt$ and $bs\le at$. A tiling with $\lfloor (bt-as)/(t^2-s^2)\rfloor$ copies of $K_{s,t}$ with the $s$-part of the $K_{s,t}$ placed in the $a$-part of the $K_{a,b}$ and $\lfloor
(at-bs)/(t^2-s^2)\rfloor$ copies placed the other way misses at most $C:=2(s+t-1)$ vertices of $K_{a,b}$.
The next lemmas, versions of the Blow-up Lemma [@KSS_bl], assert that regular pairs have almost as good tiling properties as complete bipartite graphs.
\[lem:key\] For every $d>0,\gamma \in (0,1)$ and any two graphs $R$ and $H$, there is an $\eps = \eps(H,d,\gamma) > 0$ such that the following holds for all positive integers $s$. Let $R_s$ be the graph obtained from $R$ by replacing every vertex of $R$ by $s$ vertices, and every edge of $R$ by a complete bipartite graph between the corresponding $s$-sets. Let $G$ be [any]{} graph obtained similarly from $R$ by replacing [every vertex of $R$ by $s$ vertices, and every edge of $R$ with an $\eps$-regular pair]{} of density at least $d$. If $R_s$ contains an $H$-tiling of size at least $\gamma {v(R_s)}$ then so does $G$.
\[lem:tiling2\] For every bipartite graph $H$ and every $\gamma,d>0$ there exists an $\eps = \eps(H,d,\gamma) > 0$ such that the following holds. Suppose that there is an $H$-tiling in $K_{a,b}$ of size $x$. Let $(A,B)$ be an arbitrary $\eps$-regular pair with density at least $d$, $|A|=a$, $|B|=b$. Then the pair $(A,B)$ contains an $H$-tiling of size at least $x-\gamma (a+b)$.
Finally, let us state a straightforward corollary of the König Matching Theorem.
\[bip\_matching\] Let $G = (A \dot\cup B, E)$ be a bipartite graph with color classes $A$ and $B$. If $G$ has no matching with $l+1$ edges, then $e(G) \leq l \max\{|A|, |B|\}$.
The proof
=========
In this section, we first state and prove the main technical result, Lemma \[main\_lemma\]. Then, we show how it implies Theorem \[main\_result\].
For $s,t \in \mathbb N$, we set $\mathcal{F}_1 :=
\{K_{s,t},K_{s,t-1},K_2\}$ and $\mathcal{F}_2 :=
\{K_{st,t^2},K_{st-1,(t-1)t},K_{st,(t-1)t},K_2\}$. Let us note that when $s < t$, the sizes of the two color classes of any graph from $\mathcal{F}^* :=
\mathcal{F}_1 \cup \mathcal{F}_2$ dominate $s$ and $t$.
[Let $F$ be a $K_{s,t}$-tiling in a graph $G$, $s<t$. Suppose $E_0$ and $E_1$ are matchings in $G[V(G)-V(F), V_1(F)]$ and $G[V_2(F)]$, respectively, such that each copy $K$ of $K_{s, t}$ in $F$ has at most one vertex matched by $E_0$ and at most one vertex matched by $E_1$. If any $K \in F$ which has a vertex matched by $E_0$, also has a vertex matched by $E_1$, then we call the pair $(E_0, E_1)$ an *$F$-augmentation*. Note that in this case $E_0$ and $E_1$ are vertex disjoint, as $V_1(F) \cap V_2(F)=\emptyset$.]{} The main step in our proof of Theorem \[main\_result\] is the following lemma.
\[main\_lemma\] Let $t > s \geq 1, \alpha \in (0,1)$ and $\eps > 0$. [Then there exists an $\eps' = \eps'(s,t,\alpha,\eps) > 0$ and an $h = h(s,t,\alpha,\eps) > 0$ such that the following holds.]{} Suppose $G$ is an $n$-vertex graph with $n \geq h$ and $e(G) \geq T_{s,t}(\alpha){n \choose 2}$, and $F$ is a $K_{s,t}$-tiling in $G$ [of maximum size with]{} $|F| \leq (1 - \eps)\alpha n$. Then one of the following is true:
1. there exists an $\mathcal{F}_1$-tiling $F'$ in $G$ with $|F'| \geq |F| + \eps'n$, or
2. there exists an $F$-augmentation ${(}E_0, E_1{)}$ such that $E_0$ contains at least $\eps'n$ edges.
Set $$\label{eps_def}
\eps' :=\frac14 \min\left\{\frac{\eps \alpha^2}{3t+1},
\frac{\eps s \alpha}{(3t+1)(s+t)}\right\}\;,$$ and let $h$ be sufficiently large.
Suppose for a contradiction that the assertions of the lemma are not true.
Set $L := V(G) - V(F)$ and $m := |L|$. Let $\mathcal{C} :=
\{V_1(K):K \in F\}, \mathcal{D} := \{V_2(K): K \in F\}$ and $C := \bigcup \mathcal{C}, D:=\bigcup \mathcal{D}$. We call members of $\mathcal C$ *lilliputs* while members of $\mathcal D$ are *giants*. We say that giant $V_2(K)$ $(K\in F)$ is *coupled* with lilliput $V_1(K)$.
As $F$ is a maximum [size]{} $K_{s,t}$-tiling in $G$, by we have that $$\label{eq:H0}
e(G[L]) = o(n^2)\;.$$
Let $r$ be the number of copies of $K_{s,t}$ in $F$. Then $r \leq (1-\eps)\alpha n / (s+t)$. Moreover, we have $$\label{eq:DefM}
m=n-(s+t)r\;.$$
Let us define an auxiliary graph $H=(V',E')$ as follows. The vertex-set of $H$ is $V' := \mathcal{C} \cup
\mathcal{D} \cup L$. For any $x \in L$ and $K \in F$ the edge $xV_1(K)$ belongs to $E'$ iff $\neighbor_G(x) \cap V_1(K) \neq
\emptyset$. Similarly, the edge $xV_2(K)$ belongs to $E'$ iff $\neighbor_G(x) \cap V_2(K) \neq \emptyset$. Finally, for any distinct $K,K' \in F$ the edge $V_2(K)V_2(K')$ belongs to $E'$ iff $E_G(V_2(K),V_2(K')) \neq \emptyset$. The vertices $L$ and the vertices $\mathcal C$ induce two independent sets in $H$.
As (i) does not hold, $H[L,\mathcal{D}]$ does not contain a matching with at least $\eps'n$ edges. It follows from Fact \[bip\_matching\] that $$\label{eq:H1}
e_G(L, D) \leq \eps'nt\max\{m,r\} {\leq t\eps'n^2}\;.$$
Let $M$ be a maximum matching in $H[L, \mathcal{C}]$ with $l$ edges. Obviously, $l\le r$. By Fact \[bip\_matching\], we have that $$\label{eq:H2}
e_G(L, C) \leq ls\max\{m,r\}\;.$$ Let $\mathcal C'\subseteq \mathcal C$ be the lilliputs matched by $M$. We write $\mathcal D'\subseteq\mathcal D$ for the giants coupled with $\mathcal C'$. Set $D'=\bigcup
\mathcal D'$.
Suppose for a moment that $H[\mathcal{D}'] \cup
H[\mathcal{D}',\mathcal{D}-\mathcal{D}']$ contains a matching $T$ with at least $\eps'n$ edges. Let $\mathcal{D}''$ be the giants [in $\mathcal{D}'$]{} matched by $T$ and $M'$ the set of edges in $M$ matching the lilliputs coupled with $\mathcal{D}''$. Then $M'$ and $T$ give rise to an $F$-augmentation ${(}E_0, E_1{)}$ in $G$ with $|E_0| = |M'| \geq |T| \geq \eps'n$, contradicting our assumption that (ii) does not hold.
So $H[\mathcal{D}'] \cup
H[\mathcal{D}',\mathcal{D}-\mathcal{D}']$ does not contain a matching with at least $\eps'n$ edges. Applying Theorem \[thm:ErdGall\] and passing to the graph $G$, we get $$e(G[D'] \cup G[D', D - D'])\le
t^2\ex(r,\eps'n{ \times K_2})+r{t\choose 2}\le
2t^2\eps'nr+r{t\choose 2}\;.$$ Therefore, $$\begin{aligned}
\nonumber
e(G[C\cup D])&=e(G[D'] \cup G[D', D - D'])+e(G[D-D'])+e(G[C])+e_G(C,D)\\
\label{eq:H3}
&\le 2t^2\eps'nr+r{t\choose 2}+{{(r-l)t\choose 2}}+{rs\choose 2}+r^2st.\end{aligned}$$ Summing up the bounds , , , and we get: $$\begin{aligned}
{1}
e(G) & = e(G[L]) + e_G(L,D) + e_G(L,C) + e(G[C \cup D]) \\
& \leq o(n^2) + t\eps'n^2 + ls\max\{m,r\} + 2\eps'nrt^2
+ r{t \choose 2} + {{(r-l)t\choose 2}} + {rs \choose 2} + r^2st. \\
\intertext{{Using the convexity of $f(l) := ls\max\{m, r\}+ {(r-l)t\choose 2}$ on $[0, r]$, and the fact that $rt \leq n$,
we get:}}
e(G) &\leq o(n^2) + 3t\eps'n^2 + r{t \choose 2}
+ r^2st + {rs \choose 2} + \max\left\{{rt \choose 2}, rs\max\{m, r\}\right\}. \\
\intertext{{However, $r^2s \leq {rt \choose 2} + o(n^2)$, and hence from \eqref{eq:DefM} we get:}}
e(G) &\leq o(n^2) + 3t\eps'n^2 + r{t \choose 2}
+ r^2st + {rs \choose 2} + \max\left\{{rt \choose 2}, rs(n-(s+t)r)\right\} \\
& < \max\left\{{(s+t)r \choose 2}, {rs \choose 2} + rs(n-rs)\right\} + (3t+1)\eps'n^2,\end{aligned}$$ [ where in the last inequality we have majorized the term $r\binom{t}{2}+o(n^2)$ by $\epsilon' n^2$. But $${(s+t)r \choose 2} + (3t+1)\eps'n^2
\leq \binom{(1-\eps)\alpha n}{2} + \frac{\eps \alpha^2n^2}{4}
< \left(1-\frac{\eps}{2}\right)\frac{\alpha^2n^2}{2},$$ and $$\begin{aligned}
{1}
{rs \choose 2} + rs(n-rs) + (3t+1)\eps'n^2
&< rsn - \frac{r^2s^2}{2}+\frac{\eps s\alpha n^2}{4(s+t)} \\
&\leq\frac{2s\alpha}{s+t}\left(1-\frac{\alpha s}{2(s+t)} +\frac{\eps(2-\eps)\alpha s}{2(s+t)}-\frac{3\eps}{4}\right)\frac{n^2}{2} \\
&< \frac{2s\alpha}{s+t}\left(1-\frac{\alpha s}{2(s+t)} -\frac{\eps}{4}\right)\frac{n^2}{2}.\end{aligned}$$ Consequently for large enough $n$, ]{} $$e(G) < T_{s,t}(\alpha){n \choose 2}\;,$$ a contradiction.
Suppose $G=(V,E)$ is a graph and $r \in \mathbb N$. The *$r$-expansion* of $G$ is the graph $G'=(V',E')$ defined as follows. The vertex set of $G'$ is $V \times [r]$. For $a,b \in [r]$, an edge $((u,a),(v,b))$ belongs to $E'$ iff $uv$ belongs to $E$. Note that there is a natural projection $\pi_{G'}: V' \rightarrow V$ that maps every vertex $(u,a)$ from $G'$ to the vertex $u$ in $G$. We are interested in the following property of $r$-expansions. Suppose that $K$ is a copy of any graph from $\mathcal
F^*$ in $G$. Then $\pi_{G'}^{-1}(V(K))$ contains a complete bipartite graph $B$ with color classes of sizes $s(K)r$ and $t(K)r$. By Lemma \[lem:til\] we can tile $B$ almost perfectly with copies of $K_{s,t}$. If $F$ is an $\mathcal F^*$-tiling in $G$, we can apply the above operation on each member $K\in F$ and obtain a new tiling $F'$ — which we call *retiling* — in the graph $G'$.
We are now ready to prove Theorem \[main\_result\].
Note that it suffices to prove the theorem for $H \simeq K_{s,t}$.
We first deal with the particular case $t =
s$. Set $\alpha':=(1-\epsilon/4)\alpha$. Let $\eps_1:=\frac1{{5}}(T_{s,t}(\alpha)-T_{s,t}(\alpha'))$, and $\eps_2$ be given by Lemma \[lem:tiling2\] for input parameters $H$, $d:=\eps_1$ and $\gamma:=\alpha\epsilon/8$. Suppose that $k_0$ is sufficiently large. Let $M$ be the bound from Lemma \[lem:reg\] for precision $\eps_R:=\min\{\eps_1,\eps_2\}$ and minimal number of clusters $k_0$. Let $C$ be given by Lemma \[lem:til\] for the input parameters $s,t$. Fix $n_0\gg MC$. Suppose that $G$ is an $n$-vertex graph, $n\ge n_0$, with at least $T_{s,t}(\alpha){n\choose 2}$ edges. We apply Lemma \[lem:reg\] on $G$ to obtain an $(\eps_R,d)$-reduced graph $R$ with $k$ clusters, $k_0\le k\le M$. We have that $$e(R)\ge (T_{s,t}(\alpha)-d-3\eps_1){k\choose
2}=(T_{s,t}(\alpha')+\frac1{{5}}(T_{s,t}(\alpha)-(T_{s,t}(\alpha'))){k\choose
2}{ { {\overset{\mbox{\tiny{\eqref{eq:TssVersusErdosGallai}}}}{>}} } }\ex\left(k,\frac{\alpha'
k}2\times K_2\right)\;.$$ Therefore, $R$ contains at least $\frac{\alpha'
k}2$ independent edges. These edges correspond to regular pairs in $G$ which can be tiled almost perfectly with copies of $K_{s,t}$, by means of Lemma \[lem:til\] and Lemma \[lem:tiling2\]. Elementary calculations give that in this way we get a tiling of size at least $(1- \eps)\alpha n$. Consequently we may suppose that $t > s$. We first define a handful of parameters. Set $$\label{eq:param1}
\alpha':= \frac{6-4\eps}{6-3\eps}\alpha,
\quad \gamma := (1-\eps/2)\alpha',
\quad d := \frac25(T_{s,t}(\alpha)-T_{s,t}(\alpha'))\;.$$ Note that $\gamma = (1 - 2\eps/3)\alpha$.
Let $\eps_R$ be given by Lemma \[lem:key\] for input graph $K_{s,t}$, density $d/2$ and approximation parameter $\gamma$. We may suppose that $\eps_R$ is sufficiently small such that $\gamma(1-\eps_R) > (1-\eps)\alpha$ and $\eps_R < d/2$. Let $C$ be given by Lemma \[lem:til\] for input $s,t$. Further, let $\eps'$ and $h$ be given by Lemma \[main\_lemma\] for input parameters $\alpha'$ and $\eps/4$. [We may assume that $\eps' < \eps$.]{} Set $$\label{eq:param2}
p := t^2\left\lceil\frac{4C}{\eps'}\right\rceil,
\quad q := \left\lceil\frac{2t}{\eps'}\right\rceil
$$
Let $M$ be the upper bound on the number of clusters given by Lemma \[lem:reg\] for input parameters $h$ (for the minimal number of clusters) and $\eps_Rp^{-q}/2$ (for the precision). Let $n_0 > Mp^q$ be sufficiently large.
Suppose now that $G$ is a graph with $n>n_0$ vertices and at least $T_{s,t}(\alpha){n\choose 2}$ edges. We first apply Lemma \[lem:reg\] to $G$ with parameters $\eps_Rp^{-q}/2$ and $h$. In this way we obtain an $(\eps_Rp^{-q}/2,d)$-reduced graph $R$ with at least $h$ vertices.
Let us now define a sequence of graphs $R^{(i)}$ by setting $R^{(0)} = R$ and letting $R^{(i)}$ be the $p$-expansion of $R^{(i-1)}, i = 1,2,\ldots,q$. Note that $e(R^{(i)})\ge T_{s,t}(\alpha'){{v(R^{(i)})}\choose 2}$ for every $i\in\{0,1,\ldots,q\}$.
Let $F^{(i)}$ be a maximum [size]{} $K_{s,t}$-tiling in $R^{(i)}$ for $i=0,1,\ldots,q$. We claim that $$\label{eq:Improving}
|F^{(i)}| \ge
\min\left\{
\frac{i\eps' {v(R^{(i)})}}{2t},
\left(1-\frac\eps2\right)\alpha' {v(R^{(i)})}
\right\}\;.$$ To this end it suffices to show that for any $i \geq 1$,
1. if $|F^{(i-1)}| > (1-\eps/4)\alpha'{v(R^{(i-1)})}$, then $\frac{|F^{(i)}|}{{v(R^{(i)})}} \ge
\frac{|F^{(i-1)}|}{{v(R^{(i-1)})}}-\frac{\eps\alpha'}4$, and
2. if $|F^{(i-1)}|\le
(1-\eps/4)\alpha' {v(R^{(i-1)})}$, then $\frac{|F^{(i)}|}{{v(R^{(i)})}}\ge
\frac{|F^{(i-1)}|}{{v(R^{(i-1)})}}+\frac{\eps'}{2t}$.
In the case (C1), [according to Lemma \[lem:til\]]{}, the retiling of $F^{(i-1)}$ in $R^{(i)}$ has size at least ${|F^{(i-1)}|(p-C)>}(1-\eps/2)\alpha' {v(R^{(i)})}$, thus proving the statement.
Consequently we may suppose that we are in case (C2). Apply Lemma \[main\_lemma\] to the graph $R^{(i-1)}$ and the tiling $F^{(i-1)}$, with parameters $\alpha'$ and $\eps/4$.
Suppose first that assertion (i) of the lemma holds. Then $R^{(i-1)}$ contains an $\mathcal{F}_1$-tiling $F$ with $\frac{|F|}{{v(R^{(i-1)})}}
\geq \frac{|F^{(i-1)}|}{{v(R^{(i-1)})}} + \eps'$. By retiling $F$, we get a $K_{s,t}$-tiling in $R^{(i)}$ with size at least $|F|(p-C) >
i\eps'{v(R^{(i)})}/(2t)$, thus proving the statement.
Suppose now that assertion (ii) of Lemma \[main\_lemma\] is true. Then $R^{(i-1)}$ contains an $F^{(i-1)}$-augmentation ${(}E_0, E_1{)}$ with $|E_0| \geq \eps'{v(R^{(i-1)})}$. Let $r=p/t$. We shall denote by $T$ the $t$-expansion of $R^{(i-1)}$ and by $T'$ the $r$-expansion of $T$. Note that $T'$ is isomorphic to $R^{(i)}$.
Let us build an $\mathcal{F}_2$-tiling in $T$ in the following way.
For every edge $e = (u,v) \in E_0$ with $u \in V(F^{(i-1)})$ we choose an edge $e'=(u',v')$ in $T$ with $\pi_{T}(u') = u$ and $\pi_{T}(v') = v$ . We shall denote by $w_e$ the vertex $u'$ corresponding to $u$.
For every edge $e = (u,v) \in E_1$ we choose a set $S_e$ of $t$ independent edges in $\pi_{T}^{-1}(e)$.
For every $K \in F^{(i-1)}$ we shall also choose a subgraph $K'$ of $T$. We distinguish the following cases. If $K$ has no vertex matched by $E_0$ or $E_1$, then we let $K' := T[\pi_{T}^{-1}(K)]$. If $K$ has a vertex $u$ matched by $E_1$ but no vertex matched by $E_0$, we let $K': = T[\pi_{T}^{-1}(K-u)]$. Then $K'
\simeq K_{st,(t-1)t}$. Finally, if $K$ has a vertex $u$ matched by an edge $e \in E_0$ and a vertex $v$ matched by an edge in $E_1$, we let $K' := T[\pi_{T}^{-1}(K-v)] -
w_e$. Note that in this last case $K' \simeq K_{st-1,(t-1)t}$.
It is easy to see that $$\label{eq:tiling}
F := \{e': e \in E_0\} \cup \{K' : K \in F^{(i-1)}\} \cup
\left(\bigcup_{e \in E_1}S_e\right)$$ is an $\mathcal{F}_2$-tiling in $T$. Moreover, we have that $\frac{|F|}{{v(T)}} \geq \frac{|F^{(i-1)}|}{{v(R^{(i-1)})}} +
\frac{\eps'}{t}$. So the retiling of $F$ in $T'$ has size at least $|F|(r-C) \geq i\eps'{v(R^{(i)})}/(2t)$. This proves (C2) and also .
Using Lemma \[lem:slicing\], we may subdivide every cluster corresponding to a vertex of $R$ into $p^q$ equal-sized parts, by discarding some vertices if necessary. This gives us an $(\eps_R,d/2)$-reduced graph $R'$. By construction $R' \simeq
R^{(q)}$. By , there is a $K_{s,t}$-tiling $F$ in $R'$ with size at least $(1-\eps/2)\alpha'{v(R')}$. Let $G'$ be the subgraph of $G$ induced by the clusters corresponding to the vertices of $R'$. By applying Lemma \[lem:key\] to $R'$, we see that $G'$ has a $K_{s,t}$-tiling of size at least $\gamma
{v(G')} \geq \gamma(1-\eps_R){v(G)} > (1-\eps)\alpha {v(G)}$, and so does $G$.
This finishes the proof of Theorem \[main\_result\].
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank the anonymous referee for his/her very useful comments.
[^1]: [In its original formulation, the Hajnal-Szemerédi Theorem asserts that an $n$-vertex graph $G$ with minimum-degree at least $\frac{r-1}{r}n$ contains a $K_r$-tiling missing at most $r-1$ vertices of $G$, thus giving an answer only to the question of almost perfect tilings. When the minimum-degree of $G$ is lower, we can however add auxiliary vertices which are complete to $G$ and obtain an $n'$-vertex graph $G'$ such that the Hajnal-Szemerédi Theorem applies to $G'$. The restriction of the almost perfect $K_r$-tiling of $G'$ to $G$ gives a $K_r$-tiling which is optimal in the worst case.]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Regular signed SAT is a variant of the well-known satisfiability problem in which the variables can take values in a fixed set $V \subset [0,1]$, and the literals have the form “${\mathtt{x}}\le a$” or “${\mathtt{x}}\ge a$” instead of “${\mathtt{x}}$” or “$\bar{\mathtt{x}}$”.
We answer some open question regarding random regular signed $k$-SAT formulas: The probability that a random formula is satisfiable increases with ${{{\left}\lvert{V}{\right}\rvert}}$; there is a constant upper bound on the ratio $m/n$ of clauses $m$ over variables $n$, beyond which a random formula is asypmtotically almost never satisfied; for $k=2$ and $V=[0,1]$, there is a phase transition at $m/n=2$.\
**Keywords:** Random constraint satisfaction problems; multi-valued logic; variants of SAT
address: |
DOT & CL: Fakultät für Mathematik\
Otto-von-Guericke-Universität Magdeburg\
Universitätsplatz 2\
39106 Magdeburg\
Germany. <http://dirkolivertheis.wordpress.com>
author:
- Christian Laus
- Dirk Oliver Theis
date: 'Tue Dec 6 18:05:36 CET 2011'
title: On the satisfiability of random regular signed SAT formulas
---
Introduction {#sec:intro}
============
Let $V$ be a set with at least two elements, $\mathcal S$ a set of subsets of $V$ called *signs*, and $k$ a positive integer. For the *signed $k$-satisfiability problem,* or *signed $k$-SAT,* one is given as input a set of $n$ variables $X$ and a formula in *signed conjunctive normal form.* This means that there is a list of $m$ *clauses*, each of which is a conjunction ($\land$) of $k$ *(signed) literals* of the form “${\mathtt{x}}\in
\texttt{S}$” where ${\mathtt{x}}$ is a variable in $X$ and the “sign” $\texttt{S}$ is a set in $\mathcal S$. The question is then whether there exists a satisfying *interpretation,* i.e., an assignment of values to the variables such that each of the clauses is satisfied.
Historically, signed SAT originated in the area of so-called multi-valued logic, where variables can take a certain number of truth values, not just 0 or 1. This is why the set $V$ is called the *truth-value set.* We refer the reader to the survey paper [@BeckertHaehnleManya99], and the references therein.
In the signed $k$-SAT area, the variants where the literals are inequalities have received special attention (see also [@HochbaumMorenocenteno08; @BallersteinTheis13]). One speaks of *regular* signed $k$-SAT or just *regular $k$-SAT* (*$k$-rSAT* for short), if $V$ is a (linearly) ordered set, and the allowed signs are $\{ x \mid x \le a\}$ and $\{ x \mid x \ge a\}$, $a\in V$. The literals then have the form “${\mathtt{x}}\le a$” or “${\mathtt{x}}\ge a$”. Clearly, the satisfiability of the formula depends only on $v := {{{\left}\lvert{V}{\right}\rvert}} \in \{2,3,\dots,\infty\}$ rather than on the set $V$ itself, so we always assume that $V \subset{\mathbb{R}}$ and $\min V = 0$, and $\max V = 1$.
This setting includes as a special case the classical satisfiability (SAT) problem: choose for $V$ the 2-element set $\{0,1\}$, and use the signed literals ${\mathtt{x}}\ge
1$ and ${\mathtt{x}}\le 0$ to represent the classical SAT literals ${\mathtt{x}}$ and $\bar x$, respectively.
This paper is about $k$-rSAT formulas drawn at random from all such formulas with a fixed truth-value set $V$. We allow that either $v < \infty$, or, in the limit, $V = [0,1]$. These random formulas are studied for $m=cn$, for a fixed constant $c$, in the limit $n\to \infty$. Literals ${\mathtt{x}}\le 1$ and ${\mathtt{x}}\ge 0$ are *innocuous:* they only affect the number of clauses which remain to be satisfied. We will drawn formulas uniformly at random from the set of all signed $k$-SAT formulas with $n$ literals, $m$ clauses and truth-value set $V$, which do not contain any innocuous literal.[^1] Based on computational investigations of the satisfiability of uniformly generated random 3-rSAT instances, Manyà et al. [@ManyaBejarEscaladaimaz98] have made a number of observations and conjectures. Most importantly, they observed a phase transition phenomenon similar to the one in classical SAT (see, e.g., [@Friedgut99; @AchlioptasPeres2004] and the references therein). They interpret their results as supporting the existence of a threshold $c = c_k(v)$, for $k=3$, with the following properties:
- the most computationally difficult instances tend to be found when the ratio ${{\nicefrac{m}{n}}}$ is close to $c_k(v)$;
- there is a sharp transition from satisfiable to unsatisfiable instances when the ratio ${{\nicefrac{m}{n}}}$ crosses the threshold;
- $c_k(v)$ is nondecreasing in the number of truth-values $v$.
Their results are confirmed and extended by other papers exploring uniformly random 3-rSAT instances [@BejarManya99phase; @BeckertHaehnleManya99; @BejarManyaCabiscolFernandezGomes07]. From their computational data, Béjar et al. [@BejarManya99phase; @BejarManyaCabiscolFernandezGomes07] surmise that
(i) the threshold $c_k(v)$ increases logarithmically with $v$,
and prove that for $$\label{eq:BejarManya-bound}
c > \log_{{{\nicefrac{8}{7}}}}(v)$$ a random 3-rSAT formula with ${{\nicefrac{m}{n}}} = c$ is asymptotically almost never (a.a.n., as $n\to\infty$) satisfiable. Their proof (and bound) resembles that for classical $k$-SAT ($2^k \log 2$).
Our contributions. {#our-contributions. .unnumbered}
------------------
In this paper, we prove (ii) for $k=2$ and $V=[0,1]$; establish (iii) for all $k$; improve the bound for large $v$; and falsify (iv).
To elaborate, for (iii), we show that the probability that a random formula is satisfied increases with $v$. In particular, the probability that a random formula on a finite truth-value set is satisfiable is bounded from above by the probability that a random formula with $V=[0,1]$ is satisfiable. Thus, if $c_3(v)$ increased logarithmically with $v$, then for any finite $c$, a random formula with truth-value set $[0,1]$, $n$ variables and $m=cn$ clauses would be asymptotically almost surely (a.a.s.) satisfiable.
We then prove the following.
\[thm:general-ub\] If $c > 1$ is such that $$kc{\left}( 1 - 2^{-k} {\right})^{c-1} < 1,$$ then a random $k$-rSAT formula with $n$ variables, $m=cn$ clauses, and $V = [0,1]$ is a.a.n. satisfiable.
This improves on Béjar et al.’s [@BejarManyaCabiscolFernandezGomes07] bound for large values of $v$. Most notably, it gives a finite upper bound for all $v$ and thus disproves (iv). In particular, Theorem \[thm:general-ub\] implies the following.
For all $V$, a random $3$-rSAT formula with $n$ variables and $m=cn$ clauses is a.a.n. satisfiable, if $c \ge 36.1$.
We then move on to study 2-rSAT. Here, Theorem \[thm:general-ub\] gives an upper bound of apx. 12.664 beyond which a random 2-rSAT is a.a.n. satisfiable. To prove a lower bound beneath which such a formula is satisfiable, we use a result by Chepoi et al. [@ChepoiCreignouHermannSalzer10], who prove a characterization of non-satisfiable signed 2-SAT instances based on a digraph certificate, in the spirit of Aspvall, Plass, and Tarjan’s famous result for classical 2-SAT [@AspvallPlassTarjan79]. Using Chepoi et al.’s characterization we prove the following.
\[thm:2-lb\]\[thm:2-ub\] A random $2$-rSAT formula with $n$ variables, $m=cn$ clauses, and ${{{\left}\lvert{V}{\right}\rvert}} = \infty$ is
(a) a.a.s. satisfiable, if $c < 2$, and
(b) a.a.n. satisfiable, if $c > 2$.
The improved upper bound here comes from a concentration result. This theorem shows that, as for classical random $k$-SAT, $k$-rSAT exhibits a threshold behaviour if $k=2$.
The main difficulty in the last theorem for 2-rSAT, as compared to classical 2-SAT, comes from the fact that there is an infinite number of possible literals — as opposed to the total of $2n$ possible literals for classical SAT. In our proof, we make use of the following trick. When conditioning on the number of times $R_j$ each variable ${\mathtt{x}}_j$ occurs, the structure one needs to analyze has some resemblance to the configuration model for random (multi) graphs with prescribed degrees $R_j$. This way, Chvátal and Reed’s [@ChvatalReed92] approach for classical 2-SAT can be adapted.
Organization of the paper. {#organization-of-the-paper. .unnumbered}
--------------------------
The remainder of the paper is organized as follows. In the next section, we discuss the random model, and variants of it, in the necessary details and prove the monotonicity of the probability of satisfiability mentioned above. Section \[sec:pf-general-ub\] contains the proof of Theorem \[thm:general-ub\]. Sections \[sec:pf-2-lb\] and \[sec:pf-2-ub\] hold the proof of Theorem \[thm:2-lb\]. In the final section, we discuss a few open questions.
Basics about random $k$-SAT {#sec:basics}
===========================
In this section, we discuss variants of the random model which we need. Then we will prove some basic facts about $k$-rSAT.
We will think of a random formula as being constructed as follows. First of all, we assume that the truth-value set $V$ is a subset of the unit interval $[0,1]$ which is symmetric (i.e., $V = 1-V$) and which contains both $0$ and $1$.
Now we take an “empty” formula, i.e., we have $m\times k$ empty slots. Each slot will be filled by a triple $({\mathtt{x}},\varrho,a)$ where ${\mathtt{x}}$ is one of the variables, $\varrho$ is a comparison relation “$\le$” or “$\ge$”, and $a$ is in $V\setminus\{1\}$. The interpretation of such a triple is that, if $\varrho=\mbox{``$\le$''}$, then we have the condition ${\mathtt{x}}\le a$, whereas, if $\varrho=\mbox{``$\ge$''}$, we have the condition ${\mathtt{x}}\ge 1-a$. By this construction, we exclude the cases of the inequalities ${\mathtt{x}}\ge 0$ and ${\mathtt{x}}\le 1$, which are meaningless because they do not constrain ${\mathtt{x}}$. The part $(\varrho,a)$ is referred to as the constraint part of the literal.
For each slot, the three parts of the triple are chosen independently from each other. The selection of the right hand sides and the comparison relations is done independently for all slots. For the variables, there are several possibilities. First of all, their selection may either be chosen independently for all slots (i.e., allowing a clause to contain more than one slot with the same variable), or indepdently for all slots but conditioning on the $k$ variables occuring in a clause being distinct. In the second case, the event on which we condition is asymptotically bounded away from $0$. Hence, as far as a.a.s. statements about random formulas are concerned, the two possibilities for the random selection of the variables are equivalent. We denote by $F_k(n,m,v)$ a random $k$-rSAT formula with truth-value set of cardinality $v$ in which, for each clause, the variables in the slots are distinct; by $F'_k(n,m,v)$, we denote a random formula where the variables can occur multiple times in the same clause.
Secondly, we may choose the variables conditioning on the number of times each variable occurs in the formula. For a random formula $F$, let the random variable $R_j$ denote the number of slots containing variable ${\mathtt{x}}_j$. Clearly, we have $$\label{eq:sum-var-occ}
\sum_{j=1}^n R_j = km.$$ If, when choosing the variables, we allow a clause to contain more than one slot with the same variable, then the $R := (R_j)_{j=1,\dots,n}$ has multinomial distribution, i.e., for all $r\in {\mathbb{N}}^n$ with $\sum_j r_j = km$ we have $$\label{eq:r-distrib}
\operatorname{\mathbf{P}}[ R = r ] = \frac{\binom{km}{r_1,\dots,r_n}}{n^{km}}.$$ This is same as the $R_j$ being independent Poison with mean $km/n$ conditioning on .
When constructing a random formula, we may reverse this view: We may condition on the values of $R$. This amounts to pretending that, for $j=1,\dots,n$, there are $R_j$ distinguishable copies of variable ${\mathtt{x}}_j$, and the $km$ variable copies are assigned to the $km$ slots randomly.
Monotonicity {#monotonicity .unnumbered}
------------
We now come the some basic facts about random $k$-rSAT formulas. We start the monotonicity property of rSAT formulas mentioned in the introduction. Denote by $$\label{eq:def-p}
p_k(n,m,v) := \operatorname{\mathbf{P}}[ F_k(m,n,v) \text{ is satisfiable}]$$ the probability that a random $k$-rSAT formula with $m$ clauses on $n$ variables and truth-value set of cardinality ${{{\left}\lvert{V}{\right}\rvert}}$ is satisfiable. We will habitually omit the $k$. Naively speaking, increasing ${{{\left}\lvert{V}{\right}\rvert}}$ increases the possible choices for the variables, so we would guess that $p(n,m,v)$ increases with $v$. That is in fact the case.
The easiest setting in which we can visualize this phenomenon is, if we suppose that the right hand sides are of the form $$A = \sum_{i=1}^\lambda B_i 2^{-i},$$ where the $B_i$, $i=1,2,3,\dots$ are independent Bernoulli random variables with $\operatorname{\mathbf{P}}[B_i = 1] = {{\nicefrac{1}{2}}}$, and $\lambda$ is either finite — in which case ${{{\left}\lvert{V}{\right}\rvert}} = 2^{\lambda}+1$ — or $\lambda=\infty$, in which case $V=[0,1]$. Now note that increasing $\lambda$ increases $A$. But this weakens the inequalities constraining the variables, and thus makes the formula “more satisfiable”. An only slightly more technical argument proves this monotonicity fact for general ${{{\left}\lvert{V}{\right}\rvert}}$.
\[lem:p\_monotone\] For every $k,n,m$, the following hold.
(a) For every $v$, we have $\displaystyle p_k(n,m,v) \le p_k(n,m,v+1)$.
(b) For every $v$, we have $\displaystyle p_k(n,m,v) \le p_k(n,m,\infty)$.
(c) We have $\displaystyle \lim_{v\to\infty} p_k(n,m,v) = p_k(n,m,\infty)$.
*(a).* Suppose we have $V = \{u/(v-1) \mid u=0,\dots,v-1\}$. For a random constraint, a value in $V\setminus\{1\}$ is drawn uar. We would like increase $v\leadsto v+1$. Firstly, we scale all values by the factor $(v-1)/v$. This does not influence satisfiability. Secondly, we increase $u/v$ to $(u+1)/v$ with probability $(u+1)/v$, $u=0,\dots,v-2$. This yields the uniform distribution on $V'\setminus\{1\} = \{ u/v \mid u=0,\dots,v-1\}$.
Note that increasing a in a literal $({\mathtt{x}},\rho,a)$ will never make a satisfiable formula unsatisfiable: indeed, the set of satisfying interpretations stays the same. However, if a formula is not satisfiable, whenever the value for $a$ in a literal is increased, there is a possibility that the formula becomes satisfiable. Thus, the probability that a random formula is satisfiable increases with $v\leadsto v+1$.
*(b).* To prove (b), by (a), it suffices to consider the following sets, for $\lambda = 0,1,2,\dots,\infty$: $$V_{\lambda} := \Bigl\{ \sum_{i=1}^\lambda B_i 2^{-i} \mid B \in \{0,1\}^{\mathbb{N}}\Bigr\} \cup \{1\},$$ i.e., $V_0 = \{0,1\}$, $V_1 = \{0,{{\nicefrac{1}{2}}},1\}$, $V_2 = \{0,{{\nicefrac{1}{4}}},{{\nicefrac{1}{2}}},{{\nicefrac{3}{4}}},1\}$, …, $V_\infty = [0,1]$.
To prove (b), we now use the method of deferred decisions. We draw $B=(B_1,B_2,\dots)$ randomly regardless of the value of $\lambda$. For a random formula $F_k(n,m,2^\lambda+1)$, the $B_1,\dots,B_\lambda$ have been exposed. Increasing $\lambda\leadsto \infty$ amounts to exposing all remaining $B_{i}$, $i=\lambda+1,\lambda+2,\dots$, and adjusting the literals of the formula accordingly. But this can only increase the sum, and thus, modifying the literals of a formula in this way can only turn a not satisfiable formula into a satisfiable one, and thus can only increase $p_k(n,m,\cdot)$.
*(c).* We use the same approach as in (b). Suppose that a formula $F := F_k(n,m,\infty)$ is satisfiable. We then find a finite $\lambda$ such that truncating the sums at the $2^{-\lambda}$-term already yields a satisfiable formula. First of all, we may assume that the literals $({\mathtt{x}},\rho,a)$ all have distinct values $a$. Let $\lambda_-$ be the largest number such that there are literals $({\mathtt{x}},\rho,a)$ and $({\mathtt{x}}',\rho',a')$ for which the sums in $a$ and $a'$ coincide up to the $2^{-\lambda_-}$-term. Then $\lambda_-$ is finite. Letting $\lambda := \lambda_- +1$, we see that in truncated random formula, the constraint parts of the literals have the same relative ordering as in the original formula. Hence, the truncated formula is satisfiable.
Thus, every satisfiable formula for $\lambda=\infty$ becomes satisfiable already at a finite value for $\lambda$. This proves the stated convergence.
Since the existence of a threshold is, as of now, conjectural, we let, for $v\in \{2,3,\dots,\infty\}$ $$\begin{aligned}
c^-_k(v) &:= \sup\{ c \mid F_k(n,cn,v) \text{ a.a.s.\ satisfiable} \}\text{, and}\\
c^+_k(v) &:= \inf\{ c \mid F_k(n,cn,v) \text{ a.a.n.\ satisfiable} \}.\\\end{aligned}$$ The existence of a threshold is then equivalent to $c^-_k(v) = c^+_k(v)$; cf. Fig. \[fig:transition\].
From Lemma \[lem:p\_monotone\], we immediately derive the following.
\[prop:ckv\_monotone\] For every $k$ we have the following.
(a) $c^-_k(v)$ and $c^+_k(v)$ are both nondecreasing with the cardinality $v$ of the truth-value set.
(b) $c^-_k(v) \le c^-_k(\infty)$ and $c^+_k(v) \le c^+_k(\infty)$.
(c) $\lim_{v\to\infty} c^-_k(v) = c^-_k(\infty)$ and $\lim_{v\to\infty} c^+_k(v) = c^+_k(\infty)$.
\(a) and (b) follow immediately from their counterparts in Lemma \[lem:p\_monotone\]. As for (c), let $c^+ := \lim_{v\to\infty} c^+_k(v)$ and assume that $c^+ <
c^+_k(\infty)$. Then, for $c$ with $c^+ < c < c^+_k(\infty)$, we have that $p(n,cn,\infty)$ does not converge to $0$, so there is a sequence $(n_\ell)_{\ell=1,2,\dots}$ for which $p(n_\ell,cn_\ell,\infty)$ is bounded away from 0. But $\lim_n p(n,cn,v) = 0$, contradicting part (c) of Lemma \[lem:p\_monotone\].
Similarly, let $c^- := \lim_{v\to\infty} c^-_k(v)$ and assume that $c^- < c^-_k(\infty)$. Then, for $c$ with $c^- < c < c^-_k(\infty)$, we have $\lim_n
p(n,cn,\infty) = 1$. But for all $v$, we have that $p(n,cn,v)$ is bounded away from 1. We obtain a contradiction in the same way as above.
Ancilliary {#ancilliary .unnumbered}
----------
We conclude the section with the following trivial lemma.
\[lem:prb-sat-dj\] If $V=[0,1]$, then the following holds.
(a) For every fixed $x\in[0,1]$, the probability that a random constraint defined by $(\varrho,A)$ is satisfied, is ${{\nicefrac{1}{2}}}$.
(b) For every to constraints $(\varrho,A)$, $(\varrho',A')$, the probability that there is no $x\in V$ satisfying both (i.e., the signs are disjoint) is ${{\nicefrac{1}{4}}}$.
Proof of Theorem \[thm:general-ub\] {#sec:pf-general-ub}
===================================
In this section, we prove Theorem \[thm:general-ub\]. By Lemma \[lem:p\_monotone\], it suffices to prove the statement for the case when the right hand sides of the formulas are drawn uar from $[0,1]$. In particular, we can assume that no two literals have the same right hand side.
An interpretation ${\mathtt{x}}\to x$ of a formula $F$ is called *tight,*, if for every variable ${\mathtt{x}}_j$, there one of the two literals “${\mathtt{x}}_j \le x_j$” or “${\mathtt{x}}_j \ge x_j$” occurs in $F$. In other words, only the right hand sides of the inequalities are allowed as values for the variables. The following fact is trivial.
There exists a satisfying interpretation if, and only if, there exists a satisfying tight interpretation.
For a random formula $F$, denote by $Y = Y_F$ the number of satisfying tight interpretations. To prove Theorem \[thm:general-ub\], we compute the expectation of $Y$.
When sampling a random formula, we condition on $R=(R_j)_{j=1,\dots,n}$ as discussed above. For $j=1,\dots,n$ and $\ell=1,\dots,R_j$, let $A_{j,\ell}$ be the (random) right hand side in the slot containing the $\ell$th copy of the variable ${\mathtt{x}}_j$. For every $\ell \in \prod_{j=1}^n \{1,\dots,R_j\}$, we construct an interpretation ${\mathtt{x}}\to x(\ell)$ by letting $x_j(\ell) := A_{j,\ell(j)}$. With the sum below extending over all $\ell \in \prod_{j=1}^n \{1,\dots,R_j\}$, we have $$Y
=
\sum_{\ell} {\operatorname{\mathbf{I}}}[ x(\ell) \text{ satisfies } F ].$$
For every fixed $\ell$, we can estimate the probability of the event that $x(\ell)$ satisfies $F$. Indeed, $n$ literals will be “automatically” satisfied, namely for every variable ${\mathtt{x}}_j$ the one containing the $\ell(j)$th copy of ${\mathtt{x}}_j$. Since the right hand sides are drawn independently, for each of the remaining literals, the probability of being satisfied by $x$ is ${{\nicefrac{1}{2}}}$, by Lemma \[lem:prb-sat-dj\]. The automatically satisfied literals cover at most $n$ clauses, which leaves $(c-1)n$ clauses, each of which contains exactly $k$ of the remaining literals. Conditioned on the assignment $X$ of the variable copies to the slots, the probability that all of these clauses are satisfied is thus at most $$\bigl( 1-2^{-k} \bigr)^{(c-1)n}.$$ Moreover, the event that all the remaining clauses are satisfied depends only on the constraint part of the literals and is thus independent from $R$
Now we can compute the expected number of satisfying tight interpretations. $$\begin{gathered}
\operatorname{\mathbf{E}}Y
=
\operatorname{\mathbf{E}}\biggl( \operatorname{\mathbf{E}}\Bigl( \sum_{\ell} {\operatorname{\mathbf{I}}}[ x(\ell) \text{ satisfies } F ] \Bigm| R \Bigr) \biggl)
\\
=
\operatorname{\mathbf{E}}\biggl( \operatorname{\mathbf{E}}\Bigl( \sum_{\ell} \operatorname{\mathbf{E}}( {\operatorname{\mathbf{I}}}[ x(\ell) \text{ satisfies } F ] \mid X ) \Bigm| R \Bigr) \biggl)
\\
\le
\bigl( 1-2^{-k} \bigr)^{(c-1)n} \operatorname{\mathbf{E}}\biggl( \operatorname{\mathbf{E}}\Bigl( \sum_{\ell} 1 \Bigm| R \Bigr) \biggl)
\\
=
\bigl( 1-2^{-k} \bigr)^{(c-1)n} \operatorname{\mathbf{E}}\Bigl( \prod_{j=1}^n R_j \Bigl)
\le
\bigl( 1-2^{-k} \bigr)^{(c-1)n} ({{\nicefrac{km}{n}}})^n
=
\Bigl( kc\bigl( 1-2^{-k} \bigr)^{c-1} \Bigr)^n.\end{gathered}$$ From this, Theorem \[thm:general-ub\] follows by Markov’s inequality.
Proof of Theorem \[thm:2-lb\]([)]{} {#sec:pf-2-lb}
===================================
In this section, we prove Theorem \[thm:2-lb\]. For this, we use the Aspvall-Plass-Tarjan-style [@AspvallPlassTarjan79] characterization of non-satisfiable signed 2-SAT formulas by Chepoi et al. [@ChepoiCreignouHermannSalzer10] together with Chvátal and Reed’s [@ChvatalReed92] trick of counting “bicycles”.
An *$\ell$-bicycle* contained in a 2-rSAT formula $F$ is a sequence of $2\ell$ literals $$w^f_0,w^t_1,w^f_1\dots,w^t_\ell,w^f_\ell,w^t_{\ell+1}$$ together with two (not necessarily distinct) numbers $i_0\in\{2,\dots,\ell\}$ and $i_1\in\{1,\dots,\ell-1\}$ such that
1. \[enum:bicycle:vars-dtct\] the variables in $w^t_1,w^t_2,\dots,w^t_\ell$ are all distinct;
2. \[enum:bicycle:vars-traverse\] the variables in the two literals $w^t_i,w^f_i$ are the same, for $i=1,\dots,\ell$;
3. \[enum:bicycle:vars-ends\] the variable of $w^f_0$ is the same as the one of $w^t_{i_0}$, and the variable of $w^f_{\ell+1}$ is the same as the one of $w^t_{j_1}$;
4. \[enum:bicycle:clauses\] for each $i=0,\dots,\ell$, “$w^f_{i} \lor w^t_{i+1}$” is a clause in $F$;
5. \[enum:bicycle:traverse\] for each $i=1,\dots,\ell$, the constraint parts of $w^t_i,w^f_i$ are disjoint.
The following is an immediate adaption of Chvátal and Reed’s proof in [@ChvatalReed92] to Chepoi et al.’s [@ChepoiCreignouHermannSalzer10] variant, for the signed case, of Aspvall et al.’s [@AspvallPlassTarjan79] characterization of non-satisfiable 2-SAT formulas.
Every unsatisfiable 2-rSAT formula contains an $\ell$-bicycle, for some $\ell \ge 2$.
As in the previous section, we will make use of the fact that, for ${{{\left}\lvert{V}{\right}\rvert}}=\infty$, we may assume that no two literals have the same right hand side.
As above, let $R=(R_j)_{j=1,\dots,n}$ count the occurences of the variables in the random formula $F=F'_2(n,m,\infty)$. Conditioned on $R$, we can recover the distribution of $F$ as follows. Let there be $n$ *buckets* $B_1,\dots,B_n$; bucket $B_j$ contains $R_j$ *points.* Note that there is an even number $2m$ of points. A perfect matching of the points corresponds to selecting two variables for each clause of the formula. Hence, drawing a matching at random and, independently, drawing a random constraint part for each point, gives us a random formula. The distribution is the same as that for random formulas, conditioned on $R$.
We now count the number of $\ell$-bicycles in $F$ by counting the corresponding matchings. We start with the following easy lemma.
\[lem:poisson-trick\] Let $d_1,\dots,d_n$ be nonnegative integers. Then
1. $\displaystyle \operatorname{\mathbf{E}}\prod_{j=1}^n (R_j)_{d_j} \le (2c)^{\sum_j d_j} $
2. If all $\sum_j d_j = O(1)$, then $\operatorname{\mathbf{E}}\prod_{j=1}^n (R_j)_{d_j} = (1+o(1)) (2c)^{\sum_j d_j}$.
This is a direct computation using .
We are now ready to complete the proof.
Let $b=(b_1,\dots,b_\ell) \in \{1,\dots,n\}^\ell$ be the choice of the $\ell$ distinct variables in condition (\[enum:bicycle:vars-dtct\]) of the definition of a bicycle above. For each one of these $b$, we have to choose $i_0 \in \{2,\dots,\ell\}$ and $i_1 \in \{1,\dots,\ell-1\}$.
Letting $d_{b_i} := 2$, $i=1,\dots,\ell$, and $d_j := 0$ for each $j$ not occuring in $(b_1,\dots,b_\ell)$, the number of choices for the matching edges between the buckets corresponding to the clauses “$w^f_{i} \lor w^t_{i+1}$”, $i=1,\dots,\ell-1$, conditioned on $R$, is $\prod_{j=1}^n (R_j)_{d_j}$. For the clauses “$w^f_{0} \lor w^t_{1}$” and “$w^f_{\ell} \lor w^t_{\ell+1}$”, the number of choices depend on whether $i_0=\ell$, or $i_1=0$, or both, or neither. For each choice of $i_0$ and $i_1$, if we change the $d_j$ to count the number of times the variable ${\mathtt{x}}_j$ occurs in the bicycle, the number of choices is $\prod_{j=1}^n
(R_j)_{d_j}$. There are at most $\ell^2$ choices for the $i_0$ and $i_1$, and we have $\sum_j d_j = 2(\ell+1)$, so that, by Lemma by Lemma \[lem:poisson-trick\], the expectation of the number of choices for the matching edges in the bicycle for fixed $b$ and $i_0, i_1$ is at most $(2c)^{2(\ell+1)}$, whereas the total number of choices for these matching edges equals $(2m-1)(2m-3)\dots(2m-2\ell-1)$.
There are $(n)_\ell$ possible choices of $b$, and the probability of disjointness in (\[enum:bicycle:traverse\]) is ${{\nicefrac{1}{4}}}$ for each variable (by Lemma \[lem:prb-sat-dj\]), or $4^{-\ell}$ for the whole bicycle.
Thus, denoting by $Y_\ell$ the number of $\ell$-bicycles in a random formula and by $X_{(b,i_0,i_1)}$ the indicator variable of the event that a bicycle with these parameters exists, we may compute as follows: $$\begin{gathered}
\operatorname{\mathbf{E}}Y_\ell
= \operatorname{\mathbf{E}}{\left}( \operatorname{\mathbf{E}}\Bigl( \sum_{b} \sum_{i_0,i_1} X_{(b,i_0,i_1)} \Bigm| R \Bigr) {\right})
= \sum_{b} \operatorname{\mathbf{E}}\Bigl( \sum_{i_0,i_1} \operatorname{\mathbf{E}}( X_{(b,i_0,i_1)} \mid R ) \Bigr)
\\
\le \sum_{b} 4^{-\ell} \ell^2 (2c)^{2(\ell+1)} \frac{ 1 }{{(2m-1)(2m-3)\dots(2m-2\ell-1)}} \biggr)
\\
= 4^{-\ell} \ell^2 (2c)^{2(\ell+1)} \frac{ (n)_\ell }{{(2m-1)(2m-3)\dots(2m-2\ell-1)}}.
\end{gathered}$$ For the fraction, we use ad-hoc estimates. Noting that $\ell \le n$, we have $2m-2\ell \ge 2(c-1)n$. By the monotonicity property, Lemma \[lem:p\_monotone\], it suffices to prove the theorem for $c>1$, in which case $2m-2\ell-1 =: \omega(n) \to \infty$. From Stirling’s formula, we see that $$\begin{gathered}
\frac{ (n)_\ell }{{(2m-1)(2m-3)\cdots(2m-2\ell-1)}}
=
\frac{1}{\omega\, (2c)^\ell} \, \prod_{j=0}^{\ell-1} \frac{n-j}{n- {{\nicefrac{(j+{{\nicefrac{1}{2}}})}{c}}}}
\\
\le
\frac{1}{\omega\, (2c)^\ell} \, \prod_{j=0}^{\frac{1}{(c-1)2}} \frac{n-j}{n- {{\nicefrac{(j+{{\nicefrac{1}{2}}})}{c}}}}
= (1+o(1)) \frac{1}{\omega\, (2c)^\ell}
\end{gathered}$$
Summing over $\ell$, we see that $$\sum_{\ell=2}^n Y_\ell
\le \frac{2c}{\omega} \sum_{\ell} \ell^2({{\nicefrac{c}{2}}})^\ell = O_c\Bigl( \frac{1}{\omega} \Bigr),$$ for $c < 2$ (the constant in the big-O depends on $c$). Thus, the expected number of bicycles of arbitrary length is $O_c({{\nicefrac{1}{n}}})$. From this, the statement of the theorem follows.
Proof of Theorem \[thm:2-ub\]([)]{} {#sec:pf-2-ub}
===================================
As in the previous section, to prove Theorem \[thm:2-ub\](b), we adapt the approach of Chvátal and Reed [@ChvatalReed92]: Prove the non-satisfiability by establishing the existence of an obstruction by the second moment method.
For an even integer $\ell \ge 6$, we an *$\ell$-snake* consists of a selection of $\ell$ distinct variables ${\mathtt{x}}_{b_1},\dots,{\mathtt{x}}_{b_\ell}$, and clauses “$({\mathtt{x}}_{b_i},\rho_i,a_i) \lor ({\mathtt{x}}_{b_{i+1}},\rho'_{i+1},a'_{i+1})$”, $i=0,\dots,\ell$, with $b_0 := b_{\ell/2} =: b_{\ell+1}$ such that
1. the constraint parts $(\rho'_{i},a'_{i})$ and $(\rho_{i},a_{i})$ are disjoint, for $i=1,\dots,\ell$.
2. the constraint parts $(\rho'_{\ell+1},a'_{\ell+1})$ and $(\rho_{0},a_{0})$ are disjoint.
3. the constraint parts $(\rho'_{\ell/2},a'_{\ell/2})$ and $(\rho'_{\ell+1},a'_{\ell+1})$ are disjoint; as well as the constraint parts $(\rho_{\ell/2},a_{\ell/2})$ and $(\rho_{0},a_{0})$ are disjoint.
If there exists an $\ell$-snake, then the formula is not satisiable.
A snake gives rise to a srongly connected component in the digraph of the formula defined by Chepoi et al. [@ChepoiCreignouHermannSalzer10], which contains a literal as well as its negation. Here, we give the elementary proof of the lemma.
Assume that a satisfying interpretation ${\mathtt{x}}\to x$ exists. We prove that the literal $({\mathtt{x}}_{b_{\ell/2}},\rho'_{\ell/2},a'_{\ell/2})$ can be neither satisfied nor violated by $x$. Suppose it were satisfied. Then, by disjointness of the constraint parts, $({\mathtt{x}}_{b_{\ell/2}},\rho_{\ell/2},a_{\ell/2})$ must be violated by $x$. Since there is a clause “$({\mathtt{x}}_{b_{\ell/2}},\rho_{\ell/2},a_{\ell/2}) \lor ({\mathtt{x}}_{b_{\ell/2+1}},\rho'_{\ell/2+1},a'_{\ell/2+1})$”, the later literal must be satisfied by $x$. Proceeding in this fashion, it follows that $({\mathtt{x}}_{b_{\ell+1}},\rho'_{\ell+1},a'_{\ell+1})$ is satisfied, but, since $b_{\ell+1} = b_{\ell/2}$ and by disjointness, this implies that $({\mathtt{x}}_{b_{\ell/2}},\rho'_{\ell/2},a'_{\ell/2})$ is violated, a contradiction.
Suppose that $({\mathtt{x}}_{b_{\ell/2}},\rho'_{\ell/2},a'_{\ell/2})$ is violated by $x$. Since there is a clause “$({\mathtt{x}}_{b_{\ell/2-1}},\rho'_{\ell/2+1},a'_{\ell/2-1})
\lor ({\mathtt{x}}_{b_{\ell/2}},\rho_{\ell/2},a_{\ell/2})$”, the literal $({\mathtt{x}}_{b_{\ell/2-1}},\rho'_{\ell/2+1},a'_{\ell/2-1})$ must be satisfied by $x$. Proceeding as above, we conclude that $({\mathtt{x}}_{b_0},\rho_{0},a_{0})$ is satisfied by $x$, and hence, since $b_0=b_{\ell+1}$, by disjointness, $({\mathtt{x}}_{b_{\ell+1}},\rho'_{{\ell+1}},a'_{{\ell+1}})$ is violated. Since “$({\mathtt{x}}_{b_{\ell}},\rho_{\ell+1},a_{\ell}) \lor
({\mathtt{x}}_{b_{\ell+1}},\rho'_{\ell+1},a'_{\ell+1})$” is a clause, $({\mathtt{x}}_{b_{\ell}},\rho_{\ell},a_{\ell})$ is satisfied, and hence, eventually, so is $({\mathtt{x}}_{b_{\ell/2}},\rho_{\ell/2},a_{\ell/2})$. By disjointness, then, $({\mathtt{x}}_{b_0},\rho_{0},a_{0})$ is violated, a contradiction.
Fix a $c > 2$, let $$\ell := 2 {\left}\lceil \frac{\log n}{\log({{\nicefrac{c}{2}}})} {\right}\rceil,$$ and denote by $X$ the number of $\ell$-snakes. We will compute the expectation of $X$ and of its second factorial moment $X(X-1)$, and find that $\operatorname{\mathbf{E}}X =
\Omega(n)$ and $\operatorname{\mathbf{E}}X(X-1) = O( (\operatorname{\mathbf{E}}X)^2 )$. From this, by Chebyshev’s inequality, we conclude that the probability that no $\ell$-snake exists is $O(1/\operatorname{\mathbf{E}}X) =
O({{\nicefrac{1}{n}}})$. Thus, a random formula $F_2(n,cn,\infty)$ is a.a.n. satisfiable.
The following two lemmas comtain the computations of the moments.
$\displaystyle \operatorname{\mathbf{E}}X = \Omega(n)$.
As in the previous section, we have $$\operatorname{\mathbf{E}}(X\mid R)
= \sum_{b} \frac{ (R_{b_{\ell/2}}-2)(R_{b_{\ell/2}}-3) \prod_{i=1}^\ell (R_{b_i})_2 }{ 4^{\ell+3} (2m-1)\cdots(2m-2\ell-1) },$$ where the sum extends over all possible choices $b\in\{1,\dots,n\}^{\ell}$ identifying the $\ell$ distinct variables in a snake. Thus, by Lemma \[lem:poisson-trick\], we have $$\operatorname{\mathbf{E}}X
= \Theta\Bigl( ({{\nicefrac{c}{2}}})^\ell \tfrac{1}{2m-2\ell-1} \Bigr)
= \Theta( {{\nicefrac{1}{n}}} ).$$
$\displaystyle \operatorname{\mathbf{E}}X(X-1) = O((\operatorname{\mathbf{E}}X)^2)$.
We have to compute the expectation of the sum $$\sum_S \sum_{S'\ne S} X_{S} X_{S'}$$ where the sums range over all possible snakes $S$ and $S'$, respectively, and $X_S$ denotes the indicator variable of the event that the snake $S$ exists. Taking the expectation, it can be seen that the only non-negligible terms in this sum are those for which $S$ and $S'$ are supported on disjoint sets of variables (see e.g. the computation in §9.2 of [@JansonLuczakRucinskiBk]).
Whenever $S$ and $S'$ are supported on the disjoint sets of variables, a simple calculation shows that the expectation is $O( (\operatorname{\mathbf{E}}X)^2 )$.
This completes the proof of Theorem \[thm:2-ub\](b).
Conclusion and Outlook {#sec:conclusion}
======================
What makes random $k$-rSAT intriguing is the presence of a second parameter next to $c = {{\nicefrac{m}{n}}}$: the cardinality $v$ of the truth-value set $V$. Since the probability of satisfiability $p_k(n,m,v)$ increases with $v$ (Lemma \[lem:p\_monotone\]), $c^{\pm}_k(v)$ increases with $v$, too. Based on computational experiments, Béjar et al. [@BejarManyaCabiscolFernandezGomes07] predicted that $c_3(v)$ increases logarithmically with $v$. This is clearly not the case, by Theorem \[thm:general-ub\]. However, based on Béjar et al.’s data, we conjecture that the functions $c^{\pm}_k(v)$ are strictly concave.
For all $k\ge 2$ and $v\ge 2$, we have $c^{\pm}_k(v+1)-c^{\pm}_k(v) > c^{\pm}_k(v+2)-c^{\pm}_k(v)$.
Note that $p_k(n,cn,v)$ is not in general concave in $v$.
Béjar et al. [@BejarManyaCabiscolFernandezGomes07] conjecture (for $k=3$) that $c^+_k = c^-_k$, in other words, there is a threshold behaviour. This can be rephrased in terms of the dependence on the parameter $v$:
Does there exists a $v_k^*(c)$ such that $p_k(n,cn,v)=o(1)$ if $v < v^*_k(c)$, and $p_k(n,cn,v)=1-o(1)$ if $v > v^*_k(c)$?
For example, in the case of 2-rSAT, we know that $p_2(n,\frac32 n, 2) = o(1)$ [@ChvatalReed92; @DaLaVega92; @Goerdt94], and $p_2(n,\frac32n,\infty) = 1-o(1)$ by Theorem \[thm:2-lb\], but it is not clear whether the transition happens gradually, or suddenly, at some value $v^*$ between $2$ and $\infty$. If $v^*$ exists, though, then it must be “finite” in the sense that it does not depend on $n$, cf. Fig. \[fig:transition\].
In the case of 2-rSAT, for $v=2$ there is a threshold at $c^-_2(2)=c^+_2(2)=1$, and for $v=\infty$, there is a threshold at $c^-_2(\infty)=c^+_2(\infty)=2$. It seems likely that this is true for the remaining values of $v$, too. In fact, we conjecture the following behaviour.
For all $\lambda=0,1,2,\dots,\infty$, we have $$c^-_2(2^\lambda+1)=c^+_2(2^\lambda+1) = 2 - 2^{-\lambda} = \sum_{j=0}^{\lambda} 2^{-j}.$$
[^1]: If innocuous literals are allowed in a random formula, the number of clauses which contain an innocuous literal is distributed like a binomial variable with parameters $m=\Theta(n)$ and $1/v=O(1)$, and as such is with high probability $O(\sqrt n)$, so that the ratio $c={{\nicefrac{m}{n}}}$ is unaffected for $n\to\infty$. Hence, forbidding innocuous literals does not change asymptotic results.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Rigidity theory studies the properties of graphs that can have rigid embeddings in a euclidean space $\mathbb{R}^d$ or on a sphere and which in addition satisfy certain edge length constraints. One of the major open problems in this field is to determine lower and upper bounds on the number of realizations with respect to a given number of vertices. This problem is closely related to the classification of rigid graphs according to their maximal number of real embeddings.
In this paper, we are interested in finding edge lengths that can maximize the number of real embeddings of minimally rigid graphs in the plane, space, and on the sphere. We use algebraic formulations to provide upper bounds. To find values of the parameters that lead to graphs with a large number of real realizations, possibly attaining the (algebraic) upper bounds, we use some standard heuristics and we also develop a new method inspired by coupler curves. We apply this new method to obtain embeddings in ${\mathbb{R}}^3$. One of its main novelties is that it allows us to sample efficiently from a larger number of parameters by selecting only a subset of them at each iteration.
Our results include a full classification of the 7-vertex graphs according to their maximal numbers of real embeddings in the cases of the embeddings in ${\mathbb{R}}^2$ and ${\mathbb{R}}^3$, while in the case of $S^2$ we achieve this classification for all 6-vertex graphs. Additionally, by increasing the number of embeddings of selected graphs, we improve the previously known asymptotic lower bound on the maximum number of realizations. The methods and the results concerning the spatial embeddings are part of the proceedings of ISSAC 2018 \[[@ISSAC_2018]\].
author:
- Evangelos Bartzos
- Ioannis Emiris
- Jan Legerský
- Elias Tsigaridas
bibliography:
- 'rigid\_journal.bib'
title: 'On the maximal number of real embeddings of minimally rigid graphs in ${\mathbb{R}}^2$, ${\mathbb{R}}^3$ and $S^2$'
---
Introduction
============
Rigidity theory is a very wide area of mathematical research that combines elements of graph theory and algebraic geometry. The numerous applications of rigid graphs in other domains, such as robotics [@Rob1; @Rob2; @Drone], structural bioinformatics [@Em_Ber; @Bio2], sensor network localization [@sensor] and architecture [@arch], give additional motivation to find efficient algorithms to compute them and classify their properties. One of the open problems in rigidity theory is to determine bounds on the maximal number of real embeddings of rigid graphs. We are interested in improving the currently known bounds.
An embedding of a simple graph $G=(V,E)$ in a euclidean space $\mathbb{R}^d$ is a map from the set $V$ to $\mathbb{R}^d$. We require that it satisfies certain edge constraints, namely, the distance between the images of any two adjacent vertices equals a given edge length. Let $\mathbf{p}=(p_1,p_2,\dots p_{n})$ be a configuration of $n=|V|$ points in $\mathbb{R}^d$ and $\bm{\lambda}= \left( \left\lVert p_i-p_j \right\lVert \right)_{ij \in
E}$ be the vector of edge lengths induced by $\mathbf{p}$. The graph $G$ with edge lengths $\bm{\lambda}$ is called *rigid* in $\mathbb{R}^d$ if the number of embeddings in $\mathbb{R}^d$ having the same edge lengths is finite modulo rigid motions. The graph $G$ is *generically rigid* in $\mathbb{R}^d$ iff it is rigid in $\mathbb{R}^d$ for edge lengths induced by any generic configuration. Additionally, if $G$ is *generically rigid* and removing any edge $e \in E$ yields a non-rigid graph $G-e$, then $G$ is called *generically minimally rigid* in $\mathbb{R}^d$.
In the first half of the 20th century, [@Geiringer1932; @Geiringer1927] made notable progress on understanding the properties of minimally rigid graphs, but her work was forgotten. [@Laman] rediscovered that we can fully characterize the minimally rigid graphs in $\mathbb{R}^2$ using the edge count property [@Maxwell]. Since then, these graphs are known as *Laman graphs*. In honor of Hilda Pollaczek-Geiringer, we have chosen to call minimally rigid graphs in $\mathbb{R}^3$ *Geiringer graphs*, as in [@GraKouTsiLower17]. Rigidity is defined also on spheres [@Whiteley_cone]. In the case of $S^2$, the edge count property of Laman graphs holds for minimally rigid graphs, while the distance from the origin poses an additional constraint.
Generically minimally rigid graphs are of great interest since they correspond to well-constrained algebraic systems. Given a rigid graph $G=(V,E)$ in $\mathbb{R}^d$ and edge lengths $\bm{\lambda}=\{\lambda_{ij}\}_{ij\in E}\in {\mathbb{R}}_+^{|E|}$, we denote by $r_d(G,\bm{\lambda})$ the number of embeddings of $G$ in $\mathbb{R}^d$, which are the real solutions of the corresponding algebraic systems. Let $r_d(G)$ denote the maximal number of real embeddings among all the choices of $\bm{\lambda}$ that yield a rigid conformation, i.e., when $r_d(G,\bm{\lambda})$ is finite. The total number of solutions of the corresponding algebraic system in $\mathbb{C}^d$ is the number of complex embeddings of a graph. This gives a natural upper bound for $r_d(G)$ and is denoted by $c_d(G)$. Finally, we write $c_d(n)$ and $r_d(n)$ for the maximal number of complex and real embeddings, respectively, among all $n$-vertex rigid graphs in $\mathbb{C}^d$. We will also use the notation $r_{S^2}(G,\bm{\lambda})$, $r_{S^2}(G)$ and $c_{S^2}(G)$ for the real and complex number of embeddings on $S^2$.
We can use lower and upper bounds on $r_d(G)$ to establish lower and upper bounds on $r_d(n)$ by gluing mechanisms in certain ways [@Borcea2; @GraKouTsiLower17].
#### Previous results
Asymptotic upper bounds for $r_d(n)$ were computed as complex bounds of the determinantal variety of the distance matrix in [@Borcea1; @Borcea2], while mixed volume techniques were applied in the case $d=2$ in [@Steffens]. Both bounds behave asymptotically as $\mathcal{O}(2^{dn})$, which is considered as a rather loose bound. Tighter bounds for specific classes of Laman graphs can be found in [@Jackson1].
Graph-specific approaches have been also used to compute bounds of graph embeddings in $\mathbb{R}^2$ and $\mathbb{R}^3$. Mixed volume techniques [@Emiris1] and a recent combinatorial algorithm for Laman graphs [@Joseph_lam] have treated the complex case. In [@Borcea2], it is proven that $r_2(6) = 24$ using coupler curves and some advanced stochastic methods are applied to show that $r_2(7) = 56$ in [@EM]. The latter yields $2.3003^n$ as a lower bound on maximal number of embeddings for Laman graphs; we improve this bound. The best known lower bound for $r_3(n)$ is $2.51984^n$ [@Emiris1]; we also improve this bound. In general, the maximal number of real embeddings both in $d=2$ and $d=3$ for graphs with $n\leq 6$ vertices is known.
The main question is whether we can specify edge lengths that maximize the number of real embeddings. This question is related to more general open problems in real algebraic geometry concerning possible gaps between the number of complex and real solutions of an algebraic system depending on its parameters. There exist some upper [@Sottile] and lower [@BRS-few-08; @bss-siaga-18] bounds on the number of real positive roots, which take advantage of the structure of polynomials. Regarding applied cases, there is also the famous example on the maximization of the number of real Stewart-Gough Platform configurations [@Diet], using a gradient descent method.
#### Our contribution
We extend the existing results on the maximal number of real embeddings of Laman and Geiringer graphs. We provide bounds in the previously untreated case of spherical embeddings. In both cases, we have constructed all minimally rigid graphs using the methods described in [@GraKouTsiLower17] and we classify them according to the last Henneberg step. Subsequently, we use different systems to model our problem algebraically and we compute upper bounds for all computationally feasible cases. Since our main goal is to maximize the number of real embeddings, we specify the edge lengths in each case using certain heuristics.
In the ISSAC 2018 version we have treated the case of Geiringer graphs. We have developed a new method inspired by coupler curves that can search efficiently huge parametric spaces combining local and global sampling. An open-source implementation of our method is available in [@sourceCode]. This implementation uses the polyhedral homotopy solver PHCpack [@phcpy] to find the solutions of algebraic systems. We are not aware of any other similar method. The method gave the maximal numbers of real embeddings of all 7-vertex Geiringer graphs and improved the existing lower maximal bound from $2.51984^n$ to $2.6390^{n}$ using selected 8-vertex graphs [@ISSAC_2018].
Besides the results announced in ISSAC 2018, we also improve the existing lower bounds on the number of real embeddings of selected Laman graphs on the plane and sphere. In these cases, we use some standard sampling methods to find parameters that maximize the number of embeddings. Our results give the maximal numbers of real embeddings of all 6-vertex and 7-vertex Laman graphs in $S^2$ and ${\mathbb{R}}^2$ respectively. We also specify parameters for larger graphs (up to 10 vertices for the embeddings on the plane and up to 8 vertices for spherical embeddings). These computations improve the existing lower bound on the maximal number of real embeddings from $2.3003^n$ to $2.3811^n$ for $d=2$, while they establish $2.51984^{n}$ as a lower bound for the number of embeddings in $S^2$.
#### Organization
We organize the paper as follows: in Section \[sec:alg-model\], we give a brief introduction on rigidity theory and we describe the algebraic modeling in hand. In Section \[sec:sampling\], we present the sampling methods that we use. Here we describe the method for maximizing the number of real embeddings of Geiringer graphs that is inspired by coupler curves. It previously appeared in [@ISSAC_2018]. In Section \[sec:results\], we present our results in $d=2$, $S^2$, and $d=3$. We derive a new lower bound on the maximal number of real embeddings for the first two cases and we restate the lower bound appeared in the proceedings of ISSAC 2018 for the spatial embeddings. In Section \[sec:conclusion\], we present an overview of our results and some future research problems.
Rigidity and & Algebraic Modeling {#sec:alg-model}
=================================
First, we present some standard results about minimally rigid graphs (Sec. \[sec:rigidity\]). We subsequently introduce the algebraic formulations we use to establish upper and lower bounds on the number of embeddings. In Sec. \[sec:eq-sphere\], we present a variation of the squared distance equations between adjacent vertices, while in Sec. \[sec:distance-sys\] we apply the Cayley-Menger embeddability conditions.
Rigidity and Henneberg steps {#sec:rigidity}
----------------------------
Minimally rigid graphs correspond to well-constrained algebraic systems. The following theorem provides the total number of constraints for a minimally rigid graph and an upper bound on the number of edges of each subgraph (implying that no subsystem is over-constrained).
\[thm:maxwell\] If $G=(V,E)$ is a minimally rigid graph in ${\mathbb{R}}^d$, then the total number of edges is $|E|=d\dot |V|-\binom{d+1}{2}$. Additionally, for each subgraph $G'=(V',E') \subset G$, the inequality $|E'|\leq d\dot |V'|-\binom{d+1}{2}$ holds.
This condition is also sufficient for $d=2$ and the set of these graphs coincide with Henneberg constructions starting from a single edge [@Laman; @Geiringer1927; @Geiringer1932]. On the other hand, there are counter-examples in higher dimensions. This leads to one of the most important open questions in rigidity theory, that is the quest for a combinatorial characterization of minimally rigid graphs in dimension $d\geq 3$ [@handbook1]. Despite this fact, we know that using (extended) Henneberg steps we can construct a superset of minimally rigid graphs in all dimensions.
There are two Henneberg operations that preserve minimal rigidity in any dimension, see Fig. \[fig:henneberg\] [@tay] . The first move consists of adding a new vertex of degree $d$ connecting it with $d$ existing vertices. This step is known as Hennenberg step I (H1) or vertex addition step. The second move consists of deleting an existing edge, then connecting a new vertex with the vertices of the deleted edges and $d-1$ other existing vertices. This step is known as Hennenberg step II (H2) or edge split step.
H1 and H2 steps are equivalent to the edge count property of Theorem \[thm:maxwell\] in $d=2$, so they characterize Laman graphs completely. On the other hand, for $d=3$, two extra steps are required to construct a superset of Geiringer graphs. They are known as Henneberg III (H3) or extended Henneberg steps. The graphs whose construction requires an H3 move have $n\geq 12$ vertices and they are out of the scope of this paper due to computational constraints.
[cc|cccc]{} H1 & H2 & H1 & H2 & H3x & H3v\
\(1) at (0, 0) ; (3) at (0.7, 0.3) ; (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (0.3,-0.9);
\(1) at (0, 0) ; (3) at (0.7, 0.3) ; (w) at (0.65, -0.55) ; (0.3,0.2) circle (0.6cm); (1) to (w); (3) to (w);
\(u) at (1.0,0) ; (u) at (-0.4,0) ;
&
\(1) at (0, 0) ; (2) at (0.35, 0.5) ; (3) at (0.7, 0.3) ; (1) to (2); (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (0.3,-0.9);
\(1) at (0, 0) ; (2) at (0.35, 0.5) ; (3) at (0.7, 0.3) ; (w) at (0.65, -0.55) ; (0.3,0.2) circle (0.6cm); (1) to (w); (2) to (w); (3) to (w);
\(u) at (1.0,0) ; (u) at (-0.4,0) ;
&
\(1) at (0, 0) ; (2) at (0.25, 0.4) ; (3) at (0.7, 0.3) ; (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (0.3,-0.9);
\(1) at (0, 0) ; (2) at (0.25, 0.4) ; (3) at (0.7, 0.3) ; (w) at (0.65, -0.55) ; (0.3,0.2) circle (0.6cm); (1) to (w); (2) to (w); (3) to (w);
\(u) at (1.0,0) ; (u) at (-0.4,0) ;
&
\(1) at (0, 0) ; (2) at (0.35, 0.5) ; (3) at (0.7, 0.3) ; (4) at (0.3,0.0) ; (1) to (2); (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (0.3,-0.9);
\(1) at (0, 0) ; (2) at (0.35, 0.5) ; (3) at (0.7, 0.3) ; (4) at (0.3,0.0) ; (w) at (0.65, -0.55) ; (0.3,0.2) circle (0.6cm); (1) to (w); (2) to (w); (3) to (w); (4) to (w);
\(u) at (1.0,0) ; (u) at (-0.4,0) ;
&
\(1) at (-0.05, 0.1) ; (2) at (0.35, 0.5) ; (3) at (0.7, 0.3) ; (4) at (0.1,-0.2) ; (5) at (0.0,0.45) ; (1) to (3); (2) to (4); (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (0.3,-0.9);
\(1) at (-0.05, 0.1) ; (2) at (0.35, 0.5) ; (3) at (0.7, 0.3) ; (4) at (0.1,-0.2) ; (5) at (0.0,0.45) ; (w) at (0.7, -0.5) ; (0.3,0.2) circle (0.6cm); (1) to (w); (2) to (w); (3) to (w); (4) to (w); (5) to (w);
\(u) at (1.0,0) ; (u) at (-0.4,0) ;
&
\(1) at (-0.05, 0.1) ; (2) at (0.4, 0.55) ; (3) at (0.7, 0.3) ; (4) at (0.1,-0.2) ; (5) at (0.0,0.45) ; (1) to (2); (2) to (4); (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (0.65,-0.9);
\(1) at (-0.05, 0.1) ; (2) at (0.4, 0.55) ; (3) at (0.7, 0.3) ; (4) at (0.1,-0.2) ; (5) at (0.0,0.45) ; (1) to (5); (5) to (3); (0.3,0.2) circle (0.6cm); (0.3,-0.5) – (-0.05,-0.9);
\(1) at (-0.05, 0.1) ; (2) at (0.4, 0.55) ; (3) at (0.7, 0.3) ; (4) at (0.1,-0.2) ; (5) at (0.0,0.45) ; (w) at (0.7, -0.5) ; (0.3,0.2) circle (0.6cm); (1) to (w); (2) to (w); (3) to (w); (4) to (w); (5) to (w);
\
& & & & &\
& & & & &\
Laman graphs can be embedded also on the sphere $S^2$. In that case, we need to add an additional constraint, which is the distance between the center of the sphere and the vertices. This means that spherical n-vertex minimally rigid graphs could be seen as minimally rigid graphs in ${\mathbb{R}}^3$ with $n+1$ vertices, one of which has degree $n$ [@Whiteley_cone].\
Let us notice that if there is a way to construct a graph $G_{n+1}$ by applying an H1 move to $G_{n}$, then the number of embeddings is doubled, i.e., $c_d(G_{n+1})=2c_d(G_{n})$ and $r_d(G_{n+1})=2r_d(G_{n})$. On the other hand, experiments show that the effect of other Henneberg steps on the number of embeddings varies significantly depending on a graph [@GraKouTsiLower17]. Therefore, we classify the minimally rigid graphs according to the possible last Henneberg moves. This can be translated into a minimum degree condition, since if there is a vertex of degree $d$, then the graph can be constructed by an H1 move in the last step. We consider the graphs with at least one vertex of degree $d$ as graphs, whose number of embeddings can be trivially obtained from a smaller graph, and we will use the term *H1-last* for them. We will also use the term *H2-last* for graphs with all vertices of degree at least $d+1$.
Equations of spheres {#sec:eq-sphere}
--------------------
In this section we define a set of equations to compute the embeddings of a graph. The equations are of two kinds. The first one corresponds to the squared distance between adjacent vertices. Although this set of *edge equations* suffices to find the embeddings of a graph, the mixed volume of this system is much bigger than the actual number of complex embeddings. This is not favorable for homotopy continuation polynomial solvers, which give us the fastest method to compute the embeddings. In order to overcome this problem, we use the *magnitude equations* that introduce new variables as the distance of each vertex from the origin [@Steffens; @Emiris1]. In that way, mixed volume can be significantly lowered.
\[def:magnitudeEquations\] Let $G=(V,E)$ be a graph with given edge lengths $\bm{\lambda}=(\lambda_e)_{e\in E}\in {\mathbb{R}}_+^{|E|}$ and $X_u=(x_{u1},x_{u2},\dots x_{ud})$ be the variables assigned to the coordinates of each vertex. If the graph contains a complete subgraph with $d$ vertices $v_1,v_2,\dots v_d $, then we can choose the coordinates of this $d$-simplex in a way that they satisfy the edge lengths of this subgraph. We define $S(G,\bm{\lambda},\left[v_1,v_2,\dots v_d\right])\subset
\mathbb{C}^{d\cdot |V|}$ as the solutions of the following equations $$\begin{aligned}
\lVert X_u \rVert^2 -s_u&=0 \,\,\forall u \in V\,, \\
s_u +s_v -2\langle X_u,X_v \rangle -\lambda_{uv}^2&=0 \,\, \forall uv \in E\,,
\end{aligned}$$ such that the $d$-simplex is fixed. We denote the real solutions $S(G,\bm{\lambda},\left[v_1,v_2,\dots v_d\right])\cap{\mathbb{R}}^{d\cdot |V|}$ by $S_{\mathbb{R}}(G,\bm{\lambda},\left[v_1,v_2,\dots v_d \right] )$.
Fixing the coordinates of the $d-$simplex, rotations and translations are removed from the set of solutions yielding a $0-$dimensional system. In the case of Laman graphs, we fix $v_1$ in the origin and $v_2$ in the $y-$axis with coordinates $(0,\lambda_{1,2})$. In the case of Geiringer graphs, we fix again $v_1$ in the origin and $v_2$ in the $y-$axis with coordinates $(0,\lambda_{1,2},0)$. The vertex $v_3$ is on the plane $z=0$, with $z_3\geq 0$. Finally, in the case of spherical embeddings, we consider the extension to a Geiringer graph and we use the analogous equations fixing 2 points on $(0,0,1)$ and $(0,\sqrt{(1-\cos (\theta_{1,2})^2)}, \cos (\theta_{1,2}))$, where $\theta_{1,2}$ is the angle between the vector of $v_1$ and $v_2$.
The edge equations express the geometrical constraints of the graph, while the magnitude equations are used to avoid roots at toric infinity, resulting to tighter mixed volume [@Emiris1; @Steffens]. At this point we should remark that the mixed volume of sphere equations depends on the choice of the fixed $d-$simplex, so we computed all possible combinations to get the tightest bound.
We should also comment that $|S(G,\bm{\lambda},[v_1,v_2,\dots v_d])|$ coincides with $c_d(G)$ for a generic choice of lengths $\bm{\lambda}$ and that $r_d(G,\bm{\lambda})=|S_{\mathbb{R}}(G,\bm{\lambda},\left[v_1,v_2,\dots
v_d\right])|$ for arbitrary $\bm{\lambda}$.
Distance systems {#sec:distance-sys}
----------------
A Cayley-Menger matrix is the matrix of squared distances extended by a row and column of ones (except for the diagonal which is always zero): $$CM=\begin{pmatrix}
0 & 1 & 1 & \cdots & 1\\ \vspace{-0.3em}
1& 0 & \lambda^2_{12} & \cdots & \lambda^2_{1n} \\ \vspace{-0.3em}
1& \lambda^2_{12} & 0 & \ddots & \ldots \\
\cdots & \cdots & \ddots & \ddots & \ldots \\
1& \lambda^2_{1n} & \lambda^2_{2n} & \cdots & 0
\end{pmatrix}\,,$$ where $\lambda_{ij}$ is the distance between point $i$ and $j$.\
A fundamental result in distance geometry indicates the following embeddability condition [@Blu]:\
The squared distances of a CM matrix can be embedded in ${\mathbb{R}}^d$ iff
- $\text{rank}(CM)=d+2$
- $(-1)^k \det(CM')\geq 0$, for every submatrix $CM'$ with size $k+1\leq d+2$ that includes the extending row/column.
In the case of graph embeddings, each known entry corresponds to a squared edge length, while the variables correspond to unknown edge lengths. Any solution of the semi-algebraic system is an embedding of the graph in ${\mathbb{R}}^d$ up to isometries. Considering only the solutions of the determinantal variety, we get the complex embeddings of the graph. The set of inequalities correspond to certain geometrical constraints on the edge lengths, such as positivity and triangular inequalities in dimension 2. In dimension 3, *tetrangular inequalities* (which are a generalization of triangular inequalities on the area of the triangles of a tetrahedron) should be also satisfied [@Dattorro].
The systems of equations of determinantal varieties are overconstrained. For example, there are $35$ equations in $10$ variables for 7-vertex Laman graphs, while for 7-vertex Geiringer graphs, there are $21$ equations in $6$ variables. Despite this fact, it is possible to find zero-dimensional square subsystems of these systems of equations [@Emiris1; @EM]. Notice that the zero set of the whole determinantal variety corresponds to the missing edge lengths of the complete graph. This means that the solutions of the subsystem correspond to a superset of the missing edge lengths. If the graph extended by the edges corresponding to the variables of the subsystem is *globally rigid* (a rigid graph with a unique embedding up to isometries), then the subsystem gives an upper bound on the number of embeddings of the whole graph [@Jackson2]. In dimension $2$ there is a combinatorial characterization for globally rigid graphs [@Connelly], while for arbitrary dimension we can check it using the rank of stress matrices of rigidity matroids [@Global].
In our research, it was easy to detect square subsystems, if no restriction was imposed on the number of variables. The point was to find the optimal ones in the sense that they would be 0-dimensional and serve to find the embeddings of a graph (so they should have exactly the same number of complex solutions as the whole variety) or a useful upper bound. Throughout our experiments, we found out that subsystems with $n-(d+1)$ equations can meet this requirements. Additionally, the following lemma shows that for $d=2$ and $d=3$, there is always an extension of a minimally rigid graph with $n-(d+1)$ edges that results to a globally rigid graph (the version of this lemma for $d=3$ appears also in [@ISSAC_2018]).
For every minimally rigid graph $G=(V,E)$ in dimensions $d=2$ and $d=3$, there is at least one extended graph $H=G\cup \{e_1, e_2,..,e_k\}$, with $k=n-(d+1)$ and $e_i \notin E$, which is globally rigid in $\mathbb{C}^d$.
The only 4-vertex minimally rigid graph in dimension 2 (resp. 5-vertex in dimension 3) is obtained by applying an H1 step to the triangle (resp. tetrahedron in dimension 3). If we extend this graph with the only non-existing edge, we obtain a complete graph, so the lemma holds. Let the lemma hold for all graphs with $n$ or less vertices. H2 steps are known to preserve global rigidity [@Connelly]. So we need to prove the lemma for H1 steps in both dimensions and H3 steps in dimension 3.
Let a Laman graph $G_{n+1}$ be constructed by an H1 move applied to an $n$-vertex graph $G_n$, whose extended globally rigid graph is $H_n$. Without loss of generality, this move connects a new vertex $v_{n+1}$ with vertices $v_1,v_2$. Let $u$ be a neighbour of $v_1$ in $G_{n+1}$ not such that $v_2\neq u$. The edge $uv_1$ exists also in $G_{n}$ and $H_n$. If we set $H'_{n+1}=(H_{n}\cup \{v_1v_{n+1}, v_2v_{n+1}, uv_{n+1}\})-\{v_1u\}$, then $H'_{n+1}$ is globally rigid, because it is constructed from $H_n$ by an H2 step. Hence, $H_{n+1}=H'_{n+1}\cup \{ uv_{1}\}$ is also globally rigid, proving the statement in the case of H1 steps in dimension 2. The same result holds in arbitrary dimension (see Figure \[fig:globalRigidityH1\] for $d=3$).
Both H3 steps consist of an H2 step followed by a second edge deletion in the existing graph and a new connection with $v_{n+1}$. So, if we apply an $H3$ move in $H_n$ and subsequently add the second deleted edge, then $H_{n+1}$ is globally rigid.
[cc]{}
\(1) at (-0.2, 0.2) ; (2) at (0.15, 0.5) ; (u) at (0.3,-0.15) ; (1) to (u); (0.375,0.2) ellipse (1.15cm and 0.65cm); (w) at (1.75, -0.1) ; (1) to (w); (2) to (w); (u) to (w); at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (u) [$u$]{}; at (w) [$v_{n+1}$]{};
&
\(1) at (-0.2, 0.2) ; (2) at (0.15, 0.5) ; (3) at (1.0, 0.3) ; (u) at (0.3,-0.15) ; (1) to (u); (0.375,0.2) ellipse (1.15cm and 0.65cm); (w) at (1.75, -0.1) ; (1) to (w); (2) to (w); (3) to (w); (u) to (w); at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (3) [$v_3$]{}; at (u) [$u$]{}; at (w) [$v_{n+1}$]{};
\
${\mathbb{R}}^2$ & ${\mathbb{R}}^3$
Although we proved that there are always globally rigid extentions with $n-(d+1)$ supplementary edges, it is not always possible to find a Cayley-Menger subvariety corresponding to them. We could detect such subsystems for all graphs with $n\leq 7$ vertices in both dimensions, but there exist bigger graphs for which this property does not hold.
\(6) at (0, 0); (5) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, 0) ; (2) at (1.25,2) ; (1) at (1.25,0.7) ; (7) at (1.5,-1) ;
(1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(1); (5)edge(2) (5)edge(6) (5)edge(3) (1)edge(7) (4)edge(7); (6)edge(7); (1)edge(3) (1)edge(5) (3)edge(6) (3)edge(7);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (3) [$v_3$]{}; at (4) [$v_4$]{}; at (5) [$v_5$]{}; at (6) [$v_6$]{}; at (7) [$v_7$]{};
We will give some representative examples of optimal CM subsystems in the cases of Laman, Geiringer and spherical graphs. For instance, $L_{48H2}$ is a 7-vertex Laman graph (see Figure \[fig:L48\]), which has $c_2(L_{48H2})=r_2(L_{48H2})=48$ and $c_{S^2}(L_{48H2})=r_{S^2}(L_{48H2})=64$ (See Section \[sec:results\]). There are $11$ subsystems of this CM variety in $4$ variables, which all have exactly the same number of solutions. In the following CM matrix, we present one of these choices involving the variables $x_1,x_2,x_6$ and $x_7$. $$CM_{L_{48H2}}=\begin{pmatrix}
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 0 & \lambda_{12}^2 & \Red{x_{1}} & \lambda_{14}^2 & \Red{x_{2}} & \lambda_{16}^2 & \lambda_{17}^2 \\
1 & \lambda_{12}^2 & 0 & \lambda_{23}^2 & \Blue{x_{3}} & \lambda_{25}^2 & \Blue{x_{4}} & \Blue{x_{5}} \\
1 & \Red{x_{1}} & \lambda_{23}^2 & 0 & \lambda_{34}^2 & \lambda_{35}^2 & \Red{x_{6}} & \Red{x_{7}} \\
1 & \lambda_{14}^2 & \Blue{x_{3}} & \lambda_{34}^2 & 0 & \Blue{x_{8}} & \Blue{x_{9}} & \lambda_{47}^2 \\
1 & \Red{x_{2}} & \lambda_{25}^2 & \lambda_{35}^2 & \Blue{x_{8}} & 0 & \lambda_{56}^2 & \Blue{x_{10}} \\
1 & \lambda_{16}^2 & \Blue{x_{4}} & \Red{x_{6}} & \Blue{x_{9}} & \lambda_{56}^2 & 0 & \lambda_{67} \\
1 & \lambda_{17}^2 & \Blue{x_{5}} & \Red{x_{7}} & \lambda_{47}^2 & \Blue{x_{10}} & \lambda_{67} & 0
\end{pmatrix}$$ In order to compute the number of real embeddings, we need to find the semi-algebraic set containing the positive solutions of this system that satisfy the triangular inequalities.\
We can use the same extended graph to compute the spherical embeddings of $L_{48H2}$. An additional constraint is needed in that case, which represents the distance from the origin, as a new column and row with ones. The determinantal subsystem is derived from the rank condition of 3-dimensional embeddings. Elementary matrix operations can lead to a formulation that considers the cosines of the angles between two points as matrix entries, denoted as $c_{ij}$. $$\begin{aligned}
CM_{S^2(L_{48H2})}&=\begin{pmatrix}
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 0 & \lambda_{12}^2 & \Red{x_{1}} & \lambda_{14}^2 & \Red{x_{2}} & \lambda_{16}^2 & \lambda_{17}^2 & 1\\
1 & \lambda_{12}^2 & 0 & \lambda_{23}^2 & \Blue{x_{3}} & \lambda_{25}^2 & \Blue{x_{4}} & \Blue{x_{5}} & 1\\
1 & \Red{x_{1}} & \lambda_{23}^2 & 0 & \lambda_{34}^2 & \lambda_{35}^2 & \Red{x_{6}} & \Red{x_{7}} & 1\\
1 & \lambda_{14}^2 & \Blue{x_{3}} & \lambda_{34}^2 & 0 & \Blue{x_{8}} & \Blue{x_{9}} & \lambda_{47}^2 & 1\\
1 & \Red{x_{2}} & \lambda_{25}^2 & \lambda_{35}^2 & \Blue{x_{8}} & 0 & \lambda_{56}^2 & \Blue{x_{10}} & 1\\
1 & \lambda_{16}^2 & \Blue{x_{4}} & \Red{x_{6}} & \Blue{x_{9}} & \lambda_{56}^2 & 0 & \lambda_{67} & 1\\
1 & \lambda_{17}^2 & \Blue{x_{5}} & \Red{x_{7}} & \lambda_{47}^2 & \Blue{x_{10}} & \lambda_{67} & 0 & 1\\
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 0 \\
\end{pmatrix} \\
&\sim
\begin{pmatrix}
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 2 \\
1 & 0 & c_{12} & \Red{y_{1}} & c_{14} & \Red{y_{2}} & c_{16} & c_{17} & 1\\
1 & c_{12} & 0 & c_{23} & \Blue{y_{3}} & c_{25} & \Blue{y_{4}} & \Blue{y_{5}} & 1\\
1 & \Red{y_{1}} & c_{23} & 0 & c_{34} & c_{35} & \Red{y_{6}} & \Red{y_{7}} & 1\\
1 & c_{14} & \Blue{y_{3}} & c_{34} & 0 & \Blue{y_{8}} & \Blue{y_{9}} & c_{47} & 1\\
1 & \Red{y_{2}} & c_{25} & c_{35} & \Blue{y_{8}} & 0 & c_{56} & \Blue{y_{10}} & 1\\
1 & c_{16} & \Blue{y_{4}} & \Red{y_{6}} & \Blue{y_{9}} & c_{56} & 0 & c_{67} & 1\\
1 & c_{17} & \Blue{y_{5}} & \Red{y_{7}} & c_{47} & \Blue{y_{10}} & c_{67} & 0 & 1\\
-2 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 0 \\
\end{pmatrix} \end{aligned}$$ The semi-algebraic conditions of the latter formulation, requires that any solution of the determinantal subsystem lies in the interval $[-1,1]$ and that the triangular inequalities on the sphere are satisfied. The second is equivalent to the positivity of $2c_{ij}c_{ik}c_{jk}-c_{ij}^2-c_{ik}^2-c_{jk}^2+1$ for 3 points $i,j,k$ on the sphere, where $c_{ij}$ is the cosine of the angle between points $i \text{ and } j$ and can be obtained as the determinant of a 5x5 submatrix containing both columns and rows with ones.\
We take the graph $G_{48}$, see Figure \[fig:G48\] as an example of CM subvarieties of Geiringer graphs (which was also used in [@ISSAC_2018]). The graph $G_{48}$ has the maximal number of embeddings among all 7-vertex Geiringer graphs ($c_3(G_{48})=r_3(G_{48})=48$, see Section \[sec:results\]). There are 5 different square systems in $3$ variables that completely define the embeddings. We can choose one of them involving only $x_1, x_2, x_3$: $$CM_{G_{48}}=\begin{pmatrix}
0&1&1&1&1&1&1&1 \\
1 & 0 & \lambda^2_{12}& \lambda^2_{13}& \lambda^2_{14}&\lambda^2_{15} & \lambda^2_{16} & \Red{x_{1}} \\
1 & \lambda^2_{21} & 0 & \lambda^2_{23} & \Red{x_{2}} & \Red{x_{3}} & \lambda^2_{26} & \lambda^2_{27} \\
1& \lambda^2_{31} & \lambda^2_{32} & 0 & \lambda^2_{34} & \Blue{x_{4}} & \Blue{x_{5}} & \lambda^2_{37} \\
1& \lambda^2_{41} & \Red{x_{2}}& \lambda^2_{43} & 0& \lambda^2_{45} & \Blue{x_{6}} & \lambda^2_{47}\\
1& \lambda^2_{51} & \Red{x_{3}} & \Blue{x_{4}} & \lambda^2_{54}& 0& \lambda^2_{56}& \lambda^2_{57} \\
1& \lambda^2_{61}& \lambda^2_{62} & \Blue{x_{5}} & \Blue{x_{6}} & \lambda^2_{65}& 0& \lambda^2_{67} \\
1&\Red{x_{1}}& \lambda^2_{72} & \lambda^2_{73} & \lambda^2_{74} & \lambda^2_{75} & \lambda^2_{76}& 0
\end{pmatrix}$$ The set of real embeddings in that case is given by the solutions of the subsystem that satisfy positivity, triangular and tetrangular inequalities.
Extending this graph with the edge $v_1v_7$ suffices for global rigidity. This edge corresponds to the variable $x_1$ and it is possible to get a single equation by applying resultants in the 3x3 system of determinantal equations (see Figure \[fig:G48\]).
[ (1) at (0, -1); (2) at (-1.9, 0); (3) at (-0.9, -0.3) ; (4) at (0.85, -0.3) ; (5) at (1.8,0.0) ; (6) at (-0.25, 0.25) ; (7) at (0,1) ;]{} (2)edge(3) (3)edge(4) (6)edge(2) (5)edge(4) (6)edge(5); (2)edge(1) (1)edge(4) (1)edge(3) (1)edge(5) (1)edge(6); (2)edge(7) (7)edge(4) (7)edge(3) (7)edge(5) (7)edge(6); (2)edge(4) (4)edge(6) (1)edge(7); at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (3) [$v_3$]{}; at (4) [$v_4$]{}; at (5) [$v_5$]{}; at (-0.3, 0.18) [$v_6$]{}; at (0,1) [$v_7$]{};
[ (1) at (0, -1); (2) at (-1.9, 0); (3) at (-0.9, -0.3) ; (4) at (0.85, -0.3) ; (5) at (1.8,0.0) ; (6) at (-0.25, 0.25) ; (7) at (0,1) ;]{} (2)edge(3) (3)edge(4) (6)edge(2) (5)edge(4) (6)edge(5); (2)edge(1) (1)edge(4) (1)edge(3) (1)edge(5) (1)edge(6); (2)edge(7) (7)edge(4) (7)edge(3) (7)edge(5) (7)edge(6); (1)edge(7); at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (3) [$v_3$]{}; at (4) [$v_4$]{}; at (5) [$v_5$]{}; at (-0.3, 0.18) [$v_6$]{}; at (7) [$v_7$]{}; at (0.0,-0.07) ;
Since a single edge is needed to find the whole embedding, we can use only the inequalities involving only this variable (5 triangular and 5 tetrangular inequalities instead of 35 that involve all variables).\
Increasing the number of real embeddings {#sec:sampling}
========================================
Our main goal throughout our experiments was to find the parameters that can maximize the number of real embeddings of minimally rigid graphs. One open problem in rigidity theory is whether the maximal number of real embeddings of a given graph can be the same as the number of complex embeddings. Although there exists an 8-vertex Laman graph for which it has been proven that $r_2(G)<c_2(G)$ [@Jackson2], in most cases we consider the number of complex embeddings as the upper bound we try to reach. In our research, we were mostly concentrated on the cases of graphs with the biggest number of complex embeddings, among all other minimally rigid graphs with the same number of vertices.\
Additionally to some standard sampling methods, we developed a new method that can increase efficiently the number of real embeddings for certain Geiringer graphs, which was initially introduced in ISSAC 2018. Our method is inspired by coupler curves approach and uses $G_{48}$ as a model. Taking advantage of our implementation based on this technique, we were able to increase lower bounds on $r_3(G)$ for many graphs and establish new asymptotic lower bounds on the maximal number of embeddings of Geiringer graphs.
Standard sampling methods
-------------------------
#### Finding initial configurations
We applied different heuristics to find initial configurations for our parameter sampling. First of all, we tried to compute the number of real embeddings of totally random configurations. This resulted in finding maximal numbers of real embeddings for graphs with $c_d(G)= 2^{n-(d+1)}$. For example, it took less than 20 minutes to detect parameters that attain the maximum for all 8-vertex H2-last Geiringer graphs with $c_3(G)=32$.\
We also used almost degenerate locus as starting points. In order to increase $r_2(G)$ of Laman graphs with maximal numbers of complex embeddings w.r.t. a given number of vertices, we chose lengths very close to the unit length. Similarly, in the case of Geiringer graphs, we perturbed degenerate conformations. For example, in order to find an initial point for $G_{48}$, we separate the edges into three sets with edge lengths being the same in each of them: the *ring edges* of the 5-cycle, the *top edges* that connect $v_7$ with the ring and the *bottom edges* that connect $v_1$ with the ring — see Figure\[fig:G48\]. We subsequently found edge lengths that maximized the intervals imposed by triangular and tetrangular inequalities up to scaling and we perturbed the resulting lengths.\
Finally, we also used as starting points conformations of smaller graphs with maximal numbers of embeddings. For instance, gluing $v_7$ and $v_8$ in $G_{160}$ results in $G_{48}$. Perturbing a labeling $\bm{\lambda}$ of $G_{48}$ such that $r_3(G_{48},\bm{\lambda})=48$, we could get a starting point for the sampling of $G_{160}$ that would result in a big number of real embeddings.
[ (1) at (0, -1); (2) at (-1.9, 0); (3) at (-0.9, -0.3) ; (4) at (0.85, -0.3) ; (5) at (1.8,0.0) ; (6) at (-0.25, 0.25) ; (7) at (0,1) ;]{} (2)edge(3) (3)edge(4) (6)edge(2) (5)edge(4) (6)edge(5); (2)edge(1) (1)edge(4) (1)edge(3) (1)edge(5) (1)edge(6); (2)edge(7) (7)edge(4) (7)edge(3) (7)edge(5) (7)edge(6);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (3) [$v_3$]{}; at (4) [$v_4$]{}; at (5) [$v_5$]{}; at (-0.3, 0.18) [$v_6$]{}; at (0,1) [$\Red{v_7}$]{};
[ (1) at (0, -1); (2) at (-1.9, 0); (3) at (-0.9, -0.3) ; (4) at (0.85, -0.3) ; (5) at (1.8,0.0) ; (6) at (-0.3, 0.3) ; (7) at (-0.7,1) ; (8) at (0.7,1) ;]{} (1) to (2) (2) to (7) (4) to (7) (2) to (6) (6) to (8) (4) to (5) (2) to (8) (5) to (7) (3) to (4) ; (1) to (4) (1) to (5) (1) to (3) (1) to (6) (5) to (6) (3) to (7) (7) to (8) (2) to (3) (5) to (8) ; at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (8) ;
at (1) [$v_1$]{}; at (2) [$v_2$]{}; at (3) [$v_3$]{}; at (4) [$v_4$]{}; at (5) [$v_5$]{}; at (0.37, -0.27) [$v_6$]{}; at (7) [$\Red{v_7}$]{}; at (8) [$\Red{v_8}$]{};
#### Stochastic methods
We have used stochastic methods for different graphs in order to increase the number of embeddings. Our method uses a variant of the tools suggested in [@EM]. We penalize the loss of real roots and the increase of the imaginary part of complex solutions to decide if the resulted labeling constitutes a new starting point. This method could increase the number of embeddings, but rarely attained the maximum.
#### Parametric searching with CAD method {#sec:parametric}
The methods described in the previous paragraph are local methods. In order to search globally one parameter, we used `Maple`’s subpackage `RootFinding` \[Parametric\] in Maple18. This package is an implementation of Cylindrical Algebraic Decomposition principles for semi-algebraic sets. The input consists of the equations and the inequalities of the system and the list of variables separating them from parameters. The output is a cell decomposition of the space of parameters according to the number of solutions of the semi-algebraic conditions.
In our problem, we were able to take advantage of this implementation using distance systems of 7-vertex graphs and searching for only one parameter. Sphere equations failed to give any result, while computational constraints did not let us search two or more parameters simultaneously.
In [@ISSAC_2018] we use this sampling method to increase the number of real embeddings of $G_{48}$. This sampling was also used to increase the number of spherical and planar embeddings of Laman graphs with 7 vertices. In some situations it was even possible to attend the maximal number of embeddings for a given graph.
Coupler curve {#sec:coupler}
-------------
The previous methods fail to attain tight bounds for Geiringer graphs with maximal number of embeddings efficiently. For example, using CAD, we could find 28 real embeddings for $G_{48}$, but it seems impossible to increase this number by local searching in all parameters or global sampling only one of them. Thus, we developed a new method that samples only subset of edge lengths in every iteration. This procedure is motivated by visualization of coupler curves. We remark that we already presented this method at ISSAC’18.\
Let $G=(V,E)$ be a minimally rigid graph with a triangle and an edge $uc$. If $H=(V,E\setminus {uc})$ is obtained from $G$ by removing the edge $uc$, then the set of embeddings satisfying the constraints given by generic edge lengths and fixing the triangle is 1-dimensional. The projection of this curve to the coordinates of the vertex $c$ is a so called *coupler curve*. [@Borcea2] used this idea for proving that the Desargues (3-prism) graph has 24 real embeddings in $\mathbb{R}^2$. Namely, they found edge lengths such there are 24 intersections of the coupler curve with a circle representing the removed edge. This approach can be clearly extended into $\mathbb{R}^3$ — the number of embeddings of $G$ is the same as the number of intersection of the coupler curve of $c$ with the sphere centered at $u$ with a radius $\lambda_{uc}$. Now, we define specifically a coupler curve in ${\mathbb{R}}^3$.
\[def:couplerCurve\] Let $H$ be a graph with edge lengths $\bm{\lambda}=(\lambda_e)_{e\in E_H}$ and $v_1,v_2,v_3 \in V_H$ be such that $v_1v_2,v_2v_3,v_1v_3\in E_H$. If the set $S_{\mathbb{R}}(H,\bm{\lambda},[v_1,v_2,v_3])$ is one dimensional and $c\in V_H$, then the set $$\mathcal{C}_{c,\bm{\lambda}}=\{(x_c,y_c,z_c) \colon ((x_v,y_v,z_v))_{v\in V_H} \in S_{\mathbb{R}}(H,\bm{\lambda},v_1v_2v_3)\}$$ is called a *coupler curve of $c$ w.r.t. the fixed triangle $v_1v_2v_3$*.
Assuming that a coupler curve is fixed, i.e., we have fixed lengths $\bm{\lambda}$ of the graph $H$, we can change the edge length $\lambda_{uc}$ so that the number of intersections of the coupler curve $\mathcal{C}_{c,\bm{\lambda}}$ with the sphere with the center at $u$ and radius $\lambda_{uc}$, namely, the number of real embeddings of $G$, is maximal.
The following lemma shows that we can change three more edge lengths within one parameter family without changing the coupler curve. This one parameter family corresponds to shifting the center of the sphere along a line.
\[lem:couplerCurvePreserves\] Let $G=(V,E)$ be a minimally rigid graph and $u,v,w,p,c$ be vertices of $G$ such that $pv,vw \in E$ and the neighbours of $u$ in $G$ are $v,w,p$ and $c$. Let $H$ be the graph given by $(V_H,E_H)=(V,E\setminus\{uc\})$ with generic edge lengths $\bm{\lambda}=(\lambda_e)_{e\in E_H}$. Let $\mathcal{C}_{c,\bm{\lambda}}$ be the coupler curve of $c$ w.r.t. the fixed triangle $vuw$. Let $z_p$ be the altitude of $p$ in the triangle $uvp$ with lengths given by $\bm{\lambda}$. Then the set $\{y_p \colon ((x_{v'},y_{v'},z_{v'}))_{v'\in V_H} \in S_{\mathbb{R}}(H,\bm{\lambda},vuw)\}$ has only one element $y'_p$. If the parametric edge lengths $\bm{\lambda}'(t)$ are given by $$\begin{aligned}
\lambda'_{uw}(t)&=||(x_w,y_w-t,0)||\,, \quad \lambda'_{up}(t)=||(0,y'_p-t,z_p)||\,, \\
\lambda'_{uv}(t)&= t\,, \,\text{ and }
\lambda'_{e}(t)=\lambda_{e} \text{ for all } e\in E_H\setminus\{uv,uw,up\}\,,
\end{aligned}$$ then the coupler curve $\mathcal{C}_{c,\bm{\lambda'}(t)}$ of $c$ w.r.t. the fixed triangle $vuw$ is the same for all $t\in\mathbb{R}_+$, namely, it is $\mathcal{C}_{c,\bm{\lambda}}$. Moreover, if $cw\in E$, then $\mathcal{C}_{c,\bm{\lambda}}$ is a spherical curve.
All coupler curves in the proof are w.r.t. the triangle $vuw$. Figure \[fig:couplerCurvePreserved\] illustrates the statement. Since $G$ is minimally rigid, the set $S_{\mathbb{R}}(H,\bm{\lambda},vuw)$ is 1-dimensional. The coupler curve $\mathcal{C}_{p,\bm{\lambda}}$ of $p$ is a circle whose axis of symmetry is the $y$-axis. Hence, the set $\{y_p \colon ((x_{v'},y_{v'},z_{v'}))_{v'\in V_H} \in S_{\mathbb{R}}(H,\bm{\lambda},vuw)\}$ has indeed only one element. The parametrized edge lengths $\bm{\lambda}'(t)$ are such that the position of $v$ and $w$ is the same for all $t$. Moreover, the coupler curve $\mathcal{C}_{p,\bm{\lambda'}(t)}$ of $p$ is independent of $t$. Hence, the coupler curve $\mathcal{C}_{c,\bm{\lambda'}(t)}$ is independent of $t$, because the only vertices adjacent to $u$ in $H$ are $p,v$ and $w$, Thus, the positions of the other vertices are not affected by the position of $u$.
(0,0) to (0,1.4) ; (0,0) to (2.7,0) ; (0,0) to (-1,-1) ;
at (-1.0,-0.9) [$x$]{}; at (2.7,0) [$y$]{}; at (0,1.4) [$z$]{};
\(v) at (0,0) ; (u) at (1.8,0) ; (w) at (1.5,-0.8) ; (p) at (0.8,1) ; (ut) at (1.33,0) ; (c) at (2,1) ; (s) at (0.8,0); (s) ellipse (0.353cm and 1cm);
\(s) to (p); at (0.8,-0.9) [$\mathcal{C}_{p,\bm{\lambda}}$]{};
(2.3,0) to (w); (-0.8,-0.8) to (w); (2.3,-0.05) to (2.3,0.05); (-0.8,-0.85) to (-0.8,-0.75); at (-0.8,-0.8) [$x_w$]{}; at (2.3,0) [$y_w$]{};
\(p) to (v); (v) to (u); (v) to (w); (u) to (w); (p) to (u); (ut) to (w); (p) to (ut); (v) to (ut); (u) to (c);
\(c) .. controls +(-0.2,0.1) .. (1.6,1.3); (c) .. controls +(0.2,-0.1) .. (2.4,0.6); at (2.4,0.6) [$\mathcal{C}_{c,\bm{\lambda}}$]{};
at (u) ; at (p) ;
at (v) [$v$]{}; at (u) [$u$]{}; at (ut) [$\!\!\!\!u(t)$]{}; at (w) [$w$]{}; at (p) [$p$]{}; at (s) [$y'_p$]{}; at (c) [$c$]{}; at (0.65,0.4) [$z_p$]{}; (p) – +(0.33,0.25); (p) – +(0.05,0.33); (w) – +(-0.33,-0.25); (w) – +(-0.05,-0.33); (w) – +(0.2,-0.25); (c) – +(-0.2,0.25); (c) – +(0.05,-0.33); (c) – +(-0.2,-0.25); (v) – +(-0.32,0.15); (v) – +(-0.05,0.3); (v) – +(0.2,-0.25);
Therefore, for every subgraph of $G$ induced by vertices $u,v,w,p,c$ such that $\deg(u)=4$ and $pv,vw, uv,uw,up, uc \in E$, we have a 2-parametric family of lengths $\bm{\lambda}(t,r)$ such that the coupler curve $\mathcal{C}_{c,\bm{\lambda}(t,r)}$ w.r.t. the fixed triangle $vuw$ is independent of $t$ and $r$. Recall that the parameter $r$ represents the length of $uc$, which corresponds to the radius of the sphere, and the parameter $t$ determines the lengths of $uv,uw$ and $up$. Now, we aim to find $r$ and $t$ such that $r_3(G,\bm{\lambda}(t,r))$ is maximized.
Let us clarify that whereas [@Borcea2] were changing the coupler curve, our approach is different in the sense that the coupler curve is preserved within one step of our method, while the position and radius of the sphere corresponding to the removed edge are changed in order to have as many intersections as possible. In the next step, we pick a different edge to be removed. We discuss in Section \[subsubsec:combinationOfSubgraphs\], how these steps are combined for various subgraphs.
In order to illustrate the method, let $\bm{\lambda}$ be edge lengths of $G_{48}$ given by [ $$\begin{aligned}
\lambda_{12} &= 1.99993774567597 \,, & \lambda_{27} &= 10.5360917228793 \,,& \lambda_{23} &= 0.99961432208948 \,,\\
\lambda_{13} &= 1.99476987780024 \,, & \lambda_{37} &= 10.5363171636461 \,,& \lambda_{34} &= 1.00368644488060 \,,\\
\lambda_{14} &= 2.00343646098439 \,, & \lambda_{47} &= 10.5357233031495 \,,& \lambda_{45} &= 1.00153014850485 \,,\\
\lambda_{15} &= 2.00289249524296 \,, & \lambda_{57} &= 10.5362736599978 \,,& \lambda_{56} &= 0.99572361653574 \,,\\
\lambda_{16} &= 2.00013424746814 \,, & \lambda_{67} &= 10.5364788463527 \,,& \lambda_{26} &= 1.00198771097407 \,.\\
\end{aligned}$$]{} Using `Matplotlib` by [@Matplotlib], our program [@sourceCode] can plot the coupler curve of the vertex $v_6$ of the graph $G_{48} - v_2v_6$ w.r.t. the fixed triangle $v_1v_2v_3$, see Figure \[fig:couplercurve\] for the output. There are 28 embeddings for $\bm{\lambda}$. Following Lemma \[lem:couplerCurvePreserves\] for the subgraph given by $(u,v,w,p,c)=(v_2, v_3, v_1, v_7, v_6)$, one can find a position and radius of the sphere corresponding to the removed edge $v_2v_6$ such that there are 32 intersections. Such edge lengths are obtained by taking $\lambda_{12}=4.0534\,,\, \lambda_{27}=11.1069\,,\, \lambda_{26}=3.8545\,,\, \lambda_{23}=4.0519$.
![The coupler curve $\mathcal{C}_{v_6,\bm{{\lambda}}}$ of $G_{48}$ with the edge $v_2v_6$ removed. The 28 red points are intersections of $\mathcal{C}_{v_6,\bm{{\lambda}}}$ with the sphere centered at $v_2$ with the edge lengths $\bm{{\lambda}}$, whereas the 32 green ones are for the adjusted edge lengths (illustrated by blue dashed lines). []{data-label="fig:couplercurve"}](couplerCurve.png){width="70.00000%"}
### Sampling procedure
Instead of finding suitable parameters for the position and radius of the sphere by looking at visualizations, we implemented a sampling procedure that tries to maximize the number of intersections [@sourceCode]. Whereas the version presented at ISSAC’18 worked only for a short list of predefined graphs, the current one takes an arbitrary minimally rigid graph containing a triangle. The inputs of the function `sampleToGetMoreEmbd` are starting edge lengths $\bm{\lambda}$ and vertices $u,v,w,p,c$ satisfying the assumptions of Lemma \[lem:couplerCurvePreserves\], including the extra requirement that $cw$ is an edge. For simplicity, we identify vertices with their positions in ${\mathbb{R}}^3$ and edges with the corresponding lines in the explanation of the procedure.
Let $S_u$ be the sphere centered at $u$ representing the removed edge $uc$. The extra assumption that $cw$ is an edge is useful since then the coupler curve lies in a sphere $S_w$ centered at $w$. Hence, the intersections of the coupler curve of $c$ with $S_u$ are on the intersection $S_u\cap S_w$, which is a circle. Thus, we can sample circles on the sphere $S_w$ instead of sampling the parameters $t$ and $r$.
The center of the intersection circle $S_u\cap S_w$ is on the line $uw$, which is perpendicular to the plane of the circle. Hence, the circle is determined by the angle $\varphi\in(-\pi/2,\pi/2)$ between the altitude of $w$ in the triangle $uvw$ and the line $uw$, and by the angle $\theta\in(0,\pi)$ between $uw$ and $cw$, see Figure \[fig:phiTheta\]. Clearly, the lengths of $uv, uw, up$ and $uc$ are defined uniquely by the pair $(\varphi,\theta)$ and the other edge lengths. Thus, we sample $\varphi$ and $\theta$ in their intervals instead of sampling the parameters $r$ and $t$. An advantage of this approach is that $\varphi$ and $\theta$ are in bounded intervals, whereas $t$ and $r$ are unbounded. Moreover, sampling the angles uniformly gives a more reasonable distribution of the intersection circles on the sphere $S_w$ than the uniform sampling of $t$ and $r$.
Since for every sample we have to solve a system of equations in orded to count the number of real embeddings, we exploit the following strategy to decrease the number computations: In the first phase, we sample both angles at approximately 20–24 points each and we take the pairs $(\varphi,\theta)$ attaining the maximum number. In the second phase, we sample few more points around each of these pairs to have a finer sampling in relevant areas. Of course, edge lengths with the maximum number of real embeddings are outputted.
The homotopy continuation package `phcpy` by [@phcpy] is used for solving the algebraic systems. A significant speedup of the computation is achieved by tracking the solutions from a previous system, instead of solving the system every time from scratch. Besides the fact that `phcpy` is parallelized, our implementation splits the samples into two parts and computes the numbers of embeddings for them in parallel.
(0,0) to (0,1.3) ; (0,0) to (3,0) ; (0,0) to (-1.1,-1.1) ;
at (-1,-1) [$x$]{}; at (3,0) [$y$]{}; at (0,1.3) [$z$]{};
\(v) at (0,0) ; (u) at (2.5,0) ; (w) at (0.5,-0.8) ;
\(c) at (1.92,-0.68) ;
\(f) at (1.253,0); (f) to (w); (-0.8,-0.8) to (w); (1.3,-0.05) to (1.3,0.05); (-0.8,-0.85) to (-0.8,-0.75);
\(v) to (u); (v) to (w); (u) to (w); (c) to (w);
\(u) to (c);
at (v) [$v$]{}; at (u) [$u$]{}; at (w) [$w$]{}; at (c) [$c$]{};
\(w) circle (1.5cm); (u) circle (1.2cm); (w,1.5 cm)(u,1.2 cm)
(3,0.2)[Od]{} (Od,A)(B) (-0.5,-1.2)[Of]{} (Of,B)(A)
(u,B)(A) (u,A)(B) (w,B)(A) (w,A)(B)
(fc) at ($(A)!0.5!(B)$); (fc) to (c);
at (c) ;
(v,f,w) (w,f,v)[$\cdot$]{}
(w,fc,c) (w,fc,c)[$\cdot$]{}
(u,w,f) (u,w,f)[$\varphi$]{} (c,w,u) (c,w,u)[$\theta$]{}
### More subgraphs suitable for sampling {#subsubsec:combinationOfSubgraphs}
It is likely that the sampling procedure for one choice of $(u,v,w,p,c)$ does not yield the number of real embeddings that matches the complex bound. Hence, we repeat the procedure for various choices of $(u,v,w,p,c)$, assuming that there are more subgraphs satisfying the conditions of Lemma \[lem:couplerCurvePreserves\].
If the sampling procedure produces more edge lengths with the same number of real embeddings, then we need to select starting edge lengths for sampling with a different subgraph, since it is not computationally feasible to test all of them.
We use a heuristic based on clustering of pairs $(\varphi,\theta)$ corresponding to the edge lengths by the function `DBSCAN` from `sklearn` package [@sklearn]. We take either the edge lengths belonging to the center of gravity of each cluster, or the pair $(\varphi,\theta)$ closest to this center if the edge lengths corresponding to the center have a lower number of real embeddings.
We propose two different approaches for iterating the sampling procedure for various subgraphs. The first one, called *tree search*, applies the sampling procedure using all suitable subgraphs for a given $\bm{\lambda}$. Then, the same is done recursively for all output edge lengths whose number of real embeddings increased. The state tree is traversed depth-first, until the required number of real embeddings is reached (or there are no increments). This algorithm is implemented in the function `findMoreEmbeddings_tree` in our code.
The function `findMoreEmbeddings` uses the second approach, called *linear search*. Assume an order of the suitable subgraphs. The output from the sampling procedure applied to starting edge lengths with the first subgraph is the input for the procedure with the second subgraph, etc. The output from the last subgraph is used again as the input for the first one. There is also a branching because of multiple clusters — all of them are tested in depth-first way. Again, we stop either if the required number of real embeddings is reached, or there all the subgraphs are used without increment of the number of real embeddings.
For both, the subgraphs to be used can be specified, or the program computes all suitable subgraphs by itself. Tree search is useful when one wants to find subgraphs whose application leads to the desired number of embeddings in the least number of iterations. On the other hand, linear search seems more efficient. We remark that there is also an option to relax the condition that $\deg(u)=4$. Then, such a subgraph can also be used for sampling, but the coupler curve changes during the process.
Classification and Lower Bounds {#sec:results}
===============================
The first step of our procedure was to construct Laman and Geiringer graphs by Henneberg steps. We subsequently removed isomorphic duplicates and classified them according to the last Henneberg move as described in Section \[sec:rigidity\]. Following an idea explained in [@GraKouTsiLower17], we represented every graph isomorphism class with an integer and we proceeded using a `SageMath` implementation.\
A first upper bound on the number of embeddings is the mixed volume of systems of sphere and distance equations. This bound is crucial for homotopy continuation system solving, as mentioned before. The second natural bound of graph realizations is the number of complex embeddings. The numbers of complex embeddings for all Laman graphs up to 12 vertices are known from [@Joseph_lam], while the numbers of complex embeddings of Geiringer graphs up to $10$ vertices were computed by [@GraKouTsiLower17]. We computed the complex solutions of spherical embeddings of Laman graphs up to 8 vertices. For the last part, we were motivated by a remark of Josef Schicho, who observed that the numbers of planar and spherical solutions differ for the Desargues graph.\
In order to find parameters that can maximize the number of real embeddings, we applied the methods described in Section \[sec:sampling\]. Polynomial system solving during sampling was accomplished mainly via `phcpy`. We consider an embedding being real if the absolute value of the imaginary part of every coordinate is less than $10^{-15}$. The final results were verified using Maple’s `RootFinding` \[Isolate\]. Our results ameliorate significantly what was known about the bounds of real embeddings.
Laman Graphs
------------
The numbers of realizations of all 6-vertex Laman graphs are known [@Borcea2]. There are four H2-last Laman graphs and the upper bound of real embeddings was computed in [@EM] for the graph with the maximal number of complex embeddings. Using stochastic and parametric methods, we were also able to maximize the number of embeddings for the other three 7-vertex graph with not trivial number of embeddings, completing a full classification for all 7-vertex Laman graphs according to their number of real embeddings [@sourceCode].\
[ccc]{} $L_{136}$ & $L_{344}$ & $L_{880}$\
[ (1) at (-0.8, -1); (2) at (-0.8, 3); (3) at (3.2, 3) ; (4) at (3.2, -1) ; (5) at (2,1.2) ; (6) at (2,0) ; (7) at (0.8,0.9) ; (8) at (0.5,0.1) ; ]{} (1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(7) (6)edge(4); (5)edge(6) (5)edge(2) (5)edge(3) (1)edge(8) (2)edge(7); (8)edge(7) (8)edge(4); at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (8) ;
&
[ (1) at (-1, -1); (2) at (-1, 3); (3) at (3.5, 3) ; (4) at (3.5, -1) ; (5) at (0.8,0.4) ; (6) at (2,0.5) ; (7) at (2,1.25) ; (8) at (0.6,1.75) ; (9) at (1.25,2) ; ]{} (1)edge(2) (1)edge(4) (1)edge(5); (3)edge(2) (2)edge(9) (3)edge(4); (3)edge(7) (3)edge(9) (6)edge(4); (5)edge(6) (5)edge(7) (6)edge(7); (7)edge(8) (8)edge(9) (8)edge(1); at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (8) ; at (9) ;
\(a) at (-18.00,-16.00) ; (b) at (20.00,-16.00) ; (c) at (-0.00,-2.30) ; (d) at (-0.00,13.00) ; (e) at (-9.10,-0.20) ; (f) at (9.10,-0.20) ; (g) at (-6.00,-11.00) ; (h) at (-18.00,18.00) ; (i) at (6.00,-11.00) ; (j) at (20.00,18.00) ; (a)edge(b) (a)edge(e) (a)edge(g) (a)edge(h) (b)edge(f) (b)edge(i) (b)edge(j) (c)edge(e) (c)edge(f) (c)edge(g) (c)edge(i) (d)edge(e) (d)edge(f) (d)edge(h) (d)edge(j) (g)edge(i) (h)edge(j);
\
For bigger graphs, we focused on the graphs with the maximal number of complex embeddings, see Figure \[fig:Laman\]. The following table summarizes the bound on $r_2(G)$. Notice that it shows that there exist edge lengths such that all embeddings of the 8-vertex graph $L_{136}$ and of the 9-vertex graph $L_{344}$ are real.
[c|ccc]{} $\bm{n}$ & **8** & **9** & **10**\
& $L_{136}$ & $L_{344}$ & $L_{880}$\
------------------------------------------------------------------------
MV sphere eq. & 192 & 512 & 1536\
MV distance eq. & 136 & 344 & 880\
$c_2(G)$ & 136 & 344 & 880\
$r_2(G)\geq$ & 136 & 344 &\
Now, we provide edge lengths giving the numbers of real embeddings in the table.\
We shall note that while `phcpy` gives 868 real solutions for $L_{880}$, we were able to verify only 860 of them using Maple’s `RootFinding` \[Isolate\] function for distance systems (sphere equations computation did not terminate in that case). The number of real solutions of distance systems was exactly the same as expected by the `phcpy` computation, but in some cases triangular inequalities were violated. The violation error was smaller than $10^{-8}$. Although there is a strong possibility that this error is insignificant, we take that $r_2(L_{880})\geq 860$.
### Spherical embeddings of Laman graphs {#spherical-embeddings-of-laman-graphs .unnumbered}
Maximal numbers of embeddings in $S^2$ have been not studied so far. We attempted to find edge lengths such that the number of realizations was the same as the number of complex solutions for graphs that do not have a trivial number of embeddings. We shall observe again that the $c_2(G)$ varies for certain graphs from $c_{S^2}(G)$.\
We have found parameters such that all the embeddings are real for all H2-last graphs with 6 and the 7-vertex graphs with the maximal number of complex embeddings(they can be found in [@sourceCode]). The Desargues graph has the maximal number of embeddings among 6-vertex graphs, namely, it can have $32$ realizations (instead of 24 on the plane). In the 7-vertex case, there are two H2-last graphs with $64$ realizations (instead of 48 and 56 respectively on the plane), see Figure \[fig:lamanSphere\]. Let us indicate that 64 realizations can be also achieved by the 3 graphs constructed by applying an H1 move on $L_{24}$, since H1 doubles the number of embeddings. Observe that this contrasts the situation of the complex embeddings in the plane, since it is known that for $n\leq 12$ there is always a unique Laman graph with the maximal number of complex embeddings on the plane among $n$-vertex Laman graphs. We have also found edge lengths that maximize the spherical embeddings of $L_{136}$ (see Figure \[fig:Laman\]). It has 192 real spherical embeddings. We remark that there is again another graph with 192 complex spherical embeddings, but we have found edge lengths with only 136 real spherical embeddings.
[cccccc]{} $L_{24} (Desargues)$ & $L_{48H2}$ & $L_{56}$ & $L_{48H1a}$ & $L_{48H1b}$ & $L_{48H1c}$\
\(1) at (0, -1); (2) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, -1) ; (5) at (1.25,1.75) ; (6) at (1.25,0.25) ;
(1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4); (6)edge(1) (6)edge(4) (5)edge(2) (5)edge(6) (5)edge(3) ;
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ;
&
\(6) at (0, 0); (5) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, 0) ; (2) at (1.25,2) ; (1) at (1.25,0.7) ; (7) at (1.5,-1) ;
(1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(1); (5)edge(2) (5)edge(6) (5)edge(3) (1)edge(7) (4)edge(7); (6)edge(7);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ;
&
[ (1) at (0, -1); (2) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, -1) ; (5) at (2,1.2) ; (6) at (1.25,0) ; (7) at (1,1.2) ; ]{} (1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(1) (6)edge(4); (5)edge(6) (5)edge(3) (1)edge(7) (2)edge(7); (5)edge(7);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ;
&
\(1) at (0, -1); (2) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, -1) ; (5) at (1.25,1.75) ; (6) at (1.25,0.25) ;
\(7) at (1.875,0.7) ; (1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(1); (5)edge(2) (5)edge(6) (5)edge(3) (6)edge(7) (4)edge(6); (4)edge(7);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ;
&
\(1) at (0, -1); (2) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, -1) ; (5) at (1.25,1.75) ; (6) at (1.25,0.25) ;
\(7) at (1.875,1) ; (1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(1); (5)edge(2) (5)edge(6) (5)edge(3) (3)edge(7) (4)edge(6); (4)edge(7);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ;
&
\(1) at (0, -1); (2) at (0, 3); (3) at (2.5, 3) ; (4) at (2.5, -1) ; (5) at (1.25,1.75) ; (6) at (1.25,0.25) ;
\(7) at (1.875,1) ; (1)edge(2) (2)edge(3) (3)edge(4) (1)edge(4) (6)edge(1); (5)edge(2) (5)edge(6) (5)edge(3) (5)edge(7) (4)edge(6); (4)edge(7);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ;
This table gives upper bound and the number of real spherical embeddings for all graphs with $6\leq n\leq 8$ that have the maximal number of embeddings.
[c|ccccccc]{} $\bm{n}$ & **6** & **7** & **7** & **7** & **7** &**7** & **8**\
& $L_{24}$ & $L_{48H2}$ & $L_{56}$ & $L_{48H1a}$ & $L_{48H1b}$ & $L_{48H1c}$ & $L_{136}$\
------------------------------------------------------------------------
MV sphere eq. & 32 & 64 & 64 & 64 & 64 & 64 & 192\
MV distance eq. & 32 & 64 & 64 & 64 & 64 & 64 & 192\
$c_{S^2}(G)$ & 32 & 64 & 64 & 64 & 64 & 64 & 192\
$r_{S^2}(G)$ & 32 & 64 & 64 & 64 & 64 & 64 & 192\
We present a list of lengths (using euclidean metric) that give maximal number of realizations for the non-trivial (H2-last) cases:
Geiringer graphs
----------------
The method we introduced in Section \[sec:coupler\] played a crucial role in increasing the number of embeddings of Geiringer graphs. We used our method for the only H2-last graph with 6 vertices — the cyclohexane $G_{16}$. It was known that $r_3(G_{16})=16$, a result that can be verified by our method within a few tries with random starting lengths.
[cc]{} $G_{48}$ & $G_{160}$\
\(1) at (0, -1); (2) at (-1.9, 0); (3) at (-0.9, -0.3) ; (4) at (0.85, -0.3) ; (5) at (1.8,0.0) ; (6) at (-0.2, 0.2) ; (7) at (0,1) ; (2)edge(3) (3)edge(4) (6)edge(2) (5)edge(4) (6)edge(5); (2)edge(1) (1)edge(4) (1)edge(3) (1)edge(5) (1)edge(6); (2)edge(7) (7)edge(4) (7)edge(3) (7)edge(5) (7)edge(6);
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ;
&
[ (1) at (0, -1); (2) at (-1.9, 0); (3) at (-0.9, -0.3) ; (4) at (0.85, -0.3) ; (5) at (1.8,0.0) ; (6) at (-0.3, 0.3) ; (7) at (-0.7,1) ; (8) at (0.7,1) ;]{} (1) to (2) (2) to (7) (4) to (7) (2) to (6) (6) to (8) (4) to (5) (2) to (8) (5) to (7) (3) to (4) ; (1) to (4) (1) to (5) (1) to (3) (1) to (6) (5) to (6) (3) to (7) (7) to (8) (2) to (3) (5) to (8) ;
at (1) ; at (2) ; at (3) ; at (4) ; at (5) ; at (6) ; at (7) ; at (8) ;
\
The case of $n=7$ was the first open one. There are twenty H1-last 7-vertex Geiringer graphs and six H2-last ones. We computed the mixed volumes and the number of complex embeddings for each one of them. Then, using our code we were able to find edge lengths that give a full classification of all 7-vertex Geiringer graphs according to $r_3(G)$ [@sourceCode].\
We want to remark again at this point that $G_{48}$ was the model for our coupler curve method. Using our implementation, we were able to find lengths that maximize the number of embeddings only after a few iterations. The structure of this graph fits perfectly to our method, since there are 20 subgraphs of $G_{48}$ given by vertices $(u,v,w,p,c)$ satisfying the assumption in Lemma \[lem:couplerCurvePreserves\]. Using tree search approach, we obtained edge lengths $\bm{\lambda}$ such that $r_3(G_{48},\bm{\lambda})=48$:
They can be found from the starting edge lengths given in Sec. \[sec:coupler\] with $28$ real embeddings in only 3 iterations, using the subgraphs $(v_5, v_6, v_1, v_7, v_4), (v_4, v_3, v_1, v_7, v_5)$ and $(v_3, v_2, v_1,v_ 7, v_4)$.
We repeated the same procedure for $n=8$. In that case we can use the H1 doubling property for $311$ graphs, while there are $63$ graphs with a non-trivial number of embeddings. We computed complex bounds for all H2-last graphs [@sourceCode]. We subsequently found edge lengths that increase the number of real embeddings of $G_{160}$, which is the graph with the maximal number of complex embeddings $c_3(G_{160})=160$. We were able to find parameters $\bm{\lambda}$ such that $r_3(G_{160},\bm{\lambda})=132$.
The following lengths give 132 real embeddings for $G_{160}$:
We shall remark that our results about 7-vertex graphs and $G_{160}$ appeared already in [@ISSAC_2018]. One may find a full list of Geiringer graphs with 7 and 8 vertices in [@sourceCode]. Finally, we also want to notice that for all planar (in the graph-theoretical sense) Geiringer graphs up to 10 vertices, the number of complex embeddings is always equal to the mixed volume of the sphere equations system. A possible conjecture could be that mixed volume is tight for all planar Geiringer graphs.
Lower bounds
------------
The maximal numbers of real embeddings that we found can serve to build an infinite class of bigger graphs. These frameworks can give us lower bounds on the maximum number of embeddings. To compute the lower bound, we will use the following theorem that combines caterpillar, fan and generalized fan constructions [@GraKouTsiLower17]:
Let $G=(V_G,E_G)$ be a generically rigid graph, with a generically rigid subgraph $H=(V_H,E_H)$. We construct a rigid graph using $k$ copies of $G$, where all the copies have the subgraph $H$ in common. The new graph is rigid, has $n = |V_H| + k(|V_G| - |V_H|)$ vertices, and the number of its real embeddings is at least $$2^{(n-|V_H|)\!\mod(|V_G|-|V_H|)} \cdot r_d(H) \cdot
\left(\frac{r_d(G)}{r_d(H)}\right)^{\left\lfloor\frac{n-|V_H|}{|V_G|-|V_H|}\right\rfloor}.$$
Remind that for a triangle $T$ we have that $r_2(T)=r_{S^2}(T)=2$, while $r_3(T)=1$. For Laman graphs, the best asymptotic bound is derived from $L_{880}$:
The maximum number of real embeddings on the plane among Laman graphs with $n$ vertices is bounded from below by $$2^{(n-3) \mod 7} \,\cdot 2 \cdot 430^{\lfloor (n-3)/7 \rfloor} \,.$$ The bound asymptotically behaves as $2.378^{n}$.
The previous lower bound in that case was $2.3003^n$ by [@EM].
In the case of spherical embeddings, we may use $L_{24}$:
The lower bound for the maximum number of spherical embeddings among Laman graphs with $n$ vertices is $$2^{(n-3) \mod 7} \,\cdot 2 \cdot 16^{\lfloor (n-3)/3 \rfloor} \,.$$ This bound asymptotically behaves as $2.51984^{n}$.
We remark that $L_{48H1a}$, which has the 4-vertex Laman graph as a subgraph, can give the same asymptotic lower bound. The other 7-vertex graphs with $r_{S^2}(L)=64$ can give only $2.3784^n$ as a lower bound, while the asymptotic bound from 8-vertex graph with $192$ embeddings is $2.4914^n$.
Finally, using the fact that $r_3(G_{160})\geq 132$, we obtain the following result, which appeared also in [@ISSAC_2018]:
The maximum number of real embeddings of Geiringer graphs with $n$ vertices can be bigger than $$2^{(n-3) \mod 5} \, 132^{\lfloor (n-3)/5 \rfloor} \,,$$ indicating that $r_3(n) \in \Theta (2.6553^{n})$.
The previous lower bound for Geiringer graphs was $2.51984^n$ [@Emiris1]. Using the graph $G_{48}$ yields $ r_3(n) \in \Theta (2.6321^{n})$. Notice that we use a subgraph with one embedding and not with two, as we did in the cases of Laman graphs. This happens because there is no tetrahedron as a subgraph of the 8-vertex graphs that could give a better lower bound.
Conclusion and future work {#sec:conclusion}
==========================
In this paper we have developed and used efficient methods to maximize the number of real embeddings of rigid graphs in the case of planar, spherical and spatial embeddings. We have introduced a new technique for Geiringer graphs, that exploits an invariance property of coupler curves to select the sampling parameters at each iteration. This procedure led to classification results and to an improvement of the asymptotic lower bounds.
As future work, a first goal would be to ameliorate the maximal real bounds in all cases. It is an interesting question if we can develop a similar sampling technique, as the one we introduce in this paper, for other cases and/or other structures of Geiringer graphs. Besides lower bounds, it is believed that upper bounds are really loose, so an open problem is to improve them in the general case or for specific classes of graphs.
#### Acknowledgments
This work is part of the project ARCADES that has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 675789. ET is partially supported by ANR JCJC GALOP (ANR-17-CE40-0009) and the PGMO grant GAMMA.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We present a concept for an ultra-thin solar fuel device with a nanostructured back contact. Using rigorous simulations we show that the nanostructuring significantly increases the absorption in the semiconductor, CuBi$_2$O$_4$ in this case, by 47% (5.2 mAcm$^{-2}$) through the excitation of plasmonic modes. We are able to attribute the resonances in the device to metal-insulator-metal plasmons coupled to either localised surface plasmon resonances or surface plasmon polaritons.
Rounding applied to the metallic corners leads to a blueshift in the resonance wavelength while maintaining absorption enhancement, thus supporting the possibility for a successful realization of the device. For a 2D array, the tolerance of the polarization-dependent absorption enhancement is investigated and compared to a planar structure. The device maintains an absorption enhancement up to incident angles of 75$^{\circ}$.
The study highlights the high potential for plasmonics in ultra-thin opto-electronic devices such as in solar fuel generation.
author:
- 'Phillip Manley[^1]'
- 'Fatwa F. Abdi'
- Sean Berglund
- 'A.T.M. Nazmul Islam'
- Sven Burger
- Roel van de Krol
- Martina Schmid
bibliography:
- 'main.bib'
- 'app.bib'
title:
- ' Absorption Enhancement for Ultra-Thin Solar Fuel Devices with Plasmonic Gratings'
- 'Supporting Information: Absorption Enhancement for Ultrathin Solar Fuel Devices with Plasmonic Gratings'
---
This paper was published in ACS Applied Energy Materials **1** p.5810-5815 (2018) doi: [10.1021/acsaem.8b01070](dx.doi.org/10.1021/acsaem.8b01070) and is made available as an electronic preprint with permission from the American Chemical Society.
Introduction
============
The conversion of sunlight to storable fuel is a challenge of paramount importance to modern society. Photoelectrochemical devices, consisting of semiconductor photoelectrodes immersed in aqueous solution, are a particularly interesting way to achieve this solar-to-fuel conversion. In recent years many different device designs based on various materials have seen intensive research [@main][Montoya2017,Park2006]{}. Among these, metal oxides are particularly interesting since they possess general aqueous stability and are relatively inexpensive [@main][Sivula2013,Abdi2013\_2]{}. However, they share a common drawback, that of the discrepancy between carrier transport and light absorption. Due to the poor transport properties, the carrier diffusion length in oxides is typically less than 100 nm [@main][Joly2006, Cherepy1998, Kennedy1978, Paracchino2012, Abdi2013, Berglund2016, Abdi2017]{}. On the other hand, they normally have an indirect band gap; relatively thick films ($>$500 nm) are needed to absorb enough light. This mismatch often severely limits the performance of a metal oxide photoelectrode.
In the present work we focus on the metal-oxide semiconductor CuBi$_2$O$_4$, which is an emerging p-type semiconductor for solar fuel applications. It has a band gap of around 1.8 eV, which is ideal for a top absorber in a tandem configuration [@main][Arai2007, Hahn2012, Berglund2013, Berglund2016]{}. It also has a suitable band position; the conduction and valence band edges straddle both water reduction and oxidation potentials. As a result, the photocurrent onset potential for CuBi$_2$O$_4$ has been reported to be $\sim$1 V vs reversible hydrogen electrode (RHE), which is beneficial for a tandem configuration [@main][Arai2007, Hahn2012, Berglund2013]{}. Previous reports of CuBi$_2$O$_4$ synthesis deposition have shown highly porous and irregular surface structures [@main][Berglund2016,Hahn2012,Kang2016]{}. Alternative synthesis methods such as spray-pyrolysis have been used to make dense, homogeneous CuBi2O4 thin films [@main][Wang2017]{} and new methods such as pulsed laser deposition (PLD) could be used to obtain a highly uniform CuBi2O4 ultra-thin film. PLD has been demonstrated as a viable option for other metal oxide photoelectrode materials including resonant light trapping Ti-doped alpha-Fe2O3 [@main][Dotan2013]{}.
However, the material is limited by the optical absorption vs. carrier transport mismatch as was previously mentioned. The diffusion length is in the range of $\sim$50 nm, while a thickness of more than 500 nm is needed to absorb 90% of the incident light [@main][Berglund2016]{}. This is especially true for the longer wavelengths ($>$ 450 nm), where the quantum efficiency has been reported to be very low.
In order to combat this challenge, we propose to reduce the semiconductor thickness to 100 nm which should ensure efficient carrier collection by shortening the length photo-generated carriers have to travel. Simultaneously we use light management to obtain a sufficient absorption of incident sunlight.
Various light management strategies applied to the field of solar fuels have been presented in the literature. Photonic crystal structuring of the active layer has been used to enhance absorption [@main][Jeremy2016]{}. Other approaches use dielectric particles and gratings to localize light inside the active layer [@main][Kim2014,Cheng2018]{}.Structuring the active material into nanorods can also increase absorption while maintaining short carrier diffusion lengths [@main][Pihosh2015]{}. A further proposed light management strategy is that of plasmonics [@main][Thomann2011,Abdi2014]{}. Plasmonic metallic nanoparticles act as optical antennas, allowing light to be concentrated in the vicinity of the semiconductor material, thereby enhancing absorption [@main][Atwater2010,Schmid2016]{}. Furthermore metallic particles themselves may have beneficial catalytic properties [@main][Berglund2013,Li2013]{}. Despite this, certain challenges are present for particles, such as quenching the photocatalysis process through recombination [@main][DiVece2012,Govorov2006]{}.
In this paper we circumvent the challenges of using particles by considering a 100 nm thick CuBi$_2$O$_4$ layer on a grating consisting of laterally alternating Ag and SiO$_2$ on top of an Ag layer. Ag and SiO$_2$ are less positive than CuBi$_2$O$_4$ vs. RHE. In order to circumvent a Schottky at the rear interface, an additional back contact layer or heavy doping layer of the CuBi$_2$O$_4$ may be necessary [@main][Hudait2001]{}. Since we focus on the optical device design, such a layer is not taken into account in the current work. The unit cell of this periodic structure is shown in the inset of figure \[fig:RTA\](b). We will refer to the SiO$_2$ region as a nanoslot since it forms a slot in the Ag. Plasmonic gratings have been realized for multiple applications including photovoltaic absorption enhancement [@main][Paetzold2011]{} and biosensing [@main][Iqbal2017]{}. For the current application, the nanoslot grating serves as the metallic back contact to transport photo-generated holes to the anode side as well as to enable better light management. A similar structure has been applied to infrared absorption enhancement in photovoltaics [@main][Wang2013]{}. Through careful device design, we are able to shift the operational frequency to visible wavelengths.
Results and Discussion
======================
1D Grating
----------
![The absorption in the 100 nm thick CuBi$_2$O$_4$ layer (divided into bulk \[first 95 nm\] and interface \[last 5 nm\] contributions), and the losses in Ag and Reflection for a solar fuel device with a flat Ag back contact (a) and a nanostructured Ag back contact (b). The dotted lines indicate resonance wavelengths shown in figure \[fig:NearFields\]. Insets show schematic drawings of the periodic unit cell for each case.[]{data-label="fig:RTA"}](Final/Planar_Nanostructure_Comparison_Verarbeitet_V2.pdf){width="\textwidth"}
Figure \[fig:RTA\] shows the absorption and losses for the ultra-thin photoelectrode with a planar Ag back reflector with no nanoslots (a) and a nanostructured back reflector (b). The inset shows a schematic of each structure. Both of the devices have a 100 nm thick CuBi$_{2}$O$_{4}$ layer.
The catalytic reaction, in this case water reduction, occurs at the interface between H$_{2}$O and the inorganic semiconductor photocathode, CuBi$_{2}$O$_{4}$. Light is incident through the H$_{2}$O.
The nanostructured device has a unit cell width (pitch) of 112 nm, the SiO$_{2}$ slot is 70 nm wide and 60 nm deep, it is infinitely extended in the plane perpendicular to the page, therefore defining a 1D grating. These values were obtained from an optimization (figure \[fig:Opt\]). We split the semiconductor into two regions. Firstly, the upper 95 nm of the semiconductor material (dark green), which is in contact with the water. Secondly, the final 5 nm of the semiconductor material (light green), which is in contact with the Ag back contact. These regions will be referred to as the ’bulk’ and ’interface’ regions respectively. Analogously, the absorption in each region will be referred to as ’bulk’ and ’interface’ absorption, respectively.
For the first case of the simple back reflector (no SiO$_2$ nanoslots), the interface absorption of the semiconductor is very small compared to the bulk. This is due to the exponential damping of light while traversing the first 95 nm of the material (Lambert-Beer law) and also due to the volume of the region in question being much smaller than that of the bulk region.
In order to quantify this absorption for solar fuel applications, we calculate the photocurrent density ($J_{abs}$) from the absorption curve and the solar spectrum. This assumes that all absorbed photons can contribute to water reduction. This assumption should be interpreted as the upper limit on photocurrent density for the proposed photoelectrode architecture. When modeling practical devices, the various loss mechanisms (e.g. recombination) have to be taken into account through the absorbed photon-to-current efficiency (APCE). We note that there have been reports for nanostructured metal-oxides with APCE values close to 100% [@main][Pihosh2015]{}. The proposed structure is able to reach a theoretical photocurrent density of 11.1 mAcm$^{-2}$ which is 54% of the maximum achievable short circuit current density for a material with a band-gap of 1.8 eV (20.5 mAcm$^{-2}$).
Due to the low losses provided by Ag, the absorption in the back reflector is minimal, meaning that the main loss mechanism is reflection. This loss mechanism is eliminated entirely at the wavelength of 440 nm due to the presence of a Fabry-Perot resonance which eliminates the reflection. These kinds of resonances are clearly beneficial to absorption, however they cannot provide an arbitrarily broad absorption enhancement, since the only free parameter for tuning such a resonance is the film thickness.
Absorption is seemingly still present at wavelengths up to 700 nm. The complex refractive index used for CuBi$_{2}$O$_{4}$ also contains parasitic absorption at and below the band gap (figure \[fig:nk\_data\]). However analysis of the absorption coefficient has shown that the band gap for CuBi$_{2}$O$_{4}$ lies around 700 nm. Therefore we use this wavelength as the cutoff for absorption contributing to the photocurrent density.
![The electric field in the nanostructured solar fuel device for four different wavelengths taken from the peaks in figure \[fig:RTA\]. Wavelengths are 380, 410, 470, 650 nm for parts (a-d), respectively. Light is incident normally from above and polarized in the x-z plane.[]{data-label="fig:NearFields"}](Final/NearFields_New.pdf){width="\textwidth"}
Figure \[fig:RTA\](b) shows the reflection and absorption for the nanostructured back contact (Ag with SiO$_2$ nanoslots). The polarization is oriented in the $x$ $-$ $z$ plane so that the plasmonic effects can be studied. In this case the total absorption is over 80% for most of the visible spectrum, particularly between 500 and 650 nm where the solar spectrum has peak intensity. The overall absorption increase leads to a maximum short circuit current density of 16.3 mAcm$^{-2}$ (80% of the maximum achievable). This large increase in absorption can be attributed to the excitation of resonant modes that exist at the Ag / CuBi$_2$O$_4$ interface, due to the large contribution of the interface absorption (light green). In this case the bulk absorption contributes 11.1 mAcm$^{-2}$ to $J_{abs}$ while the interface absorption contributes 5.2 mAcm$^{-2}$ to $J_{abs}$. In addition the parasitic absorption in the Ag back contact also increases due to increased interaction with the Ag arising from the resonant modes.
In order to investigate the mechanism of the absorption enhancement further, we show in figure \[fig:NearFields\] the electric near field strength of the nanostructured back contact for the peak wavelengths shown in figure \[fig:RTA\](b), namely 380, 410, 470 and 650 nm. The presence of a resonant mode is clearly shown in each case with strong localization of the electric field.
All four of the modes can be associated with metal-insulator-metal (MIM) plasmon resonances of the slot [@main][Maier2007,Kurokawa2007]{}. This can be seen in the variations of the resonance wavelength with respect to the slot length which varies with a wavelength close to the analytical MIM mode wavelength (figures \[fig:Length\_Variation\] and \[fig:Length\_Variation\_Near\_Fields\]).
The four modes can be grouped into two categories. At the shorter wavelengths of 380 and 410 nm (figures \[fig:NearFields\](a) and (b)), the resonance is mainly confined to a localized mode at the bottom corners of the slot. In contrast, the resonances at the longer wavelengths of 470 and 650 nm are mainly located at the upper corners of the slot and at the Ag / CuBi$_2$O$_4$ interface.
An analysis of the variation of these modes with the device pitch (keeping the slot size constant) reveals that at wavelengths between 450 and 700 nm, the field localized at the upper corners acts as a source of surface plasmon polariton (SPP) modes (figures \[fig:Pitch\_Variation\] and \[fig:Pitch\_Variation\_Near\_Fields\]). Since the SPP modes have significant field strength in the CuBi$_2$O$_4$ they are able to increase the absorption there. In contrast the shorter wavelength modes remain localized in the bottom of the slot and therefore contribute mainly to losses in the Ag.
The maximum absorption enhancement is observed when an antinode of the MIM mode resonance is at the slot opening and when the SPP mode which is excited constructively interferes with itself. This holds true for the device at a wavelength of 650 nm. Slot length variations (figure \[fig:Length\_Variation\]) and pitch variations (figure \[fig:Pitch\_Variation\]) which affect the MIM and SPP resonances, respectively, show a peak for 60 nm slot length and 112 nm pitch. For the peak at 470 nm wavelength, the MIM mode is at maximum enhancement while the SPP is slightly off resonance. Conversely, at 575 nm wavelength, the MIM mode is off resonance while the SPP mode shows maximum constructive interference causing the increase in interface absorption visible at this wavelength.
Due to the coupling to SPP interface modes, the total absorber layer thickness can, in principle, be reduced while maintaining a high absorption. However, the total absorption obtained will still drop with decreasing layer thickness, as the short wavelength light ($<$ 350 nm) still needs to be absorbed conventionally since the inter-band transition losses in Ag prevent any beneficial resonances from forming at these wavelengths. Therefore, although we have presented absorption curves for the device with a 100 nm thick absorbing layer, we stress that the presented nanostructured device could also conceivably provide absorption enhancements for much thinner layers.
Effect of Corner Rounding
-------------------------
A further consideration for the implementation of these nanostructures in realistic devices is the ability to fabricate precise geometries. As the near field pictures from figure \[fig:NearFields\] show, there is a strong localization of electric field in the vicinity of the sharp edges of the Ag grating. Such perfectly sharp edges are difficult to fabricate, therefore a certain amount of rounding on the edges should be expected.
In figure \[fig:CornerRounding\] we show the absorption in the 100 nm of CuBi$_{2}$O$_{4}$ for the nanostructured 1D grating for three cases of corner rounding radius ($R_{c}$) at both the upper and lower grating corners. The definition of $R_{c}$ is shown in the inset of figure \[fig:CornerRounding\]. The values of R$_{c}$ presented are 0 (same absorption curve as shown in figure \[fig:RTA\](b)), 2 nm and 10 nm. It can be seen that even for $R_{c}$ = 2 nm, the resonances are blueshifted and this becomes more pronounced with increasing corner rounding.
The blueshift seen is a combination of multiple factors. The MIM mode resonance wavelength tends to decrease with decreasing slot width. This may be more relevant to the shorter wavelength resonances since they are localised to the bottom of the slot. Furthermore, due to the inhomogeneous width of the slot, the length of slot which has the necessary width for the supporting the MIM resonance will be effectively shorter. Since the MIM mode provides an absorption enhancement when an antinode is present at the slot opening, the wavelength of MIM mode necessary for this may be shifted to shorter wavelengths. Finally, as the corner rounding increases, more of the SPP resonance will be located inside the slot which has a lower refractive index than CuBi$_{2}$O$_{4}$. This lower refractive index will tend to redshift the SPP resonance. Due to competing factors, the effect of corner rounding on the exact resonance position is difficult to predict, necessitating further numerical study for an optimum to be found.
, while the fourth has the geometrical parameters optimized for the corner rounding. Inset shows the definition of $R_{c}$. All other aspects of the geometry are the same as in figure \[fig:RTA\](b). Light is incident normally from above and polarized in the $x$-$z$ plane.[]{data-label="fig:CornerRounding"}](Final/CornerRounding.pdf){width="50.00000%"}
Despite the blueshifting the core resonant modes are all still present. We can conclude that the resonant absorption enhancement is not reliant on an unphysical singularity at the corners. This is important for the physical realization of the device.
The maximum photocurrent density obtained for the device with no corner rounding was previously shown to be 16.3 mAcm$^{-2}$. When a corner rounding of 2 nm is imposed, the photocurrent density lowers slightly to 15.8 mAcm$^{-2}$. As the corner rounding is increased to 10 nm, the photocurrent density further decreases to 14.6 mAcm$^{-2}$. All of these values show a significant improvement over the planar value of photocurrent density of 11.1 mAcm$^{-2}$. It should be further noted that the geometry can be reoptimized with corner rounding. If the corner rounding is 10 nm then a new optimum absorption enhancement can be found for a pitch of 200 nm, slot width of 100 nm and slot length of 50 nm. In this case the photocurrent density is increased to 16.5 mAcm$^{-2}$. The absorption profile for the optimized structure with corner rounding is shown in figure \[fig:CornerRounding\].
2D Grating
----------
For integration into a solar fuel device, the proposed nanostructured grating has to continue to provide a strong enhancement in the presence of unpolarized light at different angles of incidence. In order to enhance the unpolarized response, the 1D grating can be extended to a 2D grating while keeping the same dimensions as the 1D grating. The resonance conditions found previously can be maintained for either a Ag grating with SiO$_{2}$ nanoslots (1), or a SiO$_{2}$ matrix containing Ag cubic nanoparticles connected to a Ag back contact (2). A schematic of configuration (2) is shown in figure \[fig:2DSchema\]. Configuration (2) was found to be more optically beneficial than configuration (1) and was therefore chosen for the results presented in this section. For reasons of computational efficiency, corner rounding has not been used for the 2D grating.
![The absorbed photocurrent density as a function of the incident polar angle $\theta$ (a). The absorption in the semiconductor (CuBi$_2$O$_4$) for a 2D grating for polarization s (b) and p (c) as a function of both wavelength and polar incidence angle.[]{data-label="fig:2DGrating"}](Final/AngularDependenceAll.pdf){width="\textwidth"}
The periodic unit cell of the 2D grating has a 3 fold mirror symmetry. The mirror planes lie along the $x$ axis, the $y$ axis and at 45$^{\circ}$ to the $x$ and $y$ axes. Therefore it is sufficient to use azimuthal angles $0^{\circ}<\phi<45^{\circ}$ to obtain the azimuthally averaged response. We chose 5 azimuthal angles equally spaced between 0$^{\circ}$ and 45$^{\circ}$. The difference in absorption between averaging over 3 and 5 angles was smaller than $10^{-3}$, therefore it was concluded that 5 angles are sufficient to obtain an averaged response. For each polar angle ($\theta$ in figure \[fig:2DSchema\]) of incidence $> 0$ we can define two orthogonal polarizations: p polarization where the electric field orientation lies in the scattering plane and s polarization where the electric field orientation is perpendicular to the scattering plane, as shown in figure \[fig:2DSchema\].
Figure \[fig:2DGrating\](a) shows how the maximum photocurrent density $J_{abs}$ for the structure depends on the incident polar angle. For oblique incidence p polarization provides a higher current density than for s polarization. Taking the unpolarized response into account, the grating outperforms a planar stack for angles up to 75$^{\circ}$ which is highly beneficial to solar fuel applications.
Figure \[fig:2DGrating\](b-c) shows the wavelength resolved polar angular response for s and p polarization. The 2D grating, even at normal incidence, does not show the same behavior as the 1D grating. This is to be expected as they are physically different systems, however the main resonances present in figure \[fig:RTA\](b) are also present in figure \[fig:2DGrating\], namely a strong absorption enhancement between 500 and 600 nm and a slightly weaker enhancement between 600 and 700 nm. In general the absorption for the case of p polarization is higher as we move towards higher angles. This is partly due to reflection at the initial H$_{2}$O/CuBi$_{2}$O$_{4}$ interface being lower for p polarization.
The dependence of the plasmonic resonances on incident angle differs for the p and s polarizations. In figure \[fig:2DGrating\](b) we see that for s polarization, the resonance wavelengths remain constant with increasing angle. The resonance positions do not change due to the electric field remaining normal to the vertical sides of the Ag cuboids for all incident angles. For the case of p polarization shown in figure \[fig:2DGrating\](c), the resonances broaden and increase for higher angles. This is due to electric field component no longer being purely normal to the vertical sides of the Ag cuboids, thereby changing the resonance condition for excitation of MIM modes. Broader resonances will be beneficial to the broadband functioning of the device.
Conclusion
==========
We have presented a nanostructured back contact for use in ultra-thin solar fuel devices. The nanostructuring was shown to significantly increase the absorption in the absorbing semiconductor, CuBi$_{2}$O$_{4}$ in this case, by 47% (5.2 mAcm$^{-2}$) through the excitation of plasmonic modes.
By varying the length of the SiO$_2$ nanoslot and the device pitch the resonances could be classified as MIM modes which remain isolated in the nanoslot or couple to SPP modes dependent on the wavelength.
By simulating the effect of corner rounding, it could be confirmed that the presented results do not rely on unphysical singularities at material interfaces. This means that translation of the simulated results to an experimental reality is promising. Moreover, we demonstrated that the detrimental effect of corner rounding, which is unavoidable in practical systems, can be fully compensated by adjusting the pitch and slot dimensions of the SiO$_2$.
For a 2D grating based on the optimal 1D nanoslot grating, the angular tolerance of the absorption enhancement was investigated for two different polarizations. A greater angular tolerance was shown for p polarization, with the absorption even increasing at higher angles. The unpolarized response was shown to outperform planar layers up to angles of 75$^{\circ}$.
These results provide a clear pathway to overcome the mismatch between the optical absorption and carrier diffusion length, which is present in many semiconducting photoelectrodes. In the end, it is expected that higher solar-to-fuel conversion efficiency with metal oxide photoelectrodes, especially the promising photocathode material CuBi$_{2}$O$_{4}$, is to be achieved with our proposed architecture.
Methods
=======
All simulations were done using the commercial software JCMsuite [@main][Pomplun2007]{}, a finite element solver for Maxwell’s equations. All simulations modelled the upper and lower half spaces with perfectly matched layers while periodic boundary conditions were used in the $x-y$ plane. A plane wave source was incident from the upper half space. For H$_2$O and SiO$_2$ a wavelength independent refractive index of 1.33 and 1.5 was used, respectively. For Ag the data of Johnson and Christy were used [@main][Johnson1972]{}. The complex refractive index of CuBi$_2$O$_4$ was obtained via spectroscopic ellipsometry of a single crystal using an M-2000D ellipsometer (193-1000 nm, J.A. Woollam Co., Inc.). Details of crystal synthesis are given in the supporting information. By using a finite element degree of 3 and mesh side length constraint smaller or equal to one tenth of the wavelength in the material (for dielectrics) or a constant value of 5 nm (for metal) we were able to ensure an accuracy of greater than $10^{-3}$ for the absorption and reflection.
ASSOCIATED CONTENT
==================
Supporting Information Available: The complex refractive index of CuBi$_{2}$O$_{4}$ used in all simulations. Description of the procedure for obtaining absorption and photocurrent density. The effects of length and pitch variations of the nanoslot. Optimization of the nanoslot with respect to slot width, slot length and pitch ratio. The geometry of the 2D nanoslot array and the definition of the scattering angles.
Acknowledgements
================
The authors would like to thank Dr. A. Bronneberg for ellipsometry measurements. P. Manley and M. Schmid would like to acknowledge funding and support from the Initiative and Networking fund of the Helmholtz Association for the Young Investigator Group VH-NG-928. Part of the work was done at the Berlin Joint Lab for Optical Simulations for Energy Research (BerOSE). P. Manley acknowledges funding from the Helmholtz Innovation Lab HySPRINT, which is financially supported by the Helmholtz Association.
[references]{}[References]{} \#1[lbibitem\[S\][\#1]{}]{}
@affillist @authlistempty
@authlist
\[sec:SupportingMaterial\]
Figure \[fig:nk\_data\] shows the real (n) and imaginary (k) parts of the refractive index of CuBi$_{2}$O$_{4}$. The data were obtained from a spectroscopic ellipsometry measurements of a single crystal using an M-2000D ellipsometer (193-1000 nm, J.A. Woollam Co., Inc.). The crystal growth of CuBi$_{2}$O$_{4}$ was done by the floating zone technique in a four mirror type optical image furnace (Crystal Systems Corp., Japan) in the quantum materials corelabs at the Helmholtz-Zentrum Berlin. For crystal growth, high density feed rod (D= 6 mm, L= 7 cm) was prepared from a stoichiometric mixture of high purity of Bi$_{2}$O$_{3}$ (99.99%) and CuO (99.995%) by solid state reactions. Growth in the floating-zone machine was done in ambient air atmosphere at a rate of 5 mm/h.
After solving Maxwell’s equations to obtain the electric field over the whole volume of our computational domain we wish to calculate related quantities. To obtain the absorption in each material, we use the imaginary part of the electric field energy density integrated over the volume of the material [@app][Jackson1998]{},
$$A = 2\omega\int_{V_{m}} \epsilon_{m}'' |\vec{E}(\vec{r})|^{2}d\vec{r}$$
Where $\omega$ is the frequency of the electric field, $\epsilon_{m}''$ is the imaginary part of the permittivity in a material $m$, $V_{m}$ is the volume of material $m$ and $\vec{E}(\vec{r})$ is the electric field.
The associated photocurrent density may be obtained by integrating the convolution of the wavelength dependent absorption with the number of photons per wavelength from the solar flux,
$$J_{abs} = -e \int_{0}^{\lambda_{g}} A(\lambda) \Phi(\lambda) \frac{\lambda}{hc}d\lambda$$
Where $e$ is the electronic charge, $\lambda_{g}$ is the band gap of the material, $A(\lambda)$ is the wavelength dependent absorption, $\Phi(\lambda)$ is the solar flux, $h$ is Planck’s constant and $c$ is the speed of light in vacuo.
![The real (n) and imaginary (k) parts of the refractive index of CuBi$_{2}$O$_{4}$ measured using spectroscopic ellipsometry.[]{data-label="fig:nk_data"}](SupportingInformation/CuBi2O4_nk_data.pdf){width="80.00000%"}
![The absorption and reflection for the nanostructured 1D grating as a function of the SiO$_2$ slot length. The wavelength of light is 380, 410, 470, 650 nm for (a-d), respectively. The pitch and slot width are 112 and 70 nm, respectively. Arrows in (d) represent lengths for which the electric field data is shown in figure \[fig:Length\_Variation\_Near\_Fields\]. Apart from slot length the geometry is the same as in the inset of figure \[fig:RTA\](b).[]{data-label="fig:Length_Variation"}](SupportingInformation/Length_Variation_Verarbeitet.png){width="\textwidth"}
Figure \[fig:Length\_Variation\] shows the variation in the absorption / reflection of the nanostructured device with respect to the nanoslot length at the four resonance wavelengths identified in figure \[fig:RTA\](b). In all four cases the absorption in both the semiconductor and the metal oscillates for increasing slot length. The $x$ component of the electric field for the four different slot lengths labeled in figure \[fig:Length\_Variation\](d) is shown in figure \[fig:Length\_Variation\_Near\_Fields\]. These field patterns are typical for symmetric MIM modes [@main][Kurokawa2007]{}, due to the field being localized entirely in the dielectric slot and oscillating with a wavelength close to the analytical MIM mode wavelength. The wavelength of the oscillation in the near field is also equal to the distance between the extrema in figure \[fig:Length\_Variation\]. The propagation constant $\beta_{MIM}$ of a MIM mode for an infinitely long dielectric slot surrounded by two semi-infinite half spaces of the same metallic material is given by the implicit set of equations [@main][Maier2007]{}, $$\begin{aligned}
k_{1}^{2} = \beta_{MIM}^2-k_{0}^{2}\epsilon_{1}, \label{eq:MIM1} \\
k_{2}^{2} = \beta_{MIM}^2-k_{0}^{2}\epsilon_{2}, \\
\tanh \frac{k_{1}W}{2} = - \frac{k_{2}\epsilon_{1}}{k_{1}\epsilon_{2}}. \label{eq:MIM3}\end{aligned}$$ where $k$ are wavenumbers in the given medium, $\epsilon$ are permittivities in each given medium and the subscripts 0,1 and 2 refer to vacuum, the dielectric and the metal, respectively. $W$ is the slot width. The wavelength of the mode can then be obtained from $\lambda_{MIM} = \Re(2\pi/\beta_{MIM})$.
![The $x$ component of the electric field for the nanostructured 1D grating for slot lengths (a) 47 nm, (b) 159 nm, (c) 215 nm and (d) 381 nm. Light is incident normally from above with wavelength 650 nm and polarization in the $x-z$ plane.[]{data-label="fig:Length_Variation_Near_Fields"}](SupportingInformation/LengthVariation_NearFields_Cropped.png){width="\textwidth"}
In table \[tab:length\_variation\_wavelengths\] the wavelength in the slot deduced from figures \[fig:Length\_Variation\] and \[fig:Length\_Variation\_Near\_Fields\] is compared to the MIM wavelength at the same vacuum wavelength for a slot width of 70 nm. The very close agreement between the two strongly suggests that these modes can be characterized as MIM modes and is in agreement with the analysis presented in [@main][Wang2013]{}.
------------------------- ----- ----- ----- -----
$\lambda_{vacuum}$ (nm) 380 410 470 650
$\lambda_{MIM}$ (nm) 128 168 216 324
$\lambda_{slot}$ (nm) 156 189 231 344
------------------------- ----- ----- ----- -----
: Comparison of the theoretical metal-insulator-metal (MIM) plasmon resonance wavelength to the resonance wavelength of the slot modes. The MIM resonance wavelength was calculated assuming an infintely long Ag-SiO$_2$-Ag system with a SiO$_2$ thickness of 70 nm using equations \[eq:MIM1\]-\[eq:MIM3\].
\[tab:length\_variation\_wavelengths\]
Figure \[fig:Pitch\_Variation\] shows the variation in the absorption and reflection with respect to the grating pitch for the four resonance wavelengths labeled in figure \[fig:RTA\](b). For wavelengths 380 and 410 nm (a and b, respectively) there is very little dependence of the absorption and reflection on pitch. For very small pitches ($<$150 nm) the distance between the nanoslots is so small that there may be direct coupling between the modes in the slots, which can strongly affect the absorption and reflection. However for larger pitches the behavior is almost monotonic. The absorption in the metal decreases while the absorption in the semiconductor and the reflection increase. This is due to the slot having a smaller proportional effect for larger pitches. The fact that the modes do not show a strong dependence on pitch suggests that the MIM modes at these wavelengths couple to a localized mode since the only effect of changing the pitch is to monotonically reduce the interaction with the slot. It should be noted that Wood’s anomalies are present at pitches of 285 and 305 nm for the wavelengths 380 and 410 nm, respectively. These arise due to the first diffraction order becoming available at these pitches [@app][Enoch2012]{}.
![The absorption and reflection for the nanostructured 1D grating as a function of the grating pitch. The slot width and length are 70 and 60 nm, respectively. The wavelength of light is 380, 410, 470, 650 nm for (a-d), respectively. Arrows in (b,d) represent pitches where the electric field data is shown in figure \[fig:Pitch\_Variation\_Near\_Fields\]. Apart from the pitch, the geometry is the same as in the inset of figure \[fig:RTA\](b).[]{data-label="fig:Pitch_Variation"}](SupportingInformation/Pitch_Variation_Verarbeitet.png){width="100.00000%"}
![The $y$ component of the electric field for the nanostructured 1D grating for grating pitches (a) 112 nm, (b) 315 nm, (c) 525 nm, (d) 725 nm and (e) 725 nm. (a-d) are for incident wavelength 650 nm while (e) is for incident wavelength 410 nm. Light is incident normally from above and polarized in the $x-z$ plane.[]{data-label="fig:Pitch_Variation_Near_Fields"}](SupportingInformation/PitchVariation_All_NearFields.png){width="\textwidth"}
For the wavelengths of 470 and 650 nm the behavior is different. At $\lambda$ = 650 nm there are prominent oscillations in the absorption and in particular for the interface absorption with respect to the pitch. The first four peaks of this enhancement are explicitly labeled in (d). They occur at pitches of 112, 315, 525 and 725 nm. For the wavelength of 470 nm the oscillations in the interface are less obvious, nevertheless the first two peaks in the interface absorption are present at pitches of 105 and 205 nm. For both cases the distance between peaks is very close to the SPP wavelength for a SPP traveling at the CuBi$_2$O$_4$ / Ag interface.
The propagation constant ($\beta_{SPP}$) of a surface plasmon polariton at a planar interface was determined from [@main][Maier2007]{}, $$\beta_{SPP} = k_{0}\sqrt{\frac{\epsilon_{1}\epsilon_{2}}{\epsilon_{1} + \epsilon_{2}}}.
\label{eq:SPP}$$ Where $k_{0}$ is the wavenumber of light in a vacuum. $\epsilon_{1}$ and $\epsilon_{2}$ are the permittivities of Ag and CuBi$_2$O$_4$, respectively. The wavelength of the mode can then be obtained from $\lambda_{SPP} = \Re(2\pi/\beta_{SPP})$.
As in the previous case for the shorter wavelengths, the presence of Wood’s anomalies can be see at pitches of 350 and 700 nm for 470 nm wavelength and pitches of 485 and 975 nm for 650 nm wavelength.
In order to further investigate the oscillations of absorption in figure \[fig:Pitch\_Variation\](d) and the lack of any oscillation in figure \[fig:Pitch\_Variation\](b), the electric field in the device for the pitches labeled by arrows in \[fig:Pitch\_Variation\](b,d) are show in figure \[fig:Pitch\_Variation\_Near\_Fields\]. For parts a-d the presence of a mode traveling at the CuBi$_2$O$_4$ / Ag interface is clearly visible. The distance between phase fronts in figure \[fig:Pitch\_Variation\_Near\_Fields\]a-d, which matches the distance between peaks in absorption enhancement in figure \[fig:Pitch\_Variation\](d), is close to the SPP wavelength for an incident wavelength of 650 nm (see table \[tab:pitch\_variation\_wavelengths\]). Therefore the interface contribution to absorption can be well characterized by a SPP mode with its source at the grating corners that is then absorbed mainly by the CuBi$_2$O$_4$. The peaks in enhancement occur when the distance between the grating corners and the periodic boundary (determined by the pitch) are equal to a half integer number of SPP wavelengths causing constructive interference.
---------------------------- ------ ----- ----- -----
$\lambda_{vacuum}$ (nm) 380 410 470 650
$\lambda_{SPP}$ (nm) 1147 490 205 112
$\lambda_{interface}$ (nm) 200 100
---------------------------- ------ ----- ----- -----
: Comparison of the theoretical surface plasmon polariton (SPP) resonance wavelength at a Ag / CuBi$_2$O$_4$ interface to the wavelength of the modes at the Ag / CuBi$_2$O$_4$ interface in the nanostructured solar fuel device. The SPP wavelength was calculated using equation \[eq:SPP\].
\[tab:pitch\_variation\_wavelengths\]
Figure \[fig:Pitch\_Variation\_Near\_Fields\](e) shows the same pitch (725 nm) as in (d) but for the wavelength of 410 nm. In this case the resonance is clearly localized inside the slot. This agrees with the previous observation that the mode is independent of pitch variations. The localized mode is confined to the bottom corners of the grating, meaning no SPP can be emitted from the top corners.
![The theoretical maximum current density ($J_{abs}$) achievable using the nanostructured back contact for different SiO$_{2}$ slot width and pitch ratio values. Values are for a 1D grating assuming p polarization.[]{data-label="fig:Opt"}](SupportingInformation/Optimization.png){width="50.00000%"}
![(a) The 2D periodic nanostructure formed by Ag and SiO$_2$. The boundary of the periodic unit cell in the $x-y$ plane is shown by the dashed line. The azimuthal angle $\phi$ and polar angle $\theta$ define the direction of the $k$ vector of the incident plane wave and the scattering plane (white). The two linear polarizations which are parallel (p) and perpendicular (s) to the scattering plane are also shown. The angular resolved response of the full device including a 100 nm CuBi$_2$O$_4$ layer and upper H$_2$O and lower Ag half space is shown in figure \[fig:2DGrating\]). (b) Cross section of the periodic unit cell of the 2D grating.[]{data-label="fig:2DSchema"}](SupportingInformation/2D_Grating_Schematic3.png){width="100.00000%"}
Figure \[fig:Opt\] shows the maximum photocurrent density obtainable from the proposed slot waveguide back contact for different combinations of slot width ($W$) and pitch ratio $\rho = P/W$, where $P$ is the pitch. The value of $J_{abs}$ in each case has been taken using the optimal slot length found, therefore each position in figure \[fig:Opt\] has a slot length independent to those surrounding it. There are two main regions of absorption enhancement, the first is contained within the lower and left hand region of the parameter space. This corresponds to the resonance structure observed in fig \[fig:RTA\](b) with two main absorption enhancement peaks between 470 and 650 nm. The geometrical parameters for this optimum value as given in the main text are: pitch = 112 nm, slot depth = 60 nm, slot width = 70 nm.
As the slot width increases the distance between neighboring Ag regions increases, leading to the resonance at 650 nm from \[fig:RTA\](b) being redshifted out of the spectral region of interest. This can be counteracted somewhat by reducing the pitch, meaning that the region of maximum current density shifts to lower pitch ratios for higher slot widths. The second region of high current density is situated in the upper right region of figure \[fig:Opt\]. This region shows a larger number of resonances within the spectral region of interest due to higher order MIM modes arising from the larger pitch and slot width.
The region of maximum current enhancement is fulfilled with a broad range of slot width and pitch combinations, meaning that the proposed structure can be tailored to fit the constraints of the nanostructuring process used to create it.
Figure \[fig:2DSchema\](a) shows the coordinate system used for the 2D grating. Cuboid Ag nanoparticles periodically arrayed in the $x-y$ plane sitting on a Ag back reflector. The Ag nanoparticles are embedded in a SiO$_2$ matrix. The CuBi$_2$O$_4$ layer has not been shown for clarity. Also shown are the incident azimuthal angle ($\phi$) and polar angle ($\theta$) which define the $k$ vector of the incident light as well as the scattering plane. Depending on whether the electric field components are perpendicular or parallel to the scattering plane the incident light can be classified as s or p, respectively. Figure \[fig:2DSchema\](b) shows a cross section of the periodic unit cell of the 2D grating.
[references]{}[References]{}
[^1]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
David Alvarez-Melis\
CSAIL, MIT\
`[email protected]` Tommi S. Jaakkola\
CSAIL, MIT\
`[email protected]`\
bibliography:
- 'SENN.bib'
title: |
Towards Robust Interpretability\
with Self-Explaining Neural Networks
---
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Michael Felderer\
*University of Innsbruck, Austria*\
Jürgen Großmann\
*Fraunhofer FOKUS, Germany*\
Ina Schieferdecker\
*Fraunhofer FOKUS & TU Berlin, Germany*
bibliography:
- 'references.bib'
title: |
Recent Results on\
Classifying Risk-Based Testing Approaches
---
Introduction {#sec:introduction}
============
Testing of safety-critical, security-critical or mission-critical software faces the problem of determining those tests that assure the essential properties of the software and have the ability to unveil those software failures that harm the critical functions of the software. However, also for “normal” less critical software a comparable problem exists: Usually testing has to be done under severe pressure due to limited resources and tight time constraints with the consequence that testing efforts have to be focused and be driven by business risks.
Both decision problems can adequately be addressed by risk-based testing which consider risks of the software product as the guiding factor to steer all phases of a test process, i.e., test planning, design, implementation, execution, and evaluation [@gerrard2002risk; @felderer2014integrating; @felderer2014taxonomy]. Risk-based testing is a pragmatic, in companies of all sizes widely used approach [@felderer2014multiple; @felderer2015sme] which uses the straightforward idea to focus test activities on those scenarios that trigger the most critical situations of a software system [@wendland2012systematic].
Recently, the international standard ISO/IEC/IEEE 29119 Software Testing [@ISO2013SoftwareTesting] on testing techniques, processes, and documentation even explicitly specifies risk considerations to be an integral part of the test planning process. Because of the growing number of available risk-based testing approaches and its increasing dissemination in industrial test processes [@felderer2014framework], methodological support to categorize, assess, compare, and select risk-based testing approaches is required.
In this paper, we present an (updated) taxonomy of risk-based testing that provides a framework for understanding, categorizing, assessing, and comparing risk-based testing approaches and that supports the selection and tailoring of risk-based testing approaches for specific purposes. To demonstrate the application of the taxonomy and its alignment with available risk-based testing approaches, we position four recent risk-based testing approaches, i.e., the RASEN approach [@RASEN_D533], the SmartTesting approach [@ramler2015process], risk-based test case prioritization based on the notion of risk exposure [@yoon2011test] as well as risk-based testing of open source software [@yahav2014risk], in the taxonomy.
A *taxonomy* defines a hierarchy of classes (also referred to as categories, dimensions, criteria or characteristics) to categorize things and concepts. It describes a tree structure whose leaves define concrete values to characterize instances in the taxonomy. The proposed taxonomy is aligned with the consideration of risks in all phases of the test process and consists of the top-level classes *context* (with subclasses risk driver, quality property, and risk item), *risk assessment* (with subclasses factor, estimation technique, scale, and degree of automation), and *risk-based test strategy* (with subclasses risk-based test planning, risk-based test design & implementation, and risk-based test execution & evaluation). The taxonomy presented in this chapter extends and refines our previous taxonomy of risk-based testing [@felderer2014taxonomy].
The remainder of this chapter is structured as follows. Section \[sec:backgr-softw-test\] presents background on software testing and risk management. Section \[sec:taxonomy\] introduces the taxonomy of risk-based testing. Section \[sec:class-recent-risk\] presents the selected four recent risk-based testing approaches and discusses them in the context of the taxonomy. Finally, Section \[sec:summary\] summarizes this chapter.
Background on Software Testing and Risk Management {#sec:backgr-softw-test}
==================================================
*Software testing* [@istqb2012standardGlossary] is the process consisting of all lifecycle activities, both static and dynamic, concerned with planning, preparation and evaluation of software products and related work products to determine that they satisfy specified requirements, to demonstrate that they are fit for purpose and to detect defects. According to this definition it comprises static activities like reviews but also dynamic activities like classic black or white box testing. The tested software-based system is called *system under test* (SUT). As highlighted before, *risk-based testing* (RBT) is a testing approach which considers risks of the software product as the guiding factor to support decisions in all phases of the test process [@gerrard2002risk; @felderer2014integrating; @felderer2014taxonomy]. A *risk* is a factor that could result in future negative consequences and is usually expressed by its likelihood and impact [@istqb2012standardGlossary]. In software testing, the *likelihood* is typically determined by the probability that a failure assigned to a risk occurs, and the *impact* is determined by the cost or severity of a failure if it occurs in operation. The resulting *risk value* or *risk exposure* is assigned to a *risk item*. In the context of testing, a risk item is anything of value (i.e., an asset) under test, for instance, a requirement, a component or a fault.
RBT is a testing-based approach to risk management which can only deliver its full potential if a test process is in place and if risk assessment is integrated appropriately into it. A *test process* consists of the core activities test planning, test design, test implementation, test execution, and test evaluation [@istqb2012standardGlossary] (see Figure \[fig:test-process\]). In the following, we explain the particular activities and associated concepts in more detail.
{width=".25\textwidth"}
According to [@ISO2013SoftwareTesting] and [@istqb2012standardGlossary], *test planning* is the activity of establishing or updating a test plan. A test plan is a document describing the scope, approach, resources, and schedule of intended test activities. It identifies, amongst others, objectives, the features to be tested, the test design techniques, and exit criteria to be used and the rationale of their choice. *Test objectives* are reason or purpose for designing and executing a test. The reason is either to check the functional behavior of the system or its non-functional properties. *Functional testing* is concerned with assessing the functional behavior of an SUT, whereas *non-functional testing* aims at assessing non-functional requirements such as security, safety, reliability or performance. The scope of the features to be tested can be components, integration or system. At the scope of *component testing* (also referred to as unit testing), the smallest testable component, e.g., a class, is tested in isolation. *Integration testing* combines components with each other and tests those as a subsystem, that is, not yet a complete system. In *system testing*, the complete system, including all subsystems, is tested. *Regression testing* is the selective retesting of a system or its components to verify that modifications have not caused unintended effects and that the system or the components still comply with the specified requirements [@Radatz1990IEEEStandardGlossary]. *Exit criteria* are conditions for permitting a process to be officially completed. They are used to report against and to plan when to stop testing. Coverage criteria aligned with the tested feature types and the applied test design techniques are typical exit criteria. Once the test plan has been established, test control begins. It is an ongoing activity in which the actual progress is compared against the plan which often results in concrete measures.
During the *test design* phase the general testing objectives defined in the test plan are transformed into tangible test conditions and abstract test cases. *Test implementation* comprises tasks to make the abstract test cases executable. This includes tasks like preparing test harnesses and test data, providing logging support or writing test scripts which are necessary to enable the automated execution of test cases. In the *test execution* phase, the test cases are then executed and all relevant details of the execution are logged and monitored. Finally, in the *test evaluation* phase the exit criteria are evaluated and the logged test results are summarized in a test report.\
*Risk management* comprises the core activities *risk identification*, *risk analysis*, *risk treatment*, and *risk monitoring* [@ASNZS2004RiskManagement; @ISO31000RiskManaagement]. In the risk identification phase, risk items are identified. In the risk analysis phase, the likelihood and impact of risk items and, hence, the risk exposure is estimated. Based on the risk exposure values, the risk items may be prioritized and assigned to risk levels defining a risk classification. In the risk treatment phase the actions for obtaining a satisfactory situation are determined and implemented. In the risk monitoring phase the risks are tracked over time and their status is reported. In addition, the effect of the implemented actions is determined. The activities risk identification and risk analysis are often collectively referred to as *risk assessment*, while the activities risk treatment and risk monitoring are referred to as *risk control*.
Taxonomy of risk-based testing {#sec:taxonomy}
==============================
The taxonomy of risk-based testing is shown in Figure \[fig:rbt-taxonomy\]. It contains the top-level classes *contextual set up*, *risk assessment* as well as *risk-based test process* and is aligned with the consideration of risks in all phases of the test process. In this section, we explain these classes, their subclasses and concrete values for each class of the risk-based testing taxonomy in depth.
{width="100.00000%"}
Context {#sec:contextual-set-up}
-------
The *context* characterizes the overall context of the risk assessment and testing processes. It includes the subclasses *risk driver*, *quality property* and *risk item* to characterize the drivers that determine the major assets, the overall quality objectives that need to be fulfilled and the items that are subject to evaluation by risk assessment and testing.
### Risk driver {#sec:risk-driver}
A *risk driver* is the first differentiating element of risk-based testing approaches. It characterizes the area of origin for the major assets and thus determines the overall quality requirements, the direction and the general set up of the risk-based testing process. *Business* related assets are required for a successful business practice and thus often directly relate to software quality properties like functionality, availability, security and reliability. *Safety* relates to the inviolability of human health and life and thus requires software to be failsafe, robust and resilient. *Security* addresses the resilience of IT systems against threats that jeopardize confidentiality, integrity and availability of digital information and realted services. Finally, *Compliance* relates to assets that are directly derived from rules and regulations be it applicable laws, standards or other forms of governing settlements. Protection of these assets often but not exclusively relate to quality properties like security, reliability and compatibility.
### Quality property {#sec:quality-property}
A *quality property* is a distinct quality attribute [@ISO_25010] which contributes to the protection of assets and thus, is suject to risk assessment and testing. As stated in [@ISO14971], risks result from hazards. Hazards related to software-based systems stem from software vulnerabilities and from defects in software functionalities, which are critical to business cases, safety-related aspects, security of systems or applicable rules and regulations.
One needs to test that a software-based system is
- functionally suitable, i.e., able to deliver services as requested
- reliable, i.e., able to deliver services as specified over a period of time
- usable, i.e., satisfies the user expectation
- performant and efficient, i.e., able to react appropriately with respect to stated resources and time
- secure, i.e., able to remain protected against accidental or deliberate attacks
- resilient, i.e., able to recover timely from unexpected events
- safe, i.e., able to operate without harmful states
The quality properties considered determine which testing is appropriate and has to be chosen. We consider *functionality*, *security*, and *reliability* to be the dominant quality properties that are addressed for software. They together form the reliability, availability, safety, security, and resilience of a software-based system and hence constitute the options for the risk drivers in the RBT taxonomy.
As reported by different computer emergency response teams such as GovCERT-UK, software defects continue to be a major, if not the main source of incidents caused by software-based systems. The quality properties determine the test types and test techniques that are applied in a test process to find software defects or systematically provide belief in the absence of such defects. Functional testing is likewise a major test type in RBT to analyze reliability and safety aspects, see, e.g., [@Amland2000RBT]. In addition, security testing including penetration testing, fuzz testing and/or randomized testing is key in RBT [@zech2011risk; @ETSI_TR101583] to analyze security and resilience aspects. Furthermore, performance and scalability testing focusing on normal load, maximal load, and overload scenarios to analyze availability and resilience aspects, see, e.g., [@Amland2000RBT].
### Risk item {#sec:quality-item-type}
The *risk item* characterizes and determines the elements under evaluation. These risk items are the elements to which risk exposures and tests are assigned [@felderer2013experiences]. Risk items can be of type *test case* [@yoon2011test], i.e., directly test cases themselves as in regression testing scenarios, *runtime artifact* like deployed services, *functional artifact* like requirements or features, *architectural artifact* like component, or *development artifact* like source code file. The risk item type is determined by the test level. For instance, functional or architectural artifacts are often used for system testing, and generic risks for security testing. In addition, we use the term *artifact* to openly refer to other risk items used in requirements capturing, design, development, testing, deployment, and/or operation and maintenance, which all might relate to the identified risks.
Risk assessment {#sec:risk-assessment}
---------------
The second differentiating element of RBT approaches is the way risks are being determined. According to [@istqb2012standardGlossary], *risk assessment* is the process of *identifying* and subsequently *analyzing* the identified risk to determine its level of risk, typically by assigning likelihood and impact ratings. Risk assessment itself has multiple aspects, so that one needs to differentiate further into the *factors* influencing risks, the risk *estimation technique* used to estimate and/or evaluate the risk, the *scale* type that is used to characterize the risk exposure, and the *degree of automation* for risk assessment.
### Factor {#sec:factors}
The risk factors quantify identified risks [@bai2012risk]. *Risk exposure* is the quantified potential for loss. It is calculated by the likelihood of risk occurrence multiplied by the potential loss, also called the impact. The risk exposure considers typically aspects like liability issues, property loss or damage, and product demand shifts. RBT approaches might also consider the specific aspect of *likelihood* of occurrence, e.g., for test prioritization or selection or the specific aspect of *impact rating* to determine test efforts needed to analyze the countermeasures in the software.
### Estimation technique {#sec:estimation-technique}
The estimation technique determines how the risk exposure is actually estimated and can be *list-based* or *formal model* [@Jorgensen2009EffortEstimation]. The essential difference between formal-model-based and list-based estimation is the quantification step-that is, the final step that transforms the input into the risk estimate. Formal risk estimation models are based on a complex, multi-valued quantification step such as a formula or a test model. On the other hand, list-based estimation methods are based on a simple quantification step-for example, what the expert believes is riskiest. List-based estimation processes range from pure ’gut feelings’ to structured, historical data including failure history and checklist-based estimation processes.
### Scale
Any risk estimation uses a scale to determine the risk “level”. This risk scale can be *quantitative* or *qualitative*. Quantitative risk values are numeric and allow computations, qualitative risk values can only be sorted and compared. An often used qualitative scale for risk levels is low, medium, and high [@wendland2012systematic].
### Degree of automation
Risk assessment can be supported by automated methods and tools. For example, risk-oriented metrics can be measured *manually* or *automatically*. The manual measurement is often supported by strict guidelines and the automatic measurement is often performed via static analysis tools. Other examples for automated risk assessment include the derivation of risk exposures from formal risk models, see, for instance, [@FredriksenKGSOD02].
Risk-based testing strategy {#sec:risk-based-test-process}
---------------------------
Based on the risks being determined and characterized, RBT follows the fundamental test process [@istqb2012standardGlossary] or variations thereof. The notion of risk can be used to optimize already existing testing activities by introducing risk-based strategies for prioritization, automation, selection, resource planning etc. Dependent on the approach, nearly all activities and phases in a test process may be impacted by taking a risk-based perspective. This taxonomy aims for highlighting and characterizing the RBT specifics by relating them to the major phases of a normal test process. For the sake of brevity, we have focused on the phases *risk-based test planning*, *risk-based test design & implementation*, and *risk-based test execution & evaluation* that are outlined in the following subsections.
### Risk-based test planning {#sec:risk-based-test}
The main outcome of test planning is a test strategy and a plan that depicts the staffing, the required resources, as well as a schedule for the individual testing activities. *Test planning* establishes or updates the scope, approach, resources, and schedule of intended test activities. Amongst others, *test objectives*, *test techniques*, and *test completion criteria*, which impact risk-based testing [@redmill2005theory], are determined.
#### Test objective & technique.
*Test objectives & techniques* are relevant parts of a test strategy. They determine what and how to test a test item. The reason to design or execute a test, i.e., a *test objective*, can be related to the risk item to be tested, to the thread scenarios of a risk item or to the counter measures established to secure that risk item, see also Section \[RBTdesign\]. The selection of adequate *test techniques* can be done of basis of the *quality properties* as well as from information related to defects, vulnerabilities and threat scenarios coming from risk assessment.
#### Test completion criterion.
Typical exit criteria for testing that are used to report against and to plan when to stop testing, include all tests ran successfully, all issues have been retested and signed off, or all acceptance criteria have been met. Specific RBT-related exit criteria [@Amland2000RBT] add criteria on the residual risk in the product and coverage-related criteria: all risk items, their threat scenarios and/or counter measures being covered. Risk-based metrics are used to quantify different aspects in testing such as the minimum level of testing, extra testing needed because of high number of faults found, the quality of the tests and the test process. They are used to manage the RBT process and optimize it with respect to time, efforts, and quality [@Amland2000RBT].
#### Resource planning & scheduling.
RBT requires focusing the testing activities and efforts based on the risk assessment of the particular product or of the project, in which it is developed. In simple words: if there is high risk, then there will be serious testing. If there is no risk, then there will be rather little testing. For example, products with high complexity, new technologies, many changes, many defects found earlier, developed by personnel with less experiences or lower qualification, or developed along new or renewed development processes may have a higher probability to fail and need to be tested more thoroughly. Within this context, information from risk assessment can be used to roughly identify high-risk areas or features of the system under test (SUT) and thus determine and optimize the respective test effort, the required personnel and their qualification and the scheduling and prioritization of the activities in a test process.
### Risk-based test design & implementation {#RBTdesign}
*Test design* is the process of transforming test objectives into test cases. This transformation is guided by the coverage criteria, which are used to quantitatively characterize the test cases and often used for exit criteria. Furthermore, the technique of transformation depends on the test types needed to realize a test objective. These test types directly relate to the *quality property* defined in Section \[sec:contextual-set-up\]. *Test implementation* comprises tasks like preparing test harnesses and test data, providing logging support or writing automated test scripts to enable the automated execution of test cases [@istqb2012standardGlossary]. Risk aspects are especially essential for providing *logging support* and for *test automation*.
#### Coverage item determination.
RBT uses coverage criteria specific to the risk artifacts and test types specific to the risk drivers on functionality, security, and safety. The classical code-oriented and model-based coverage criteria like path coverage, condition-oriented coverage criteria like modified condition decision coverage, requirements-oriented coverage criteria like requirements or use case coverage are extended with coverage criteria to cover selected or all assets, threat scenarios, and counter measures [@Stallbaum:2008:ATR:1370042.1370057]. While *asset coverage* rather belongs to requirements-oriented coverage [@wendland2012systematic], *threat scenario & vulnerability coverage*, and *counter measure coverage* can be addressed by code-oriented, model-based, and/or condition-oriented coverage criteria [@hosseingholizadeh2010source].
#### Test or feature prioritization & selection.
In order to optimize the costs of testing and/or the quality and fault detection capability of testing, techniques for prioritizing, selecting, and minimizing tests as well as combinations thereof have been developed and are widely in use [@Yoo:2012:RTM:2284811.2284813]. In the ranges of intolerable risk and “As Low As Reasonably Practicable (ALARP)”[^1] risks, these techniques are used to identify tests for the risk-related test objectives determined before. For example, design-based approaches for test selection [@Briand:2009:ART:1465742.1466092] and coverage-based approaches [@Amland2000RBT] for test prioritization are well-suited for RBT. Dependent on the approach prioritization & selection can take place during different phases of the test process. A risk-based feature or requirement prioritization & selection selects the requirements or features to be tested. This activity is usually started during test planning and continued during test design. Test case prioritization & selection requires existing test specifications or test cases. It is thus either carried out before test implementation to determine the test case to be implemented or in the preparation of test execution or regression testing to determine the optimal test sets to be executed.
#### Test case derivation/generation.
Risk assessment often comprises information about threat scenarios, faults, vulnerabilities that can be used to derive the test data, the test actions, probably the expected results and other testing artifacts. Especially when addressing publicly known threat scenarios, these scenarios can be used to directly refer to predefined and reusable test specification fragments i.e., so called test pattern. These test patterns already contain test actions and test data that are directly applicable to either test specification, test implementation or test execution [@Botella2014].
#### Test automation.
Test automation is the use of special software (separate from the software under test) to control the execution of tests and the comparison of actual outcomes with predicted outcomes [@huizinga2007automated]. Experiences from test automation [@graham2012experiences] show possible benefits like improved regression testing or a positive return on investment, but also caveats like high initial investments or difficulties in test maintenance. Risks may therefore be beneficial to guide decisions where and to what degree testing should be automated.
### Risk-based test execution & evaluation {#sec:risk-based-test-2}
*Test execution* is the process of running test cases. In this phase, risk-based testing is supported by *monitoring* and *risk metrics measurement*. *Test evaluation* comprises decisions on the basis of exit criteria and logged test results compiled in a test report. In this respect, risks are *mitigated* and may require a *re-assessment*. Furthermore, risks may guide *test exit decisions* and *reporting*.
#### Monitoring & risk metrics measurement.
Monitoring is run concurrently with a system under test and supervises, records or analyzes the behavior of the running system [@Radatz1990IEEEStandardGlossary; @istqb2012standardGlossary]. Differing from software testing, which actively stimulates the system under test, monitoring only passively observes a running system. For risk-based testing purposes, monitoring enables additional complex analysis, e.g., of the internal state of a system for security testing, as well as tracking the project’s progress toward resolving its risks and taking corrective action where appropriate. *Risk metrics measurement* determines risk metrics defined in the test planning phase. A measured risk metric could be the number of observed critical failures for risk items where failure has high impact [@felderer2013using].
#### Risk reporting.
Test reports are documents summarizing testing activities and results [@istqb2012standardGlossary] that communicate risks and alternatives requiring a decision. They typically report progress of testing activities against a baseline (such as the original test plan) or test results against exit criteria. In *risk reporting*, assessed risks which are monitored during the test process, are explicitly reported in relation to other test artifacts. Risk reports can be descriptive summarizing relationships of the data or predictive using data and analytical techniques to determine the probable future risk. Typical descriptive risk reporting techniques are risk burn down charts which visualize the development of the overall risk per iteration as well as traffic light reports providing a high level view on risks using colors red for high risks, yellow for medium risks and green for low risks. A typical predictive risk reporting technique is residual risk estimation, for instance, based on software reliability growth models [@goel1985reliability-models].
#### Test & risk re-assessment.
The *re-assessment of risks* after test execution may be planned in the process or triggered by a comparison of test results against the assessed risks. This may reveal deviations between the assessed and the actual risk level and require a re-assessment to adjust them. Test results can explicitly be integrated into a formal risk analysis model [@stallbaum2007employing] or just trigger the re-assessment in an informal way.
#### Test exit decision.
The *test exit decision* determines if and when to stop testing [@felderer2013experiences], but may also trigger further risk mitigation measures. This decision may be taken on the basis of a test report matching test results and exit criteria or ad hoc, for instance, solely on the basis of the observed test results.
#### Risk mitigation.
*Risk mitigation* covers efforts taken to reduce either the likelihood or impact of a risk [@571734]. In the context of risk-based testing, the assessed risks and their relationship to test results and exit criteria (which may be outlined in the test report), may trigger additional measures to reduce either the likelihood or impact of a risk to occur in the field. Such measures may be bug fixing, re-design of test cases or re-execution of test cases.
Classification of Recent Risk-Based Testing Approaches {#sec:class-recent-risk}
======================================================
In this section, we present four recent risk-based testing approaches, i.e., the RASEN approach (Section \[sec:the-rasen-approach\]), the SmartTesting approach (Section \[sec:smarttesting-approach\]), risk-based test case prioritization based on the notion of risk exposure (Section \[sec:riskexposure-approach\]), as well as risk-based testing of open source software (Section \[sec:rbtoss-approach\]), and position them in the risk-based testing taxonomy presented in the previous section.
The RASEN approach {#sec:the-rasen-approach}
------------------
### Description of the approach
The RASEN project (www.rasen-project.eu) has developed a process for combining compliance assessment, security risk assessment and security testing based on existing standards like ISO-31000 and ISO-29119. The approach is currently extended in the PREVENT project (http://www.prevent-project.org) to cover business driven security risk and compliance management for critical banking infrastructures. Figure \[fig:rasen-approach\] shows an overview of the RASEN process.
{width="70.00000%"}
The process covers three distinguishable workstreams that each consist of a combination of typical compliance assessment, security risk assessment activities and/or security testing activities emphasizing the interplay and synergies between these former independent assessment approaches.
1. The test-based security risk assessment workstream starts like a typical risk assessment workstream and uses testing results to guide and improve the risk assessment. Security testing is used to provide feedback on actually existing vulnerabilities that have not been covered during risk assessment or allows risk values to be adjusted on the basis of tangible measurements like test results. Security testing should provide a concise feedback whether the properties of the target under assessment have been really met by the risk assessment.
2. The risk-based compliance assessment workstream targets the identification and treatment of compliance issues. It relies on security risk assessment results to identify compliance risk and thus systematize the identification of compliance issues. Moreover, legal risk assessment may be used to prioritize the treatment of security issues.
3. The risk-based security testing workstream starts like a typical testing workstream and uses risk assessment results to guide and focus the testing. Such a workstream starts with identifying the areas of risk within the target’s business processes and building and prioritizing the testing program around these risks. In this setting risks help focus the testing resources on the areas that are most likely to cause concern or support the selection of test techniques dedicated to already identified threat scenarios.
According ISO 31000, all workstreams start with a preparatory phase called *Establishing the Context* that includes preparatory activities like understanding the business and regulatory environment as well as the requirements and processes. During this first phase the high-level security objectives are identified and documented and the overall process planning is done. Moreover, the process shows additional support activities like *Communication & Consult* and *Monitoring and Review* that are meant to set up the management perspective, thus to continuously control, react, and improve all relevant information and results of the process. From a process point of view, these activities are meant to provide the contextual and management related framework. The individual activities covered in these phases might differ in detail dependent on whether the risk assessment or testing activities are the guiding activities. The main phase, namely the *Security Assessment* phase covers the definition of the integrated compliance assessment, risk assessment and a security testing workstreams.
#### The risk assessment workstream.
The overall risk assessment workstream is decomposed into the three main activities *Risk Identification*, *Risk Estimation*, and *Risk Evaluation*. RASEN has extended the risk identification and risk estimation activities with security testing activities in order to improve the accuracy and efficiency of the overall workstream.
Risk identification is the process of finding, recognizing and describing risks. This consists og identifying sources of risk (e.g., threats and vulnerabilities), areas of impacts (e.g., the assets), malicious events, their causes and their potential impact on assets. In this context, security testing is used to obtain information that eases and supports the identification of threats and threat scenarios. Appropriate are testing and analysis techniques that yield information about the interfaces and entry points (i.e., the attack-surface) like automated security testing, network discovery, web-crawling, and fuzz testing.
Following risk identification, risk estimation is the process of expressing the likelihood, intensity, and magnitude of the identified risks. In many cases, the relevant information on potential threats are often imprecise and insufficient, so that estimation often relies on expert judgment only. This, amongst others, might result in a high degree of uncertainty related to the correctness of the estimates. Testing or test-based risk estimation may increase the amount of information on the target of evaluation. Testing might in particular provide feedback regarding the resilience of systems, i.e., it can support the estimation of the likelihood that an attack will be successful if initiated. Information from testing on the presence or absence of potential vulnerabilities have direct impact on the likelihood values of the associated threat scenarios. Similar to test-based risk identification, penetrating testing tools, model-based security testing tools, static and dynamic code analysis tools, and vulnerability scanners are useful to obtain this kind of information.
#### The compliance assessment workstream.
The risk-based compliance assessment workstream consists of three major steps. The compliance risk identification step provides a systematic and template based approach to identify and select compliance requirements that imply risk. These requirements are transformed into obligations and prohibitions that are the basis for further threat and risk modelling using the CORAS tool. The second step, the compliance risk estimation step is dedicated to understanding and documenting the uncertainty that originates from compliance requirements interpretation. Uncertainty may arise from unclear compliance requirements or from uncertainty about the consequences in case of non-compliance. During compliance risk evaluation compliance requirements are evaluated and prioritized based on their level of risk so that during treatment compliance resources may be allocated efficiently based on their level of risk. In summary, combining security risk assessment and compliance assessment helps prioritizing compliance measures based on risks and helps to identify and deal with compliance requirements that directly imply risk.
#### The security testing workstream.
The risk-based security testing workstream is structured like a typical security testing process. It starts with a test planning phase, followed by a test design & implementation phase and ends with test execution, analysis and summary. The result of the risk assessment, i.e., the identified vulnerabilities, threat scenarios and unwanted incidents, are used to guide the test planning, test identification and may complement requirements engineering results with systematic information concerning the threats and vulnerabilities of a system.
Factors like probabilities and consequences can be additionally used to weight threat scenarios and thus help identifying which threat scenarios are more relevant and thus identifying the ones that need to be treated and tested more carefully. From a process point of view, the interaction between risk assessment and testing could be best described following the phases of a typical testing process.
1. Risk-based security test planning deals with the integration of security risk assessment in the test planning process.
2. Risk-based security test design and implementation deals with the integration of security risk assessment in the test design and implementation process.
3. Risk-based test execution, analysis and summary deals with a risk-based test execution as well as with the systematic analysis and summary of test results.
### Positioning in the risk-based testing taxonomy.
#### Context.
The overall process [@ETSI_EG203251; @Grossmann2015] is directly derived from ISO-31000 and slightly extended to highlight the integration with security testing and compliance assessment. The approach explicitly addresses *compliance* but also *business* and in a limited way *safety* as major *risk drivers*. It is defined independent from any application domain and independent from the level, target or depth of the security assessment itself. It could be applied to any kind of technical assessment process with the potential to target the full number of *quality properties* that are defined in Section \[sec:quality-property\]. Moreover, it addresses legal and compliance issues related to data protection and security regulations. Looking at risk-based security testing, the approach emphasizes executable *risk items*, i.e., *runtime artifacts*. Considering risk-based compliance assessment, the approach also addresses the other risk items mentioned in the taxonomy.
#### Risk assessment.
The test-based risk assessment workstream uses test results as explicit input to various activities of the risk assessment. Risk assessment in RASEN has been carried out on basis of the CORAS method and language. Thus, risk estimation is based on *formal* models that support the definition of *likelihood* values for events and *impact* values to describe the effect of incidents on assets. Both, likelihood and impact values are used to calculate the overall *risk exposure* for unwanted incidents, i.e., the events that directly harm assets. CORAS is flexible with respect to the calculation scheme as well as to the scale for defining risk factors. It generally supports values with *qualitative scale* as well as with *quantitative scale*.
#### Risk-based test strategy.
Security is a non-functional property and thus requires dedicated information that addresses the (security) context of the system. While functional testing is more or less guided directly by the system specification (i.e., features, requirements, architecture), security testing often is not. The RASEN approach to *risk-based security test planning* especially addresses the risk-based selection of *test objectives & test techniques* as well as risk-based *resource planing & and scheduling*. Security risk assessment is serving this purpose and can be used to roughly identify high-risk areas or features of the system under test (SUT) and thus determine and optimize the respective test effort. Moreover, a first assessment of the identified vulnerabilities and threat scenarios may help to select test strategies and techniques that are dedicated to deal with the most critical security risks. Considering *security test design & implementation*, especially the selection and prioritization of the feature to test, the concrete tests design and the determination of *test coverage items* are critical. A recourse to security risks, potential threat scenarios and potential vulnerabilities provide a good guidance to improve *item prioritization & selection*. Security risk related information support the selection of features and test conditions that require testing. It helps in identifying which coverage items should be covered in which depth and how individual test cases and test procedures should look. The RASEN approach to risk-based security test design & implementation uses information on expected threats and potential vulnerabilities to systematically determine and identify coverage items (besides others *asset coverage*, *threat scenario & vulnerabilities coverage* ), test conditions (testable aspects of a system) and test purposes. Moreover, the security risk assessment provides quantitative estimations on the risks, i.e., the product of frequencies or probabilities and estimated consequences. This information is used to select and prioritize either the test conditions or the actual tests when they are assembled into test sets. Risks as well as their probabilities and consequence values are used to set priorities for the test selection, test case generation as well as for the order of test execution expressed by risk-optimized test procedures. Risk-based test execution allows the prioritization of already existing test cases, test sets or test procedures during regression testing. *Risk-based security test evaluation* aims for improving *risk reporting* and the *test exit decision* by introducing the notion of risk coverage and remaining risks on basis of the intermediate test results as well as on basis of the errors, vulnerabilities or flaws that have been found during testing. In summary we have identified the three activities that are supported through results from security risk assessment.
The SmartTesting Approach {#sec:smarttesting-approach}
-------------------------
### Description of the approach
Figure \[fig:smarttesting-process\] provides an overview of the overall process. It consists of different steps, which are either directly related to the risk-based test strategy development (shown in bold font) or which are used to establish the preconditions (shown in normal font) for the process by linking test strategy development to the related processes (drawn with dashed lines) of defect management, requirements management and quality management. The different steps are described in detail in the following subsections.
{width=".7\textwidth"}
#### Definition of risk items.
In a first step, the risk items are identified and defined. The risk items are the basic elements of a software product that can be associated with risks. Risk items are typically derived from the functional structure of the software system, but they can also represent non-functional aspects or system properties. In the context of testing it should be taken into account that the risk items need to be mapped to test objects [@istqb2012standardGlossary], i.e., testable objects such as sub-systems, features, components, modules or functional as well as non-functional requirements.
#### Probability estimation.
In this step the probability values (for which an appropriate scale has to be defined) are estimated for each risk item. In the context of testing the probability value expresses the likelihood of defectiveness of a risk item, i.e., the likelihood that a fault exists in a specific product component due to an error in a previous development phase that may lead to a failure. There are several ways to estimate or predict the likelihood of a component’s defectiveness. Most of these approaches rely on historical defect data collected from previous releases or related projects. Therefore, defect prediction approaches are well suited to support probability estimation [@ramler2016defectsrbt].
#### Impact estimation.
In this step the impact values are estimated for each risk item. The impact values express the consequences of risk items being defective, i.e., the negative effect that a defect in a specific component has on the user or customer and, ultimately, on the company’s business success. The impact is often associated with the cost of failures. The impact is closely related to the expected value of the components for the user or customer. The value is usually determined in requirements engineering when eliciting and prioritizing the system’s requirements. Thus, requirements management may be identified as main source of data for impact estimation.
#### Computation of risk values.
In this step risk values are computed from the estimated probability and impact values. Risk values can be computed according to the definition of risk as $R = P \times I$ where P is the probability value and I is the impact value. Aggregating the available information to a single risk value per risk item allows the prioritization of the risk items according to their associated risk values or ranks. Furthermore, the computed risk values can be used to group risk items, for example, according high, medium and low risk. Nevertheless, for identifying risk levels it is recommended to consider probability and impact as two separate dimensions of risk.
#### Determination of risk levels.
In this step the spectrum of risk values is partitioned into risk levels. Risk levels are a further level of aggregation. The purpose of distinguishing different risk levels is to define classes of risks such that all risk items associated to a particular class are considered equally risky. As a consequence, all risk items of the same class are subject to the same intensity of quality assurance and test measures.
#### Definition of test strategy.
In this step the test strategy is defined on the basis of the different risk levels. For each risk level the test strategy describes how testing is organized and performed. Distinguishing different levels allows testing to be performed with differing levels of rigor in order to adequately address the expected risks. This can either be by achieved by applying specific testing techniques (e.g., unit testing, use case testing, beta testing, reviews) or by applying these techniques with more or less intensity according to different coverage criteria (e.g., unit testing at the level of 100% branch coverage or use case testing for basic flows and/or alternative flows).
#### Refinement of test strategy.
In the last step the test strategy is refined to match the characteristics of the individual components of the software system (i.e., risk items). The testing techniques and criteria that have been specified in the testing strategy for a particular risk level can be directly mapped to the components associated with that risk level. However, the test strategy is usually rather generic. It does not describe the technical and organizational details that are necessary for applying the specified techniques to a concrete software component. For each component, thus, a test approach has to be developed that clarifies how the test strategy should be implemented.
### Positioning in the risk-based testing taxonomy
#### Context.
SmartTesting provides a lightweight process for development and refinement of a risk-based test strategy. It does not explicitly address risk drivers, but - as in every risk-based testing process - implicitly it is assumed that a risk driver and a quality property to be improved are available. The risk drivers of the broad range of companies involved in the accompanying study [@ramler2015process] cover all types, i.e., business, safety and compliance. Also different quality properties of interest are covered, mainly as impact factors. For instance, in the involved companies considered performance and security besides functionality as impact factors.
#### Risk assessment.
SmartTesting explicitly contains a step to define risk items, which can in principle be of any type from the taxonomy. In the case companies, risk items were typically derived from the system’s component structure. Via the process step computation of risk values, SmartTesting explicitly considers *risk exposure*, which is *qualitatively* estimated by a mapping of risk values to risk levels in the process step determination of risk levels. The risk value itself is measured based on a *formal model* in the process step computation of risk values, which combines values from probability and impact estimation. Probability estimation takes defect data into account, and impact estimation is based on impact factors, which are typically *assessed manually*.
#### Risk-based test strategy.
The process steps definition and refinement of the test strategy comprises *risk-based test planning* resulting in the assignment of concrete techniques, resource planning and scheduling, prioritization and selection strategies, metrics as well as exit criteria to the risk levels and further to particular risk items.
Risk-based test case prioritization based on the notion of risk exposure {#sec:riskexposure-approach}
------------------------------------------------------------------------
### Description of the approach
Choi et. al. present different test case prioritization strategies based on the notion of *risk exposure*. In [@yoon2011test], test case prioritization is described as an activity with the aim “to find the most important defects as early as possible against the lowest costs” [@redmill2005theory]. Choi et. al. claim that their risk-based approach to test case prioritization performs well against this background. They empirically evaluate their approach in a setting where various versions of a Traffic Conflict Avoidance System (TCAS) are tested and show how their approach performs well compared to the prioritization approach of others. In [@Hettiarachchi20161] the approach is extended using an improved prioritization algorithm and towards an automated risk estimation process using fuzzy expert systems. A fuzzy expert system is an expert system that uses fuzzy logic instead of Boolean logic to reason about data. Conducting risk estimation with this kind of expert system, Choi et al. aim to replace the human actor during risk estimation and thus avoid subjective estimation results. The second approach has been evaluated by prioritizing test cases for two software products, the electronic health record software *iTrust*, an open source product, and an industrial software application called *Capstone*.
### Positioning in the risk-based testing taxonomy
#### Context.
Both approaches do not explicitly mention one of the risk drivers from Section \[sec:risk-driver\] nor do they provide exhaustive information on the addressed quality properties. However, in [@yoon2011test] the authors evaluate their approach in context of a safety critical application. Moreover, the authors emphasize that they refer to risks that are identified and measured during the product risk assessment phase. Such a phase is typically prescribed for safety critical systems. Both facts indicate that *safety* seems to be the major*risk driver* and the safety relevant attributes like *functionality*, *reliability* and *performance* are the major quality properties that are addressed by testing. In contrast, the evaluation in [@Hettiarachchi20161] is carried out with *business* critical software considering quality properties like *functionality* and *security*. Both approaches have in common that they do not address *compliance* as a risk driver.
#### Risk assessment.
The risk assessment process for both approaches aim for calculating *risk exposure*. The authors define risk exposure as a value with *quantitative* scale that express the magnitude of a given risk. While in [@yoon2011test] the authors explicitly state that they are intentionally not using their own testing related equivalent for expressing risk exposure but directly refer to risk values coming from a pre-existing risk assessment, risk estimation in [@Hettiarachchi20161] is done automatically and tailored towards testing. The authors calculate risks on basis of a number of indicators that are harvested from *development artifacts* like requirements. They use properties like requirements modification status and frequency as well as requirements complexity and size to determine the risk likelihood and risk impact for each requirement. In addition indicators on potential security threats are used to address and consider the notion of *security*. In contrast to [@yoon2011test], [@Hettiarachchi20161] addresses the *automation* of the risk estimation process using an expert system that is able to aggregate the risk indicators and thus to automatically compute the overall risk exposure. While [@yoon2011test] does not explicitly state whether the initial risk assessment relies on formal models or not, the approach in [@Hettiarachchi20161] is completely *formal*. However, since [@yoon2011test] refers to safety critical systems, we can assume that the assessment is not just a list-based assessment.
#### Risk-based test strategy.
With respect to testing, both approaches aim for an efficient *test prioritization & selection* algorithm. Thus, they are mainly applicable in situations where test cases or at least test case specifications are already available. This addresses first of all regression testing but as well decision problems during test management, e.g., when test cases are already specified and the prioritization of test implementation efforts is required.
To obtain an efficient test prioritization strategy, both approaches aim for deriving risk related weights for individual test cases. In [@yoon2011test], the authors propose two different strategies. The first strategy aims for simple *risk coverage*. Test cases that cover a given risk, obtain a weight that directly relates to the risk exposure for that risk. If a test case covers multiple risks, the risk exposure values are summed. The second strategy additionally tries to consider the fault revealing capabilities of the test cases. Thus, the risk related weight for a test case is calculated by means of the risk exposure for a given risk correlated with the number of risk-related faults that are detectable by that test case, so that test cases with a higher fault revealing capabilities are rated higher. Fault revealing capabilities of test cases are derived through mutation analysis, i.e., this strategy requires that the test cases already exist and that they are executable.
In [@Hettiarachchi20161], test cases are prioritized on basis of their relationship to risk-rated requirements. Risk rating for requirements is determined by an automated risk rating conducted by the fuzzy expert system and an additional analysis of fault classes and their relation to the individual requirements. In short, a fault class is considered to have more impact if it relates to requirements with a higher risk exposure. In addition, a fault of given fault class is considered to occur more often if that fault class relates to a larger number of requirements. Both values determine the overall risk rating for the individual requirements and thus provide the prioritization criteria for requirements. Test cases finally are ordered by means of their relationship to the prioritized requirements. During the evaluation of the approach, the authors obtained the relationship between test cases and requirements from existing traceability information.
While both approaches provide strong support for *risk-based item selection*, they do not support other activities during *risk-based test design & implementation* nor do they establish dedicated activities in the area of *risk-based test execution & evaluation*.
Risk-based testing of Open Source Software {#sec:rbtoss-approach}
------------------------------------------
### Description of the approach
Yahav et al. [@yahav2014risk; @yahav2014data] provide an approach to risk-based testing of open source software (OSS) to select and schedule dynamic testing based on software risk analysis. Risk levels of open source components or projects are computed based on communication between developers and users in the open source software community. Communication channels usually include mail, chats, blogs and repositories of bugs and fixes. The provided data-driven testing approach therefore builds on three repositories, i.e., a social repository which stores the social network data from the mined OSS community, a bug repository which links the community behavior and OSS quality, as well as a test repository which traces test (scripts) to OSS projects. As a preprocessing step, OSS community analytics is performed to construct a social network of communication between developers and users. In a concrete case study [@yahav2014data], the approach predicts the expected number of defects for a specific project with logistic regression based on the email communication and the time since the last bug.
### Positioning in the risk-based testing taxonomy
#### Context.
The approach does not explicitly mention one of the risk drivers from Section \[sec:risk-driver\] nor of the quality properties. However, the authors state that the purpose for risk-based testing is the experienced significant failures in product quality, timeliness and delivery cost when adapting OSS components in commercial software packages [@yahav2014risk]. Therefore, risk drivers may be *business* to guarantee the success of a system, where the tested OSS component is integrated, or even safety, if the OSS component would be integrated into a *safety*-critical system. Due to the testing context, i.e., selection or prioritization of available tests of OSS components, the main quality property is supposed to be *functionality*. The risk item type are OSS components (developed in OSS projects), i.e., *architectural artifacts*.
#### Risk Assessment.
The risk assessment approach quantifies the *likelihood*. For this purpose, a *formal model* is created to predict the number of bugs based on the communication in different communities and the time since the last bug. The scale is therefore *quantitative* as the approach tries to predict the actual number of bugs. The approach implements an *automatic assessment* as it uses monitors to automatically store data in repositories and then applies machine learning approaches, i.e., logistic regression, to predict the risk level.
#### Risk-based test strategy.
The approach explicitly supports risk-based test planning in terms of *test prioritization & selection* and *resource planning & scheduling*. The approach mainly addresses the allocation of available test scripts for dynamic testing and highlights that exhaustive testing is infeasible and that therefore selective testing techniques are needed to allocate test resources to the most critical components. Therefore, risk-based test design is not explicitly addressed. As *risk metrics* the number of communication metrics as well as the time since the last defect are computed. The approach uses specific *logging support* to log and trace community and defect data. For *risk reporting* confusion matrices are used, which contrast the actual and predicted number of defects.
Summary {#sec:summary}
=======
In this chapter, we presented a taxonomy of risk-based testing. It is aligned with the consideration of risks in all phases of the test process and consists of three top-level classes, i.e., contextual set up, risk assessment, and risk-based test strategy. The contextual set up is defined by risk drivers, quality properties and risk items. Risk assessment comprises the subclasses factors, estimation technique, scale, and degree of automation. The risk-based test strategy then takes the assessed risks into account to guide test planning, test design & implementation, as well as test execution & evaluation. The taxonomy provides a framework to understand, categorize, assess and compare risk-based testing approaches to support their selection and tailoring for specific purposes. To demonstrate its application and alignment with available risk-based testing approaches, we positioned four recent risk-based testing approaches, i.e., the RASEN approach, the SmartTesting approach, risk-based test case prioritization based on the notion of risk exposure as well as risk-based testing of Open Source Software, in the taxonomy.
[^1]: The ALARP principle is typically used for safety-critical, but also for mission-critical systems. It says that the residual risk shall be as low as reasonably practical.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The sensitivity of inertial confinement fusion implosions of the type performed on the National Ignition Facility (NIF) [@lindl:339] to low-mode flux asymmetries has been investigated numerically. It is shown that large-amplitude, low-order mode shapes (Legendre polynomial $P_4$), resulting from associated low order flux asymmetries, cause spatial variations in capsule & fuel momentum that prevent the DT “ice” layer from being decelerated uniformly by the hot spot pressure. This reduces the transfer of kinetic to internal energy of the central hot spot, thus reducing neutron yield. Furthermore, synthetic gated x-ray images of the hot spot self-emission indicate that $P_4$ shapes may be unquantifiable for DT layered capsules. Instead the positive $P_4$ asymmetry “aliases” itself as an oblate $P_2$ in the x-ray self emission images. Correction of this apparent $P_2$ distortion can further distort the implosion while creating a round x-ray image. Long wavelength asymmetries may be playing a significant role in the observed yield reduction of NIF DT implosions relative to detailed post-shot 2D simulations.'
author:
- 'R.H.H. Scott'
- 'D.S. Clark'
- 'D.K. Bradley'
- 'D.A. Callahan'
- 'M.J. Edwards'
- 'S.W. Haan'
- 'O.S. Jones'
- 'B.K. Spears'
- 'M.M. Marinak'
- 'R.P.J. Town'
- 'P.A. Norreys'
- 'L.J. Suter'
title: 'Numerical Modeling of the Sensitivity of X-Ray Driven Implosions to Low-Mode Flux Asymmetries'
---
[^1]
Indirect-drive inertial confinement fusion (ICF) [@Basov:1991uq; @NUCKOLLS:1972fk; @lindl:339] uses lasers to heat the inside of a cavity (or hohlraum). The absorbed laser energy is re-emitted as approximately black-body radiation in the soft x-ray regime. These x-rays heat the outer surface of a hollow, spherical, $\sim 2$ mm diameter, shell that contains a $\sim 70$ m thick layer of cryogenically frozen Deuterium and Tritium fuel (“DT fuel” or “DT layered capsules”). The heated outer shell ablates, which creates a reaction force, accelerating the remaining shell spherically inwards at extremely high velocity ($\sim 350$ km/s). During the implosion, spherical convergence causes the pressure in the gaseous void (or hot spot) within the shell to rise rapidly. This pressure decelerates the shell, simultaneously compressing the solid fuel and converting the shell’s kinetic energy into hot spot internal energy. If this conversion rate exceeds loss rates due to thermal conduction and bremsstrahlung radiation, the hot spot will heat, initiating DT fusion reactions. Provided the hot spot areal density is sufficient, $\alpha$-particles created by the fusion reactions will redeposit their energy locally, further heating the hot spot, resulting in bootstrap heating, ignition, and propagation of burn into the surrounding cold fuel. Numerical modeling indicates that the National Ignition Facility (NIF) can, for the first time, initiate inertial fusion ignition in the laboratory [@haan:051001; @clark:052703; @1742-6596-112-2-022021].
In this Letter, the effects of large, low-mode asymmetries in the x-ray drive are examined numerically. The non-uniformity of the x-ray flux incident upon the shell and the resultant non-spherical shell shapes can be described mathematically as a series of Legendre polynomials [@Abramowitz:1972kx]. It is shown that a large-amplitude $P_4$ implosion asymmetry, that might result from low-order hohlraum generated flux asymmetries, causes spatial variations in the capsule & fuel momentum. This can inhibit the DT fuel from being decelerated uniformly by the hot spot pressure, reducing the efficient transfer of implosion kinetic energy to hot spot internal energy thus significantly reducing the capsule performance. Furthermore, simulated gated x-ray images of the hot spot self-emission show reduced sensitivity to the $P_4$ mode, instead the images appear to have a pronounced oblate $P_2$ shape. Reducing the amplitude of the oblate $P_2$ shape (as measured from the x-ray image) further reduces the sensitivity to the $P_4$ mode such that no quantitative evaluation of the hot spot $a_4$ (where $a_4$ is the amplitude of the $P_4$ mode) can be made, furthermore the x-ray images are circular despite the capsule shape being highly distorted. Comparisons are made between key physical properties of the implosion, synthetically generated experimental observables, and NIF experimental data.
![(a) Applied radiation temperature vs time. Colors depict when during the drive pulse the various flux asymmetries were applied as shown on subsequent plots. (b) The relationship between applied flux asymmetry and hot spot $a_4$ at x-ray bangtime (the time at which peak emission occurs). Each point represents a 2D Hydra run. (c) Burn averaged hot spot + fuel + ablator $\rho r$ vs hot spot $a_4$ at x-ray bangtime: large dots are spatially averaged $\rho r$, while the smaller points with the same color and $a_4$ are the maxa and minima of the spatially averaged value; large variations in $\rho r$ occur due to $P_4$. (d) The burn averaged energy partition as a function of hot spot $a_4$; increasing $P_4$ perturbations prevent kinetic energy (black) from being converted to both hot spot (red) and solid fuel (blue) internal energy during stagnation. (e) Burn averaged hot spot pressure as a function of hot spot $a_4$. (f) Total thermonuclear neutron yield as a function of hot spot $a_4$; yield varies by a factor of 15. []{data-label="fig2"}](./trad_time.png "fig:") ![(a) Applied radiation temperature vs time. Colors depict when during the drive pulse the various flux asymmetries were applied as shown on subsequent plots. (b) The relationship between applied flux asymmetry and hot spot $a_4$ at x-ray bangtime (the time at which peak emission occurs). Each point represents a 2D Hydra run. (c) Burn averaged hot spot + fuel + ablator $\rho r$ vs hot spot $a_4$ at x-ray bangtime: large dots are spatially averaged $\rho r$, while the smaller points with the same color and $a_4$ are the maxa and minima of the spatially averaged value; large variations in $\rho r$ occur due to $P_4$. (d) The burn averaged energy partition as a function of hot spot $a_4$; increasing $P_4$ perturbations prevent kinetic energy (black) from being converted to both hot spot (red) and solid fuel (blue) internal energy during stagnation. (e) Burn averaged hot spot pressure as a function of hot spot $a_4$. (f) Total thermonuclear neutron yield as a function of hot spot $a_4$; yield varies by a factor of 15. []{data-label="fig2"}](./DTp4_bt_fluxmyC.png "fig:") ![(a) Applied radiation temperature vs time. Colors depict when during the drive pulse the various flux asymmetries were applied as shown on subsequent plots. (b) The relationship between applied flux asymmetry and hot spot $a_4$ at x-ray bangtime (the time at which peak emission occurs). Each point represents a 2D Hydra run. (c) Burn averaged hot spot + fuel + ablator $\rho r$ vs hot spot $a_4$ at x-ray bangtime: large dots are spatially averaged $\rho r$, while the smaller points with the same color and $a_4$ are the maxa and minima of the spatially averaged value; large variations in $\rho r$ occur due to $P_4$. (d) The burn averaged energy partition as a function of hot spot $a_4$; increasing $P_4$ perturbations prevent kinetic energy (black) from being converted to both hot spot (red) and solid fuel (blue) internal energy during stagnation. (e) Burn averaged hot spot pressure as a function of hot spot $a_4$. (f) Total thermonuclear neutron yield as a function of hot spot $a_4$; yield varies by a factor of 15. []{data-label="fig2"}](./DTrhor_DTp4myC.png "fig:") ![(a) Applied radiation temperature vs time. Colors depict when during the drive pulse the various flux asymmetries were applied as shown on subsequent plots. (b) The relationship between applied flux asymmetry and hot spot $a_4$ at x-ray bangtime (the time at which peak emission occurs). Each point represents a 2D Hydra run. (c) Burn averaged hot spot + fuel + ablator $\rho r$ vs hot spot $a_4$ at x-ray bangtime: large dots are spatially averaged $\rho r$, while the smaller points with the same color and $a_4$ are the maxa and minima of the spatially averaged value; large variations in $\rho r$ occur due to $P_4$. (d) The burn averaged energy partition as a function of hot spot $a_4$; increasing $P_4$ perturbations prevent kinetic energy (black) from being converted to both hot spot (red) and solid fuel (blue) internal energy during stagnation. (e) Burn averaged hot spot pressure as a function of hot spot $a_4$. (f) Total thermonuclear neutron yield as a function of hot spot $a_4$; yield varies by a factor of 15. []{data-label="fig2"}](./DTkeshell_DTp4myC.png "fig:") ![(a) Applied radiation temperature vs time. Colors depict when during the drive pulse the various flux asymmetries were applied as shown on subsequent plots. (b) The relationship between applied flux asymmetry and hot spot $a_4$ at x-ray bangtime (the time at which peak emission occurs). Each point represents a 2D Hydra run. (c) Burn averaged hot spot + fuel + ablator $\rho r$ vs hot spot $a_4$ at x-ray bangtime: large dots are spatially averaged $\rho r$, while the smaller points with the same color and $a_4$ are the maxa and minima of the spatially averaged value; large variations in $\rho r$ occur due to $P_4$. (d) The burn averaged energy partition as a function of hot spot $a_4$; increasing $P_4$ perturbations prevent kinetic energy (black) from being converted to both hot spot (red) and solid fuel (blue) internal energy during stagnation. (e) Burn averaged hot spot pressure as a function of hot spot $a_4$. (f) Total thermonuclear neutron yield as a function of hot spot $a_4$; yield varies by a factor of 15. []{data-label="fig2"}](./DTprhs_DTp4myC.png "fig:") ![(a) Applied radiation temperature vs time. Colors depict when during the drive pulse the various flux asymmetries were applied as shown on subsequent plots. (b) The relationship between applied flux asymmetry and hot spot $a_4$ at x-ray bangtime (the time at which peak emission occurs). Each point represents a 2D Hydra run. (c) Burn averaged hot spot + fuel + ablator $\rho r$ vs hot spot $a_4$ at x-ray bangtime: large dots are spatially averaged $\rho r$, while the smaller points with the same color and $a_4$ are the maxa and minima of the spatially averaged value; large variations in $\rho r$ occur due to $P_4$. (d) The burn averaged energy partition as a function of hot spot $a_4$; increasing $P_4$ perturbations prevent kinetic energy (black) from being converted to both hot spot (red) and solid fuel (blue) internal energy during stagnation. (e) Burn averaged hot spot pressure as a function of hot spot $a_4$. (f) Total thermonuclear neutron yield as a function of hot spot $a_4$; yield varies by a factor of 15. []{data-label="fig2"}](./DTnyield_DTp4myC.png "fig:")
The indirect-drive approach to ICF smooths high mode spatial non-uniformities in the x-ray flux incident on the capsule, however the spatial distribution of the cones of laser beams which illuminate the hohlraum means that low mode x-ray flux non-uniformities can occur [@lindl:339], these are considerably lower mode than those recently examined by Thomas *et al* [@PhysRevLett.109.075004]. The growth of Legendre polynomial $P_4$ capsule shapes was investigated using the radiation-magnetohydrodynamics code Hydra [@marinak:2275]. A frequency and time dependent x-ray source was developed to drive these capsule-only simulations. The initial x-ray drive was taken from an integrated hohlraum simulation and then adjusted to match the shock timing data obtained using the VISAR diagnostic [@barker:4669] from NIF shot N110521, and the capsule implosion trajectory [@hicks:102703] measured on NIF shot N110625. A $90^{\circ}$ ‘wedge’ of the capsule ($2\pi$ Sr) was modeled using two-dimensional (2D) cylindrically-symmetric geometry with $256\times312$ cells. Doubling and quadrupling the cell resolution demonstrated convergence. In all runs the Quotidian Equation of State [@more:3059] was used with tabular opacities and multi-group radiation diffusion. The effects of Legendre polynomial $P_4$ hohlraum flux asymmetries were investigated by perturbing the tuned x-ray drive with spatially varying flux asymmetries of the form: $fds(\theta,t)=(a_0P_0 + a_4P_4(\theta))*fds(t)$ where $fds$ is the energy density of the tuned photon frequency dependent x-ray drive source, $a_n$ is the amplitude of the $n^{th}$ Legendre polynomial, $a_0=1$, $a_4= (\pm0.10, \pm0.05, \pm0.025, \pm0.01, \pm0.005, \pm0.0025)$, $\theta$ the angle between the equatorial plane and polar axis, and $t$ time. Hydra modeling of the hohlraum & capsule for nominal implosions suggests the flux asymmetry incident on the capsule would be expected to vary by $<3$% except for in the first $\sim 2$ ns of the laser pulse were it can vary by up to 10% [@Jones:2012]. The flux asymmetries were applied 100 m from the capsule ablation front during discrete time intervals (see fig. \[fig2\](a)), creating a database of $>200$ 2D modeling runs of both DT layered implosions and DHe$^3$ gas filled capsules with a surrogate inner CH layer of equal mass to a DT fuel layer (symmetry capsules). Time resolved synthetic gated x-ray images of the hot spot self-emission $>6$ keV, including its attenuation by the compressed fuel and ablator, were created from both polar and equatorial directions by post processing each Hydra run. The images were blurred in order to reproduce the $11$ m resolution of the diagnostic. The key implosion performance metrics (neutron yield, hot spot pressure, mass, volume, density, ion & electron temperatures, the effective ion temperature computed from the FWHM of the DT neutron spectrum, fuel and ablator areal density ($\rho r$) and kinetic energy) were extracted from the simulations. The hot spot shape was evaluated as a function of time by performing a Legendre polynomial decomposition (modes 1-10) of the appropriate contour. For DT layered capsules the hot spot contour is defined for each angular ‘strip’ of cells $j$ as the minimum radius where $T_{e_j} > \frac{1}{2}T_{e_{j_{max}}}$ and $\rho_j < \frac{1}{2}\rho_{j_{max}}$ where $T_{e}$ is the electron temperature and $\rho$ the mass density, ‘max’ denotes the maximum value within the $j^{th}$ strip. This has been found to produce a robust definition of the hot spot even for highly distorted implosions. The 17% contour of the gated x-ray diagnostic (GXD) is used both for the synthetic GXD and experimentally, as previous studies have shown this provides a faithful representation of the hot spot shape for small departures from sphericity.
![Axis of rotational symmetry is horizontal at Radius $=0$ m. (a) DT layered capsule density plot at x-ray bangtime showing a positive Legendre polynomial $P_4$ shape. This simulation had a 10% flux asymmetry applied from 11.5-14 ns. Black arrows indicate the mass flows which occur during stagnation. After bangtime ‘fingers’ of fuel continue to flow inwards (red arrows). White dots depict the hot spot contour. (b) Synthetic gated x-ray image of the hot spot self emission from (a), white dots show the 17% contour. (c) Fig. (a) 100 ps later: due to burn truncation with large $a_4$ this is the neutron bangtime for an equivalent spherical implosion. (d) The synthetic GXD from (c), showing a large negative $P_2$ and almost zero $a_4$ despite the obvious $P_4$ in (c).[]{data-label="picture"}](./density_paper.png "fig:") ![Axis of rotational symmetry is horizontal at Radius $=0$ m. (a) DT layered capsule density plot at x-ray bangtime showing a positive Legendre polynomial $P_4$ shape. This simulation had a 10% flux asymmetry applied from 11.5-14 ns. Black arrows indicate the mass flows which occur during stagnation. After bangtime ‘fingers’ of fuel continue to flow inwards (red arrows). White dots depict the hot spot contour. (b) Synthetic gated x-ray image of the hot spot self emission from (a), white dots show the 17% contour. (c) Fig. (a) 100 ps later: due to burn truncation with large $a_4$ this is the neutron bangtime for an equivalent spherical implosion. (d) The synthetic GXD from (c), showing a large negative $P_2$ and almost zero $a_4$ despite the obvious $P_4$ in (c).[]{data-label="picture"}](./btgxd_paper.png "fig:") ![Axis of rotational symmetry is horizontal at Radius $=0$ m. (a) DT layered capsule density plot at x-ray bangtime showing a positive Legendre polynomial $P_4$ shape. This simulation had a 10% flux asymmetry applied from 11.5-14 ns. Black arrows indicate the mass flows which occur during stagnation. After bangtime ‘fingers’ of fuel continue to flow inwards (red arrows). White dots depict the hot spot contour. (b) Synthetic gated x-ray image of the hot spot self emission from (a), white dots show the 17% contour. (c) Fig. (a) 100 ps later: due to burn truncation with large $a_4$ this is the neutron bangtime for an equivalent spherical implosion. (d) The synthetic GXD from (c), showing a large negative $P_2$ and almost zero $a_4$ despite the obvious $P_4$ in (c).[]{data-label="picture"}](./density2_paper.png "fig:") ![Axis of rotational symmetry is horizontal at Radius $=0$ m. (a) DT layered capsule density plot at x-ray bangtime showing a positive Legendre polynomial $P_4$ shape. This simulation had a 10% flux asymmetry applied from 11.5-14 ns. Black arrows indicate the mass flows which occur during stagnation. After bangtime ‘fingers’ of fuel continue to flow inwards (red arrows). White dots depict the hot spot contour. (b) Synthetic gated x-ray image of the hot spot self emission from (a), white dots show the 17% contour. (c) Fig. (a) 100 ps later: due to burn truncation with large $a_4$ this is the neutron bangtime for an equivalent spherical implosion. (d) The synthetic GXD from (c), showing a large negative $P_2$ and almost zero $a_4$ despite the obvious $P_4$ in (c).[]{data-label="picture"}](./gxd+100ps_paper.png "fig:")
The applied Legendre $P_4$ flux asymmetries induce $P_4$ hot spot shapes at stagnation (see Figs. \[picture\] (a) & (c)), the sign of which is dependent on the timing of the applied flux asymmetry. If the asymmetry is present only during the shock compression phase (the first $\sim 18$ ns), shocks created in regions of the capsule exposed to higher flux propagate faster, these faster shocks break out of the inner DT ice layer earlier, causing these regions to move ahead of those exposed to less flux. This also causes ablator mass to flow laterally, away from the high flux region. Consequently during peak drive the regions initially exposed to high flux are at smaller radii, meaning they are accelerated less efficiently by the hohlraum flux and gain less total momentum. They can also have less ablator aerial density. The net effect is that the regions experiencing high flux during shock compression will protrude outwards at stagnation. Conversely if the flux asymmetry is applied during peak drive, the regions of the capsule exposed to more flux gain more momentum, and protrude inwards at stagnation. Regardless of the timing of the applied asymmetry, during the stagnation phase of the implosion, pressure within the lower density hot spot decelerates the higher density fuel from peak velocity, making this interface Rayleigh-Taylor unstable [@Rayleigh:1900kx; @Taylor:1950vn]. The instability will accentuate any shape imperfections during deceleration, as indicated by the significant simulated growth shown in figs. \[picture\](a) & (c).
The scalings of some important DT layered capsule implosion parameters as a function of hot spot $a_4$ measured at x-ray bangtime are summarized in figure \[fig2\]. Fig. \[fig2\](b) shows the relationship between applied $P_4$ flux perturbation amplitude and the resulting shape $a_4$ at x-ray bangtime. Fig. \[fig2\](c) depicts the ‘burn averaged’ $\rho r$ (the burn average of a quantity $Q_b = (\sum_{t=0}^{t=\infty} Q_t Eprodr dt)/\int_{t=0}^{t=\infty}Eprodr\ dt$ where $Q_t$ is Q at time $t$ and $Eprodr$ the thermonuclear energy production rate in time $dt$) as a function of hot spot $a_4$. Although the spatially averaged $\rho r$ is relatively constant, the lateral mass flows caused by the $P_4$ can create large spatial variations in $\rho r$. The regions with higher momentum continue to propagate radially inwards; fig. \[fig2\](d) depicts the remaining capsule kinetic energy as a function of $a_4$ and the partition of that energy into hot spot and solid fuel internal energy. For large $a_4$ less of the implosion kinetic energy is converted into hot spot internal energy and the hot spot pressure is reduced (see fig. \[fig2\](e)). The reduction in neutron yield can be as large as 15$\times$ for hot spot $a_4=20$ m (flux asymmetry $\sim 10\%$) as shown in fig. \[fig2\](f)).
![(a) DT layered capsule hot spot $a_4$ plotted against the synthetic GXD $a_4$; particularly for large positive $a_4$ the GXD is unable to effectively measure the amplitude of the $P_4$ mode. (b) DT layered capsule hot spot $a_4$ plotted against the synthetic GXD $a_2$; the GXD measures a significant $P_2$ mode amplitude despite the DT layered capsule hot spot $a_2$ being $0\pm1$ m (not shown).[]{data-label="fig3"}](./DTp4myC_DTp4gxd17pc.png "fig:") ![(a) DT layered capsule hot spot $a_4$ plotted against the synthetic GXD $a_4$; particularly for large positive $a_4$ the GXD is unable to effectively measure the amplitude of the $P_4$ mode. (b) DT layered capsule hot spot $a_4$ plotted against the synthetic GXD $a_2$; the GXD measures a significant $P_2$ mode amplitude despite the DT layered capsule hot spot $a_2$ being $0\pm1$ m (not shown).[]{data-label="fig3"}](./DTp4myC_DTp2gxd17pc.png "fig:")
Analysis of synthetic GXD images created from the 2D Hydra runs suggest that the $a_4$ measured experimentally with the GXD is not a true representation of the hot spot $a_4$, particularly for large positive $a_4$ amplitudes. Fig. \[fig3\](a) depicts the relationship between the DT layered capsule “hot spot $a_4$” (as previously defined) and that from the 17% contour of the synthetic GXD (the synthetic GXD $a_4$), both were extracted at x-ray bangtime (the principal value used for analysis of experimental data). $a_4$ measured from the synthetic GXD is consistently lower than that of the hot spot. The insensitivity to positive hot spot $a_4$ is caused by lateral ablator mass flows which accumulate at $\sim 45 ^{\circ}$ at the expense of ablator material near the equator and poles (see Fig. \[picture\] (a)). The ablator material is rotationally symmetric about the horizontal axis, so the accumulated material absorbs the x-rays emitted from the polar-lobes of the hot spot (left and right), while allowing x-rays to more readily pass through the equatorial regions (top and bottom). Consequently the polar-lobes of the hot spot which are visible in the density plots of Fig. \[picture\] as dark regions (the hot spot is the central region of low density) are almost completely invisible in the GXD plots compared to the emission through the equator. This causes the x-ray image to have a pronounced negative $P_2$ shape (oblate or “pancaked”). As the hot spot $a_2=0\pm 1$ m ($a_2$ is the amplitude of the $P_2$ mode) for all these pure $P_4$ modelling runs, the $P_2$ inferred from the x-ray image is a “false” negative $P_2$ mode. This suggests that a negative $P_2$ mode measured from the self-emission x-ray image may in fact be a signature of a positive $P_4$ mode, although it does not, of course, preclude the presence of a true $P_2$ mode. This is potentially important for interpretation of x-ray images from DT implosions, which often exhibit oblate (negative) $P_2$ modes [@Glenzer05032010].
NIF experiments also use low convergence, DHe$^3$ gas filled “symmetry capsules” which have a surrogate CH fuel mass. Using Hydra, symmetry capsule & DT layered capsule pairs of runs were created by applying identical x-ray drives to both capsules. When realistic experimental noise is applied to the synthetic GXD images, symmetry capsules, which have better signal to noise ratios and larger stagnation diameters, enable a far better measurement of the $P_4$ mode than equivalent DT layered capsules. Nevertheless, these calculations indicate symmetry capsules also show reduced sensitivity to $a_4$, exhibit a “false” $P_2$, and are quantitatively very similar to those shown in Figs. \[fig3\] (a) & (b) respectively.
![(a) Density plot of a DT layered capsule run with both $P_2$ and $P_4$ flux modes applied. Axis of rotational symmetry is horizontal at Radius $=0$ m. (b) The equatorial synthetic GXD image of (a) at the same time, dotted line shows 17% contour. Despite the highly non-spherical density distribution, the equatorial GXD image is almost perfectly round. (c) As the $P_2$ flux amplitude is increased in order to make the GXD image look round (reducing $a_2$) the sensitivity to $a_4$ is reduced towards zero.[]{data-label="fig5"}](./10pcP2_10pcP4_DT_density_at_xray_bt.png "fig:") ![(a) Density plot of a DT layered capsule run with both $P_2$ and $P_4$ flux modes applied. Axis of rotational symmetry is horizontal at Radius $=0$ m. (b) The equatorial synthetic GXD image of (a) at the same time, dotted line shows 17% contour. Despite the highly non-spherical density distribution, the equatorial GXD image is almost perfectly round. (c) As the $P_2$ flux amplitude is increased in order to make the GXD image look round (reducing $a_2$) the sensitivity to $a_4$ is reduced towards zero.[]{data-label="fig5"}](./10pcP2_10pcP4_DTequatorial_gxd_at_xray_bt.png "fig:") ![(a) Density plot of a DT layered capsule run with both $P_2$ and $P_4$ flux modes applied. Axis of rotational symmetry is horizontal at Radius $=0$ m. (b) The equatorial synthetic GXD image of (a) at the same time, dotted line shows 17% contour. Despite the highly non-spherical density distribution, the equatorial GXD image is almost perfectly round. (c) As the $P_2$ flux amplitude is increased in order to make the GXD image look round (reducing $a_2$) the sensitivity to $a_4$ is reduced towards zero.[]{data-label="fig5"}](./p2p4DTgxdSensitivity.png "fig:")
DT implosions on the NIF currently have yields $\sim 3 - 10 \times$ below detailed 2D post-shot Hydra simulations [@Clark:2012fk] that match the measured shock timing, implosion velocity, and capsule and ice surface roughnesses. In comparison to experimentally measured or inferred values [@Springer:2011uq], these simulations have similar hot spot temperatures, however the hot spot volumes are reduced while the hot spot mass is increased, causing a $2-3\times$ discrepancy in the hotspot density. $P_4$ shape perturbations offer one mechanism which may explain these experimental observations in particular bringing the yield and ion temperature relationship into better agreement. In these simulations, the DT fuel and hot spot do not mix; clear boundaries still exist (note these simulations use smooth capsules, but when nominal realistic capsule surface roughness [@clark:082701] was employed and modes up to 200 resolved, no significant implosion degradation occurred for the full range of $a_4$). Consequently unlike high mode ‘mix’ [@lindl:339] (where the hot spot can be radiatively cooled by high $Z$ impurities), the simulated ion temperature inferred from the neutron spectrum remains unaffected at $3.9\pm 0.05$ keV for all $a_4$. The large $a_4$ does however truncate the thermonuclear burn, moving both the neutron and x-ray bangtimes earlier in time, therefore as the capsule is still converging at bangtime, the hot spot size and volume are increased. The hot spot mass decreases with positive $a_4$, bringing Hydra simulations approximately in line with experimental data, as shown in Table \[table1\]. This compares NIF experimental data [@Springer:2011uq] with two Hydra implosions; one is perfectly spherical while the other has a hot spot $a_4$ of +20 m. Notable features are the significantly reduced yield, reduced pressure, reduced hot spot mass, unchanged ion temperature and increased hot spot volume. We must emphasize, however, that this should not be interpreted as conclusive evidence that a $P_4$ asymmetry is responsible for the observed reduced NIF capsule performance. Although this study has concentrated on the $P_4$ mode, it is likely that all low modes would reduce the conversion of capsule kinetic energy into hot spot pressure, and may result in similar ambiguity in the shape of the x-ray emission from the hot spot [@B.K.:2012uq]. To explore the issue of low mode asymmetries further, experiments using x-ray backlighters are currently being conducted on NIF to measure the implosion shape both in-flight [@hicks:102703] and at stagnation using Compton radiography [@tommasini:056309].
-------------------------------- ------------------------- --------------- ---------------- -- -- -- -- --
NIF expt. Hydra Hydra
range[@Springer:2011uq] ($a_4 = 0$ m) ($a_4 = 20$ m)
Pressure (GBar) 57-81 348 115
Mass (g) 2-6.4 8 5.5
Density (gcm$^{-3}$) 22-35 136 69
Volume ($\times 10^{-7}cm^3$) 0.9-1.9 0.6 1.0
Tion (keV) 3.3-4.4 3.9 3.9
Fuel $\rho r$ (gcm$^{-2}$) 0.77-0.98 0.7 0.72
Yield (neut. $\times 10^{14}$) 1.9-6.0 74 5.3
-------------------------------- ------------------------- --------------- ---------------- -- -- -- -- --
: A comparison of NIF DT layered capsule experimental data from 4 shots N110608-N110908 with two Hydra implosions, one spherical ($a_4=0$ m, and another with $a_4 = 20$ m. Large positive $P_4$ brings the modeled implosion observables approximately in line with the experimental data.[]{data-label="table1"}
As discussed, implosions with a significant $P_4$ asymmetry can have a very apparent but “false” $P_2$ asymmetry in GXD images. We find that attempting to correct this “false” $P_2$ by increasing laser power to the hohlraum waist (capsule equator) [@Glenzer05032010] can lead to a round GXD image even though the correction actually produces a more distorted DT fuel ice layer. This is depicted in fig. \[fig5\] for the case of a DT layered capsule where we applied and empirically adjusted a $P_2$ flux asymmetry, in addition to the original $P_4$, in order to make the synthetic GXD image appear round. Fig. \[fig5\](c) quantifies a related effect; as the applied $P_2$ flux is increased in order to reduce the “false” GXD $a_2$ towards zero, there is a marked additional reduction in sensitivity to $a_4$ (relative to that shown in Fig. \[fig3\]). This suggests that attempts to tune the hohlraum to eliminate a “false” $P_2$ can have the unintended consequence of exacerbating overall asymmetry. Other information, such as comparison of the widths of images taken from both the polar and equatorial lines of sight [@benedetti:2012] need to be taken into consideration. These simulations show that when a hotspot has a positive but pure $P_4$ asymmetry the equatorial image width is larger than the width in the polar image (for negative $P_4$ this is reversed). This could be used to identify an implosion where the measured $P_2$ may be caused by a dominant $P_4$ asymmetry. However, our simulations also show us that the empirically but incorrectly tuned implosion of fig. \[fig5\] would have a polar image width that is equal to the equatorial image width, further misleading us into thinking that we had engineered an approximately spherical implosion. A corollary of figure \[fig5\], is that it is possible to create imploded configurations which appear to be symmetric in the GXD but, in fact, are significantly asymmetric and have greatly reduced performance in comparison to equivalent spherical implosions because a large fraction of the imploding shell’s kinetic energy remains unstagnated.
In summary, numerical simulations have been used to examine the sensitivity of implosions similar to those currently taking place on NIF to low-mode flux asymmetries. It is shown that Legendre polynomial $P_4$ flux modes induce $P_4$ shape modes at the time of capsule stagnation. The largest $P_4$ amplitudes can cause up to 50% of the capsule kinetic energy to remain unconverted to hot spot and DT ice internal energy, in turn reducing the neutron yield by up to $15\times$. Simulated x-ray images of the hot spot self-emission show reduced sensitivity to the positive $P_4$ mode, instead the images appear to have a pronounced oblate $P_2$ shape. Attempting to correct for this apparent $P_2$ distortion can further distort the implosion while creating x-ray images which appear round & self-consistent from both equatorial and polar directions. This also further reduces the sensitivity to the $P_4$ mode such that that no quantitative evaluation of the hot spot $a_4$ can be made. Long wavelength asymmetries may be playing a significant role in the observed yield reduction of NIF DT implosions relative to detailed post-shot 2D simulations.
[10]{}
J. D. Lindl *et al.*, Physics of Plasmas **11**, 339 (2004).
N. Basov, Edward teller lectures (1991).
J. NUCKOLLS *et al.*, Nature **239**, 139 (1972).
S. W. Haan *et al.*, Physics of Plasmas **18**, 051001 (2011).
D. S. Clark *et al.*, Phys. Plamas **17**, 052703 (2010).
D. A. Callahan *et al.*, Journal of Physics: Conference Series **112**, 022021 (2008).
M. Abramowitz and I. A. Stegun, *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables*, volume pp. 331-339 and 771-802, Dover (1972).
V. A. Thomas and R. J. Kares, Phys. Rev. Lett. **109**, 075004 (2012).
M. M. Marinak *et al.*, Physics of Plasmas **8**, 2275 (2001).
L. M. Barker and R. E. Hollenbach, Journal of Applied Physics **43**, 4669 (1972).
D. G. Hicks *et al.*, Physics of Plasmas **17**, 102703 (2010).
R. M. More *et al.*, Physics of Fluids **31**, 3059 (1988).
O. S. Jones, Private Communication, (2012).
L. Rayleigh, *Scientific Papers*, Cambridge University Press, Cambridge, England (1900).
G. I. Taylor, *The Instability of Liquid Surfaces when Accelerated in a Direction Perpendicular to their Planes. I*, Proc. R. Soc. London, Ser. A (1950).
S. H. Glenzer *et al.*, Science **327**, 1228 (2010).
D. Clark *et al.*, Physics of Plasmas *(Submitted)* (2012).
P. Spinger and C. Cerjan, in [*IFSA Proceedings*]{} Bordeaux, France (2011).
B.K. Spears, Private Communication, (2012).
R. Tommasini *et al.*, Physics of Plasmas **18**, 056309 (2011).
L.R. Benedetti, [*APS DPP Proceedings*]{}, Providence, RI, United States (2012).
[^1]: Also visiting scientist at Lawrence Livermore National Laboratory, Livermore, CA 94551, United States of America, & Department of Physics, The Blackett Laboratory, Imperial College London, Prince Consort Road, London, SW7 2AZ, United Kingdom. The authors thank M.H. Key and C. Cerjan for useful discussions, the staff of NIF and Livermore Computing.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We have used the CLEO II detector to study the multiplicity of charged particles in the decays of $B$ mesons produced at the $\Upsilon(4S)$ resonance. Using a sample of $1.5{\times}10^6$ $B$ meson pairs, we find the mean inclusive charged particle multiplicity to be $10.71\pm{0.02}_{-0.15}^{+0.21}$ for the decay of the pair. This corresponds to a mean multiplicity of $5.36\pm{0.01}_{-0.08}^{+0.11}$ for a single $B$ meson. Using the same data sample, we have also extracted the mean multiplicities in semileptonic and nonleptonic decays. We measure a mean of $7.82\pm{0.05}_{-0.19}^{+0.21}$ charged particles per $B\bar{B}$ decay when both mesons decay semileptonically. When neither $B$ meson decays semileptonically, we measure a mean charged particle multiplicity of $11.62\pm{0.04}_{-0.18}^{+0.24}$ per $B\bar{B}$ pair.'
author:
- CLEO Collaboration
title: Charged Track Multiplicity in $B$ Meson Decay
---
6.5 in 9.0 in -0.50in 0.00in 0.00in
G. Brandenburg,$^{1}$ A. Ershov,$^{1}$ Y. S. Gao,$^{1}$ D. Y.-J. Kim,$^{1}$ R. Wilson,$^{1}$ T. E. Browder,$^{2}$ Y. Li,$^{2}$ J. L. Rodriguez,$^{2}$ H. Yamamoto,$^{2}$ T. Bergfeld,$^{3}$ B. I. Eisenstein,$^{3}$ J. Ernst,$^{3}$ G. E. Gladding,$^{3}$ G. D. Gollin,$^{3}$ R. M. Hans,$^{3}$ E. Johnson,$^{3}$ I. Karliner,$^{3}$ M. A. Marsh,$^{3}$ M. Palmer,$^{3}$ C. Plager,$^{3}$ C. Sedlack,$^{3}$ M. Selen,$^{3}$ J. J. Thaler,$^{3}$ J. Williams,$^{3}$ K. W. Edwards,$^{4}$ R. Janicek,$^{5}$ P. M. Patel,$^{5}$ A. J. Sadoff,$^{6}$ R. Ammar,$^{7}$ P. Baringer,$^{7}$ A. Bean,$^{7}$ D. Besson,$^{7}$ R. Davis,$^{7}$ S. Kotov,$^{7}$ I. Kravchenko,$^{7}$ N. Kwak,$^{7}$ X. Zhao,$^{7}$ S. Anderson,$^{8}$ V. V. Frolov,$^{8}$ Y. Kubota,$^{8}$ S. J. Lee,$^{8}$ R. Mahapatra,$^{8}$ J. J. O’Neill,$^{8}$ R. Poling,$^{8}$ T. Riehle,$^{8}$ A. Smith,$^{8}$ S. Ahmed,$^{9}$ M. S. Alam,$^{9}$ S. B. Athar,$^{9}$ L. Jian,$^{9}$ L. Ling,$^{9}$ A. H. Mahmood,$^{9,}$[^1] M. Saleem,$^{9}$ S. Timm,$^{9}$ F. Wappler,$^{9}$ A. Anastassov,$^{10}$ J. E. Duboscq,$^{10}$ K. K. Gan,$^{10}$ C. Gwon,$^{10}$ T. Hart,$^{10}$ K. Honscheid,$^{10}$ H. Kagan,$^{10}$ R. Kass,$^{10}$ J. Lorenc,$^{10}$ H. Schwarthoff,$^{10}$ M. B. Spencer,$^{10}$ E. von Toerne,$^{10}$ M. M. Zoeller,$^{10}$ S. J. Richichi,$^{11}$ H. Severini,$^{11}$ P. Skubic,$^{11}$ A. Undrus,$^{11}$ M. Bishai,$^{12}$ S. Chen,$^{12}$ J. Fast,$^{12}$ J. W. Hinson,$^{12}$ J. Lee,$^{12}$ N. Menon,$^{12}$ D. H. Miller,$^{12}$ E. I. Shibata,$^{12}$ I. P. J. Shipsey,$^{12}$ Y. Kwon,$^{13,}$[^2] A.L. Lyon,$^{13}$ E. H. Thorndike,$^{13}$ C. P. Jessop,$^{14}$ K. Lingel,$^{14}$ H. Marsiske,$^{14}$ M. L. Perl,$^{14}$ V. Savinov,$^{14}$ D. Ugolini,$^{14}$ X. Zhou,$^{14}$ T. E. Coan,$^{15}$ V. Fadeyev,$^{15}$ I. Korolkov,$^{15}$ Y. Maravin,$^{15}$ I. Narsky,$^{15}$ R. Stroynowski,$^{15}$ J. Ye,$^{15}$ T. Wlodek,$^{15}$ M. Artuso,$^{16}$ R. Ayad,$^{16}$ E. Dambasuren,$^{16}$ S. Kopp,$^{16}$ G. Majumder,$^{16}$ G. C. Moneti,$^{16}$ R. Mountain,$^{16}$ S. Schuh,$^{16}$ T. Skwarnicki,$^{16}$ S. Stone,$^{16}$ A. Titov,$^{16}$ G. Viehhauser,$^{16}$ J.C. Wang,$^{16}$ A. Wolf,$^{16}$ J. Wu,$^{16}$ S. E. Csorna,$^{17}$ K. W. McLean,$^{17}$ S. Marka,$^{17}$ Z. Xu,$^{17}$ R. Godang,$^{18}$ K. Kinoshita,$^{18,}$[^3] I. C. Lai,$^{18}$ P. Pomianowski,$^{18}$ S. Schrenk,$^{18}$ G. Bonvicini,$^{19}$ D. Cinabro,$^{19}$ R. Greene,$^{19}$ L. P. Perera,$^{19}$ G. J. Zhou,$^{19}$ S. Chan,$^{20}$ G. Eigen,$^{20}$ E. Lipeles,$^{20}$ M. Schmidtler,$^{20}$ A. Shapiro,$^{20}$ W. M. Sun,$^{20}$ J. Urheim,$^{20}$ A. J. Weinstein,$^{20}$ F. Würthwein,$^{20}$ D. E. Jaffe,$^{21}$ G. Masek,$^{21}$ H. P. Paar,$^{21}$ E. M. Potter,$^{21}$ S. Prell,$^{21}$ V. Sharma,$^{21}$ D. M. Asner,$^{22}$ A. Eppich,$^{22}$ J. Gronberg,$^{22}$ T. S. Hill,$^{22}$ D. J. Lange,$^{22}$ R. J. Morrison,$^{22}$ T. K. Nelson,$^{22}$ J. D. Richman,$^{22}$ D. Roberts,$^{22}$ R. A. Briere,$^{23}$ B. H. Behrens,$^{24}$ W. T. Ford,$^{24}$ A. Gritsan,$^{24}$ H. Krieg,$^{24}$ J. Roy,$^{24}$ J. G. Smith,$^{24}$ J. P. Alexander,$^{25}$ R. Baker,$^{25}$ C. Bebek,$^{25}$ B. E. Berger,$^{25}$ K. Berkelman,$^{25}$ F. Blanc,$^{25}$ V. Boisvert,$^{25}$ D. G. Cassel,$^{25}$ M. Dickson,$^{25}$ S. von Dombrowski,$^{25}$ P. S. Drell,$^{25}$ K. M. Ecklund,$^{25}$ R. Ehrlich,$^{25}$ A. D. Foland,$^{25}$ P. Gaidarev,$^{25}$ R. S. Galik,$^{25}$ L. Gibbons,$^{25}$ B. Gittelman,$^{25}$ S. W. Gray,$^{25}$ D. L. Hartill,$^{25}$ B. K. Heltsley,$^{25}$ P. I. Hopman,$^{25}$ C. D. Jones,$^{25}$ D. L. Kreinick,$^{25}$ T. Lee,$^{25}$ Y. Liu,$^{25}$ T. O. Meyer,$^{25}$ N. B. Mistry,$^{25}$ C. R. Ng,$^{25}$ E. Nordberg,$^{25}$ J. R. Patterson,$^{25}$ D. Peterson,$^{25}$ D. Riley,$^{25}$ J. G. Thayer,$^{25}$ P. G. Thies,$^{25}$ B. Valant-Spaight,$^{25}$ A. Warburton,$^{25}$ P. Avery,$^{26}$ M. Lohner,$^{26}$ C. Prescott,$^{26}$ A. I. Rubiera,$^{26}$ J. Yelton,$^{26}$ and J. Zheng$^{26}$
$^{1}$[Harvard University, Cambridge, Massachusetts 02138]{}\
$^{2}$[University of Hawaii at Manoa, Honolulu, Hawaii 96822]{}\
$^{3}$[University of Illinois, Urbana-Champaign, Illinois 61801]{}\
$^{4}$[Carleton University, Ottawa, Ontario, Canada K1S 5B6\
and the Institute of Particle Physics, Canada]{}\
$^{5}$[McGill University, Montréal, Québec, Canada H3A 2T8\
and the Institute of Particle Physics, Canada]{}\
$^{6}$[Ithaca College, Ithaca, New York 14850]{}\
$^{7}$[University of Kansas, Lawrence, Kansas 66045]{}\
$^{8}$[University of Minnesota, Minneapolis, Minnesota 55455]{}\
$^{9}$[State University of New York at Albany, Albany, New York 12222]{}\
$^{10}$[Ohio State University, Columbus, Ohio 43210]{}\
$^{11}$[University of Oklahoma, Norman, Oklahoma 73019]{}\
$^{12}$[Purdue University, West Lafayette, Indiana 47907]{}\
$^{13}$[University of Rochester, Rochester, New York 14627]{}\
$^{14}$[Stanford Linear Accelerator Center, Stanford University, Stanford, California 94309]{}\
$^{15}$[Southern Methodist University, Dallas, Texas 75275]{}\
$^{16}$[Syracuse University, Syracuse, New York 13244]{}\
$^{17}$[Vanderbilt University, Nashville, Tennessee 37235]{}\
$^{18}$[Virginia Polytechnic Institute and State University, Blacksburg, Virginia 24061]{}\
$^{19}$[Wayne State University, Detroit, Michigan 48202]{}\
$^{20}$[California Institute of Technology, Pasadena, California 91125]{}\
$^{21}$[University of California, San Diego, La Jolla, California 92093]{}\
$^{22}$[University of California, Santa Barbara, California 93106]{}\
$^{23}$[Carnegie Mellon University, Pittsburgh, Pennsylvania 15213]{}\
$^{24}$[University of Colorado, Boulder, Colorado 80309-0390]{}\
$^{25}$[Cornell University, Ithaca, New York 14853]{}\
$^{26}$[University of Florida, Gainesville, Florida 32611]{}
INTRODUCTION
============
Measurements of the charged track mutliplicity distribution in $B$ meson decay are used to constrain unmeasured or poorly measured branching fractions in Monte Carlo simulations so that generated event samples more closely represent actual data. The CLEO Monte Carlo parameterization of $B$ meson decays has been tuned to agree with our measurements and our model is used by other experimental groups [@CDF]. Charged particle multiplicity in heavy meson decay has been studied by several groups [@CLEOI][@MARKIII][@ACCMOR]. In this paper we present a measurement of the charged particle multiplicity in inclusive $B\bar{B}$ decays that is an improvement over our previous result [@CLEOI]. We also present improved measurements of the charged particle multiplicities in semileptonic and nonleptonic decays.
For clarity, we use the term “observed multiplicity” to denote the number of well reconstructed charged particle tracks in a given event. We use the term “decay multiplicity” to denote the number of $e^{\pm}$, $\mu^{\pm}$, $\pi^{\pm}$, $K^{\pm}$ and $p^{\pm}$ that come from the decay of the primary $B$ mesons and also from the subsequent decays of any secondary or tertiary particles other than neutrons, $K_L$, or $\pi^0$. The decay multiplicity excludes any tracks produced through interactions with the detector or surrounding material. Not all charged decay products will result in reconstructed tracks, and not all observed tracks come from the primary decay, so the observed multiplicity may be less than, equal to, or greater than the decay multiplicity for a given event.
INCLUSIVE MULTIPLICITY MEASUREMENT
==================================
The CLEO detector is located at the Cornell Electron Storage Ring, a high luminosity $e^+e^-$ collider operated at or near the $\Upsilon(4S)$ resonance. The results presented here are derived from a sample of 1.4 fb$^{-1}$, corresponding to $1.5{\times}10^6$ $B$ meson pairs, collected with the CLEO II detector [@CD]. Charged particle tracks are measured by cylindrical wire drift chambers inside a 1.5 T superconducting solenoid. A CsI crystal calorimeter is also inside the magnet, and energy deposition information from both the calorimeter and the drift chamber is used for particle identification. Muon counters are layered in the steel yoke surrounding the coil.
To obtain a clean sample of candidate $B\bar{B}$ events, we select hadronic events by requiring that an event have three or more reconstructed tracks, energy deposition in the calorimeter greater than 15% of the center of mass energy and an event vertex consistent with the interaction region. For additional background suppression, the total reconstructed event energy, including charged and neutral particles, is required to be between 4 GeV and 12 GeV, and the total reconstructed vector momentum of the event is required to have a magnitude less than 3 GeV/$c$. This hadronic event sample contains events from both $B\bar{B}$ and continuum processes such as $q\bar{q}$ and $\tau^+\tau^-$ production. We remove the continuum contribution by rescaling and subtracting the observed multiplicity distribution of a separate 0.7 fb$^{-1}$ data sample collected 65 MeV/$c^2$ below the $\Upsilon(4S)$ resonance.
To be counted in our observed multiplicity, drift chamber tracks are required to be well reconstructed and consistent with having originated from the event vertex. Tracks must not be within 25.8 degrees of the $e^+e^-$ beam axis. Once the event selection, continuum subtraction, and track selection are completed, we count the selected tracks in each event to obtain the [*observed*]{} charged track multiplicity distribution in Fig. \[fig1\]. There are events with fewer than three selected tracks because not all reconstructed tracks pass the track selection criteria.
To obtain the true decay multiplicity distribution from the observed multiplicity distribution, we must account for detector effects. If the number of events with observed multiplicity $j$ is $O_j$ and the number of events with decay multiplicity $i$ is $D_i$, these quantities are related by Eqn. \[E:migration\].
$$O_j = \sum_{i=2,4,...}^{n}\epsilon_{ij} D_i
\label{E:migration}$$
where $\epsilon_{ij}$ is the probability that an event with decay multiplicity $i$ will be reconstructed with observed multiplicity $j$. We have assumed that charge is conserved, so the index $i$ can only take even values. In principle, there can be events with decay multiplicity of zero, where two neutral $B$ mesons decay to all neutral final states. However, we do not include zero decay multiplicity events in our analysis both because of the very low branching ratio for such events and also because our event selection criteria make detection of such events extremely unlikely. The upper bound in Eqn. \[E:migration\] is, in principle, the maximum decay multiplicity in a $B\bar{B}$ event, which is not known. We vary the maximum decay multiplicity in our analysis as described below. The fact that the sum of $D_i$ and the sum of $O_j$ are both equal to the total number of events is used to constrain the values of the $D_i$, as expressed in Eqn. \[E:constraint\].
$$\sum_{i=2,4,...}^{n} D_i = \sum_{j=0}^{m} O_j
\label{E:constraint}$$
where the upper bound $m$ is the maximum value of our observed multipicity which is 20.
The coefficients $\epsilon_{ij}$ in Eqn. \[E:migration\] are obtained from Monte Carlo simulation and depend primarily on the detector’s track finding efficiency and also on the probability of producing extra charged particles that pass the track selection cuts and that are counted. While these coefficients depend on accurate simulation of detector response and processes such as photon conversion and decays in flight, they do not depend significantly on the exact tuning of the branching fractions or the decay multiplicity distribution in the simulation.
The parameters $D_i$ in Eqn. \[E:migration\] are determined by a $\chi^2$ fit. This fit unfolds the detector effects to give the decay multiplicity distribution of events that pass the event selection criteria. These selection criteria are biased against very low decay multiplicity events, particularly because of the requirement of three or more reconstructed tracks. We remove the event selection bias using Monte Carlo simulation to determine the probability for events of a given decay multiplicity to pass the event selection cuts. After unfolding detector and reconstruction effects, we obtain the decay multiplicity distribution in Fig. \[fig2\]. The error bars represent the statistical uncertainty in both the $\chi^2$ fit and the event bias correction, but do not include systematic errors. The dashed lines represent the high multiplicity and low multiplicity statistical fluctuations in the fit. The large error bar on the $i~=~2$ point is due to the event bias correction.
From the distribution in Fig. \[fig2\], we obtain a mean of $10.71\pm{0.02}$ charged particles for inclusive $B\bar{B}$ decay, where the error is statistical only. The points with $i~>~20$ show no evidence for events with such high multiplicities. We have tested our fitting procedure by varying the maximum decay multiplicity included in the fit between 20 and 28. The fit is stable and the unfolded mean decay multiplicity does not change significantly when decay multiplicities of 22 and higher are included or excluded from the fit.
The most important systematic effect in this analysis is the accuracy of modeling the detector’s track finding efficiency. Our studies indicate that the overall efficiency for finding a single track is known to within $\pm{1\%}$ for tracks with momentum greater than $250$ MeV/$c$ and with decreasing accuracy as track momentum decreases. The uncertainty in single track finding efficiency gives a $_{-0.9}^{+1.6}\%$ uncertainty in the measured mean decay multiplicity. Removal of the event selection bias shifts the measured mean decay multiplicity by $+0.7\%$, which we also take as the systematic uncertainty for event selection bias. Because the track finding efficiency depends on the momentum of the track, we account for the uncertainty due to this dependence. This analysis uses all tracks without regard to momentum. When we add a track selection cut requiring a reconstructed momentum of at least $150$ MeV/$c$, we observe a shift in the decay multiplicity of $-0.6\%$. We assign an additional $\pm{0.6\%}$ uncertainty due to this momentum dependence. Charged pions produced by the decays of $K_S$ will have lower track finding efficiency than average charged pions, so our result depends on the rate of $K_S$ production in our simulation. Based on our studies of inclusive $K_S$ production in $B$ meson decay, we assign a systematic uncertainty of $^{+0.5}_{-0.3}\%$ from this source. Interactions of neutral and charged decay products with the detector material produce additional charged tracks, some of which satisfy the track selection criteria. We study the effect of these extra tracks by varying the rates of photon conversion and hadronic interactions in our Monte Carlo sample. Misidentification of these extra tracks contributes an additional $\pm{0.3\%}$ systematic uncertainty. Extra tracks also come from the decay of particles in the detector volume, adding another $\pm{0.1\%}$ uncertainty. Additional uncertainty comes from contamination by non-$B\bar{B}$ events, most notably from beam-gas interactions. We study the effect of non-$B\bar{B}$ events by varying the size of the continuum subtraction by $1\%$, which is the uncertainty in our measured luminosity, and we find a $\pm{0.2\%}$ uncertainty in our mean decay multiplicity.
We have looked for other potential systematic effects by varying our track selection cuts but do not observe any significant change in mean decay multiplicity. When all systematic errors are added in quadrature, we obtain a total systematic uncertainty of $_{-1.4}^{+1.9}\%$, which gives a final result of $10.71\pm{0.02}_{-0.15}^{+0.21}$ for the mean inclusive charged particle decay multiplicity in the decay of a $B\bar{B}$ pair. Systematic errors are summarized in Table \[tab1\].
The results presented here are a significant improvement over the previous CLEO result [@CLEOI]. The earlier analysis found a mean multiplicity of $11.5\pm{0.2}\pm{0.4}$. In addition to using a much larger sample of $B\bar{B}$ events, the present analysis uses a substantially different method. The former analysis treated the coefficients $\epsilon_{ij}$ in Eqn. \[E:migration\] as a matrix and used matrix inversion to solve for the $D_i$. Such matrix inversion is very sensitive to singularities and can produce unstable results. In the current analysis, we do not attempt to invert the $\epsilon_{ij}$ matrix, but instead we perform a more stable $\chi^2$ fit for the parameters $D_i$.
--------------------------------------------- -----------------------------------
Systematic Error Source Effect on Mean Decay Multiplicity
(Charged Tracks)
Overall Track Finding Efficiency $_{-0.09}^{+0.17}$
Event Selection Bias $\pm{0.07}$
Low Momentum Tracking Efficiency $\pm{0.06}$
$K_S$ Modeling $_{-0.03}^{+0.05}$
Hadronic Interactions and Photon Conversion $\pm{0.03}$
Contamination by Non $B\bar{B}$ Events $\pm{0.02}$
Decays in Flight $\pm{0.01}$
All Other Sources $\pm{0.01}$
Total Systematic Error $_{-0.15}^{+0.21}$
--------------------------------------------- -----------------------------------
: Systematic Errors for the Inclusive Decay Multiplicity Measurement for $B\bar{B}$ Pairs.[]{data-label="tab1"}
SEMILEPTONIC AND NONLEPTONIC MULTIPLICITY MEASUREMENT
=====================================================
We have further analyzed the same data sample to measure multiplicities separately for semileptonic and nonleptonic decays. We define semileptonic to include only decays of $B$ mesons into an electron or muon plus a neutrino and any number of hadrons. All other decays are classified as nonleptonic. Decays involving tau leptons are counted as nonleptonic because approximately 65% of taus decay hadronically [@TAUBR] and such events cannot be distinguished easily from purely hadronic $B$ decays.
We use the same event and track selection criteria as in the inclusive analysis and sort the events by the number of leptons identified. We identify electrons by combining information on energy deposition in the drift chamber, the shape of the shower observed in the calorimeter, and the ratio of the calorimeter energy to the track momentum. Muons are required to traverse at least three pion nuclear interaction lengths of material. We also require all lepton candidates to have momentum between $1.4$ GeV/$c$ and $2.5$ GeV/$c$ to suppress false lepton identification and secondary leptons.
Once we have sorted the events by the number of leptons found, we correct for misidentified and secondary leptons. We move these misidentified events from the sample with one identified lepton to the sample of events with no identified lepton by using the observed multiplicity distributions of Monte Carlo generated events that are similarly misidentified in the reconstruction. The fake and secondary observed multiplicity distributions are scaled using our best estimates of the fake and secondary lepton rates and are added to the sample with no identified lepton and subtracted from the sample with one identified lepton. Uncertainties in the rates and observed multiplicity distributions of these events are treated as systematic errors.
We use the same methods as in the inclusive analysis to extract the mean decay multiplicity from the corrected observed multiplicity distributions for the samples with zero or one identified lepton. For events with no reconstructed leptons, we obtain a mean decay multiplicity of $11.02\pm{0.01}$, and for the sample with one reconstructed lepton, we find a mean of $9.28\pm{0.02}$ charged tracks, where the errors are statistical only. Not all semileptonic decays will produce a detected lepton because not all electrons and muons will enter the fiducial tracking volume of the detector and pass the track selection and lepton identification criteria. To obtain the true decay multiplicities for semileptonic and nonleptonic $B\bar{B}$ decays, we use simulation to unfold the migration between the number of leptons generated by the decay of the $B$ mesons and the number of leptons actually identified. If the migration probability for a given event were independent of the event’s decay multiplicity, we could write:
$$M^i = a^i M_{n-n} + b^i M_{s-s},
\label{E:lepmigration}$$
where the $M^{i}$ are the mean decay multiplicities of the sample with $i$ identified leptons, $M_{n-n}$ and $M_{s-s}$ are the mean decay multiplicities of events where both $B$ mesons decay nonleptonically and semileptonically, respectively, and $a^i$ and $b^i$ are migration coefficients determined from Monte Carlo simulation. These multiplicities include the lepton tracks. There are many factors that affect the number of leptons that will be reconstructed for a given event. Lepton identification depends strongly on particle momentum not only because of the explicit momentum requirements but also because a particle’s momentum determines the probability that it will penetrate the muon chambers sufficiently to be classified as a muon candidate. Because of phase space limitations, events with one or more high momentum (above $1.4$ GeV/$c$) particles will tend to have fewer charged tracks than events without a high momentum particle, so events with an identified lepton tend to have lower multiplicities than those where the lepton was not identified.
In addition to assuming that migration probability is independent of decay multiplicity, Eqn. \[E:lepmigration\] also assumes that the mean decay multiplicity of events where only one $B$ meson decays semileptonically is simply the average of $M_{n-n}$ and $M_{s-s}$. This assumption would be valid if the mean true decay multiplicities of $B^-$ and $B^0$ mesons were equal, but this may not be the case. To account for the possibility of unequal $B^-$ and $B^0$ multiplicities and the effect of multiplicity dependent migration, we can introduce another term into Eqn. \[E:lepmigration\]:
$$M^i + \Delta^i= a^i M_{n-n} + b^i M_{s-s}.
\label{E:fixedmigration}$$
The correction terms $\Delta^i$ and migration coefficients are determined from Monte Carlo simulation, and we can solve Eqn. \[E:fixedmigration\] for $M_{n-n}$ and $M_{s-s}$. We obtain mean true decay multiplicities of $11.62\pm{0.04}$ charged particles for events where neither $B$ meson decays semileptonically and $7.82\pm{0.05}$ charged particles for events where both $B$ mesons decay semileptonically, where the errors are statistical and include analytic propagation of the uncertainty in each of the parameters in Eqn. \[E:fixedmigration\].
All of the systematic uncertainties that affect the inclusive analysis are also present in these semileptonic and nonleptonic results. This $^{+1.9}_{-1.4}\%$ uncertainty in unfolding the decay multiplicity from the observed multiplicity is the largest systematic effect for these results. The next largest uncertainty comes from lepton identification, including inefficiencies, false particle identification, and acceptance of secondary leptons not directly produced in the $B$ meson decay. We have studied these effects by varying all of our lepton identification criteria. We have also varied the rates of fake and secondary lepton identification directly in simulation. All of the lepton identification systematics combine to give a $\pm 0.06$ charged track uncertainty in the nonleptonic result and a corresponding uncertainty of $\pm 0.11$ charged tracks in the mean decay multiplicity of semileptonic decays. Our result is also sensitive to the modeling of semileptonic decays in our simulation, including the inclusive rate of semileptonic decays and the exclusive branching fractions of semileptonic decays that produce the various resonances of the $D$ meson. We have varied both the inclusive and exclusive branching fractions within their uncertainties and find $\pm 0.05$ uncertainty in the nonleptonic mean decay multiplicity and $\pm 0.08$ uncertainty for the semileptonic result. Uncertainty in the observed multiplicity of events with fake and secondary leptons contributes $\pm 0.06$ charged tracks uncertainty to the semileptonic result. The systematic uncertainties are summarized in Table \[Ta:lsystematic\].
Combining all errors, we find mean multiplicities of $11.62\pm{0.04}^{+0.24}_{-0.18}$ charged tracks when both $B$ mesons decay nonleptonically and $7.82\pm{0.05}^{+0.21}_{-0.19}$ charged tracks when both $B$ mesons decay semileptonically. As a further check on our systematic errors, we have repeated this analysis by using only electron identification or only muon identification. In both cases, the results are statistically consistent with the means stated above and with each other. Our present results are a significant improvement over the previous CLEO results [@CLEOI], in which mean multiplicities of $12.6\pm{0.4}\pm{0.4}$ and $8.2\pm{0.7}\pm{0.4}$ were obtained for events where both $B$ mesons decay nonleptonically and semileptonically, respectively.
---------------------------- -------------------- ---------------------
Systematic Error Source Nonleptonic Events Semileptonic Events
(Charged Tracks) (Charged Tracks)
Inclusive Unfolding $^{+0.23}_{-0.16}$ $^{+0.15}_{-0.11}$
Lepton Identification $\pm 0.06$ $\pm 0.11$
Branching Fractions $\pm 0.05$ $\pm 0.08$
Fake and Secondary Leptons $\pm 0.01$ $\pm 0.06$
All Other Sources $<\pm 0.01$ $\pm 0.01$
Total Uncertainty $^{+0.24}_{-0.18}$ $^{+0.21}_{-0.19}$
---------------------------- -------------------- ---------------------
: Systematic Error for the Charged Track Mutliplicity in Nonleptonic and Semileptonic Events. []{data-label="Ta:lsystematic"}
ACKNOWLEDGEMENTS
================
We gratefully acknowledge the effort of the CESR staff in providing us with excellent luminosity and running conditions. J.R. Patterson and I.P.J. Shipsey thank the NYI program of the NSF, M. Selen thanks the PFF program of the NSF, M. Selen and H. Yamamoto thank the OJI program of DOE, J.R. Patterson, K. Honscheid, M. Selen and V. Sharma thank the A.P. Sloan Foundation, M. Selen and V. Sharma thank the Research Corporation, F. Blanc and S. von Dombrowski thank the Swiss National Science Foundation, and H. Schwarthoff and E. von Toerne thank the Alexander von Humboldt Stiftung for support. This work was supported by the National Science Foundation, the U.S. Department of Energy, and the Natural Sciences and Engineering Research Council of Canada.
[99]{}
See, for example, CDF Collaboration, F. Abe [*et al.*]{}, Phys. Rev. D [**50**]{}, 2966 (1994), and D0 Collaboration, B. Abbot [*et al.*]{}, hep-ex/9905024 (1999).
CLEO Collaboration, M.S. Alam [*et al.*]{}, Phys. Rev. Lett. [**49**]{}, 357 (1982).
MARK III Collaboration, D. Coffman [*et al.*]{}, Phys. Lett. B [**263**]{}, 135 (1991).
ACCMOR Collaboration, S. Barlag [*et al.*]{}, Z. Phys. C [**55**]{}, 383 (1992).
CLEO Collaboration, Y. Kubota [*et al.*]{}, Nucl. Instrum. Methods Phys. Res., Sect. A [**320**]{}, 66 (1992).
C. Caso [*et al.*]{}, European Phys. J. [**C3**]{}, 1 (1998).
[^1]: Permanent address: University of Texas - Pan American, Edinburg TX 78539.
[^2]: Permanent address: Yonsei University, Seoul 120-749, Korea.
[^3]: Permanent address: University of Cincinnati, Cincinnati OH 45221
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this letter, we demonstrate a strong dependence of the electrostatic deformation of doubly-clamped single-walled carbon nanotubes on both the field strength and the tube length, using molecular simulations. Metallic nanotubes are found to be more sensitive to an electric field than semiconducting ones of the same size. For a given electric field, the induced deformation increases with tube length but decreases with tube radius. Furthermore, it is found that nanotubes can be more efficiently bent in a center-oriented transverse electric field.'
author:
- Zhao Wang
- Laetitia Philippe
title: 'Deformation of doubly-clamped single-walled carbon nanotubes in an electrostatic field'
---
Carbon nanotubes (CNTs) can be used in nanoelectromechanical systems (NEMS) for uses in sensing, actuation, vibration, and laboratory-on-a-chip applications [@Anantram2006]. Particularly, doubly-clamped (suspended) structures of CNTs have been reported to be used as key components in a number of nanodevices [@Javey2002; @Javey2003; @Keren2003; @Sapmaz2003; @Sazonova2004; @Jonsson2005]. In most of these recent NEMS, nanotubes are usually suspended between two electrodes with an applied voltage. However, as to be demonstrated in this paper, even without applying electric current and excess charges, doubly-clamped CNTs can still be significantly bent in a static transversal electric field, as a result of electric polarization.
Cantilevered CNTs have long been known to deform in an electric field [@Poncharal1999; @Wei2001a; @Wang2007a]. The mechanism of this deformation relies on the fact that CNTs can be bent by a bending moment induced by interactions between the electric field and the molecular dipoles due to electric polarization. This property has been exploited in the design of nanorelays [@Kinaret2003; @lee-04]. Compared with cantilevered CNTs, doubly-clamped ones can have many advantages in electronic devices [@Dai2002] (e.g. they can be integrated in nanodevices with a well-defined spatial structure and resonance frequency). This letter investigates the deflection of neutral single-walled CNTs (SWCNTs) by an electrostatic field. Two semiclassical theories have been combined for characterizing chemical potential and electrostatic interactions in both metallic and semiconducting CNTs. Considering the general correlation between the conductivity and polarizabilities of CNTs, metallic CNTs are supposed to be more sensitive to electric fields than semiconducting ones are, and hence can be expected to play a more important role in NEMS.
When an electric field is applied to a metallic CNT, an induced dipole can be created at each atomic site with a quantity of free charge, by shifting the electrons and the nuclei. In our calculation, each atom is therefore modeled as an interacting polarizable point with a free electric charge, while the chemical bonds are described by using a many-body potential function. Motionless equilibrium positions of carbon atoms in an electric field are computed by minimizing the total potential energy of systems $U^{tot}$, which is the sum of two terms: $U^{tot}=U^{elec}+U^{p}$, where $U^{p}$ is the interatomic potential due to the C-C chemical bonds in absence of an external field, including the long-range interactions. $U^{elec}$ stands for the electrostatic energy from the interaction between charges, dipoles and external fields. $U^{elec}$ is calculated using a Gaussian-regularized charge-dipole [@Wang2007a] and a dipole-only [@zhaowang-06-01] model for metallic and semiconducting CNTs, respectively. $U^{p}$ is computed using an adaptive interatomic reactive empirical bond order (AIREBO) potential [@Stuart2000], which is an extension of the REBO potential [@Brenner1990b] which includes a Lennard-Jones potential. Compared to first-principle and semi-empirical methods, an important feature of our combined models is their ability to deal with large systems (up to $5000$ atoms). This is particularly important for studying geometric effects on the electrostatic bending of CNTs, since the axial periodic condition can hardly be applied in this issue.
The strength of electric field $E$ used in this work is in the order of V/nm, because the tubes used in our calculation are too short ($18 \backsim 27$ nm) to be bent in a weak electric field. As will be shown, the required field strength decreases with increasing tube length $L$ for a given deformation. Considering that the CNTs used in experiments are usually of $\mu$m-length, field emission effects [@Jo2003] and conductance switching due to strong transversal fields [@Son2005] are both neglected. These field strengths are not large enough to cause a significant change in the chemical bond strength [@Guo2003a]. Moreover, it can expected that an armchair CNT can exhibit more deformation by an electric field than a metallic zigzag tube of the same size, since previous studies concluded that bending deformation would have a negligible effect on the electronic transport properties of armchair CNTs [@Rochefort1999; @Nardelli1999], while the quantum conductance of zigzag CNTs can significantly decrease under large bending deformations [@Maiti2002]. In this work, the tubes are assumed to be suspended between two electrically insulated supports.
The mechanism of electrostatic bending of a SWCNT is depicted in Fig. \[fig:schemadef\] (a). When a CNT (initially electrically neutral) is submitted to a transversal electric field $\bm{E}^{ext}$, the field tends to shift negative and positive charges in opposite directions (in this figure we can see that positive charges move to the top and negative ones move to the bottom of the tube), two induced molecular dipoles ($\bm{p}_{1}$ and $\bm{p}_{2}$) are hence created in the tube. The tube is then bent by two bending moments ($\tau_{1}$ and $\tau_{2}$), which is induced by the interaction between the molecular dipoles and $\bm{E}^{ext}$ ($\tau = \bm{p}^{m} \otimes \bm{E}^{ext}$).
![\[fig:schemadef\] (Color online) Schematics of electrostatic deformation of a SWCNT. (a) Two arrows $\bm{p}_{1}$ and $\bm{p}_{2}$ stand for two induced molecular dipoles. $\bm{E}^{ext}$ is represented by the arrows with dashed lines. $\tau_{1}$ and $\tau_{2}$ stand for the induced bending moments acting on the tube. The color scale of atoms is proportional to the density of induced charges. The vectors stand for the induced atomic dipoles. (b) Two possible deflection directions in bending (upward or downward).](ebbschema.jpg){width="14cm"}
Molecular dynamics (MD) simulations were performed using the AIREBO potential, for generating arbitrary initial configurations of CNTs at room temperature. The energy of these MD-generated configurations is minimized in order to calculate their equilibrium positions in an electric field. MD simulations showed that CNTs slightly oscillate around their initial positions due to thermal fluctuations, which have been observed in experiments [@Treacy1996; @Babic2003]. We found that the orientation of electrostatic deformation of CNTs strongly depends on their initial position at the moment when $\bm{E}^{ext}$ is applied, e.g., as shown in Fig. \[fig:schemadef\] (b), if $\bm{E}^{ext}$ is applied when the tube is at position $1$, its equilibrium state in electric field will be at position $3$; conversely if the tube is initially in position $2$, it will be bent to position $4$. These results also imply that no change of the deformation direction will take place if the field direction is totally reversed, due to the system symmetry. For measuring the amplitude of deflection, we define $u$ as the displacement of the center of the tube middle from its initial position. For a small electric field, the bending of the CNT is elastically reversible, i.e., the tube will come back and oscillate around its initial position once $\bm{E}^{ext}$ is removed [@Sazonova2004; @Wei2001]. In our MD simulations, the average frequency of thermal oscillation (without electric field) of a (5, 5) tube (length $L\approx20.0$ nm) is found to be $60 \pm 10$ GHz for the several first harmonics. This value is comparable to oscillation frequency of doubly-clamped CNTs recently reported in Ref. [@Li2003].
![\[fig:strength\_break\] (Color online) Electrostatic deformation $u$ versus field strengths $E=\left| \bm{E}^{ext} \right|$ for two SWCNTs: a metallic (4, 4) ($L \approx 24.5$ nm) and a semiconducting (5, 3) with almost the same length and radius. The tube (4, 4) begins to break down when $E > 11.0$ V/nm, then failure occurs very soon thereafter ($u_{max} > 5.1$ nm). The inset shows the position of fracture and the distribution of von-Mises stress at the bottom (before the fracture occurs).](ebbstrbreak.jpg){width="12cm"}
We plot in Fig. \[fig:strength\_break\] computed values of $u$ versus external field strength $E$ for both a (4, 4) and a (5, 3) CNTs. It can be seen that the metallic tube is clearly more sensitive to $E^{ext}$ than the semiconducting one due to their different polarizabilities [@Joselevich2002]. When $u \approx 5.5$ nm, it is observed that the (4, 4) tube begins to fracture near the fixed end, in which the maximal tensile strain is localized (inset in Fig. \[fig:strength\_break\]). We can also see that there is no deformation when $E$ remains small, since in such cases the moment of electric force are not large enough to exceed the mechanical resistance barrier of the tube. These results also imply that semiconducting tubes can sustain higher field strengths.
![\[fig:lengthdep\] (a) $u/L$ versus the tube length $L$ for (4, 4) SWCNTs, in 4 different transverse $E^{ext}$ (which is perpendicular to the initial axis of the tubes). (b) $E$ versus $L$ on a logarithmic scale for two given $u$. The symbols present the calculated points and the lines stand for the extrapolation curves.](ebblength.jpg){width="11cm"}
The CNTs used in experiments are usually microns in length. However, their field-induced deformations can hardly be addressed by the calculations using atomic models in a direct way. Hence, it is necessary to address the influence of the tube length $L$. In Fig. \[fig:lengthdep\] (a), it can be seen that the relative displacement $u/L$ increases with $L$, i.e., a weaker electric field is required for a given $u$ in a longer tube. This increased displacement depends on the field strength. An extrapolation was done from the data $E$ versus $L$ for tube lengths up to $1$ $\mu$m, in order to make experimental verification possible (Fig. \[fig:lengthdep\] (b)). It was found that $\ln(E)$ decreases almost linearly with $\ln(L)$ for two given $u$. This is related to the fact that the electric capacitance per unit length of a metallic cylinder is roughly proportional to $\ln(L)$. Thus, we can conclude that, for the $\mu$m-length tubes used in experiments, the required field strength is on the order of V$/\mu$m.
![\[fig:radiuseb\] $u$ versus $E$ for 6 individual armchair SWCNTs ($L \approx 24.5$ nm) with different radii. The direction of $E^{ext}$ is perpendicular to the initial axis of tubes. Inset shows the force balance between an electric moment and internal stresses in a half of a deformed tube.](ebbradius.jpg){width="12cm"}
To explore further geometry effects, we plot $u$ versus $E$ for the tubes of different radii $R$ in Fig. \[fig:radiuseb\]. It is found that small tubes are more sensitive to $E^{ext}$ than large ones. This looks abnormal since it is known that the polarizability of a CNT increases with its radius. However, from a mechanical point of view, the tube becomes more difficult to be bent due to the increase of its area moment of inertia ($I=0.5m(R^{2}_{int }+R^{2}_{out})$ for a thick-walled cylindrical tube, where $m$ is the mass). From our results it is clear that the later effect plays a more important role. The ratio $u/E$ is roughly the same for all of these tubes in case of large deformation ($u>1$ nm).
In Figs. \[fig:strength\_break\] and \[fig:radiuseb\], we can see a threshold field of a few V/nm for each $u$ versus $E$ curve before one obtains a nonzero deformation. The existence of this threshold field relies on the fact that the electric bending moments $\tau$ remains weaker than the tube resistance when $E$ is small. With increasing $E$, one reaches the limit of the threshold field until a value of $u$ allows a moment balance in the nanotube (inset in Fig. \[fig:radiuseb\]). It is found that the threshold field increases with $R$ but decreases with $L$.
![\[fig:directioneb\] (Color online) (a): Schematic (in *y-z* plane) of the deformation of a CNT subjected to an $\nearrow\,\nwarrow$-like electric field $E^{ext}$ (dashed lines). The field angle $\theta$ is defined as the angle between the field lines and the axis $z$. The field vector $\bm{E}=(E_{x},\,\,E_{y},\,\,E_{z}) = (0,\,\,E\sin(\theta),\,\,+E\cos(\theta))$ when $z<0$ and $\bm{E}=(0,\,\,E\sin(\theta),\,\,-E\cos(\theta))$ when $z>0$. (b): $u$ versus $E$ for a (4, 4) SWCNT ($L\approx24.5$ nm) in $E^{ext}$ with different $\theta$. The simulation data are represented by the symbols with eye-guide lines.](ebbdirection.jpg){width="12cm"}
In this study, the applied external field is assumed to be uniform as a common theoretical simplification. However, in experimental situations, the external fields are usually not homogeneous and strongly depend on the experimental set-up. By changing the field direction, we found that CNTs can be more efficiently bent in a center-oriented ($\nearrow\,\nwarrow$-like) electric field (see Fig. \[fig:directioneb\] (a)). Results of the deformation of a SWCNT by this field is plotted in Fig.\[fig:directioneb\] (b). We can see that values of $u$ roughly follow a linear relationship with $E$ when $u>1$ nm. It was also found that, when $\theta=\pi/4$, the tube can be most efficiently bent and its $u$ versus $E$ curve has straightest shape.
In summary, deformation of doubly-clamped SWCNTs in an electrostatic field has been simulated using a charge-dipole polarization model combined with an empirical potential. The interplay between the mechanical resistance and electric polarization of CNTs was investigated for both metallic and semiconducting tubes. Metallic CNTs are found to be more sensitive to an electric field than semiconducting ones. This study reveals that the field-induced deflection increases with tube length while it decreases with tube radius. It was also found that CNTs can be more efficiently bent in a center-oriented electric field. From a theoretical point of view, graphene nanoribbons and other metallic nanowires/tubes should have similar properties, depending on their dielectric constants. This electrostatic deformation of doubly-clamped nanostructures can be expected to open a path to designs of novel nanodevices.
We gratefully thank S. J. Stuart for his help to the implementation of our computational code. M. William, A. Mayer, R. Langlet, M. Devel and W. Ren are acknowledged for useful discussions.
[27]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , , , ****, ().
, ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , , , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate a known problem whether a Sobolev homeomorphism between domains in ${\mathbb R}^n$ can change sign of the Jacobian. The only case that remains open is when $f\in W^{1,[n/2]}$, $n\geq 4$. We prove that if $n\geq 4$, and a sense-preserving homeomorphism $f$ satisfies $f\in W^{1,[n/2]}$, $f^{-1}\in W^{1,n-[n/2]-1}$ and either $f$ is Hölder continuous on almost all spheres of dimension $[n/2]$, or $f^{-1}$ is Hölder continuous on almost all spheres of dimensions $n-[n/2]-1$, then the Jacobian of $f$ is non-negative, $J_f\geq 0$, almost everywhere. This result is a consequence of a more general result proved in the paper. Here $[x]$ stands for the greatest integer less than or equal to $x$.'
address:
- 'Paweł Goldstein, Institute of Mathematics, Faculty of Mathematics, Informatics and Mechanics, University of Warsaw, Banacha 2, 02-097 Warsaw, Poland'
- 'Piotr Hajłasz, Department of Mathematics, University of Pittsburgh, 301 Thackeray Hall, Pittsburgh, PA 15260, USA'
author:
- Paweł Goldstein
- Piotr Hajłasz
title: 'Jacobians of $W^{1,p}$ homeomorphisms, case $p=[n/2]$'
---
*In memoriam: Bogdan Bojarski (1931-2018)*
Introduction and results
========================
A diffeomorphism $f$ between domains in $\mathbb{R}^n$ has either positive or negative Jacobian $J_f=\det Df$. Recall that domains are open and connected. We say that a diffeomorphism is [*sense preserving*]{} ([*sense reversing*]{}) if its Jacobian is positive (negative). More generally, we say that a homeomorphism between domains is [*sense preserving*]{} ([*sense reversing*]{}) if it has local topological degree $1$ ($-1$) at every point of its domain, see Section \[Sec2.5\]. Every homeomorphism is either sense preserving or sense reversing. It easily follows from the topological properties of the degree that if $f$ is a sense preserving (reversing) homeomorphism of domains in ${\mathbb R}^n$, $f$ is differentiable at $x$ and $J_f(x)\neq 0$, then $J_f(x)>0$ ($J_f(x)<0$). In particular, if a homeomorphism is differentiable almost everywhere, then $J_f\geq 0$ a.e. or $J_f\leq 0$ a.e. However, it may happen that a homeomorphism is differentiable a.e., but its Jacobian equals zero a.e. For elementary constructions, see [@takacs] and references therein.
Let us assume now that a homeomorphism $f$ between domains in ${\mathbb R}^n$ is in the Sobolev space $W^{1,p}_{\rm loc}$, $p\geq 1$. If $p>n-1$, then $f$ is differentiable a.e., [@HK Corollary 2.25], and therefore $J_f\geq 0$ a.e. or $J_f\leq 0$ a.e. Another approach, based on the topological degree, allows one to extend this result to $p\geq n-1$. However, the method completely fails when $1\leq p<n-1$. Note also that, if $1\leq p<n$, then there are very pathological examples of Sobolev homeomorphisms with the Jacobian equal zero a.e. The first such example was constructed by Hencl [@Hencl1], see also [@Cerny; @DHS].
In 2001, Hajłasz asked a question whether a Sobolev $W^{1,p}_{\rm loc}$, $1\leq p<n-1$, homeomorphism between domains in ${\mathbb R}^n$ can change sign of the Jacobian. That is, whether there is a homeomorphism such that $J_f>0$ on a set of positive measure and $J_f<0$ on a set of positive measure.
Hencl and Malý [@henclm1] proved the following two results:
\[T1\] Let $\Omega\subset{\mathbb R}^n$, $n\leq 3$, be a domain and let $f\in W^{1,1}_{\rm loc}(\Omega,{\mathbb R}^n)$ be a sense preserving homeomorphism. Then $J_f\geq 0$ a.e.
\[T2\] Let $\Omega\subset{\mathbb R}^n$, $n\geq 4$, be a domain and let $f\in W^{1,p}_{\rm loc}(\Omega,{\mathbb R}^n)$, $p>[n/2]$ be a sense preserving homeomorphism. Then $J_f\geq 0$ a.e.
Here $[x]$ stands for the greatest integer less than or equal to $x$. Also, by a homeomorphism $f:\Omega\to{\mathbb R}^n$ we mean a homeomorphism onto the image. Since for $p>[n/2]$ we have the embedding into the Lorentz space $L_{\rm loc}^p\subset L^{[n/2],1}_{\rm loc}$, Theorem \[T2\] is a special case of a more general result of Hencl and Malý:
\[T2.5\] Let $\Omega\subset{\mathbb R}^n$, $n\geq 4$, be a domain and let $f:\Omega\to{\mathbb R}^n$ be a sense preserving homeomorphism such that $Df\in L^{[n/2],1}_{\rm loc}$. Then $J_f\geq 0$ a.e.
On the other hand, Hencl and his collaborators, [@CHT; @henclv], constructed the following surprising example:
\[T3\] If $n\geq 4$ and $1\leq p<[n/2]$, then there is a homeomorphism $f\in W^{1,p}((-1,1)^n,{\mathbb R}^n)$ such that $J_f>0$ on a set of positive measure and $J_f<0$ on a set of positive measure. Moreover, $f$ has the Lusin property.
Recall that the Lusin property means that the sets of Lebesgue measure zero are mapped to sets of Lebesgue measure zero.
The result of [@henclv] provides such a homeomorphism for $n\geq 4$ with $p=1$, and the general case is obtained in [@CHT]. See also [@GH1; @GH2] for related examples of approximately differentiable homeomorphisms.
Theorems \[T1\], \[T2\], and \[T3\] leave only the borderline case open.
\[Q1\] Let $\Omega\subset{\mathbb R}^n$, $n\geq 4$, be a domain. Does there exist a homeomorphism ${f\in W^{1,[n/2]}_{\rm loc}(\Omega,{\mathbb R}^n)}$ such that $J_f>0$ on a set of positive measure and $J_f<0$ on a set of positive measure?
The main result of the paper answers this question in the negative under some additional assumptions.
\[main\] Let $\Omega\subset{\mathbb R}^{n}$, $n\ge 4$, be a domain and let $f\in W^{1,[n/2]}_{\rm loc}(\Omega,{\mathbb R}^{n})$ be a sense preserving homeomorphism such that $f^{-1}\in W^{1,n-[n/2]-1}_{\rm loc}(f(\Omega),{\mathbb R}^{n})$. Assume also that one of the two conditions is satisfied:
1. $f$ maps almost all spheres of dimension $[n/2]$ to sets of $\mathcal{H}^{[n/2]+1}$-measure zero,
2. $f^{-1}$ maps almost all spheres of dimension $n-[n/2]-1$ to sets of $\mathcal{H}^{n-[n/2]}$-measure zero.
Then $J_f\ge 0$ a.e.
Here and in what follows, by $\mathcal{H}^k$ we shall denote the $k$-dimensional Hausdorff measure.
The space of $k$-dimensional spheres in ${\mathbb R}^n$ can be parameterized by the product $G(k+1,n)\times{\mathbb R}^n\times (0,\infty)$, where $G(k+1,n)$ is the Grassmannian of $(k+1)$-dimensional subspaces in ${\mathbb R}^n$. Indeed, $(V,x,r)\in G(k+1,n)\times{\mathbb R}^n\times (0,\infty)$ defines a sphere centered at $x$, of radius $r$ and parallel to $V$. Since there is a natural measure on $G(k+1,n)\times{\mathbb R}^n\times (0,\infty)$, it makes sense to talk about almost every $k$-dimensional sphere in ${\mathbb R}^n$.
The classes of bi-Sobolev homeomorphisms, i.e., homeomorphisms such that $f$ and $f^{-1}$ belong to Sobolev spaces, have been investigated for example in [@CSM; @DHS; @HK2; @HKM; @HMPS; @Oliva; @onninen; @Pratelli].
The corollaries listed below show particular situations when the condition (a) or (b) is satisfied.
\[T13\] Let $\Omega\subset{\mathbb R}^{n}$, $n\ge 4$, be a domain and let $f\in W^{1,[n/2]}_{\rm loc}(\Omega,{\mathbb R}^{n})$ be a sense preserving homeomorphism such that $f^{-1}\in W^{1,n-[n/2]-1}_{\rm loc}(f(\Omega),{\mathbb R}^{n})$. Assume also that $f$ or $f^{-1}$ is Hölder continuous. Then $J_f\ge 0$ a.e.
In fact, it suffices to assume $f$ is Hölder continuous on almost all $[n/2]$-dimensional spheres or $f^{-1}$ is Hölder continuous on almost all $n-[n/2]-1$ dimensional spheres; the proof remains the same.
\[T14\] Let $\Omega\subset{\mathbb R}^{2m}$, $m\geq 2$, be a domain in the even dimensional space and let $f\in W^{1,m}_{\rm loc}(\Omega,{\mathbb R}^{2m})$ be a sense preserving homeomorphism such that $f^{-1}\in W^{1,m-1+{\varepsilon}}_{\rm loc}(f(\Omega),{\mathbb R}^{m})$. Then $J_f\ge 0$ a.e.
\[T15\] Let $\Omega\subset{\mathbb R}^{2m}$, $m\geq 2$, be a domain in the even dimensional space and let $f\in W^{1,m}_{\rm loc}(\Omega,{\mathbb R}^{2m})$ be a sense preserving homeomorphism such that $Df^{-1}\in L^{m-1,1}_{\rm loc}$. Then $J_f\ge 0$ a.e.
In Corollaries \[T14\] and \[T15\], we restrict the setting to even dimensions, because a corresponding result in odd dimensions would be a consequence of Theorems \[T2\] and \[T2.5\] respectively.
The corollaries easily follow from Theorem \[main\].
If $f\in W^{1,k}({\mathbb S}^k,{\mathbb R}^m)$, then there is a set $E\subset {\mathbb S}^k$ of measure zero such that the complement of this set is the union of the sets such that on each of these sets $f$ is Lipschitz continuous (see a discussion around below) and hence the Hausdorff dimension of $f({\mathbb S}^k\setminus E)$ is at most $k$. According to a theorem of Malý and Martio [@MalyM Theorem C], [@zap Theorem 1], if $f\in W^{1,k}({\mathbb S}^k,{\mathbb R}^m)$ is Hölder continuous, then it maps sets of measure zero to sets of $\mathcal{H}^k$-measure zero so $\mathcal{H}^k(E)=0$ and hence the Hausdorff dimension of $f({\mathbb S}^k)$ is at most $k$. Let now $f$ be as in Corollary \[T13\]. Assume that $f$ is Hölder continuous. According to the Fubini theorem for Sobolev functions (Lemma \[T5\]), $f$ restricted to almost all spheres $[n/2]$-dimensional spheres is a Hölder continuous map in $W^{1,[n/2]}$ so the image of almost every such sphere has Hausdorff dimension at most $[n/2]$ and hence its $\mathcal{H}^{[n/2]+1}$-measure is zero, so condition (a) from Theorem \[main\] is satisfied and the result follows. Similarly, if $f^{-1}$ is Hölder continuous, the condition (b) is satisfied and the result follows.
Since $L^{m-1+{\varepsilon}}_{\rm loc}\subset L^{m-1,1}_{\rm loc}$, Corollary \[T14\] follows from Corollary \[T15\]. According to [@HKM Theorem C], mappings $f:{\mathbb S}^{k}\to{\mathbb R}^m$ with the weak derivative in $L^{k,1}$ map sets of measure zero to sets of $\mathcal{H}^k$-measure zero, so exactly the same argument as in the proof of Corollary \[T13\] yields that $\mathcal{H}^{k+1}(f({\mathbb S}^k))=0$ and then, again as in the proof of Corollary \[T13\], the result follows.
The main idea in the proofs of Theorems \[T1\] and \[T2\] is to use the linking number. If $n\geq 4$ and $p>[n/2]$, one can find linked spheres in $\Omega$ of dimensions less than $p$. This allows one to use the Sobolev embedding theorem on the linked spheres to control the topological linking number in terms of the Sobolev norm of the mapping. Since a sense preserving homeomorphism maps linked spheres onto linked topological spheres with the same linking number, one can use this fact to prove that the Jacobian of a sense preserving map cannot be negative on a set of positive measure. A similar argument is used when $n\leq 3$.
The proof of our Theorem \[main\] is based on a similar idea. However, we cannot use the Sobolev embedding theorem on spheres, because now $p=k=[n/2]$ equals to the dimension of one of the linked spheres. This causes many technical problems and in order to handle them, we need to assume Sobolev regularity of the inverse map.
Although Theorem \[main\] gives an answer to Question \[Q1\] only in a very special case, the main motivation behind Theorem \[main\] was to modify the technique of the linking number so it could be used in the limiting case, in which we do not have the Sobolev embedding on spheres. We believe that if the answer to Question \[Q1\] is in the negative, the proof should be based on the linking number technique and we hope that, with further modifications, our new technique can lead to the negative answer to Question \[Q1\] in full generality, for all $n\geq 4$. However, we do not know yet how to do it and we are not even sure what the final answer to Question \[Q1\] is.
The paper is structured as follows. In Section \[Prelim\] we collect basic tools that are used in the proof of Theorem \[main\]. Some of the tools collected there are known, but some other are new and of independent interest. In Section \[Proof1and2\] we recall the proof of Theorems \[T1\] and \[T2\]. This helps to understand the main idea of our proof and to see what are the additional difficulties we have to face. In the last Section \[Proofof5\] we prove Theorem \[main\]. We put a lot of effort to make the paper self-contained and accessible to those who are new to this area of research.
Notation in the paper is quite standard. The Lebesgue measure of a set $A\subset{\mathbb R}^n$ is denoted by $|A|$. By $\mathcal{H}^k$ we denote the $k$-dimensional Hausdorff measure. A $k$-dimensional open ball centered at a point $x$, with radius $r$, is denoted by ${\mathbb B}^k(x,r)$, and ${\mathbb B}^k$ denotes the open unit ball in $k$ dimensions. Similarly, ${\mathbb S}^k$ denotes the unit $k$-dimensional sphere. The surface measure on ${\mathbb S}^k$ is denoted by $d\sigma(x)$. Open half-space will be denoted by ${\mathbb R}^{n+1}_+={\mathbb R}^n\times (0,\infty)$. $W^{1,p}$ is the Sobolev space of functions $f\in L^p$ with $\nabla f\in L^p$. The Lorentz space is denoted by $L^{p,q}$. We do not recall the definition of this space since it does not play any role in our proofs. The Jacobian of a mapping $f\colon{\mathbb R}^n\to{\mathbb R}^n$ is denoted by $J_f=\det Df$. A domain is an open and connected set. The integral average is denoted by $$\mvint_E f\, dx =\frac{1}{|E|}\int_E f\, dx.$$ By $C$ we denote a generic constant whose value may change in a single string of estimates. Writing $C=C(n,m)$ we will indicate that the constant $C$ depends on $n$ and $m$ only.
[**Acknowledgement.**]{} We would like to thank Jan Malý for providing us with a beautiful proof of Proposition \[T10.5\].
A few days before completion of this work we learned the sad news that Professor Bogdan Bojarski had passed away. He was the PhD advisor of Piotr Hajłasz and an inspiration for both of us. We mourn his passing, and we dedicate this paper with deep respect to his memory.
Preliminaries {#Prelim}
=============
In this section we collect some basic facts that are used in the proof of the main result. We present the results in a slightly more general form than we actually need, because they might be useful for some other applications.
Chain rule
----------
The main result of [@DHS] (see also [@Oliva]) provides an example of a surjective homeomorphism $f:(0,1)^n\to (0,1)^n$, $n\geq 3$, such that $f\in W^{1,1}$, $f^{-1}\in W^{1,1}$ and $J_f=0$ a.e., $J_{f^{-1}}=0$ a.e. Note that $f^{-1}\circ f=\operatorname{Id}$, but the chain rule $$Df^{-1}(f(x))Df(x)=\operatorname{Id}$$ cannot be satisfied on a set of positive measure because of the vanishing Jacobians. In fact, $f$ maps the set of full measure to the set of measure zero where $Df^{-1}$ is not defined. The situation is different if we assume that $J_f\neq 0$. Namely, we have
\[chain\] Let $U,V\subset{\mathbb R}^n$ be open sets. Assume that $f\in W^{1,1}_{\rm loc}(U,V)$, $g\in W^{1,1}_{\rm loc}(V,{\mathbb R})$ and $g\circ f\in W^{1,1}_{\rm loc}(U,{\mathbb R})$. Then $$\label{eq1}
D(g\circ f)(x)=Dg(f(x))Df(x)$$ for almost all points $x$ in the set $\{x\in U:\, J_f(x)\neq 0\}$.
In particular, the result says that $Dg(f(x))$ is well defined at almost all points $x$ such that $J_f(x)\neq 0$.
The set where the Jacobian is different than zero splits into two sets where the Jacobian is positive and negative, respectively. Thus it suffices to show that is satisfied almost everywhere in the set where the Jacobian is positive, $$X=\{x\in U:\, J_f(x)>0 \},$$ because a similar argument can be applied to the set where the Jacobian is negative.
It is well known [@acerbif; @bojarskih; @hajlasz1] that $u\in W^{1,1}({\mathbb R}^n)$ satisfies the pointwise inequality $$\label{eq11}
|u(x)-u(y)|\leq C(n)|x-y|({\mathcal{M}}|D u|(x)+{\mathcal{M}}|D u|(y))
\quad\text{a.e.,}$$ where ${\mathcal{M}}|Du|(x)=\sup_{r>0}\mvint_{{\mathbb B}^n(x,r)}|D u|(y)\, dy$ is the Hardy-Littlewood maximal function. Hence for each $t>0$, $u$ restricted to the set $\{{\mathcal{M}}|Du|\leq t\}$ is Lipschitz continuous. This implies that ${\mathbb R}^n$ can be decomposed into a set of measure zero and countably many sets such that on each of these sets $u$ is Lipschitz continuous. This fact and a partition of unity argument implies that $U$ can be decomposed into Borel sets $$\label{eq3}
U=N_o\cup\bigcup_{i=1}^\infty K_i$$ such that $|N_o|=0$ and $f|_{K_i}$ is Lipschitz continuous. We need to use here a partition of unity argument, because $f$ is defined in $U$, while applies to functions defined on ${\mathbb R}^n$.
It remains to show that is satisfied at almost all points of the set $X\cap K_i$ for each $i=1,2,\ldots$ Let $f_i$ be a Lipschitz extension of $f|_{X\cap K_i}$ to all of ${\mathbb R}^n$ (see [@EG Theorem 3.1]). According to the Rademacher theorem, [@EG Theorem 3.2], $D f_i$ exists a.e. Also $Df_i=Df$ a.e. in $X\cap K_i$. Indeed, $f-f_i=0$ in $X\cap K_i$ so $D(f-f_i)=0$ a.e. in $X\cap K_i$ by [@EG Theorem 4.4(iv)]. Let $$W_i=\{x\in X\cap K_i:\, \text{$Df_i(x)$ exists and $Df_i(x)=Df(x)$}\}.$$ Since $|(X\cap K_i)\setminus W_i|=0$, it remains to show that is satisfied at almost all points of the set $W_i$. Note that $J_{f_i}>0$ on $W_i$. According to [@EG Lemma 3.3] we can decompose the set $W_i$ into a family of pairwise disjoint Borel sets $$W_i=\bigcup_{j=1}^\infty S_j$$ such that $f_i|_{S_j}$ is bi-Lipschitz for each $j=1,2,\ldots$, and it remains to prove at almost all points of $S_j$. If $|S_j|=0$, the result is obvious, so we can assume that $|S_j|>0$ and hence $f(S_j)=f_i(S_j)$ has positive measure, too.
Since $g\in W^{1,1}_{\rm loc}$, we have a decomposition $$f(S_j)=M_o\cup\bigcup_{k=1}^\infty E_k,
\quad
|M_o|=0,
\quad
\text{$g|_{E_k}$ is Lipschitz continuous.}$$ Indeed, we have a decomposition of $V$ similar to and then we take intersections with the set $f(S_j)$. We can also assume that $|E_k|>0$ for all $k$, as otherwise we could add sets $E_k$ of measure zero to the set $M_o$. Since the mapping $f$ is bi-Lipschitz on $S_j$, we have that $|(f|_{S_j})^{-1}(M_o)|=0$ and it remains to prove that is satisfied at almost all points of $Z_{jk}=(f|_{S_j})^{-1}(E_k)$ for $k=1,2,\ldots$ Let $g_k$ be a Lipschitz extension of $g|_{E_k}$ to all of ${\mathbb R}^n$. Then $g_k$ is differentiable a.e. and $Dg_k=Dg$ almost everywhere in $E_k$. Since $f|_{S_j}$ is bi-Lipschitz, the preimage $(f|_{S_j})^{-1}$ of the set of points in $E_k$ where $g_k$ is not differentiable has measure zero. This and the classical chain rule for differentiable functions imply that $g\circ f=g_k\circ f_i$ in $Z_{jk}$ and $$D(g_k\circ f_i)(x)=Dg_k(f_i(x))Df_i(x)=Dg(f(x))Df(x)
\quad
\text{a.e. in $Z_{jk}$.}$$ Since $g_k\circ f_i$ is Lipschitz continuous and it coincides with $g\circ f$ in $Z_{jk}$, it follows that $D(g_k\circ f_i)=D(g\circ f)$ almost everywhere in $Z_{jk}$.
As an immediate corollary we obtain the following result that will be used in the proof of Theorem \[main\], see also [@DSS Theorem 1.1 and 1.3], [@FMS Lemma 2.1], [@HK].
\[lem:inverse\] Assume that $\Omega\subset{\mathbb R}^n$ is a domain and $f\in W^{1,1}_{\rm loc}(\Omega,{\mathbb R}^n)$ is a homeomorphism such that ${f^{-1}\in W^{1,1}_{\rm loc}(f(\Omega),{\mathbb R}^n)}$. Then $$\label{eq2}
(Df(x))^{-1}=Df^{-1}(f(x))$$ almost everywhere in the set where $J_f\neq 0$.
In particular, Corollary \[lem:inverse\] applies to homeomorphisms described in Theorem \[main\].
\[T6\] Let $U\subset{\mathbb R}^n$ be open and let $f\in W^{1,1}_{\rm loc}(U,{\mathbb R}^n)$ be continuous. If a compact set $K\subset \{x\in U: J_f(x)\neq 0\}$ has positive measure, then the set $f(K)$ has positive measure.
In general, continuous mappings (even homeomorphisms) may map measurable sets to non-measurable sets. This is why we assume that $K$ is compact to guarantee measurability of the set $f(K)$.
This is a corollary of the proof of Lemma \[chain\] and we assume the same notation as in the proof of Lemma \[chain\]. In particular we assume that the sets $W_i$ and $S_j$ are defined in the same way.
Let $K\subset\{x\in U:\, J_f(x)\neq 0\}$ be compact and of positive measure. Since $f$ is continuous, $f(K)$ is compact and hence measurable. One of the sets $K\cap\{ J_f>0\}$ or $K\cap \{J_f<0\}$ has positive measure. Without loss of generality we may assume that the set $K\cap\{J_f>0\}$ has positive measure.
The sets $W_i$ constructed in the proof of Lemma \[chain\] cover almost all points of the set $\{ J_f>0\}$ so $|K\cap W_i|>0$ for some $i$. Since $W_i$ is the union of sets $S_j$, $|K\cap S_j|>0$ for some $j$. The mapping $f|_{S_j}$ is bi-Lipschitz and it follows that $f(K\cap S_j)$ is measurable and of positive measure. Hence also $f(K)$ has positive measure.
Blow-up technique
-----------------
In this section we describe a [*blow-up technique*]{} (Lemma \[T4\]) that is often used in the study of partial differential equations. This technique has also been used in [@henclm1]. Later, we generalize the blow-up technique to a simultaneous blow-up for a homeomorphism and its inverse, Lemma \[T9\]. This result will be used in the proof of Theorem \[main\].
[*All*]{} results of this section are local in nature, so they are true for functions and mappings defined on domains in ${\mathbb R}^n$ and not necessarily on all of ${\mathbb R}^n$. However, for simplicity of notation we decided to formulate the results on ${\mathbb R}^n$.
We will need the following two classical lemmata, the first of which is due to Lebesgue.
\[lem:LDT\] If $f\in L^p_{\rm loc}({\mathbb R}^n)$, $1\leq p<\infty$, then $$\label{LDT}
\mvint_{{\mathbb B}^n(x,r)} |f(y)-f(x)|^p dy\xrightarrow{r\to 0} 0
\quad
\text{for a.e. $x \in {\mathbb R}^n$.}$$
The points $x\in {\mathbb R}^n$ where holds true are called *$p$-Lebesgue points* of $f$.
The second, due to Calderón and Zygmund, [@CZ Theorem 12], is also an immediate consequence of [@EG Theorem 6.2].
\[lem:CZ\] If $f\in W^{1,p}_{\rm loc}({\mathbb R}^n)$, $1\le p<\infty$, then $$\label{eq:CZ}
\mvint_{{\mathbb B}^n(x,r)}\left|\frac{f(y)-f(x)-Df(x)\cdot(y-x)}{r}\right|^p\,dy\xrightarrow{r\to 0} 0
\quad
\text{for a.e. $x\in{\mathbb R}^n$.}$$
Note that the above lemmata immediately generalize to the case of vector valued functions $f\in W^{1,p}_{\rm loc}({\mathbb R}^n,{\mathbb R}^k)$, since it suffices to apply them to components of $f$. In particular we have that if $f\in W^{1,p}_{\rm loc}({\mathbb R}^n,{\mathbb R}^k)$, then $$\label{eq12}
\mvint_{{\mathbb B}^n(x,r)} |Df(y)-Df(x)|^p dy\xrightarrow{r\to 0} 0
\quad
\text{for a.e. $x \in {\mathbb R}^n$.}$$
\[D12\] Let $f\in W^{1,p}_{\rm loc}({\mathbb R}^n,{\mathbb R}^k)$, $1\leq p<\infty$. We say that $x\in{\mathbb R}^n$ is a [*$p$-good point*]{} for $f$ if both of the integrals and converge to zero.
Clearly, almost all points of ${\mathbb R}^n$ are $p$-good points for $f\in W^{1,p}_{\rm loc}$.
The basic [*blow-up technique*]{} is described by the following lemma. It allows us to regard $f$ almost as a linear map near any $p$-good point.
\[T4\] Let $f\in W^{1,p}_{\rm loc}({\mathbb R}^n,{\mathbb R}^k)$. For a $p$-good point $x_o\in{\mathbb R}^n$ and $r>0$ we define $$f_r(x)=\frac{f(x_o+rx)-f(x_o)}{r}
\quad
\text{and}
\quad
f_0(x)=Df(x_o)x.$$ Then $f_r$ converges to the linear map $f_0$ in the norm of $W^{1,p}({\mathbb B}^n,{\mathbb R}^k)$ as $r\to 0$, where ${\mathbb B}^n={\mathbb B}^n(0,1)$ is the unit ball.
Let $x_o\in{\mathbb R}^n$ be a $p$-good point for $f$. Note that $Df_0(x)=Df(x_o)$. We have $$\mvint_{{\mathbb B}^n}|f_r(x)-f_0(x)|^p\, dx =
\mvint_{{\mathbb B}^n(x_o,r)}\left|\frac{f(y)-f(x_o)-Df(x_o)\cdot(y-x_o)}{r}\right|^p\, dy\to 0$$ as $r\to 0$, and $$\mvint_{{\mathbb B}^n}|Df_r(x)-Df_0(x)|^p\, dx =
\mvint_{{\mathbb B}^n(x_o,r)} |Df(y)-Df(x_o)|^p\, dy\to 0
\quad
\text{as $r\to 0$.}$$
The rest of the section is devoted to a simultaneous blow-up for a homomorphism and its inverse.
\[T7\] Let $f\in W^{1,p}_{\rm loc}({\mathbb R}^n,{\mathbb R}^n)$, $1\leq p<\infty$, be a homeomorphism such that $f^{-1}\in W^{1,q}_{\rm loc}({\mathbb R}^n,{\mathbb R}^n)$, $1\leq q<\infty$. Then almost all points of the set $\{x\in{\mathbb R}^n:\, J_f(x)\neq 0\}$ have the following three properties satisfied simultaneously
1. $x$ is a $p$-good point for $f$,
2. $f(x)$ is a $q$-good point for $f^{-1}$,
3. $(Df(x))^{-1}=(Df)^{-1}(f(x))$.
A homeomorphic image of a Lebesgue measurable set need not be measurable, but a homeomorphic image of a Borel set is Borel, so we need to work with Borel sets.
Let $A\subset{\mathbb R}^n$ be a Borel set of $q$-good points for $f^{-1}$ such that $|{\mathbb R}^n\setminus A|=0$. Then $f^{-1}(A)$ is Borel and hence measurable. Almost all points of the set $\{J_f\neq 0\}\cap f^{-1}(A)$ have properties (a) and (b) and in order to show that almost all points of the set $\{J_f\neq 0\}$ have properties (a) and (b) it suffices to show that the set $$X=\{x\in{\mathbb R}^n:\, J_f(x)\neq 0\}\setminus f^{-1}(A)$$ has measure zero. Suppose to the contrary that $|X|>0$. Let $K\subset X$ be a compact set of positive measure. Then $f(K)\subset{\mathbb R}^n\setminus A$, and according to Corollary \[T6\], $f(K)$ has positive measure. This is, however, impossible, since ${\mathbb R}^n\setminus A$ has measure zero.
We proved that almost all points of the set $\{J_f\neq 0\}$ have properties (a) and (b). Now it follows from Corollary \[lem:inverse\] that almost all points of the set $\{J_f\neq 0\}$ have all three properties (a), (b) and (c).
The next lemma is easy to prove.
\[T8\] Let $f$ be as in Lemma \[T7\] and let $A\in GL(n)$ be a non-degenerate linear transformation on ${\mathbb R}^n$. If a point $x_o\in \{J_f\neq 0\}$ satisfies conditions (a), (b) and (c), then $x_o$ also satisfies conditions (a), (b), (c) for a homeomorphism $g=A\circ f$.
\[R1\] Whether a point $x_o$ satisfies conditions (a), (b) and (c) for the mapping $f$ depends on the choice of representatives of $Df$ and $Df^{-1}$. More precisely, it depends on how the values of $Df(x_o)$ and $Df^{-1}(f(x_o))$ are defined. However, we proved that no matter how we choose representatives of $Df$ and $Df^{-1}$, almost all points will satisfy (a), (b), (c). If a point $x_o$ satisfies conditions (a), (b) and (c) for the mapping $f$, then we will prove that $x_o$ satisfies the same conditions for $g$, provided the representatives of $Dg$ and $Dg^{-1}$ are such that $$Dg(x_o)=ADf(x_o),
\quad
Dg^{-1}(g(x_o))=Df^{-1}(f(x_o))A^{-1},$$ but we can make a choice of such representatives without any harm being done.
The proof that $x_o$ is good for $g$ is straightforward. The proof that it is also good for $g^{-1}=f^{-1}\circ A^{-1}$ follows from the change of variables $\Phi(y)=A^{-1}(y)$. Then the averages over the balls will became averages over scaled and translated ellipsoids $A^{-1}{\mathbb B}^n$ and it remains to observe (a well known fact) that if averages at and (for $g^{-1}$ in place of $f$) over the balls ${\mathbb B}^n$ converge to zero, then also the averages over the ellipsoids $A^{-1}{\mathbb B}^n(g(x_o),r)$ converge to zero. Indeed, the average over an ellipsoid can be estimated from above by the average over a larger ball that contains $A^{-1}{\mathbb B}^n(g(x_o),r)$, with a uniform constant that does not depend on the diameter of the ellipsoid. Finally the condition (c) for $g=A\circ f$ is a consequence of linear algebra and the choice of representatives of $Dg$ and $Dg^{-1}$ (see Remark \[R1\]). We leave details to the reader.
Let $f$ and $x_o$ are as in Lemma \[T8\]. If $J_f(x_o)>0$ and $A=Df(x_o)^{-1}$, then $$Dg(x_o)=Dg^{-1}(g(x_o))=\operatorname{Id}.$$ If $J_f(x_o)<0$ and $A={\mathcal R}Df(x_o)^{-1}$, where $${\mathcal R}=\operatorname{Diag}(1,1,\ldots,1,-1)=
\begin{pmatrix}1&0&\hdots&0&0\\0&1&\hdots&0&0\\\vdots&\vdots&\ddots&\vdots&\vdots\\0&0&\hdots&1&0\\0&0&\hdots&0&-1\end{pmatrix},$$ then $$Dg(x_o)=Dg^{-1}(g(x_o))={\mathcal R},$$ because ${\mathcal R}={\mathcal R}^{-1}$.
In both cases the linear transformation $A$ has positive determinant.
The next lemma describes the simultaneous blow-up in the case of negative Jacobian. This is what we will need in the proof of Theorem \[main\]. In the case of positive Jacobian one can easily formulate a similar result with $\mathcal{R}$ replaced by $\operatorname{Id}$.
\[T9\] Let $f\in W^{1,p}_{\rm loc}({\mathbb R}^n,{\mathbb R}^n)$, $1\leq p<\infty$, be a homeomorphism such that $f^{-1}\in W^{1,q}_{\rm loc}({\mathbb R}^n,{\mathbb R}^n)$, $1\leq q<\infty$. Then for almost every point $x_o$ of the set $\{x\in{\mathbb R}^n:\, J_f(x)<0\}$, there is a linear transformation $A$ with positive determinant such that the homeomorphism $g=A\circ f$ satisfies $$\lim_{r\to 0^+}\Vert g_r-\mathcal{R}\Vert_{W^{1,p}({\mathbb B}^n,{\mathbb R}^n)}=
\lim_{r\to 0^+}\Vert (g_r)^{-1}-\mathcal{R}\Vert_{W^{1,q}({\mathbb B}^n,{\mathbb R}^n)}=0,$$ where $$g_r(x)=\frac{g(x_o+rx)-g(x_o)}{r},
\quad
r>0.$$
Almost every point of the set $\{ J_f<0\}$ satisfies conditions (a), (b) and (c) of Lemma \[T7\] for $f$. Fix such a point $x_o$. Then, by Lemma \[T8\], $x_o$ satisfies conditions (a), (b) and (c) for $g=A\circ f$. Choosing $A={\mathcal R}(Df(x_o))^{-1}$, we have that $$Dg(x_o)=Dg^{-1}(g(x_o))=\mathcal{R}.$$ Since $J_f(x_o)<0$, it follows that $\det A>0$. We identify $\mathcal{R}$ with the linear transformation $x\mapsto\mathcal{R}x$. Since $x_o$ is a $p$-good point for $g$ and $g(x_o)$ is a $q$-good point for $g^{-1}$, Lemma \[T4\] implies that $$\lim_{r\to 0^+}\Vert g_r-\mathcal{R}\Vert_{W^{1,p}({\mathbb B}^n,{\mathbb R}^n)}=
\lim_{r\to 0^+}\Vert (g^{-1})_r-\mathcal{R}\Vert_{W^{1,q}({\mathbb B}^n,{\mathbb R}^n)}=0,$$ where $(g^{-1})_r$ is the blow up of $g^{-1}$ at $g(x_o)$. It remains to observe that $(g^{-1})_r=(g_r)^{-1}$, which easily follows from the definitions.
Fubini’s theorem for Sobolev spaces
-----------------------------------
The fact that almost all slices $f_x$, $f_{i,x}$ in the lemma below belong to the Sobolev space $W^{1,p}$ is well known. It is often called Fubini’s theorem for Sobolev spaces. On the other hand, facts and are not so well known.
\[T5\] Let $f,f_i\in W^{1,p}\big((0,1)^n\big)$, $1\leq p<\infty$, with $f_i\to f$ as $i\to\infty$. Denote the points of the cube $(0,1)^n$ by $$(x,y)\in(0,1)^{n-\ell}\times(0,1)^\ell=(0,1)^n$$ and define $$f_x(y)=f(x,y),
\quad
f_{i,x}(y)=f_i(x,y).$$ Then for almost all $x\in (0,1)^{n-\ell}$ we have $f_x,f_{i,x}\in W^{1,p}\big((0,1)^\ell\big)$ and there is a subsequence $f_{i_j}$ such that for almost all $x\in (0,1)^{n-\ell}$ $$\label{eq4}
\lim_{j\to\infty}\Vert f_{i_j,x}-f_x\Vert_{W^{1,p}((0,1)^\ell)}=0.$$ Moreover, for any ${\varepsilon}>0$, there is a compact set $K\subset (0,1)^{n-\ell}$ such that $|(0,1)^{n-\ell}\setminus K|<{\varepsilon}$ and $$\label{eq5}
\lim_{j\to\infty}\sup_{x\in K}\Vert f_{i_j,x}-f_x\Vert_{W^{1,p}((0,1)^\ell)}= 0.$$
The fact that $f_x\in W^{1,p}\big((0,1)^\ell\big)$ for almost all $x\in (0,1)^{n-\ell}$ is an easy consequence of the classical Fubini theorem applied to a sequence of smooth functions approximating $f$ in $W^{1,p}\big((0,1)^n\big)$. We leave details to the reader. Similarly, we prove that for every $i\in{\mathbb N}$, $f_{i,x}\in W^{1,p}\big((0,1)^\ell\big)$ for almost all $x\in (0,1)^{n-\ell}$. Since we have countably many functions $f$, $f_i$, $i\in{\mathbb N}$, there is a set $A\subset (0,1)^{n-\ell}$ of full measure such that $f_x,f_{i,x}\in W^{1,p}$ for all $x\in A$ and all $i\in{\mathbb N}$. We have $$\begin{split}
&
\Vert f_i-f\Vert_{W^{1,p}((0,1)^n)}^p =\\
&
\int_{(0,1)^{n-\ell}}\int_{(0,1)^\ell} |f_i(x,y)-f(x,y)|^p +|Df_i(x,y)-Df(x,y)|^p\, dy\, dx\\
&=
\int_{(0,1)^{n-\ell}}\underbrace{\Vert f_{i,x}-f_x\Vert_{W^{1,p}((0,1)^\ell)}^p}_{F_i(x)}\, dx
\rightarrow 0
\quad
\text{as $i\to\infty$.}
\end{split}$$ In other words, $F_i\to 0$ in $L^1\big((0,1)^{n-\ell}\big)$. Therefore, there is a subsequence $F_{i_j}$ such that $F_{i_j}(x)\to 0$ for almost all $x\in (0,1)^{n-\ell}$, which is . Moreover, according to Egoroff’s theorem [@EG Theorem 1.16], for any ${\varepsilon}>0$ there is a compact set $K\subset (0,1)^{n-\ell}$ such that $|(0,1)^{n-\ell}\setminus K|<{\varepsilon}$ and $F_{i_j}\to 0$ uniformly on $K$, which is .
The above result allows for a lot of flexibility and instead of applying Fubini’s theorem to the products of cubes, we can apply it for example to ${\mathbb B}^{n-\ell}\times{\mathbb S}^{\ell}$, as described in the next result.
\[T10\] Let $f_r,f_0\in W^{1,p}({{\mathbb B}}^{n-\ell}\times{\mathbb S}^\ell,{\mathbb R}^n)$, $1\leq p<\infty$, $0<r<r_0$, be a family of mappings such that $f_r\to f_0$ in $W^{1,p}$ as $r\to 0^+$. If $r_i\searrow 0$ is a sequence decreasing to zero, then there is a subsequence $r_{i_j}$ such that $f_j=f_{r_{i_j}}$ satisfies:
For almost all $x\in {\mathbb B}^{n-\ell}$ $$f_j|_{\{ x\}\times{\mathbb S}^\ell}\to f_0|_{\{ x\}\times{\mathbb S}^\ell}
\quad
\text{in $W^{1,p}(\{ x\}\times{\mathbb S}^\ell)$ as $j\to\infty$,}$$ and for any ${\varepsilon}>0$ there is a compact subset of the unit ball $K\subset {\mathbb B}^{n-\ell}$ such that $|{\mathbb B}^{n-\ell}\setminus K|<{\varepsilon}$ and $$\sup_{x\in K}\Vert f_j-f_0\Vert_{W^{1,p}(\{ x\}\times{\mathbb S}^\ell)}\to 0
\quad
\text{as $j\to\infty$.}$$
Traces and extensions
---------------------
The following lemma is known, but not very well known.
\[T10.5\] For $n\geq 1$ and $p>1$, there is a bounded linear extension operator $$E:W^{1,p}({\mathbb R}^{n})\to W^{1,q}\cap C^\infty({\mathbb R}^{n+1}_+),
\quad
\text{where $q=\frac{(n+1)p}{n}$.}$$ In other words, $W^{1,p}({\mathbb R}^{n})$ continuously embeds into the trace space $W^{1-\frac{1}{q},q}({\mathbb R}^{n})$ of $W^{1,q}({\mathbb R}^{n+1}_+)$.
\[T11\] For $n\geq 2$, there is a bounded linear extension operator $$E:W^{1,n}({\mathbb R}^{n})\to W^{1,n+1}\cap C^\infty({\mathbb R}^{n+1}_+).$$ In other words, $W^{1,n}({\mathbb R}^{n})$ continuously embeds into the trace space $W^{1-\frac{1}{n+1},n+1}({\mathbb R}^{n})$ of $W^{1,n+1}({\mathbb R}^{n+1}_+)$.
Corollary \[T11\] fails when $n=1$, see [@BIN], [@leoni Exercise 14.36] and [@SW Proposition 4].
Proposition \[T10.5\] was proved in [@BIN]. It also follows from Theorem 14.32, Remark 14.35 and Proposition 14.40 in [@leoni] (first edition). Corollary \[T11\] was also proved in [@GH Lemma 14] as a consequence of Theorem 2.5.6, Theorem 2.7.1, Proposition 2.3.2.2(8), Theorem 2.5.7 and 2.5.7(9) in [@triebel]. All of the arguments listed here are difficult. Below we present an elementary and unpublished proof of Proposition \[T10.5\] due to Jan Malý.
In the proof we need the following result.
\[T12\] If $F\in L^1_{\rm loc}({\mathbb R}^{n+1}_+)$, $f\in L^p({\mathbb R}^{n})$, $p>1$, $n\geq 1$, and $$\label{eq6}
|F(x,t)|\leq C\mvint_{{\mathbb B}^{n}(x,t)}|f(y)|\, dy
\quad
\text{for $(x,t)\in{\mathbb R}^{n}\times (0,\infty)={\mathbb R}^{n+1}_+$,}$$ then $$\Vert F\Vert_{L^q({\mathbb R}^{n+1}_+)}\leq C\Vert f\Vert_{L^p({\mathbb R}^{n})},
\quad
\text{where $q=\frac{(n+1)p}{n}$.}$$
Since the right-hand side of is bounded, up to the constant $C$, by the maximal function ${\mathcal{M}}f(x)$, we have $$\label{eq7}
\int_0^r |F(x,t)|^q\, dt \leq
Cr({\mathcal{M}}f)^q(x)
\quad
\text{for $r>0$.}$$ On the other hand, the inequality $$|F(x,t)|\leq C\left(\,\mvint_{{\mathbb B}^{n}(x,t)}|f(y)|^p\, dy\right)^{1/p}\leq
C t^{-\frac{n}{p}} \Vert f\Vert_p$$ yields $$\label{eq8}
\int_r^\infty |F(x,t)|^q\, dt\leq C\Vert f\Vert_p^q r^{1-\frac{nq}{p}}=
C\Vert f\Vert_p^q r^{-n}
\quad
\text{for $r>0$.}$$ We used here the fact that $$\frac{nq}{p}=n+1>1.$$ Now if we choose $r>0$ such that $$r({\mathcal{M}}f)^q(x)=\Vert f\Vert_p^qr^{-n},
\quad
\text{i.e.,}
\quad
r=\left(\frac{\Vert f\Vert_p}{{\mathcal{M}}f(x)}\right)^{\frac{p}{n}},$$ then the right-hand sides of and are equal to $C\Vert f\Vert_p^{p/n}({\mathcal{M}}f(x))^p$. Adding integrals at and yields $$\int_0^\infty |F(x,t)|^q\, dt \leq C\Vert f\Vert_p^{p/n}({\mathcal{M}}f(x))^p$$ and Fubini’s theorem along with boundedness of the maximal operator in $L^p$, $p>1$, give $$\int_{{\mathbb R}^{n+1}_+} |F(x,t)|^q\, dt\, dx\leq
C\Vert f\Vert_p^{p/n}\Vert f\Vert_p^p=C\Vert f\Vert_p^q.$$
Now we can complete the proof of Proposition \[T10.5\].
Let $\varphi\in C_0^\infty({\mathbb B}^{n})$, $\varphi\geq 0$, $\int_{{\mathbb B}^{n}}\varphi(x)\, dx =1$, and $\varphi_t(x)=t^{-n}\varphi(x/t)$.
For $f\in L^1_{\rm loc}({\mathbb R}^{n})$ we define $$(Ef)(x,t)=(f*\varphi_t)(x)=\int_{{\mathbb B}^{n}}f(x-ty)\varphi(y)\, dy,
\quad
(x,t)\in{\mathbb R}^{n+1}_+.$$ Clearly, properties of the convolution guarantee that $Ef\in C^\infty({\mathbb R}^{n+1}_+)$ and a simple change of variables yields $$|(Ef)(x,t)|=\left|\,\int_{{\mathbb B}^{n}(x,t)}f(y)\varphi_t(x-y)\, dy\right|\leq C \mvint_{{\mathbb B}^{n}(x,t)}|f(y)|\, dy,$$ where the constant $C$ depends on $\varphi$ only. Therefore, if $f\in L^{p}({\mathbb R}^{n})$, Lemma \[T12\] gives the estimate $$\label{eq9}
\Vert Ef\Vert_{L^q({\mathbb R}^{n+1}_+)}\leq C\Vert f\Vert_{L^{p}({\mathbb R}^{n})}.$$ Assume now that $u\in W^{1,p}({\mathbb R}^{n})$. We have $$\label{eq10}
|\nabla (Eu)(x,t)|=|\nabla_{(x,t)}(Eu)(x,t)|\leq
C\mvint_{{\mathbb B}^{n}(x,t)}|\nabla u(y)|\, dy.$$ Indeed, $$|\nabla_{x}(Eu)(x,t)|=\left|\,\int_{{\mathbb B}^{n}}(\nabla u)(x-ty)\varphi(y)\, dy\right|\leq
C\mvint_{{\mathbb B}^{n}(x,t)}|\nabla u(y)|\, dy,$$ and $$\left|\frac{\partial}{\partial t}(Eu)(x,t)\right| =
\left|\,\int_{{\mathbb B}^{n}}(\nabla u)(x-ty)\cdot(-y)\varphi(y)\, dy\right|\leq
C\mvint_{{\mathbb B}^{n}(x,t)}|\nabla u(y)|\, dy,$$ because $|-y|\leq 1$ for $y\in{\mathbb B}^{n}$. Now and Lemma \[T12\] imply that $$\Vert\nabla (Eu)\Vert_{L^q({\mathbb R}^{n+1}_+)}\leq C\Vert\nabla u\Vert_{L^{p}({\mathbb R}^{n}).}$$ This estimate and for $f=u$ complete the proof.
A localized version of Corollary \[T11\] gives the following result:
\[L26\] Let $n\geq 2$. Then there is an extension operator $$E:W^{1,n}(\partial({\mathbb S}^n\times [0,1]))\to
W^{1,n+1}\cap C^\infty ({\mathbb S}^n\times(0,1))$$ such that if $h=f_0$ on ${\mathbb S}^n\times\{0\}$ and $h=f_1$ on ${\mathbb S}^n\times\{1\}$, then traces (in the Sobolev sense) satisfy $\Tr Eh(\cdot,0)=f_0$, $\Tr Eh(\cdot,1)=f_1$, and $$\Vert Eh\Vert_{W^{1,n+1}({\mathbb S}^n\times(0,1))}\leq
C(n)\left(\Vert f_0\Vert_{W^{1,n}({\mathbb S}^n)}+\Vert f_1\Vert_{W^{1,n}({\mathbb S}^n)}\right).$$
The next result seems new. Its proof uses some ideas from the proof of [@hajlaszm Proposition 3.3].
\[L27\] Let $n\geq 2$, $m\geq n+1$ and let $f\in W^{1,n}\cap C^0({\mathbb S}^n,{\mathbb R}^m)$, $g\in C^\infty({\mathbb S}^n,{\mathbb R}^m)$. Then there is a function $H\in C^0({\mathbb S}^n\times [0,1],{\mathbb R}^m)\cap C^\infty({\mathbb S}^n\times [0,1))$ such that $H(x,0)=g(x)$, $H(x,1)=f(x)$ and $$\mathcal{H}^{n+1}(H({\mathbb S}^n\times [0,1)))\leq
C\Vert f-g\Vert_{W^{1,n}({\mathbb S}^n)}
\left(\Vert f-g\Vert_{W^{1,n}({\mathbb S}^n)}+\Vert Dg\Vert_{L^{n+1}({\mathbb S}^n)}\right)^n,$$ where the constant $C$ depends on $n$ and $m$ only and ${\mathcal H}^{n+1}$ denotes the Hausdorff measure.
\[rem1\] Fix $g\in C^\infty({\mathbb S}^n,{\mathbb R}^m)$. Let $f,f_k\in W^{1,n}\cap C^0({\mathbb S}^n,{\mathbb R}^m)$ and let $H,H_k$ be the homotopies for $f$ and $f_k$ constructed as in Proposition \[L27\]. If $f_k\to f$ uniformly on ${\mathbb S}^n$ as $k\to\infty$ (we do not require convergence in the Sobolev norm), then $H_k\to H$ uniformly on ${\mathbb S}^n\times [0,1]$. Indeed, the homotopies are defined by the formula and the extension operator $Eh$ is defined through the averaging and multiplication by a cut-off function, and such a construction is continuous in the uniform norm.
The proposition has a clear geometric interpretation. The image of a continuous mapping $f\in W^{1,n}\cap C^0({\mathbb S}^n,{\mathbb R}^m)$ can be very large. It can even fill a ball in ${\mathbb R}^m$. However, if $f$ is very close in the $W^{1,n}$ norm to a fixed smooth map $g$, then there is a homotopy between $f$ and $g$ such that the $\mathcal{H}^{n+1}$-volume of the image of the homotopy, except the endpoint where the homotopy equals $f$, is very small. We hope that this result might be useful for other applications.
In the proof we will need the following estimate.
\[L28\] If $A$ and $B$ are two $n\times n$-matrices, then $$\label{eq:l28}
|\det (A+B)|\leq C(n)(|A|^n+|B|^n),$$ where $|\cdot |$ stands for the Hilbert-Schmidt norm of a matrix.
Since the determinant is continuous and homogeneous of order $n$, we immediately get that $$|\det A|=|A|^n\left|\det\left(\frac{A}{|A|}\right)\right|\leq\Lambda |A|^n,
\quad
\text{where}
\quad
\Lambda=\sup\{ |\det M|:\, |M|\leq 1\}.$$ Then follows from the triangle inequality and the standard convexity estimate $(a+b)^n\leq 2^{n-1}(a^n+b^n)$ for $a,b\geq 0$.
Let $$h(x)=
\begin{cases}
f-g & \text{on ${\mathbb S}^n\times\{ 1\}$},\\
0 & \text{on ${\mathbb S}^n\times\{ 0\}$},
\end{cases}$$ and define $$\label{eq15}
H(x,t)=(Eh)(x,t)+g(x)
\quad
\text{for $(x,t)\in{\mathbb S}^n\times [0,1]$,}$$ where $E$ is the extension operator from Lemma \[L26\]. Since the extension $Eh$ is continuous (smooth) up to the boundary if the function on the boundary is continuous (smooth), we conclude that $H\in C^0({\mathbb S}^n\times [0,1],{\mathbb R}^m)\cap C^\infty({\mathbb S}^n\times [0,1))$ and $H(x,0)=g(x)$, $H(x,1)=f(x)$.
According to the area formula [@EG] we have that $$\label{eq13}
\mathcal{H}^{n+1}(H({\mathbb S}^n\times [0,1)))=\int_0^1\int_{{\mathbb S}^n} J_H(x,t)\, d\sigma(x)\, dt,$$ where $$J_H(x,t)=\sqrt{\det\big((DH)^T(DH)\big)}.$$ Denote the derivative in ${\mathbb S}^n\times [0,1]$ by $D=(D_x,\partial_t)$, where $D_x$ is the derivative on ${\mathbb S}^n$. Then $$DH(x,t)=(D_x(Eh)+D_xg,\partial_t(Eh)).$$ The Cauchy-Binet formula [@EG Sect. 3.2.1, Theorem 4], the Laplace expansion along the last column $(\partial_t(Eh))_I$, and Lemma \[L28\] yield the following estimate, where the sum is taken over all $I=(i_1,\ldots,i_{n+1})$, $1\leq i_1<\ldots< i_{n+1}\leq m$: $$J_H
=
\sqrt{\sum_{I}\Big(\det\big((D_x(Eh))_I+(D_xg)_I, (\partial_t(Eh))_I\big)\Big)^2}\\
\leq
C |\partial_t(Eh)|(|D_x(Eh)|^n+|D_x g|^n).$$ Therefore gives $$\begin{aligned}
\mathcal{H}^{n+1}(H({\mathbb S}^n&\times [0,1)))
\leq
C\Vert Eh\Vert_{W^{1,n+1}({\mathbb S}^n\times(0,1))}^{n+1}\\
&+
C\left(\int_0^1\int_{{\mathbb S}^n} |D_x g|^{n+1}\right)^{\frac{n}{n+1}}
\left(\int_0^1\int_{{\mathbb S}^n}|\partial_t(Eh)|^{n+1}\right)^\frac{1}{n+1}\\
&\leq C\Vert f-g\Vert^{n+1}_{W^{1,n}({\mathbb S}^n)}+
\Vert Dg\Vert_{L^{n+1}({\mathbb S}^n)}^n\Vert f-g\Vert_{W^{1,n}({\mathbb S}^n)}.\end{aligned}$$
The local degree and the linking number {#Sec2.5}
---------------------------------------
To keep the paper self contained, we present here a short introduction to the theory of local degree. The standard references are the books of Fonseca and Gangbo [@FG] and Outerelo and Ruiz [@OR]. Then, at the end of the section, we discuss the linking number.
Throughout this section, we assume that $\Omega\subset{\mathbb R}^n$ is a domain. By $C^1({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ we denote the set of all these mappings from ${\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}$ to ${\mathbb R}^n$ which admit an extension to a $C^1$ mapping on some open $U\supset{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}$.
We begin by defining the local degree for $C^1$ mappings at their regular values.
\[defn12\] Assume $\phi\in C^1({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$. Let $p\in {\mathbb R}^n$ be a regular value of $\phi$ and $p\not\in \phi(\partial \Omega)$. We define the local degree of $\phi$ at $p$ with respect to $\Omega$ as $$\deg(\phi,\Omega,p)=\sum_{x\in \phi^{-1}(p)} \sgn J_\phi(x).$$ Moreover, if $p\not \in \phi({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu})$, we set $\deg(\phi,\Omega,p)=0$.
It turns out that $\deg(\phi,\Omega,\cdot)$ is constant on connected components of ${{\mathbb R}^n\setminus \phi(\partial\Omega)}$.
\[prop18\] Let $V$ be a connected component of ${\mathbb R}^n\setminus \phi(\partial\Omega)$ and assume $p_1,p_2\in V$ are regular values of $\phi$. Then $\deg(\phi,\Omega,p_1)=\deg(\phi,\Omega,p_2)$.
Proposition \[prop18\] allows us to define the local degree also at critical points of $\phi$, as long as they are not in the image of $\partial\Omega$.
Assume $\Omega$ and $\phi$ are as in Definition \[defn12\] and let $p\in \phi(\Omega)\setminus \phi(\partial \Omega)$ be a critical value of $\phi$. Let $p_1$ be any regular value of $\phi$ such that $p$ and $p_1$ lie in the same connected component of ${\mathbb R}^n\setminus \partial\Omega$. We set $\deg(\phi,\Omega,p)=\deg(\phi,\Omega,p_1)$.
Note that by Sard’s Lemma such $p_1$ always exists; Proposition \[prop18\] shows that the above definition does not depend on the choice of $p_1$.
The local degree is a $C^1$ homotopy invariant:
\[thm112\] If $H:{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}\times[0,1]\to{\mathbb R}^n$ is a $C^1$ mapping such that $H(\cdot,0)=\phi(\cdot)$, $H(\cdot,1)=\psi(\cdot)$ and $p\not\in H(\partial\Omega\times[0,1])$, then $\deg(\phi,\Omega,p)=\deg(\psi,\Omega,p)$.
Proposition \[thm112\] allows us to extend the notion of local degree to continuous mappings:
\[defn118\] If $\phi\in C({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ and $p\not \in \phi(\partial\Omega)$, we set $\deg(\phi,\Omega,p)=\deg(\psi,\Omega,p)$, where $\psi$ is any mapping in $ C^1({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ such that $\sup_{x\in {\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}} |\phi(x)-\psi(x)|<\mathrm{dist}(p,\phi(\partial \Omega))$.
One easily checks that the above definition is independent on the choice of $\psi$: if we choose $\psi_1, \psi_2\in C^1({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ satisfying $\sup_{x\in{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}}|\phi(x)-\psi_i(x)|<\mathrm{dist}(p,\phi(\partial \Omega))$, $i=1,2$, then $p\not\in H(\partial \Omega\times[0,1])$, where $H$ is the standard homotopy between $\psi_1$ and $\psi_2$, $H(x,t)=(1-t)\psi_1(x)+t\psi_2(x)$, and thus, by Proposition \[thm112\], the local degrees of $\psi_1$ and $\psi_2$ at $p$ are the same.
\[R2\] In fact, a standard smoothing argument shows that the local degree for continuous maps at a point $p$, given by the above definition, is a homotopy invariant (without the $C^1$ assumption), as long as the homotopy $H$ does not map any points of $\partial \Omega$ into $p$, i.e. $p\not\in H(\partial \Omega\times [0,1])$.
We shall need the following deep facts on the local degree.
\[propo210\] Assume $\Omega\subset{\mathbb R}^n$ is a domain, $\phi\in C({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$, $V\subset {\mathbb R}^n$ is a domain containing $\phi({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu})$ and $\psi\in C({\mkern 1.7mu\overline{\mkern-1.7muV\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$. Let $D=V\setminus \phi(\partial\Omega)$ and denote by $D_i$ the connected components of $D$. Then for any $p\not\in (\psi\circ\phi)(\partial \Omega)\cup \psi(\partial V)$ we have $p\not\in \psi(\partial D_i)$, and the following formula holds: $$\label{pro39f}
\deg(\psi\circ\phi,\Omega,p)=\sum_{i}\deg(\psi,D_i,p)\deg(\phi,\Omega,q_i)$$ for arbitrary $q_i\in D_i$.
\[propo40\] If $\psi$ and $\phi$ are as in Proposition \[propo210\] and additionally $\psi$ and $\phi$ are homeomorphisms, then
- for any $p\not\in (\psi\circ\phi)(\partial \Omega)\cup \psi(\partial V)$ $$\label{pro40f}
\deg(\psi\circ\phi,\Omega,p)=\deg(\psi,\phi(\Omega),p)\deg(\phi,\Omega,\psi^{-1}(p)),$$
- for every $q\in \phi(\Omega)$ we have either $$\deg(\phi,\Omega,q)=\deg(\phi^{-1},\phi(\Omega),\phi^{-1}(q))=1$$ or $$\deg(\phi,\Omega,q)=\deg(\phi^{-1},\phi(\Omega),\phi^{-1}(q))=-1.$$
If $p\not\in \psi(\phi(\Omega))$, then both sides of are zero. Assume thus $p\in \psi(\phi(\Omega))$.
By Brouwer’s Invariance of Domain Theorem [@FG Theorem 3.30], $\phi(\partial\Omega)=\partial\phi(\Omega)$ and there is only one component $D_i$ of $V\setminus \phi(\partial\Omega)$ such that $\psi(D_i)$ contains $p$, namely $D_i=\phi(\Omega)$, thus all the terms of the sum in are zero, except the one with $D_i=\phi(\Omega)$. Also, we may choose $q_i=\psi^{-1}(p)$, which gives .
To prove b), take any domain $\Omega'$ such that $q\in \phi(\Omega')$ and ${\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}'\subset \Omega$. Applying a) to $\phi|_{{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}'}:{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}'\to{\mathbb R}^n$ and $\psi=\phi^{-1}:\phi({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu})\to{\mathbb R}^n$, we see that $$1=\deg({\mathrm{Id}},\Omega',\phi^{-1}(q))=\deg(\phi^{-1},\phi(\Omega'),\phi^{-1}(q))\deg(\phi,\Omega',q),$$ thus $\deg(\phi,\Omega',q)=\deg(\phi^{-1},\phi(\Omega'),\phi^{-1}(q))=\pm 1$. What remains to prove is that $\deg(\phi,\Omega',q)=\deg(\phi,\Omega,q)$ and $\deg(\phi^{-1},\phi(\Omega'),\phi^{-1}(q))=\deg(\phi^{-1},\phi(\Omega),\phi^{-1}(q))$. Let $\zeta\in C^1({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ be such that $$\sup_{{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}}|\phi-\zeta|<\dist(q,\phi(\partial\Omega'))<\dist(q,\phi(\partial\Omega)).$$ Then, by Definition \[defn118\], $\deg(\phi,\Omega,q)=\deg(\zeta,\Omega,q)$ and $\deg(\phi,\Omega',q)=\deg(\zeta,\Omega',q)$. However, for any $z\in \Omega\setminus\Omega'$, $$|\zeta(z)-q|\geq
|q-\phi(z)|-|\phi(z)-\zeta(z)|>
\dist(q,\phi(\partial\Omega'))-\dist(q,\phi(\partial\Omega'))=0,$$ thus $q\not\in \zeta(\Omega\setminus \Omega')$ and applying Definition \[defn12\] we see that $\deg(\zeta,\Omega,q)=\deg(\zeta,\Omega',q)$. Calculations for $\phi^{-1}$ follow the same steps. This concludes the proof of b).
A homeomorphism $h:\Omega\to{\mathbb R}^n$ with degree $+1$ at all its values is called *sense-* or *orientation-preserving*; if the degree is $-1$ at all values, we call it *sense-reversing*.
As an immediate corollary of Proposition \[propo40\], b) and Proposition \[prop18\], we obtain that every homeomorphism of a domain is either sense-preserving or sense-reversing.
Assume $\Omega\subset{\mathbb R}^n$ is open and connected and $h:{\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu}\to{\mathbb R}^n$ is a homeomorphism. Then either for every ${p\in h(\Omega)}$ we have ${\deg(h,\Omega,p)=1}$ or for every ${p\in h(\Omega)}$ we have ${\deg(h,\Omega,p)=-1}$.
If $h:\Omega\to{\mathbb R}^n$ is a sense-preserving (reversing) homeomorphism and $\Omega'\subset\Omega$, then $h|_{\Omega'}:\Omega'\to{\mathbb R}^n$ is also sense-preserving (reversing).
This is a corollary from the proof of Proposition \[propo40\], where we showed in the last step that $\deg(\phi,\Omega,q)=\deg(\phi,\Omega',q)$.
The terms *sense-preserving* and *sense-reversing* are justified by the following fact.
\[propo31\] Assume $\Omega\subset{\mathbb R}^n$ is a domain, $\phi\in C({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ and $h:{\mkern 1.7mu\overline{\mkern-1.7muU\mkern-1.5mu}\mkern 1.5mu}\to{\mathbb R}^n$ is a homeomorphism of a domain $U\supset \phi({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu})$.
- If $h$ is sense-preserving, then for every $p\in U\setminus \phi(\partial\Omega)$ we have $$\deg(\phi,\Omega ,p)=\deg(h\circ\phi,\Omega,h(p)).$$
- If $h$ is sense-reversing, then for every $p\in U\setminus \phi(\partial\Omega)$ we have $$\deg(\phi,\Omega ,p)={-\deg(h\circ\phi,\Omega,h(p))}.$$
Assume $h$ is sense-preserving and apply Proposition \[propo210\] to $h\circ \phi$: let $D=U\setminus \phi(\partial \Omega)$ and denote by $D_i$ the connected components of $D$. Then $p\in D_j$ for exactly one $j\in{\mathbb N}$ and we have for any $q_i\in D_i$, $i\neq j$, $$\begin{aligned}
\deg(h\circ\phi,\Omega,h(p))
&=\deg(h,D_j,h(p))\deg(\phi,\Omega,p)+
\sum_{i\neq j} \deg(h,D_i,h(p))\deg(\phi,\Omega,q_i)\\
&=\deg(\phi,\Omega,p),\end{aligned}$$ because $\deg(h,D_j,h(p))=+1$, while $\deg(h,D_i,h(p))=0$ for $i\neq j$, since $h(p)\not\in h(D_i)$.
The case of sense-reversing $h$ is proved in exactly the same way.
If $M$ and $N$ are compact, connected, oriented, smooth $n$-dimensional manifolds without boundary and $\phi:M\to N$ is $C^\infty$ smooth, then we define $$\deg(\phi,y)=
\sum_{x\in\phi^{-1}(y)}\sgn J_\phi(x),$$ where $y\in N$ is a regular value of $\phi$, and $J_\phi(x)$ is the determinant of the derivative $D\phi(x):T_xM\to T_{\phi(x)}N$. It turns out that $\deg(\phi,y)$ does not depend on the choice of a regular value $y$. The common value of all $\deg(\phi,y)$ is denoted by $\deg\phi$ and is called the [*degree of $\phi$*]{}.
One can prove that homotopic mappings have equal degrees. Since every continuous mapping is homotopic to a smooth one, one can extend the notion of degree to the class of all continuous mappings $\phi:M\to N$. For more details, see [@Milnor].
The following result relates the local degree of $\phi$ with the topological degree of $\phi$ restricted to the boundary.
\[propo46\] Let $\partial \Omega$ be a connected, compact and smooth manifold oriented by the outward normal vector ([@OR Section II.7.7]) and assume $\phi\in C({\mkern 1.7mu\overline{\mkern-1.7mu\Omega\mkern-1.5mu}\mkern 1.5mu},{\mathbb R}^n)$ is such that $\phi|_{\partial\Omega}:\partial\Omega\to {\mathbb S}^{n-1}$. Then $$\deg \phi|_{\partial \Omega}=\deg(\phi,\Omega,0).$$
Our main purpose for introducing the local degree is to justify the properties of yet another invariant, the linking number.
The linking number is an important and well studied invariant in the theory of knots. It was introduced by Gauss in a short note of 1833 [@Gauss] (see also [@ricca] for a nice historical account and modern interpretation): if $\gamma_1$, $\gamma_2$ are two parameterized, non-intersecting, oriented curves in ${\mathbb R}^3$, $\gamma_1,~\gamma_2:{\mathbb S}^1\to{\mathbb R}^3$, then the linking number $\ell(\gamma_1,\gamma_2)$ is defined (in modern notation) as the integral $$\label{Gauss int}
\ell(\gamma_1,\gamma_2)=\frac{1}{4\pi}\int_{{\mathbb S}^1\times{\mathbb S}^1}\frac{\det(\gamma_1'(s),\gamma_2'(t),\gamma_1(s)-\gamma_2(t))}{|\gamma_1(s)-\gamma_2(t)|^3}\,ds\, dt.$$ The Gauss map for $\gamma_1$, $\gamma_2$ is defined as $$\label{Gauss map}
{\mathbb S}^1\times{\mathbb S}^1\ni(s,t)\xmapsto{~~\psi~~}\frac{\gamma_1(s)-\gamma_2(t)}{|\gamma_1(s)-\gamma_2(t)|}\in {\mathbb S}^2.$$ It turns out (see e.g. [@ricca]) that the linking number given by is equal to the degree of the Gauss map: $\ell(\gamma_1,\gamma_2)=\deg(\psi)$. In colloquial terms, the linking number tells us how many times (counting directions) one curve winds around the other.
These definitions have been later generalized to pairs of non-intersecting manifolds in higher dimensions (see the paper by M. Kervaire, [@Kervaire], who attributes the idea to A. Shapiro). Here, we shall restrict ourselves to the case when the manifolds in question are two oriented spheres ${\mathbb S}^k$, ${\mathbb S}^l$, where $k+l=n-1$, continuously embedded in ${\mathbb R}^n$ in such a way that the embedded spheres do not intersect. Then, if we denote the embeddings as before, in the case of curves, by $\gamma_1:{\mathbb S}^k\to{\mathbb R}^n$ and $\gamma_2:{\mathbb S}^l\to{\mathbb R}^n$, we can define the Gauss map of the two embeddings by the formula , (this time, however, $\psi:{\mathbb S}^k\times{\mathbb S}^l\to {\mathbb S}^{n-1}$), and the linking number of the two embedded (oriented) spheres, identified here with the embedding maps $\gamma_1$ and $\gamma_2$, is defined again as $\ell(\gamma_1,\gamma_2)=\deg \psi$.
More generally, we can define, by the same formula, the linking number $\ell(\gamma_1,\gamma_2)$ between any two continuous maps $\gamma_1:{\mathbb S}^k\to{\mathbb R}^n$ and $\gamma_2:{\mathbb S}^l\to{\mathbb R}^n$, $k+l=n-1$, provided $\gamma_1({\mathbb S}^k)\cap\gamma_2({\mathbb S}^l)=\varnothing$.
The linking number is a homotopy invariant in the sense that if $\gamma_1, \tilde{\gamma}_1:{\mathbb S}^k\to{\mathbb R}^n\setminus \gamma_2({\mathbb S}^l)$ are two mappings of ${\mathbb S}^k$, which are homotopic in ${\mathbb R}^n\setminus \gamma_2({\mathbb S}^l)$ (i.e. neither the two images $\gamma_1({\mathbb S}^k)$, $\tilde{\gamma}_1({\mathbb S}^k)$, nor the image of the homotopy between them intersects $\gamma_2({\mathbb S}^l)$), then $\ell(\gamma_1,\gamma_2)=\ell(\tilde{\gamma}_1,\gamma_2)$. Indeed, if $\Gamma:{\mathbb S}^k\times [0,1]\to {\mathbb R}^n$ is the homotopy between $\gamma_1$ and $\tilde{\gamma}_1$, the image of which is disjoint with the image of $\gamma_2$, then $\Psi:{\mathbb S}^k\times {\mathbb S}^l\times[0,1]\to{\mathbb S}^n$ given by the formula $$(s,t,r)\xmapsto{\Psi}\frac{\Gamma(s,r)-\gamma_2(t)}{|\Gamma(s,r)-\gamma_2(t)|}$$ is a homotopy between $$\psi(s,t)=\frac{\gamma_1(s,t)-\gamma_2(s,t)}{|\gamma_1(s,t)-\gamma_2(s,t)|}
\quad
\text{and}
\quad
\tilde{\psi}(s,t)=\frac{\tilde{\gamma_1}(s,t)-\gamma_2(s,t)}{|\tilde{\gamma_1}(s,t)-\gamma_2(s,t)|},$$ and thus the linking numbers $\ell(\gamma_1,\gamma_2)=\deg \psi$ and $\ell(\tilde{\gamma_1},\gamma_2)=\deg \tilde{\psi}$ are the same.
The following known invariance result is very hard to find in the literature, thus we present the proof here (essentially the same argument, in a less general situation, was given in [@henclm1]).
\[Propo inv\] Assume $k+l=n-1$ and let $\gamma_1:{\mathbb S}^k\to{\mathbb R}^n$ and $\gamma_2:{\mathbb S}^l\to{\mathbb R}^n$ be continuous maps such that $\gamma_1({\mathbb S}^k)\cap \gamma_2({\mathbb S}^l)=\varnothing$. Let $h:{\mathbb R}^n\to{\mathbb R}^n$ be a homeomorphism. Then
- if $h$ is sense-preserving, then $\ell(h\circ\gamma_1,h\circ\gamma_2)=\ell(\gamma_1,\gamma_2)$;
- if $h$ is sense-reversing, then $\ell(h\circ\gamma_1,h\circ\gamma_2)=-\ell(\gamma_1,\gamma_2)$.
We begin by fixing some notation. Let $\bar{\gamma}_1\in C({\mathbb B}^{k+1},{\mathbb R}^n)$ be any extension of $\gamma_1$, i.e. $\bar{\gamma}_1|_{\partial{\mathbb B}^{k+1}}=\gamma_1$. Let $A={\mathbb B}^{k+1}\times{\mathbb S}^l$ be a full torus, embedded smoothly in ${\mathbb R}^n$ in a way that the orientation of the boundary of embedded $A$ by the outward normal vector is consistent with the orientation of ${\mathbb S}^k\times{\mathbb S}^l$, and define $F:A\to{\mathbb R}^n$, $F(x,y)={\bar{\gamma}}_1(x)-\gamma_2(y)$.
In the proof, we shall need a simple lemma, connecting the linking number with the degree of the non-normalized Gauss map $F$.
\[PL1\] With the above notation, $$\ell(\gamma_1,\gamma_2)=\deg(F,A,0).$$
The claim would follow from Proposition \[propo46\], if $F$ mapped $\partial A$ to ${\mathbb S}^{n-1}$, not to ${\mathbb R}^n$ (because $F=\gamma_1-\gamma_2$ on $\partial A$). Note, however, that $F(\partial A)\subset {\mathbb R}^n\setminus\{0\}$. Set $d=\dist (\gamma_1({\mathbb S}^k),\gamma_2({\mathbb S}^l))=\dist (F(\partial A),0)$, and take $\phi:[0,\infty)\to[0,\infty)$ to be a smooth, positive function such that $$\phi(s)=\begin{cases}1&\text{ for }s\leq d/2,\\
s &\text{ for }s\geq d.
\end{cases}$$ Then $$h_t(z)=\frac{z}{\phi(t|z|)}, \quad t\in[0,1],$$ is a homotopy connecting ${\mathrm{Id}}:{\mathbb R}^n\to {\mathbb R}^n$ with a mapping which is identity on ${\mathbb B}^n(0,d/2)$ and a projection $z\mapsto z/|z|$ outside ${\mathbb B}^n(0,d)$. Obviously, $0\not\in (h_t\circ F)(\partial A)$ for any $t\in [0,1]$, thus $$\deg(F,A,0)=\deg(h_0\circ F, A, 0)=\deg(h_1\circ F, A, 0).$$ However, $$(h_1\circ F)|_{\partial A}(x,y)=\frac{\gamma_1(x)-\gamma_2(y)}{|\gamma_1(x)-\gamma_2(y)|}$$ is the Gauss map whose degree, by definition, equals $\ell(\gamma_1,\gamma_2)$. Thus Proposition \[propo46\] yields $$\ell(\gamma_1,\gamma_2)=\deg(h_1\circ F)|_{\partial A}=
\deg(h_1\circ F,A,0)=\deg (F,A,0).$$
Assume now $h$ is sense-preserving (the case of $h$ sense-reversing is treated in the same way).
Let $G_t:A\times[0,1]\to{\mathbb R}^n$, $$G_t(x,y)=h(t{\bar{\gamma}}_1(x))-h(\gamma_2(y)-(1-t){\bar{\gamma}}_1(x)).$$ Then $G_1(x,y)=(h\circ{\bar{\gamma}}_1)(x)-(h\circ\gamma_2)(y)$, thus by Lemma \[PL1\] applied to $h\circ\gamma_1$ and $h\circ\gamma_2$ in place of $\gamma_1$ and $\gamma_2$, $$\ell(h\circ\gamma_1,h\circ\gamma_2)=\deg(G_1,A,0).$$ We have $$G_0(x,y)=h(0)-h(\gamma_2(y)-{\bar{\gamma}}_1(x))=h(0)-h(-F(x,y))=h(0)+(-{\mathrm{Id}})\circ h\circ(-{\mathrm{Id}})\circ F(x,y).$$ Note also that $G_t(x,y)=0$ if and only if ${\bar{\gamma}}_1(x)=\gamma_2(y)$, which is not possible if $(x,y)\in\partial A$, thus $0\not\in G_t(\partial A)$ for any $t\in[0,1]$, and Remark \[R2\] yields $$\begin{split}
\ell(h\circ\gamma_1,h\circ\gamma_2)&=\deg(G_1,A,0)=\deg(G_0,A,0)\\
&=\deg(h(0)+(-{\mathrm{Id}})\circ h\circ(-{\mathrm{Id}})\circ F, A,0)\\
&=\deg(F,A,0)=\ell(\gamma_1,\gamma_2),
\end{split}$$ because $z\mapsto h(0)+(-{\mathrm{Id}})\circ h\circ(-{\mathrm{Id}})(z)$ is a sense-preserving homeomorphism of ${\mathbb R}^n$ which maps $0$ to $0$ and we can apply Proposition \[propo31\].
Proofs of Theorems \[T1\] and \[T2\] {#Proof1and2}
====================================
For $n=1$, the claim of Theorem \[T1\] is obvious: a sense-preserving homeomorphism of an open subset of a real line is an increasing function, thus it is differentiable a.e. and its derivative is non-negative. In dimension $n=2$, every $W^{1,1}$ homeomorphism is again a.e. differentiable (see [@Menchoff; @GehringL]) and its weak Jacobian coincides with its classical one a.e. The sign of the latter reflects whether the homeomorphism preserves or reverses the local orientation, and thus for a sense preserving homeomorphism $f$ we have $J_f\geq 0$ a.e.
Assume $n\geq 3$. To simplify the notation we shall write $\nu=n-1-[n/2]$. We will argue by contradiction: assume that the set $\{J_f<0\}$ has positive measure.
Pick a $p$-good point $x_o\in \{J_f<0\}$ for $f$ (in the sense of Definition \[D12\]) and consider the blow-up $f_r$ of $f$ at $x_o$: $$f_r(x)=\frac{f(x_o+rx)-f(x_o)}{r}.$$ According to Lemma \[T4\], $f_r\to f_0$ in $W^{1,p}({\mathbb B}^n,{\mathbb R}^n)$ as $r\to 0$, where $f_0(x)=Df(x_o)x$. Note that $f_0$ is a linear, orientation reversing isomorphism.
We pick in ${\mathbb B}^n$ two solid tori, i.e. smooth embeddings $\iota_1:{\mathbb B}^{[n/2]+1}\times{\mathbb S}^\nu\to{\mathbb B}^n$ and $\iota_2:{\mathbb B}^{\nu+1}\times{\mathbb S}^{[n/2]}\to{\mathbb B}^n$, such that for any $x\in {\mathbb B}^{[n/2]+1}$ and $y\in {\mathbb B}^{\nu+1}$ the embedded spheres ${\mathbb S}_x^{\nu}=\iota_1(\{x\}\times {\mathbb S}^{\nu})$ and ${\mathbb S}_y^{[n/2]}=\iota_2(\{y\}\times {\mathbb S}^{[n/2]})$ are linked, with linking number $\ell({\mathbb S}_x^{\nu},{\mathbb S}_y^{[n/2]})=+1$ (see the next section for a particular construction).
By Lemma \[T10\], we can choose a sequence $r_j\searrow 0$ and particular $x\in {\mathbb B}^{[n/2]+1}$ and $y\in {\mathbb B}^{\nu+1}$ such that $f_j=f_{r_j}$ converge to $f_0$ in $W^{1,p}$ on ${\mathbb S}_x^{\nu}\cup{\mathbb S}_y^{[n/2]}$. If $n>3$, then $p>[n/2]\geq \nu$, and by the Sobolev-Morrey embedding theorem $f_j$ converge to $f_0$ uniformly on ${\mathbb S}_x^{\nu}\cup{\mathbb S}_y^{[n/2]}$. If $n=3$ and $p=1$, the $W^{1,1}$ convergence of $f_k$ on the sum of circles ${\mathbb S}_x^{\nu} \cup {\mathbb S}_y^{[n/2]}$ implies uniform convergence as well.
Since $f_j$ converge to $f_0$ uniformly on ${\mathbb S}_x^{\nu}\cup{\mathbb S}_y^{[n/2]}$, $f_j$ are homotopic to $f_0$ for $j$ sufficiently large, and the image of each sphere in that homotopy does not intersect the image of the other (the image of ${\mathbb S}_x^{\nu}$ in that homotopy stays in a small tubular neighborhood of $f_0({\mathbb S}_x^{\nu})$, and, similarly, the image of ${\mathbb S}_y^{[n/2]}$ in that homotopy stays near $f_0({\mathbb S}_y^{[n/2]})$). Thus $$\ell(f_j({\mathbb S}_x^{\nu}),f_j({\mathbb S}_y^{[n/2]}))=\ell(f_0({\mathbb S}_x^{\nu}),f_0({\mathbb S}_y^{[n/2]}))=-\ell({\mathbb S}_x^{\nu},{\mathbb S}_y^{[n/2]})=-1,$$ since $f_0$ is a linear, orientation reversing homeomorphism.
However, each $f_j$, as a translation and rescaling of an orientation preserving homeomorphism $f$, is again an orientation preserving homeomorphism, thus Proposition \[Propo inv\] yields $$\ell(f_j({\mathbb S}_x^{\nu}),f_j({\mathbb S}_y^{[n/2]}))=\ell({\mathbb S}_x^{\nu},{\mathbb S}_y^{[n/2]})=+1,$$ which gives the desired contradiction.
Note that the proofs of Theorems \[T1\] and \[T2\] required only a few results from Section \[Prelim\], namely Lemma \[T4\], Lemma \[T10\] and Proposition \[propo46\] (i.e. the whole Section \[Sec2.5\]). However, the proof of Theorem \[main\] will require the whole content of Section \[Prelim\], that is, in addition to results needed for the proofs of Theorems \[T1\] and \[T2\], we will need Lemma \[T9\] and Proposition \[L27\].
Proof of Theorem \[main\]. {#Proofof5}
==========================
Throughout the proof, we shall assume that the assumption a) holds, i.e. $f$ maps almost all $[n/2]$-dimensional spheres into sets of $[n/2]+1$ dimensional Hausdorff measure zero. The case when b) holds is treated in the same way. We simply need to exchange $f$ and $f^{-1}$ in the proof below.
We argue by contradiction: assume that the set $\{J_f<0\}$ has positive measure.
To simplify the notation we shall write $\nu=n-1-[n/2]$. According to a local version of Lemma \[T9\] for homeomorphisms on domains instead of ${\mathbb R}^n$, we can find $x_o$ and a linear transformation $A\in GL(n)$ with $\det A>0$ such that the sense preserving homeomorphism $g=A\circ f$ satisfies $$\lim_{r\to 0^+}\Vert g_r-\mathcal{R}\Vert_{W^{1,[n/2]}({\mathbb B}^n,{\mathbb R}^n)}=
\lim_{r\to 0^+}\Vert (g_r)^{-1}-\mathcal{R}\Vert_{W^{1,\nu}({\mathbb B}^n,{\mathbb R}^n)}=0.$$ Let ${\mathbb B}_1={\mkern 1.7mu\overline{\mkern-1.7mu{\mathbb B}\mkern-1.5mu}\mkern 1.5mu}^{[n/2]+1}$, ${\mathbb B}_2={\mkern 1.7mu\overline{\mkern-1.7mu{\mathbb B}\mkern-1.5mu}\mkern 1.5mu}^{\nu+1}$ be closed unit balls. Note that the manifolds ${\mathbb B}_1\times{\mathbb S}^\nu$ and ${\mathbb B}_2\times{\mathbb S}^{[n/2]}$ have dimension $n$. Let $\iota_1:{\mathbb B}_1\times{\mathbb S}^\nu\hookrightarrow{\mathbb B}^n$ and $\iota_2:{\mathbb B}_2\times{\mathbb S}^{[n/2]}\hookrightarrow{\mathbb B}^n$ be smooth embeddings, smooth up to the boundary. According to Lemma \[T10\] applied to the family $g_r\circ\iota_2 $ and then for the second time to the family $(g_r)^{-1}\circ\iota_1$, we can find a sequence $r_k\searrow 0$ such that the mappings $g_k:=g_{r_k}$ satisfy:
There are compact sets $K_1\subset {\mathbb B}_1$ and $K_2\subset {\mathbb B}_2$ of positive measure such that $$\label{eq16}
\lim_{k\to\infty}
\sup_{x\in K_1}\Vert (g_k)^{-1}\circ\iota_1-\mathcal{R}\circ\iota_1\Vert_{W^{1,\nu}(\{x\}\times{\mathbb S}^\nu)}=0$$ and $$\label{eq24}
\lim_{k\to\infty}
\sup_{y\in K_2}
\Vert g_k\circ\iota_2-\mathcal{R}\circ\iota_2\Vert_{W^{1,[n/2]}(\{y\}\times{\mathbb S}^{[n/2]})}
=0.$$ We shall define the embeddings $\iota_1$ and $\iota_2$ explicitly, to make sure that the embedded spheres $\iota_1(\{x\}\times{\mathbb S}^\nu)$ and $\iota_2(\{y\}\times{\mathbb S}^{[n/2]})$ are linked with the linking number $1$.
- for $x\in {\mathbb B}_1\subset{\mathbb R}^{[n/2]+1}$, $\sigma\in {\mathbb S}^\nu\subset{\mathbb R}^{\nu+1}$, we set $$\iota_1(x,\sigma)=(\frac{5+x_1}{10}\sigma_1,\ldots,\frac{5+x_1}{10}\sigma_\nu,\frac{5+x_1}{10}\sigma_{\nu+1}-\frac{1}{4},\frac{x_2}{10}\ldots,\frac{x_{[n/2]+1}}{10}),$$ thus the image of $\iota_1$ is the full torus with its core sphere lying in the hyperplane of the first $\nu+1$ coordinates,
- for $y\in {\mathbb B}_2\subset{\mathbb R}^{\nu+1}$ and $\rho\in{\mathbb S}^{[n/2]}\subset {\mathbb R}^{[n/2]+1}$, we set $$\iota_2(y,\rho)=
(\frac{y_1}{10},\ldots,\frac{y_\nu}{10},\frac{5+y_{\nu+1}}{10}\rho_1+\frac{1}{4},\frac{5+y_{\nu+1}}{10}\rho_2,\ldots,\frac{5+y_{\nu+1}}{10}\rho_{[n/2]+1}),$$ thus the image of $\iota_2$ is the full torus with its core sphere lying in the hyperplane of the last $[n/2]+1$ coordinates.
To simplify the notation, as in the previous section we shall write ${\mathbb S}^\nu_x=\iota_1(\{x\}\times{\mathbb S}^\nu)$, ${\mathbb S}^{[n/2]}_y=\iota_2(\{y\}\times{\mathbb S}^{[n/2]}))$.
Since the assumption a) holds, i.e., $f$ maps almost every sphere of dimension $[n/2]$ to a set of $([n/2]+1)$-Hausdorff measure zero, it easily follows that $g_k$, for every $k$, has the same property. We may thus assume, possibly shrinking $K_2$, that for every $y\in K_2$ and any $k$ the set $g_k({\mathbb S}_y^{[n/2]})$ has $([n/2]+1)$-Hausdorff measure zero. (Similarly, if b) holds, i.e., $f^{-1}$ maps almost every sphere of dimension $\nu$ to a set of $(\nu+1)$-Hausdorff measure zero, we can assume that for every $x\in K_1$ and any $k$ the set $g^{-1}_k({\mathbb S}_x^{\nu})$ has $(\nu+1)$-Hausdorff measure zero.)
The images of $\iota_1$ and $\iota_2$ are two full, disjoint, linked tori, and for each $x\in {\mathbb B}_1$ and $y\in {\mathbb B}_2$, ${\mathbb S}^\nu_x$ and ${\mathbb S}^{[n/2]}_y$ are two linked spheres. We choose orientations of the spheres ${\mathbb S}^\nu$ and ${\mathbb S}^{[n/2]}$ so that the linking number equals $$\ell({\mathbb S}^\nu_x,{\mathbb S}^{[n/2]}_y):=\ell\big(\iota_1|_{\{x\}\times{\mathbb S}^\nu},\iota_2|_{\{y\}\times{\mathbb S}^{[n/2]}}\big)=+1.$$ The first equality mans that for all $x\in{\mathbb B}_1$ and all $y\in{\mathbb B}_2$ we equip ${\mathbb S}_x^\nu$ and ${\mathbb S}_y^{[n/2]}$ with orientations so that the diffeomorphisms $$\iota_1|_{\{x\}\times{\mathbb S}^\nu}:\{x\}\times{\mathbb S}^\nu \to {\mathbb S}_x^\nu
\quad
\text{and}
\quad
\iota_2|_{\{y\}\times{\mathbb S}^{[n/2]}}:\{y\}\times{\mathbb S}^{[n/2]}\to {\mathbb S}_y^{[n/2]}$$ are orientation preserving.
The linear transformation ${\mathcal R}$ is the reflection in the last coordinate. Since the spheres ${\mathbb S}_y^{[n/2]}$ are centered at a point lying in ${\mathbb R}^{\nu+1}\times\{0\}$, the reflection ${\mathcal R}$ preserves the center and hence ${\mathcal R}({\mathbb S}^{[n/2]}_y)={\mathbb S}^{[n/2]}_y$. However, the reflection ${\mathcal R}$ changes the orientation of the sphere ${\mathbb S}^{[n/2]}_y$. More precisely, the mapping $$\label{eq21}
{\mathcal R}:{\mathbb S}_y^{[n/2]}\to{\mathbb S}_y^{[n/2]}
\quad
\text{has degree $-1$.}$$ By ${\mathcal R}({\mathbb S}_y^{[n/2]})$ we denote the sphere ${\mathbb S}^{[n/2]}_y$ with the opposite orientation, so the mapping ${\mathcal R}:{\mathbb S}_y^{[n/2]}\to{\mathcal R}({\mathbb S}_y^{[n/2]})$ is orientation preserving. In particular, implies that $$\label{eq22}
\ell({\mathbb S}_x^\nu,{\mathcal R}({\mathbb S}_y^{[n/2]}))=-1.$$
On the other hand, the spheres ${\mathbb S}^\nu_x$ are centered at points lying in $\{0\}\times{\mathbb R}^{[n/2]+1}$ and the last coordinate of the center is $x_{[n/2]+1}/10$, where $x=(x_1,\ldots,x_{[n/2]+1})$, so ${\mathcal R}({\mathbb S}^\nu_x)={\mathbb S}^\nu_{\tilde{x}}$ where $\tilde{x}=(x_1,\ldots,x_{[n/2]},-x_{[n/2]+1})$. Also, the orientation of ${\mathcal R}({\mathbb S}^\nu_x)$ is the same as that of ${\mathbb S}^\nu_{\tilde{x}}$. More precisely, the mapping ${\mathcal R}:{\mathbb S}_x^\nu\to {\mathbb S}_{\tilde{x}}^\nu$ is orientation preserving, so ${\mathcal R}({\mathbb S}_x^\nu)$ denotes the sphere ${\mathbb S}_{\tilde{x}}^\nu$ with the original orientation. In particular, $$\label{eq23}
\ell({\mathcal R}({\mathbb S}_x^\nu),S_y^{[n/2]})=\ell({\mathbb S}_{\tilde{x}}^\nu,{\mathbb S}_y^{[n/2]})=+1.$$
By Proposition \[L27\] and Remark \[rem1\], for each $k\in{\mathbb N}$ we may define a continuous map $$H_{1,k}:K_1\times{\mathbb S}^\nu\times [0,1]\to{\mathbb R}^n$$ such that for each $x\in K_1$, $H_{1,k}(x,\cdot,\cdot)$ is a homotopy between $$H_{1,k}(x,\cdot,1)=g^{-1}_k\circ\iota_1|_{\{ x\}\times{\mathbb S}^\nu}
\quad
\text{and}
\quad
H_{1,k}(x,\cdot,0)={\mathcal R}\circ\iota_1|_{\{ x\}\times{\mathbb S}^\nu}.$$ Moreover, Proposition \[L27\] along with yields $$\sup_{x\in K_1}{\mathcal H}^{\nu+1}(H_{1,k}(\{x\}\times{\mathbb S}^\nu\times [0,1))\to 0
\quad
\text{as $k\to\infty$.}$$ Therefore by taking a suitable subsequence of $g_k^{-1}$ (still denoted by $g_k^{-1}$) we may require that $$\label{eq17}
\sup_{x\in K_1}\mathcal{H}^{\nu+1}(H_{1,k}(\{x\}\times {\mathbb S}^\nu \times [0,1)))< \frac{1}{2^k},
\quad
\text{for all $k$.}$$ Likewise, we define a continuous map $$H_{2,k}:K_2\times{\mathbb S}^{[n/2]}\times [0,1]\to{\mathbb R}^n$$ such that for all $y\in K_2$, $H_{2,k}(y,\cdot,\cdot)$ is a homotopy between $$H_{2,k}(y,\cdot,1)=g_k\circ\iota_2|_{\{ y\}\times{\mathbb S}^{[n/2]}}
\quad
\text{and}
\quad
H_{2,k}(y,\cdot,0)={\mathcal R}\circ\iota_2|_{\{ y\}\times{\mathbb S}^{[n/2]}}.$$ Again, Proposition \[L27\] along with yields $$\sup_{y\in K_2}{\mathcal H}^{[n/2]+1}(H_{2,k}(\{y\}\times{\mathbb S}^{[n/2]}\times [0,1))\to 0
\quad
\text{as $k\to\infty$.}$$ Since for every $y\in K_2$ we have $${\mathcal H}^{[n/2]+1}(H_{2,k}(\{y\}\times{\mathbb S}^{[n/2]}\times\{ 1\})=
{\mathcal H}^{[n/2]+1}(g_k({\mathbb S}_y^{[n/2]}))=0,$$ by taking a suitable subsequence we may require that $$\label{eq18}
\sup_{y\in K_2}{\mathcal H}^{[n/2]+1}(H_{2,k}(\{ y\}\times{\mathbb S}^{[n/2]}\times [0,1]))<\frac{1}{2^k}
\quad
\text{for all $k$.}$$ Note that this is a stronger condition than in the sense that now we have the estimate for the whole interval $[0,1]$, while in we only have the estimate for the interval $[0,1)$.
We prove the following:
\[lem:no inters\] For every $y\in K_2$, for almost every $x\in K_1$ $$\label{nointer1}
\exists_{m\in {\mathbb N}}\ \forall_{k\geq m}
\quad
{\mathbb S}_x^\nu\cap H_{2,k}(\{y\}\times{\mathbb S}^{[n/2]}\times[0,1])=\varnothing.$$ Also, for every $x\in K_1$, for almost every $y\in K_2$ $$\label{nointer2}
\exists_{m\in {\mathbb N}}\ \forall_{k\geq m}
\quad
{\mathbb S}_y^{[n/2]}\cap H_{1,k}(\{x\}\times{\mathbb S}^{\nu}\times[0,1))=\varnothing.$$
We will only prove since the proof of follows from the same reasoning. Fix $y\in K_2$. We want to show that for almost all $x\in K_1$, is satisfied.
We define a projection of the embedded full torus $\iota_1({\mathbb B}_1\times{\mathbb S}^\nu)$ onto ${\mathbb B}_1$ by $$\pi:\iota_1({\mathbb B}_1\times{\mathbb S}^\nu)\to {\mathbb B}_1,
\qquad
\pi(\iota(x,\sigma))=x.$$ It is easy to see that $\pi$ is $10$-Lipschitz and hence it increases the Hausdorff measure ${\mathcal H}^{[n/2]+1}$ of a set at most by a constant factor $C(n)=10^{[n/2]+1}$. Let $$T_k(y)=H_{2,k}(\{ y\}\times{\mathbb S}^{[n/2]}\times [0,1])\cap\iota_1({\mathbb B}_1\times{\mathbb S}^\nu)$$ be a part of the image of the homotopy $H_{2,k}(y,\cdot,\cdot)$ that is contained in the domain of the projection $\pi$. Estimate and the fact that $\pi$ increases the Hausdorff measure by at most $C(n)$ imply $${\mathcal H}^{[n/2]+1}\left(\bigcup_{k=m}^\infty \pi(T_k(y))\right)<
\sum_{k=m}^\infty C(n) 2^{-k}=C(n)2^{-m+1}\to 0
\quad
\text{as $m\to\infty$,}$$ so $${\mathcal H}^{[n/2]+1}\left(\bigcap_{m=1}^\infty\bigcup_{k=m}^\infty \pi(T_k(y))\right)=0.$$ To complete the proof of it suffices to show that is satisfied by all $$\label{eq19}
x\in K_1\setminus \bigcap_{m=1}^\infty\bigcup_{k=m}^\infty \pi(T_k(y))=
\bigcup_{m=1}^\infty\bigcap_{k=m}^\infty (K_1\setminus \pi(T_k(y))).$$ Note that if $x\in K_1\setminus \pi(T_k(y))$, then $$\pi({\mathbb S}_x^\nu\cap T_k(y))\subset \pi({\mathbb S}_x^\nu)\cap \pi(T_k(y))=
\{ x\}\cap \pi(T_k(y))=\varnothing,$$ so $$\label{eq20}
{\mathbb S}_x^\nu\cap H_{2,k}(\{ y\}\times{\mathbb S}^{[n/2]}\times [0,1])=
{\mathbb S}_x^\nu\cap T_k(y)=\varnothing.$$ Therefore if $x$ belongs to the set , then $$\exists_{m\in {\mathbb N}}\ \forall_{k\geq m} \quad
x\in K_1\setminus \pi(T_k(y)).$$ Since the condition $x\in K_1\setminus \pi(T_k(y))$ implies , claim follows.
To finish the proof of Theorem \[main\], we want to choose $x\in K_1$ and $y\in K_2$ in such a way that
- there exists $m$ such that $${\mathbb S}_x^\nu\cap \bigcup_{k=m}^\infty
H_{2,k}(\{y\}\times{\mathbb S}^{[n/2]}\times [0,1])=\varnothing,$$ i.e., the sphere ${\mathbb S}_x^\nu$ avoids, for all sufficiently large $k$, the image of the homotopy ${H_{2,k}(y,\cdot,\cdot)\circ\iota_2^{-1}}$ joining $g_k|_{{\mathbb S}_y^{[n/2]}}$ with ${\mathcal R}|_{{\mathbb S}_y^{[n/2]}}$,\
and simultaneously
- there exists $m$ such that $${\mathbb S}_y^{[n/2]}\cap \bigcup_{k=m}^\infty
H_{1,k}(\{ x\}\times{\mathbb S}^\nu\times [0,1))=\varnothing,$$ i.e., the sphere ${\mathbb S}_y^{[n/2]}$ avoid, for all sufficiently large $k$, the image of the homotopy $H_{1,k}(x,\cdot,\cdot)\circ\iota_1^{-1}$ joining $g^{-1}_k|_{{\mathbb S}_x^\nu}$ with ${\mathcal R}|_{{\mathbb S}_x^\nu}$, except possibly at the endpoint: we do not rule out yet that ${\mathbb S}_y^{[n/2]}\cap g^{-1}_k({\mathbb S}_x^\nu)\neq \varnothing$.
Assume we have chosen $x$ and $y$ satisfying conditions i) and ii) above. Then, for all sufficiently large $k$, ${\mathbb S}_x^\nu\cap g_k({\mathbb S}_y^{[n/2]})=\varnothing$, and since $g_k^{-1}$ is a homeomorphism, this immediately implies that ${\mathbb S}_y^{[n/2]}\cap g^{-1}_k({\mathbb S}_x^\nu)= \varnothing$. Therefore, for all sufficiently large $k$, the sphere ${\mathbb S}_y^{[n/2]}$ avoids the image of the whole homotopy joining $g^{-1}_k|_{{\mathbb S}_x^\nu}$ with ${\mathcal R}|_{{\mathbb S}_x^\nu}$, including the endpoint:
- there exists $m$ such that $${\mathbb S}_y^{[n/2]}\cap \bigcup_{k=m}^\infty
H_{1,k}(\{ x\}\times{\mathbb S}^\nu\times [0,1])=\varnothing.$$
Denote by $A_1\subset K_1\times K_2$ the set of all $(x,y)$ satisfying the condition i) above. By Lemma \[lem:no inters\], , for every $y_o\in K_2$ the section $A_1\cap \{(x,y_o)~~:~~a\in K_1\}$ is of full measure, and thus, by Fubini’s theorem, $A_1$ is of full measure in $K_1\times K_2$, provided that $A_1$ is a measurable set. Similarly, the set $A_2\subset K_1\times K_2$ of these $(x,y)$, which satisfy the condition ii), if measurable, is of full measure in $K_1\times K_2$. We shall leave the issue of measurability of $A_1$ and $A_2$ and address it at the end of the proof. Since $A_1$ and $A_2$ are of full measure, their intersection is not empty and we can find $x$ and $y$ simultaneously satisfying the conditions i) and ii) and hence conditions i) and ii’).
In particular, there is $x\in K_1$, $y\in K_2$ and $k\in{\mathbb N}$ such that $${\mathbb S}_x^\nu\cap
H_{2,k}(\{y\}\times{\mathbb S}^{[n/2]}\times [0,1])=
{\mathbb S}_y^{[n/2]}\cap
H_{1,k}(\{ x\}\times{\mathbb S}^\nu\times [0,1])=
\varnothing.$$ We fix such a point $(x,y)\in K_1\times K_2$ and we look at linking numbers of spheres and their images in $g_k$, $g_k^{-1}$ and ${\mathcal R}$.
The mappings $g_k^{-1}|_{{\mathbb S}_x^\nu}$ and ${\mathcal R}|_{{\mathbb S}_x^\nu}$ are homotopic (with $H_{1,k}(x,\cdot,\cdot)\circ\iota_1^{-1}$ providing the homotopy), and the image of the homotopy does not intersect ${\mathbb S}_y^{[n/2]}$. This and yield $$+1=\ell({\mathbb S}_{\tilde{x}}^\nu,{\mathbb S}_y^{[n/2]})=\ell({\mathcal R}({\mathbb S}_x^\nu),{\mathbb S}_y^{[n/2]})=
\ell(g_k^{-1}({\mathbb S}_x^\nu),{\mathbb S}_y^{[n/2]})$$ and, since $g_k$ is a sense preserving homeomorphism, $$+1=\ell(g_k^{-1}({\mathbb S}_x^\nu),{\mathbb S}_y^{[n/2]})=\ell({\mathbb S}_x^\nu,g_k({\mathbb S}_y^{[n/2]})).$$ Next, using the homotopy between ${\mathcal R}|_{{\mathbb S}_y^{[n/2]}}$ and $g_k|_{{\mathbb S}_y^{[n/2]}}$, given by $H_{2,k}(y,\cdot,\cdot)\circ\iota_2^{-1}$, we have $$+1=\ell({\mathbb S}_x^\nu,g_k({\mathbb S}_y^{[n/2]}))=\ell({\mathbb S}_x^\nu,{\mathcal R}({\mathbb S}_y^{[n/2]}))=-1,$$ where the last equality follows from . This gives the desired contradiction and finishes the proof, except for the set aside problem of measurability of the sets $A_1$ and $A_2$.
Since the proof of measurability of both sets follows exactly the same scheme, we shall prove only that $A_1$ is measurable.
If we write $$W_k=\{(x,y)\in K_1\times K_2~~:~~{\mathbb S}_x^\nu \cap H_{2,k}(\{ y\}\times{\mathbb S}^{[n/2]}\times [0,1])=\varnothing\},$$ then $A_1=\bigcup_{m=1}^\infty \bigcap_{k=m}^\infty W_k$, thus to prove measurability of $A_1$, it suffices to prove it for $W_k$.
Let $$F_k=(\iota_1, H_{2,k})\colon K_1\times {\mathbb S}^\nu\times K_2\times {\mathbb S}^{[n/2]}\times [0,1]\to {\mathbb R}^n\times {\mathbb R}^n$$ and denote by $\Delta=\{(x,x)~~:~~x\in{\mathbb R}^n\}$ the diagonal in ${\mathbb R}^n\times {\mathbb R}^n$. Then, since $F_k$ is continuous, $K_1\times {\mathbb S}^\nu\times K_2\times {\mathbb S}^{[n/2]}\times [0,1]$ is compact and $\Delta$ is closed, the set $F_k^{-1}(\Delta)\subset {\mathbb R}^{2n+1}$ is compact. Let now $$\Pi:K_1\times {\mathbb S}^\nu\times K_2\times {\mathbb S}^{[n/2]}\times [0,1]\to K_1\times K_2$$ be the projection on the first and third factors. The set $\Pi(F_k^{-1}(\Delta))$ is a compact subset of $K_1\times K_2$. We have $$\begin{split}
(x,y)&\in \Pi(F_k^{-1}(\Delta))\\
&\Leftrightarrow \text{ there exist }\sigma\in {\mathbb S}^\nu,\,\rho\in {\mathbb S}^{[n/2]},\, t\in [0,1] \text{ with } (x,\sigma,y,\rho,t)\in F_k^{-1}(\Delta)\\
& \Leftrightarrow \exists_{\sigma,\rho,t} \quad F_k(x,\sigma,y,\rho,t)\in\Delta\\
& \Leftrightarrow \exists_{\sigma,\rho,t}\quad \iota_1(x,\sigma)=H_{2,k}(y,\rho,t)\\
&\Leftrightarrow {\mathbb S}_x^\nu\cap H_{2,k}(\{y\}\times{\mathbb S}^{[n/2]}\times [0,1])\neq \varnothing,
\end{split}$$ which shows that $W_k=(K_1\times K_2)\setminus \Pi(F_k^{-1}(\Delta))$, and that $W_k$ is an open (and thus measurable) subset of $K_1\times K_2$. This concludes the proof of measurability of $A_1$ and the proof of Theorem \[main\].
Under the assumptions of Corollaries \[T14\] and \[T15\], the proof simplifies greatly. Recall that in these corollaries we assume $n=2m$, thus $\nu=n-[n/2]-1=m-1$. If we assume $f^{-1}\in W^{1,m-1+\varepsilon}$, then $g_k^{-1}\to {\mathcal R}$ in $W^{1,m-1+\varepsilon}$, and by the Morrey-Sobolev imbedding, on almost every sphere ${\mathbb S}_x^{\nu}$ this convergence is uniform (the same conclusion holds if we assume $Df^{-1}\in L^{m-1,1}_{loc}$). We can set the homotopy between ${\mathcal R}|_{{\mathbb S}_x^{\nu}}$ and $g_k^{-1}|_{{\mathbb S}_x^{\nu}}$ to be $H_1,k(x,\sigma,t)=tg_k^{-1}(\iota_1(x,\sigma)+{(1-t){\mathcal R}(\iota_1(x,\sigma))}$; then for a.e. $x\in {\mathbb B}_1$ and sufficiently large $k$ the whole image of the homotopy $H_1(x,\cdot,\cdot)$ lies in the torus $\iota_1({\mathbb B}_1\times {\mathbb S}^{[n/2]})$, thus for any $y\in{\mathbb B}_2$ this image does not intersect ${\mathbb S}_y^{[n/2]}$: $$\label{rem:noint1}
\exists_l\ \forall_{k>l}\quad {\mathbb S}_y^{[n/2]}\cap H_1(\{x\}\times {\mathbb S}^\nu\times [0,1])=\varnothing.$$ Fix any $y\in {\mathbb B}_2$. Denoting by $\tilde{Y}_k(y)$ the projection of $\tilde{T}_k(y)$ onto the cross section $\iota_1({\mathbb B}_1\times \{\sigma_o\})$ (in analogy to $Y_k(y)$), we see that the $(m+1)$-dimensional Hausdorff measure of $\tilde{Y}_k(y)$ tends to zero. We can thus find $x\in {\mathbb B}_1$ such that holds and $\iota_1(x,\sigma_o)\cap \tilde{Y}_k(y)=\varnothing$ for some $k>l$. Then ${\mathbb S}_x^\nu$ does not intersect the image of the homotopy joining ${\mathcal R}|_{{\mathbb S}_y^{[n/2]}}$ with $g_k|_{{\mathbb S}_y^{[n/2]}}$, except possibly at the endpoint – we have not ruled out that ${\mathbb S}_x^\nu\cap g_k({\mathbb S}_y^{[n/2]})\neq\varnothing$. We have thus found $x$ and $y$ satisfying conditions i) and ii) in the proof of Theorem \[main\] (although with $x$ and $y$ exchanged) and we may conclude the proof as it is done there.
[99]{}
An approximation lemma for $W^{1,p}$ functions. In: Material instabilities in continuum mechanics (Edinburgh, 1985–1986), pp. 1–5, Oxford Sci. Publ., Oxford Univ. Press, New York, 1988.
Vol. II. Scripta Series in Mathematics. Edited by Mitchell H. Taibleson. V. H. Winston & Sons, Washington, D.C.; Halsted Press \[John Wiley & Sons\], New York-Toronto, Ont.-London, 1979.
Pointwise inequalities for Sobolev functions and some applications. [*Studia Math.*]{} 106 (1993), 77–92.
Approximation of $W^{1,p}$ Sobolev homeomorphism by diffeomorphisms and the signs of the Jacobian. [*Adv. Math.*]{} 331 (2018), 748–829.
Homeomorphism with zero Jacobian: sharp integrability of the derivative. [*J. Math. Anal. Appl.*]{} 373 (2011), 161–174.
Homeomorphisms in the Sobolev space $W^{1,n-1}$. [*J. Reine Angew. Math.*]{} 644 (2010), 221–235.
Local properties of solutions of elliptic partial differential equations. [*Studia Math.*]{} 20 (1961), 171–225.
Bi-Sobolev homeomorphism with zero Jacobian almost everywhere. [*Calc. Var. Partial Differential Equations*]{} 51 (2014), 139–170.
On the approximate differentiability of inverse maps. [*J. Fixed Point Theory Appl.*]{} 15 (2014), 473–499.
Revised edition. Textbooks in Mathematics. CRC Press, Boca Raton, FL, 2015.
Oxford Lecture Series in Mathematics and its Applications, 2. Oxford Science Publications. The Clarendon Press, Oxford University Press, New York, 1995.
The limit of $W^{1,1}$ homeomorphisms with finite distortion. [*Calc. Var. Partial Differential Equations*]{} 33 (2008), 377–390.
in: Werke, Band V, Georg Olms Verlag, Hildesheim, 1973 (reprint of the 1967 original).
On the total differentiability of functions of a complex variable. [*Ann. Acad. Sci. Fenn. Ser. A I No.*]{} 272 (1959) 1–9.
Topological obstructions to continuity of Orlicz-Sobolev mappings of finite distortion. [*Ann Mat. Pura Appl.*]{} 198 (2019), 243–262.
Modulus of continuity of orientation preserving approximately differentiable homeomorphisms with a.e. negative Jacobian. [*Ann. Acad. Sci. Fenn. Math.*]{} 43 (2018), 147–170.
A measure and orientation preserving homeomorphism with approximate Jacobian equal $-1$ almost everywhere. [*Arch. Ration. Mech. Anal.*]{} 225 (2017), 65–88.
Sobolev spaces on an arbitrary metric space. [*Potential Anal.*]{} 5 (1996), 403–415.
Approximation in Sobolev spaces of nonlinear expressions involving the gradient. [*Ark. Mat.*]{} 40 (2002), 245–274.
Sobolev homeomorphism with zero Jacobian almost everywhere. [*J. Math. Pures Appl.*]{} 95 (2011), 444–458.
Lecture notes in mathematics, 2096. Springer, Cham, 2014.
Regularity of the inverse of a planar Sobolev homeomorphism. [*Arch. Ration. Mech. Anal.*]{} 180 (2006), 75–95.
Regularity of the inverse of a Sobolev homeomorphism in space. [*Proc. Roy. Soc. Edinburgh Sect. A*]{} 136 (2006), 1267–1285.
Jacobians of Sobolev homeomorphisms. [*Calc. Var. Partial Differential Equations*]{} 38 (2010), 233–242.
Bi-Sobolev mappings and elliptic equations in the plane. [*J. Math. Anal. Appl.*]{} 355 (2009), 22–32.
Sobolev homeomorphism that cannot be approximated by diffeomorphisms in $W^{1,1}$. [*Arch. Ration. Mech. Anal.*]{} 219 (2016), 183–202.
On functions with derivatives in a Lorentz space. [*Manuscripta Math.*]{} 100 (1999), 87–101.
An Interpretation of G. Whitehead’s Generalization of H. Hopf’s Invariant. [*Ann. of Math.*]{} 69 (1959) 345–365.
Graduate Studies in Mathematics, 105. American Mathematical Society, Providence, RI, 2009.
Lusin’s condition (N) and mappings of the class $W^{1,n}$. [*J. Reine Angew. Math.*]{} 458 (1995), 19–36.
Sur les différentielles totales des fonctions univalentes. [*Math. Ann.*]{} 105 (1931), 75–85.
Based on notes by David W. Weaver. Revised reprint of the 1965 original. Princeton Landmarks in Mathematics. Princeton University Press, Princeton, NJ, 1997.
Bi-Sobolev homeomorphisms $f$ with $Df$ and $Df^{-1}$ of low rank using laminates. [*Calc. Var. Partial Differential Equations*]{} 55 (2016), no. 6, Art. 135, 38 pp.
Regularity of the inverse of spatial mappings with finite distortion. [*Calc. Var. Partial Differential Equations*]{} 26 (2006), 331–341.
Graduate Studies in Mathematics, 108. American Mathematical Society, Providence, RI; Real Sociedad Matemática Española, Madrid, 2009.
On the bi-Sobolev planar homeomorphisms and their approximation. [*Nonlinear Anal.*]{} 154 (2017), 258–268.
Springer-Verlag, Berlin-Göttingen-Heidelberg, 1955.
Gauss’ linking number revisited. [*J. Knot Theory Ramifications*]{} Vol. 20, No. 10 (2011), 1325-–1343.
On embeddings between $BV$ and $W^{s,p}$. Lehrstuhl I für Mathematik, RWTH Aachen Preprint no. 6, 2000.
An increasing continuous singular function. [*Amer. Math. Monthly*]{} 85 (1978), 35–37.
(Reprint of 1983 edition.) Modern Birkhüser Classics. Birkhäuser/Springer Basel AG, Basel, 2010.
Hölder continuous Sobolev mappings and the Lusin N property. [*Illinois J. Math.*]{} 58 (2014), 585–591.
Graduate Texts in Mathematics 120, Springer Verlag 1989.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
\
Sorbonne Universités, UPMC Univesité Paris 6, UMR 7095, Institut d’Astrophysique de Paris, 98 bis Bd Arago, 75014 Paris, France\
CNRS, UMR 7095, Institut d’Astrophysique de Paris, 98 bis Bd Arago, 75014 Paris, France E-mail:
bibliography:
- 'master.bib'
title: 'Testing the strong equivalence principle with gravitational-wave observations of binary black holes'
---
Introduction
============
The existence of gravitational waves (GWs) is an unavoidable consequence of the hyperbolic structure of the field equations of General Relativity (GR). GWs are also a generic prediction of other (relativistic) gravitational theories extending GR, since their field equations must retain, at least in the spin-2 sector and in the infrared limit, the hyperbolic structure of the Einstein equations in order to provide a continuous limit to GR and to pass existing experimental tests [@Berti:2015itd].
Among the latter, the timing of the orbital period of binaries comprised of at least one pulsar (henceforth, “binary pulsars” for brevity’s sake) has long provided very convincing indirect evidence of the existence of GWs [@damour-taylor]. Indeed, GR predicts that GWs should carry energy and angular momentum away from binary systems with a specific rate given (at leading order) by the quadrupole formula, which in turn causes the binary to spiral in with shorter and shorter orbital revolution periods. This damping of the orbital period of binary pulsars has been measured (by timing the pulsar component of the binary) in a number of systems over a time span of several years or even decades, and matches the predictions of GR’s quadrupole formula to within the observational errors (i.e. to percent-level accuracy or better, depending on the system).
All of the observed binary pulsars have small orbital velocities compared to the speed of light, $v/c\lesssim 10^{-2}$, hence these systems only probe the mildly relativistic regime of the Einstein equations. In this sense binary pulsars are on par with solar-system experiments, which probe a similar range of velocities. Nevertheless, an important and often under-appreciated point is that since pulsars are self-gravitating (i.e. the gravitational potential inside the star is strong, $U\sim c^2$), these systems probe the mildly relativistic *but strong-field* regime of GR.
In 2015, the LIGO interferometers also detected GWs directly for the first time, by observing two signals from binary systems with (source-frame) masses respectively $m_1\approx 36 M_\odot $, $m_2\approx 29 M_\odot $, and $m_1\approx 14 M_\odot $, $m_2\approx 7.5 M_\odot $ [@TheLIGOScientific:2016pea]. These binaries were observed in their late inspiral and in their merger-ringdown phase – i.e. their maximum relative velocities are comparable with the speed of light (“highly relativistic regime”) –, and are believed to be comprised of two black holes (BHs), which are of course self-gravitating objects (“strong-field regime”). Indeed, the inspiral signal from these systems is indistinguishable from that of a binary of point masses down to very small separations. This fact, coupled with the measured values of the masses of the two bodies, excludes that these binaries could be comprised of two neutron stars (NSs), or a NS and a BH. Moreover, the fact that the merger-ringdown part of the signal is consistent with the predictions of GR for a BH binary [@TheLIGOScientific:2016pea] (and in particular the absence of any sign of matter mode excitations in the waveforms [@Yunes:2016jcc]) disfavor the hypothesis that these systems may be comprised of more exotic compact objects such as boson stars. In short, not only have the LIGO detections provided the first direct evidence of the existence of GWs and started probing, for the first time, gravitation in the strong-field highly relativistic regime, but they have also confirmed the existence of BHs.
The importance of these detections for understanding gravity and for testing extensions of GR cannot be overstated. One of the cornerstone predictions of GR is the equivalence principle, which implies that all bodies, irrespective of their nature and composition, should move along exactly the same trajectory in an external gravitational field. This “universality of free fall” no longer holds in theories that modify or extend GR. Indeed, the most natural and generic way to extend GR is to add extra fields to the gravitational sector [@Berti:2015itd]. These additional gravitons are typically coupled minimally to the matter fields at tree level, to avoid the appearance of spurious “fifth forces” in particle-physics experiments. Nevertheless, in general these extra gravitons will also couple non-minimally to the spin-2 metric field of GR. This non-minimal coupling appears for instance in all non-trivial theories in which the extra graviton is scalar (“scalar-tensor theories”), and in theories with spin-1 or spin-2 extra gravitons (where it follows simply from general covariance). As result, classical interactions between the extra gravitational fields and matter, even though absent at tree level, re-appear at higher perturbative orders, mediated by the metric perturbations. Therefore, fifth forces are expected to appear in regimes where the metric perturbations are large, i.e. in strong-gravity systems such as those involving NSs or BHs. This causes deviations from the universality of free fall, as these fifth forces will depend in general on the nature of the body (e.g. whether it is a NS or a BH) and its composition (e.g. the NS equation of state).
This phenomenon is usually referred to as “Nördtvedt effect”, or violation of the strong equivalence principle (where “strong” signifies that the universality of free fall is only broken for strongly gravitating objects) [@eardley; @Nordtvedt:1968qr], and has profound implications for GW emission from binaries in GR extensions. Indeed, these strong-field fifth forces are in general both conservative and dissipative, the dissipative ones being caused by the transfer of energy and angular momentum from “matter” (the binary) to the extra gravitons. This additional gravitational emission channel, absent in GR, causes the binary to shrink faster, an effect that would leave a characteristic imprint on the frequency evolution of GW signals, which would chirp faster to higher frequencies. The conservative interactions between the binary and the extra fields will also affect the gravitational waveforms, by influencing for instance the binary’s precession and thus the waveform modulation. The transfer of the binary’s energy and angular momentum to the non-GR gravitons will also cause the appearance of additional GW polarizations, in principle observable with a network of interferometers [@Hayama:2012au].
In this contribution, I will briefly review the physics behind these violations of the strong equivalence principle, focusing, as an instructive example, on a class of theories that imply no deviations away from the GR predictions in systems of two NSs, but which would produce gravitational fifth forces in systems involving at least a BH [@Barausse:2015wia; @Yagi:2015oca]. I will review the constraints that can be placed on the dissipative components of these extra forces based on the LIGO detections of BH binaries, and the bounds that will be obtained by the space-borne interferometer LISA [@Barausse:2016eii].
Strong equivalence principle violations and dipole emission
===========================================================
Let us consider, as an example, the case of scalar-tensor theories, in which gravity is described by the spin-2 metric field of GR and by a scalar field $\phi$. Let us also assume that the scalar field couples minimally to matter and that the action is shift-symmetric, i.e. invariant under shifts $\phi\to\phi+$ const. The most generic action that gives second-order field equations (thus avoiding Ostrogradsky ghosts) is given by the “Horndeski action” [@Horndeski:1974wa] (also known as “generalized Galileon action”) $$\begin{aligned}
\label{action}
&S=\frac{1}{16 \pi G} \int d^4x \sqrt{-g} \Big\{R+ K(X) -G_3(X)\Box\phi + G_{4}(X)R
+G_{4X}(X)\left[
\left(\Box\phi\right)^2-\left(\nabla_\mu\nabla_\nu\phi\right)^2
\right]\\\nonumber
&\!\!\!\!\!\!\!+G_5(X) G_{\mu\nu}\nabla^\mu\nabla^\nu\phi
-\frac{G_{5X}(X)}{6}\Bigl[
\left(\Box\phi\right)^3
-3\left(\Box\phi\right)\left(\nabla_\mu\nabla_\nu\phi\right)^2
+2\left(\nabla_\mu\nabla_\nu\phi\right)^3
\Bigr]+\chi \phi {\cal G}\Big\}+S_m(g_{\mu\nu},\psi)\,.\end{aligned}$$ Here, $\nabla$, $R$ and $G_{\mu\nu}$ are respectively the Levi-Civita connection, the Ricci scalar and the Einstein tensor; $K$, $G_3$, $G_4$, and $G_5$ are functions of $X\equiv-\nabla_\mu\phi \nabla^\mu\phi/2$ (which we assume to be analytic about $X=0$, with $K(X)=X+{\cal O}(X)^2$ and $G_3, G_4, G_5={\cal O}(X)$, c.f. [@Barausse:2015wia]); $\chi=$ const; ${\cal G}\equiv R^{\mu\nu\lambda\kappa}R_{\mu\nu\lambda\kappa}-4 R^{\mu\nu}R_{\mu\nu}+R^2$ is the Gauss-Bonnet scalar; $S_m$ is the matter action; $\psi$ collectively represents the matter fields; and we have introduced the shortcuts $G_{iX}\equiv \partial G_i/\partial X$, $\Box \equiv \nabla^\mu\nabla_\mu$, $\left(\nabla_\mu\nabla_\nu\phi\right)^2 \equiv \nabla_\mu\nabla^\nu \phi\nabla_\nu\nabla^\mu\phi$, $\left(\nabla_\mu\nabla_\nu\phi\right)^3 \equiv \nabla_\mu\nabla^\rho \phi\nabla_\rho\nabla^\nu\phi\nabla_\nu\nabla^\mu\phi$. The term $\chi \phi {\cal G}$ is shift-symmetric because ${\cal G}$ is (locally) a total divergence, and may also be obtained by setting $G_5\propto\ln\vert X\vert$ [@Kobayashi:2011nu].
To study binaries of compact objects in GR, one can (effectively) describe the two bodies by point particles. This simple approach is also possible in theories that extend GR, but care should be used because the masses of the point particles can no longer be assumed to be constant. Indeed, because of the aforementioned “Nördtvedt effect”, the point-particle masses should be allowed to depend on the local value of the non-GR gravitational fields that appear in the theory, i.e. in our case the scalar field $\phi$. An elementary example of this fact is provided by Fierz-Jordan-Brans-Dicke theory, a scalar-tensor theory (not invariant under shifts) in which the scalar $\phi$ renormalizes the gravitational constant $G$ by multiplying the Einstein-Hilbert term $R/(16 \pi G)$ in the action (c.f. e.g. the review [@Berti:2015itd], and references therein, for the full action of the theory). As a result, the gravitational binding energy becomes $\phi$-dependent. Since the binding energy represents a non-negligible fraction of the total mass of strongly gravitating objects, these latter can only be effectively described by point particles if the masses are allowed to depend on $\phi$ [@eardley; @Nordtvedt:1968qr; @damour_esposito_farese; @zaglauer].
Another example is provided by Lorentz-violating gravitational theories (Hořava gravity and Einstein-æther theory, c.f. e.g. [@Yagi:2013qpa; @Yagi:2013ava] for details), in which the structure and gravitational binding energy of a compact object depends on the body’s velocity relative to a preferred reference frame defined by a non-GR, Lorentz-violating gravitational degree of freedom (a scalar field in Hořava gravity, and a vector field in Einstein-æther theory). This dependence arises exactly because different reference frames are not equivalent, since Lorentz invariance is violated. Thus, the masses of the point particles representing self-gravitating objects must depend on the body’s velocity relative to the preferred frame (i.e. on the local value of the Lorentz-violating field [@Yagi:2013qpa; @Yagi:2013ava]).
In the case of the action , let us then describe a system of two strongly gravitating objects (e.g. NSs or BHs) by replacing the matter action $S_m$ with an effective point-particle action with $\phi$-dependent masses $\tilde{m}_A(\phi)$ ($A=1,2$): $$\label{action_pp}
S_{\rm pp}=-\sum_A \int\tilde{m}_A(\phi) d\tau_A \,.$$ One can then derive the field equations of the theory by taking variations with respect to $g_{\mu\nu}$ and $\phi$. To solve these equations, one may then perform an expansion in the relative velocity of the binary over the speed of light, $v/c$, i.e. a post-Newtonian (PN) expansion. This is indeed a good approximation at least in the early inspiral phase, where $v\ll c$.
In the PN framework, it is convenient to Taylor expand the full point-particle action around the background, flat-space scalar field configuration $\phi_0=$ const as $$\begin{aligned}
&S_{\rm pp}=-\sum_A\int m_A [1+\alpha_A \delta\phi + {\cal O} (\delta\phi)^2 ] d\tau_A \,,\\\label{defs}
& \delta\phi\equiv \phi-\phi_0 \,,\quad m_A \equiv \tilde{m}_A(\phi_0)=\; \mbox{const},\quad \alpha_A\equiv \frac{1}{m_A}\frac{\partial \tilde{m}_A}{\partial \phi}(\phi_0)\Big\vert_{N_A,\Sigma_A}\,.\end{aligned}$$ Here, the “scalar charges” $\alpha_A$ (also referred to as “sensitivities” or “hairs” in the literature) are computed as partial derivatives of the masses while keeping the total entropy $\Sigma$ and baryon number $N$ of each body fixed, i.e. they describe the response of a strongly gravitating object to a small change $\delta \phi$ of the scalar field value at the object’s position, away from the background value $\phi_0$. By using this parametrization and solving the field equations at the leading PN order, one then finds that these scalar charges modify the GR GW emission total power as [@Barausse:2015wia; @damour_esposito_farese; @zaglauer] $$\label{flux}
\dot{E}_{\rm GW}=\dot{E}_{\rm GR} \left[1+B \left(\frac{v}{c}\right)^{-2}+{\cal O}\left(\alpha_1,\alpha_2\right)+{\cal O}\left(\frac{v}{c}\right)^2\right]\,,$$ where $\dot{E}_{\rm GR}$ is the GR GW emission power, and $B=5 (\alpha_1-\alpha_2)^2/96$.
Several comments are in order about this equation: *(I)* The corrections to GR are enhanced by $(v/c)^{-2}$. Therefore, the non-GR term will be the dominant one in the early inspiral (where $v\ll c$) if $|\alpha_1-\alpha_2|\sim {\cal O}(1)$. This will in turn have dramatic consequences for the emitted GW signal, which will chirp to higher frequencies faster (as a result of the more rapid dissipation of the binary’s potential and kinetic energy through GWs). *(II)* One can show that the non-GR term is due to the dipole emission of waves of the scalar graviton field $\phi$ [@damour_esposito_farese; @zaglauer]. Note that in GR, monopole and dipole emission of gravitational radiation is forbidden by the covariant conservation of the stress-energy tensor of matter (i.e. by the conservation of the total energy and linear momentum of matter and gravitational field). However, once the strong equivalence principle is violated in theories that modify or extend GR, the effective coupling between matter and the extra gravitons that appears beyond tree level, and which is parametrized by the charges $\alpha_1$ and $\alpha_2$, implies that the stress-energy tensor of a binary of compact objects is *not*, in general, covariantly conserved. Physically, this means that energy and linear momentum can be transferred from the binary to the extra gravitational field ($\phi$ in our case). As a result, both monopole and dipole GW emission become in principle possible, although monopole emission is typically suppressed for quasi-circular binaries [@damour_esposito_farese; @zaglauer]. These additional emission mechanisms also give rise to extra polarizations in the detector (besides the tensor polarizations of GR). However, these polarizations are typically difficult to observe directly, because one would in any case need a network of interferometers [@Hayama:2012au], and because they are weakly coupled to the experimental apparatus (essentially because the interferometer mirrors or test masses are weakly gravitating objects, hence their coupling to the scalar waves, which arises beyond tree level as a result of the Nördtvedt effect, is highly suppressed). *(III)* Quadrupole emission of scalar waves is also present, and in general modifies the quadrupole formula of GR. This is represented in Eq. by the term ${\cal O}\left(\alpha_1,\alpha_2\right)$. *(IV)* Eq. is also valid in a larger class of theories than the shift-symmetric scalar-tensor theories described by the action . For instance, it remains valid in more general scalar-tensor theories (not invariant under shifts of the scalar field) [@damour_esposito_farese; @zaglauer], and in Lorentz-violating gravity [@Yagi:2013qpa; @Yagi:2013ava].
It should be noted, however, that Eq. is, by itself, quite uninformative, because the magnitude of the dipole GW emission depends critically on the scalar charges $\alpha_{1,2}$. These quantities will be $\approx 0$ for weakly gravitating objects, for which the gravitational binding energy gives a negligible contribution to the total mass, but may be $\sim 1$ for self-gravitating bodies such as NSs or BHs. Moreover, let us stress that in general the scalar charges will depend on the exact nature of the body, e.g. they will generally be different for BHs and NSs, or for NSs with different masses or equations of state, exactly because the strong equivalence principle is violated.
To compute the charges in a given theory, note that for a binary in the PN regime $v\ll c$ the distance between the two compact objects is much larger than the object sizes. Thus, to first approximation we may focus on one object at a time, and solve its full field equations (and derive its mass) by assuming it is *in isolation* (i.e. neglecting the presence of the other object altogether). We may then compute the object’s mass with different “boundary” values of the scalar field (say $\phi_0$ and $\phi_0+\Delta \phi$) at a large distance $r$ from the object. (This “boundary” distance must be much larger than the object’s size $R$, but still much smaller than the binary separation $r_{12}$ in order to allow neglecting the other body’s effect, i.e $R\ll r\ll r_{12}$). The charge may then be computed from its own definition, as $\alpha\approx {\Delta \tilde{m}}/(m \Delta \phi)$. However, one can also note that the charge of each object is actually already encoded in the fall-off of the scalar in the region $R\ll r\ll r_{12}$. In more detail, in that region, since we can neglect the presence of the other body, by varying the action and adopting coordinates comoving with the object we obtain the following scalar-field equation $$\Box \phi= \frac{\partial \tilde{m}}{\partial \phi}(\phi_0) \delta^{(3)}(\boldsymbol{x}) \left[1+{\cal O}\left(\frac1r\right)\right]=
m \alpha \delta^{(3)}(\boldsymbol{x})\left[1+{\cal O}\left(\frac1r\right)\right]\,.$$ The solution to this equation, in the region $R\ll r\ll r_{12}$, therefore becomes $$\label{exp_phi}
\phi = \phi_0- \frac{\alpha m}{4 \pi r} +{\cal O}\left(\frac1r\right)^2\,,$$ i.e. we can extract the scalar charge of each object by solving its field equations in isolation, and then by looking at the $1/r$ fall-off of the scalar field at large distances [@Barausse:2015wia; @damour_esposito_farese; @zaglauer]. Again, the fact that one can extract the extra graviton charges by looking at the fall-off of those fields away from isolated objects remains true in more general situations, e.g. in scalar-tensor theories not invariant under shifts [@damour_esposito_farese; @zaglauer], and in Lorentz-violating gravity [@Yagi:2013qpa; @Yagi:2013ava].
Let us now prove that for the action , the scalar charges are *exactly zero for all stars*, but not for BHs. This result – proven in [@Barausse:2015wia] by introducing a canonical expression for the mass of a star, and then computing the derivative in the definition by using the field equations and the Stokes theorem – will be re-derived here in a simpler way, resembling a similar proof for dilatonic Gauss-Bonnet gravity [@Yagi:2015oca]. In more detail, let us consider an isolated, spherically symmetric and static star and note that the field equation for the scalar can be written as [@Sotiriou:2013qea] $$\label{eq_phi}
\nabla_\mu J^\mu
=-\chi {\cal G}\,,$$ where $J^\mu$ is a current vector. Staticity implies $J^t=0$, and one can also see that at large distances from the star, $J^\mu = \partial^\mu \phi [1+{\cal O}(1/r)]$ [@Sotiriou:2013qea]. Therefore, by integrating Eq. over a four-dimensional volume, and using Stokes theorem and Eq. , for the left-hand side we obtain $$\int J^\mu \sqrt{-g} dS_\mu=\int J^i \sqrt{-g} dt d\Sigma_i = \int dt m \alpha\,,$$ where $dS_\mu$ and $d\Sigma_i$ are respectively three- and two-dimensional (coordinate) surface elements. The integral of the right-hand side of Eq. vanishes, since ${\cal G}$ is a topological invariant and the spacetime of a star is homeomorphic to Minkowski, and we thus obtain $\alpha=0$.
The calculation proceeds *almost* unchanged for BHs, but when integrating the left-hand side there is an extra boundary at the horizon (and not just at spatial infinity), while the integral of the right-hand side is not zero because a BH spacetime is *not* topologically equivalent to Minkowski (i.e. it has a “hole” due to the horizon-enclosed singularity). In fact, one can show [@Yagi:2015oca; @Sotiriou:2013qea] that unless $\chi=0$, BHs in shift-symmetric scalar-tensor theories do *not* have vanishing scalar charge.
Discussion
==========
As mentioned above, the most evident effect of a dipole emission channel for a binary system in its early inspiral is to shed the system’s energy and angular momentum faster than in GR, thus making the GW frequency increase (“chirp”) and the separation decrease more rapidly than predicted by GR’s quadrupole formula. The perfect systems to study this effect are in principle binary pulsars, which have $v/c\sim 10^{-3}-10^{-2}$ and which would therefore be driven mainly by dipole emission for scalar charges $\sim {\cal} O(1)$. Indeed, binary pulsars constrain the parameter $B$ in Eq. to $|B|\lesssim 10^{-9}$. However, for the shift-symmetric scalar-tensor theories described by the action , we have just shown that the scalar charges are exactly zero, so the pulsar bound on $B$, no matter how strong, cannot constrain these theories at all. In other words, all shift-symmetric scalar-tensor theories are in perfect agreement with binary-pulsar tests (although the situation may be more problematic for the theories in this class that have a screening mechanism with Vainshtein radius of cosmological size, c.f. discussion in [@Barausse:2015wia]). To test these theories, one has to look at systems involving at least a BH, since BHs are the only objects for which $\alpha\neq0$. This is in stark contrast with Fierz-Jordan-Brans-Dicke-like scalar-tensor theories (e.g. those studied in [@damour_esposito_farese; @zaglauer]), for which the BH scalar charges are exactly zero and stellar scalar charges do *not* vanish. Unfortunately, since the BH binaries detected by LIGO have $v\sim c$, the bounds they provide on $B$ are very loose, $|B|\lesssim 10^{-2}$, even weaker than the bound $|B|\lesssim 2\times 10^{-3}$ from low-mass X-ray binaries [@kent-LMXB].
Nevertheless, the future European-led space-borne interferometer LISA [@2017arXiv170200786A] will be able to constrain $|B|\lesssim 10^{-7}-10^{-6}$ by observing binaries of supermassive BHs, and $|B|\lesssim 10^{-8}-10^{-7}$ by observing extreme or intermediate mass-ratio inspirals (i.e. systems comprised of a supermassive BH surrounded by either a stellar-mass compact object or an intermediate-mass BH) [@Barausse:2016eii]. However, the best LISA sources to constrain BH dipole emission will be binaries of BHs with masses $\sim 30-100 M_\odot$. These are comparable to the masses of the BHs of the first LIGO detection, GW150914. Remarkably, GW150914-like sources would be detectable by LISA months or even years in advance, before disappearing for a few days/weeks and then re-appearing close to merger in the LIGO band [@Sesana:2016ljz]. Moreover, LISA observations of these sources would allow one to predict the coalescence time in the LIGO band to within ten seconds or less, several days/weeks in advance [@Sesana:2016ljz]. Clearly, if an additional dipole emission channel dominates the evolution of these systems in the early inspiral observable by LISA, it would produce not only a significant change in the frequency evolution in the LISA band, but also a completely different coalescence time in the LIGO band than predicted by GR. Therefore, it is not at all surprising that these sources will constrain $|B|\lesssim 10^{-8}$ when observed with LISA alone, and $|B|\lesssim 10^{-9}-10^{-8}$ when observed jointly by LISA and by a LIGO-like (or better) ground-based interferometer [@Barausse:2016eii]. Interestingly, this bound is comparable with the one from binary pulsars, but unlike that, it will constrain *BH dipole emission*, thus allowing constraints to be placed on a class of theories (shift-symmetric scalar-tensor ones) that are otherwise untestable with binary pulsars.
I thank Kent Yagi and Leo Stein for many invaluable insights and discussions. I also acknowledge support from the H2020-MSCA-RISE-2015 Grant No. StronGrHEP-690904.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we present a method based on Radial Basis Functions (RBFs) for numerically solving the fully nonlinear 1D Serre Green-Naghdi equations. The approximation uses an RBF discretization in space and finite differences in time; the full discretization is obtained by the method of lines technique. For select test cases the approximation achieves spectral (exponential) accuracy. Complete <span style="font-variant:small-caps;">matlab</span> code of the numerical implementation is included in this paper (the logic is easy to follow, and the code is under 100 lines).'
author:
- 'Maurice S. Fabien'
bibliography:
- 'biblo.bib'
date: June 2014
title: 'A Radial Basis Function (RBF) Method for the Fully Nonlinear 1D Serre Green–Naghdi Equations'
---
Acknowledgments {#acknowledgments .unnumbered}
===============
I would like to acknowledge useful discussions concerning this work with Professor Randall J. LeVeque the University of Washington.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A subspace or subgroup is isotropic under a bilinear map if the restriction of the map on it is trivial. We study maximal isotropic subspaces or subgroups under skew-symmetric maps, and in particular the isotropy index—the maximum dimension of an isotropic subspace or maximum rank of an isotropic subgroup. For a smooth closed orientable manifold $M$, we describe the geometric meaning of the isotropic subgroups of the first cohomology group with different coefficients under the cup product. We calculate the corresponding isotropy index, as well as the set of ranks of all maximal isotropic subgroups, for the connected sum and the direct product of manifolds. Finally, we study the relationship of the isotropy index with the first Betti number and the co-rank of the fundamental group. We also discuss applications of these results to the topology of foliations.'
author:
- 'I. Gelbukh'
title: Isotropy index for the connected sum and the direct product of manifolds
---
Introduction
============
Let $M$ be a smooth closed orientable connected $n$-dimensional manifold. We study *isotropic* subgroups (subspaces) $H$ of its cohomology group (space) $H^1(M;R)$, where $R$ is a field or the ring of integers, under the cup-product $$\begin{aligned}
{\smile}\colon H^1(M;R)\times H^1(M;R)\to H^2(M;R),
\label{eq:intro-cupprod}\end{aligned}$$ i.e., $H\subseteq H^1(M;R)$ such that $H\smile H=0$. To avoid duplication of terminology, such as “rank (dimension),” we will refer to $H^i(M;R)$ as modules over the ring $R$.
Specifically, we study the set ${{\cal H}}(M;R)$ of ranks of maximal isotropic submodules of $H^1(M;R)$ and the corresponding *isotropy index* $$\begin{aligned}
h(M;R)=\max{{\cal H}}(M;R),
\label{eq:h=maxH}\end{aligned}$$ the maximum rank of an isotropic submodule. We study the structure of (maximal) isotropic submodules for finite connected sums and direct products of manifolds. In particular, we show (Theorems \[theor:conn\_sum\] and \[theor:dir\_prod\]) that $$\begin{aligned}
&\begin{alignedat}1
{{\cal H}}(M_1{\mathrel\#}M_2;R)&={{\cal H}}(M_1;R)+{{\cal H}}(M_2;R),\\ {{\cal H}}(M_1 \times M_2;R)&={\{\,1\,\}}\cup {{\cal H}}(M_1;R)\cup {{\cal H}}(M_2;R), \end{alignedat}
\label{eq:intro-sum}
\shortintertext{and thus}
&\begin{alignedat}1
\hphantom{{{\cal H}}}\makebox[0pt][r]{$h$}(M_1 {\mathrel\#}M_2;R)&=h(M_1;R)+h(M_2;R),\\
\hphantom{{{\cal H}}}\makebox[0pt][r]{$h$}(M_1 \times M_2;R)&=\max{\{\,h(M_1;R), h(M_2;R)\,\}}
\end{alignedat}
\label{eq:intro-prod}\end{aligned}$$ under certain conditions and with certain exceptions described in the corresponding theorems (here the sum of sets is understood element-wise).
Isotropy index bounds the co-rank ${{b_1'(M)}}$ of the fundamental group, i.e., the maximum rank of a free homomorphic image of $\pi_1(M)$ [@Gelb10] and, obviously, is bounded by the first Betti number $b_1(M)={\mathop{\mathrm{rk}}\nolimits}H_1(M;{{\mathbb{Z}}})$: $$\begin{aligned}
{{b_1'(M)}}\le h(M;{{\mathbb{Z}}})\le b_1(M).
\label{eq:intro-b'<h<b_1}\end{aligned}$$ In [@Meln3], for a field $F$ upper and lower bounds on $h(M;F)$ were given in terms of Betti numbers; see Proposition \[prop:h-bounds\]. Using and , for a given $R$ we describe all possible sets ${{\cal H}}(M;R)$ (Proposition \[prop:hS\]) and all possible values of $h(M,R)$ with different $M$ in terms of $b_1(M;R)={\mathop{\mathrm{rk}}\nolimits}H_1(M;R)$ (Theorem \[theor:h,b\]), as well as extend to fields of characteristic zero and show that in this case these bounds are exact (Proposition \[prop:b’<h<b1\]).
The notion of isotropy has been studied in the context of algebraic geometry. For instance, isotropic subspace theorems by Catanese [@Catanese] and Bauer [@Bauer] establish relations between isotropic subspaces of $H^1(M;{{\mathbb{C}}})$ for a smooth quasi-projective variety $M$ and certain irrational pencils. These theorems have been studied in [@Dimca; @Dimca-Su].
The isotropy index has numerous applications to the topological study of manifolds and foliations. As we show, isotropy for manifolds has a clear geometric meaning: (maximal) isotropic subgroups of $H^1(M;{{\mathbb{Z}}})$ of rank $k$ correspond to (maximal) systems of $k$ homologically independent, homologically non-intersecting closed orientable codimension-one submanifolds, $h(M;{{\mathbb{Z}}})$ being the maximum number of such submanifolds (Theorem \[theor:h(M)\_geometry\]). While this geometric meaning is defined for $R={{\mathbb{Z}}}$, we show that the relevant aspects of isotropy coincide for $R={{\mathbb{Z}}}$ and $R={{\mathbb{Q}}}$ (Lemma \[lem:h(G)=h(L)\]), which enables the use of simpler, vector space-based techniques in geometric applications of isotropy.
Consider a foliation defined on $M$ by a Morse form ${\omega}$, i.e., a closed one-form that is locally the differential of a Morse function. Such foliations have important applications in modern physics, for example, in supergravity [@Bab-Laz-conf; @Bab-Laz]. A Morse form foliation defines a decomposition of $M$ into a finite number $m({\omega})$ of minimal components and a finite number $M({\omega})$ of maximal components, i.e., connected components of the union of compact leaves, which are cylinders over a compact leaf. These two numbers are bounded by $h(M;{{\mathbb{Z}}})$: $
M({\omega})+m({\omega})\le h(M;{{\mathbb{Z}}})+|{\mathop{\mathrm{Sing}}\nolimits}{\omega}|-1,
$ where ${\mathop{\mathrm{Sing}}\nolimits}{\omega}$ is the singular set, which is finite [@Gelb09]. In homological terms, for the number $c({\omega})$ of homologically independent compact leaves it holds $
c({\omega})+m({\omega})\le h(M;{{\mathbb{Z}}})
$ [@Gelb09].
A sufficient condition of existence of a minimal component has been given in [@Meln2] in terms of ${\mathop{\mathrm{rk}}\nolimits}{\omega}$, the rank of the group of the periods: if $
{\mathop{\mathrm{rk}}\nolimits}{\omega}> h(M;{{\mathbb{Z}}}),
$ then the foliation has a minimal component, i.e., $m({\omega})\ge1$. Also, in case of strong inequality in the upper bound in , i.e., if $
h(M;{{\mathbb{Z}}})<b_1(M),
$ the foliation of a Morse form in general position has a minimal component [@Gelb08].
If the subgroup $H_{\omega}\subseteq H_{n-1}(M)$ generated by the homology classes of all compact leaves of a Morse form foliation is maximal isotropic, then the foliation has no minimal components, i.e., $m({\omega})=0$ [@Meln4]. Subgroups of $H_{n-1}(M)$ are related, by Poincaré duality, with those of $H^1(M;{{\mathbb{Z}}})$. In particular, if the homology classes of some compact leaves of a Morse form foliation generate a maximal isotropic subgroup, then $m({\omega})=0$ [@Meln4]. Obviously, this is the case when the foliation has $h(M;{{\mathbb{Z}}})$ homologically independent compact leaves. However, if $M=M_1\times M_2$, then in some cases our results allow us to conclude that $m({\omega})=0$ by examening only one leaf (Examples \[ex:M22xS1-foliation\] and \[ex:H(M2gxM2g)\]).
Isotropic submodules of $L$ can be defined for arbitrary bilinear map ${\varphi}{{\,:\;}}L\times L\to V$, where $L,V$ are finitely generated groups or finite-dimensional vector spaces. This gives the corresponding notions of ${{\cal H}}({\varphi})$ and the isotropy index $h({\varphi})$ as in . In order to establish our main results and , we study the behavior of isotropic submodules for skew-symmetric maps under operations of extension of scalars (Proposition \[prop:h\_over-diff-rings\]) and direct sum $L_1\oplus L_2$ (Lemmas \[lem:sum\] and \[lem:prod\]).
By Poincaré duality, ${{\cal H}}(M;{{\mathbb{Z}}})$ and $h(M;{{\mathbb{Z}}})$ can be equivalently defined in terms of the intersection map $
\circ\colon H_1(M)\times H_1(M)\to{{\mathbb{Z}}}$ instead of the cup product . For a closed one-form ${\omega}$ on $M$, the isotropy index $h({\omega})$ is defined by the restriction $\circ|_{\ker[{\omega}]\times \ker[{\omega}]}$ of the intersection map to the group $\ker[{\omega}]\subseteq H_1(M)$, where $[{\omega}]$ is the integration map. This notion has been extensively used to study the structure of Morse form foliations on closed orientable surfaces $M^2_g$ of genus $g$. For example, $
c({\omega})\le h({\omega});
$ if the foliation has no minimal components, then $
h({\omega})=g
$ [@Gelb13]. For so-called weakly generic forms, $
m({\omega})\ge g-\frac12k({\omega})-h({\omega})
$, where $k({\omega})$ is the number of singularities surrounded by a minimal component [@Gelb13a]. Since $h({\omega})\le h(M;{{\mathbb{Z}}})$, these inequalities hold also for $h(M;{{\mathbb{Z}}})$.
The paper is organized as follows. In Section \[sec\_def-spaces\], we give basic facts on isotropy in finite-dimensional vector spaces and finitely generated abelian groups. In Section \[sec\_def\_manif\], we introduce the isotropy index for manifolds and consider its properties and geometric meaning. In Section \[sec\_conn-sum\], we calculate the isotropy index of the connected sum of two manifolds. In Section \[sec\_dir-prod\], we calculate the isotropy index of the direct product of two manifolds and describe the possible sets ${{\cal H}}(M;R)$. In Section \[sec\_realization:h,b\], we completely characterize the relation between $h(M;R)$ and $b_1(M;R)$. Finally, in Section \[sec\_corank\] we consider the relations between $h(M;R)$ and $b_1'(M)$.
\[sec\_def-spaces\]Isotropy index for vector spaces and abelian groups
======================================================================
In this section, we will define the isotropy index and discuss how it changes from groups to vector fields or between vector fields with different scalars.
We will deal with finite-dimensional vector spaces and finitely generated abelian groups. To avoid duplication of terminology, such as “any subspace or subgroup” or “its dimension or rank”, we will use terminology from $R$-modules, where $R$ will be a field or ${{\mathbb{Z}}}$, correspondingly. In particular, *submodule* will stand for subspace or subgroup; *rank* will stand for the dimension of a space or the rank of a group.
Definitions
-----------
Let $L$, $V$ be finitely generated $R$-modules and $\varphi\colon L\times L\to V$ a bilinear map; $R$ is a field or $R={{\mathbb{Z}}}$.
\[def:isotropic-space\] A submodule $H\subseteq L$ is called [*isotropic*]{} under the map $\varphi$ if $\varphi|_{H\times H}=0$, i.e., $\varphi (l_1,l_2)=0$ for any $l_1,l_2\in H$.
If $R$ is a field, $R$-modules $L$ and $V$ are finite-dimensional vector spaces, so we deal with [*isotropic subspaces*]{}; if $R={{\mathbb{Z}}}$, then $R$-modules $L,V$ are finitely generated abelian groups, so we have [*isotropic subgroups*]{}.
Since $L$ is Noetherian, every isotropic submodule is contained in some maximal isotropic submodule, not necessarily unique. Denote by ${{\cal H}}(\varphi)$ the set of ranks of maximal isotropic submodules under the map $\varphi$: $${{\cal H}}(\varphi) ={\{\,{\mathop{\mathrm{rk}}\nolimits}H\mid H\mbox{ is a maximal isotropic submodule of } L\,\}}.$$ Obviously, ${{\cal H}}({\varphi})$ is a finite set of non-negative integers such that $$\begin{aligned}
0\notin{{\cal H}}({\varphi})\text{\quad or\quad}{{\cal H}}({\varphi})={\{\,0\,\}}.
\label{eq:0notinHorH=0}\end{aligned}$$ Proposition \[prop:hS\] below shows that these are the only restrictions on ${{\cal H}}({\varphi})$.
The [*isotropy index*]{} $h(\varphi)$ is the maximum rank of the isotropic submodules of $L$: $$h(\varphi)=\max{{\cal H}}(\varphi).$$
Consider the skew-symmetric map $\varphi\colon{{\mathbb{R}}}^3\times{{\mathbb{R}}}^3\to{{\mathbb{R}}}^3$, $\varphi(x,y)=[[x,y],l]$, where $l$ is a fixed vector and $[\;,\;]$ is the vector product. For any vector $x\not\perp l$, for example $x=l$, the subspace $L_1=\langle x\rangle$, $\dim L_1=1$, is maximal isotropic, and so is $L_2=l^{\perp}$, $\dim L_2=2$. Thus ${{\cal H}}(\varphi)={\{\,1,2\,\}}$, and $h(\varphi)=2$.
For skew-symmetric maps, usually $h(\varphi)\ge1$ and thus $0\notin{{\cal H}}({\varphi})$:
\[lem:h(L)=0\] Let ${\varphi}$ be skew-symmetric. Then $h(\varphi)=0$, i.e., ${{\cal H}}={\{\,0\,\}}$, iff either
- $L=0$ or
- $L=R$, ${\mathop{\mathrm{char}}\nolimits}R=2$, and $\varphi\not\equiv 0$.
Let $h(\varphi)=0$, then ${\varphi}(l,l)\ne0$ for any $0\ne l\in L$. Unless $L=0$, for a skew-symmetric map this implies ${\mathop{\mathrm{char}}\nolimits}R=2$. Suppose ${\mathop{\mathrm{rk}}\nolimits}L\ge 2$. Consider independent $l_1,l_2\in L$; $\varphi(l_i,l_i)=1$. Then for $l=l_1+l_2\ne0$, we have $\varphi(l,l)=0$, a contradiction.
The [*kernel*]{} of a bilinear map ${\varphi}{{\,:\;}}L\times L\to V$ is $$\ker\varphi={\{\,l\in L\mid\varphi (l,l')=0\;\mbox{for any}\; l'\in L\,\}}.$$ Obviously, $\ker\varphi$ is an isotropic submodule; moreover, any maximal isotropic submodule contains $\ker\varphi$, so $h(\varphi)\ge\dim\ker\varphi$.
Isotropy index for different coefficients
-----------------------------------------
Generally speaking, the isotropy index depends on the coefficients. Namely, let $L$, $V$ be finitely generated abelian groups and ${{\varphi}}\colon L\times L\to V$ a skew-symmetric bilinear map. For a field $F$, consider the corresponding vector spaces $$\begin{aligned}
L_F&=F\otimes L,\\
V_F&=F\otimes V \end{aligned}$$ and the induced skew-symmetric bilinear map $${{\varphi}}_F\colon L_F\times L_F\to V_F,\quad {\varphi}_F({\alpha}_1\otimes x_1, {\alpha}_2\otimes x_2)={\alpha}_1{\alpha}_2\otimes{\varphi}(x_1, x_2).$$ The isotropy index depends on the field $F$, and generally $h({{\varphi}})\ne h({{\varphi}}_F)$:
Consider ${\varphi}{{\,:\;}}{{\mathbb{Z}}}\times{{\mathbb{Z}}}\to{{\mathbb{Z}}}_2={{\mathbb{Z}}}/2\,{{\mathbb{Z}}}$, ${\varphi}(1,1)=1$. It has an isotropic subgroup is $2\,{{\mathbb{Z}}}$, thus $h({\varphi})=1$. Similarly, for $F={{\mathbb{Q}}}$, we have ${\varphi}_{{\mathbb{Q}}}{{\,:\;}}{{\mathbb{Z}}}\times{{\mathbb{Z}}}\to0$, thus again $h({\varphi})=1$, in accordance with below. However, for $F={{\mathbb{Z}}}_2$, we have $L_F={{\mathbb{Z}}}_2\otimes{{\mathbb{Z}}}={{\mathbb{Z}}}_2$, $V_F={{\mathbb{Z}}}_2\otimes{{\mathbb{Z}}}_2={{\mathbb{Z}}}_2$, so ${{\varphi}}_{{{\mathbb{Z}}}_2}\colon {{\mathbb{Z}}}_2\times{{\mathbb{Z}}}_2\to{{\mathbb{Z}}}_2$, ${\varphi}(1,1)=1$; obviously, ${{\varphi}}_{{{\mathbb{Z}}}_2}^{-1}(0)=0$ and thus $h({{\varphi}}_{{{\mathbb{Z}}}_2})=0$: $$\begin{aligned}
h({\varphi}_{{{\mathbb{Z}}}_2})<h({\varphi})=h({\varphi}_{{\mathbb{Q}}}).\end{aligned}$$ On the other hand, consider ${{\varphi}}{{\,:\;}}{{\mathbb{Z}}}^2\times{{\mathbb{Z}}}^2\to{{\mathbb{Z}}}$ defined by the matrix [$\begin{pmatrix*}[r]
0&k\\
\!-k&0
\end{pmatrix*}
\!,
$ ]{} $k\ge2$. Then $h({\varphi})=h_{{\mathbb{Q}}}({\varphi})=1$, but ${\varphi}_{{{\mathbb{Z}}}_p}{{\,:\;}}{{\mathbb{Z}}}_p^2\times{{\mathbb{Z}}}_p^2\to0$ (thus $h({\varphi}_{{{\mathbb{Z}}}_p})=2$) iff $p\mid k$. So for $p\mid k$ and $q\nmid k$ we have: $$\begin{aligned}
h({\varphi})=h({\varphi}_{{\mathbb{Q}}})=h({\varphi}_{{{\mathbb{Z}}}_q})<h({\varphi}_{{{\mathbb{Z}}}_p}).\end{aligned}$$
However, extension of scalars for vector spaces does not decrease $h({\varphi}_F)$:
\[lem:h\_ext-of-scalars\] Let $L_F,V_F$ be finite-dimensional vector spaces over a filed $F$ and ${{\varphi}_F}\colon L_F\times L_F\to V_F$ be a skew-symmetric bilinear map. Let $F'$ be a field, $F\subseteq F'$, $$\begin{aligned}
L_{F'}&=F'\otimes_F L,\\
V_{F'}&=F'\otimes_F V\end{aligned}$$ vector spaces obtained from $L_F$, $V_F$ by extension of scalars, and ${\varphi}_{F'}$ the induced map: $$\begin{aligned}
{{\varphi}_{F'}}\colon L_{F'}\times L_{F'}\to V_{F'},\quad {\varphi}_{F'}({\alpha}'_1\otimes x_1, {\alpha}'_2\otimes x_2)={\alpha}'_1{\alpha}'_2\otimes{\varphi}_F(x_1, x_2).
\label{eq:phi_F'}\end{aligned}$$ Then $$h({\varphi}_F)\le h({\varphi}_{F'}).$$
Consider an isotropic subspace $H_F\subseteq L_F$, $\dim H_F=h({\varphi}_F)=k$. A basis $\left\langle e_1,\dots,e_k\right\rangle=H_F$ can be be extended to a basis for $L_F$; thus $H_F= F^k$. Extension of scalars from $F$ to $F'$ gives $H_{F'}={F'}\otimes_F H_F={F'}\otimes_F F^k={(F')}^k$, so $\dim H_{F'}=\dim H_F$. By , the subspace $H_{F'}$ is isotropic, i.e., $k=\dim H_{F'}\le h({\varphi}_{F'})$. We obtain $h({\varphi}_F)\le h({\varphi}_{F'})$.
A stronger fact holds for groups and ${{\mathbb{Q}}}$:
\[lem:h(G)=h(L)\] Let $L$, $V$ be finitely generated abelian groups and ${{\varphi}}\colon L\times L\to V$ be a skew-symmetric bilinear map. Denote by $$\begin{aligned}
{2}
L_{{\mathbb{Z}}}&=L,\qquad &L_{{\mathbb{Q}}}&={{\mathbb{Q}}}\otimes L,\\
V_{{\mathbb{Z}}}&=V,\qquad &V_{{\mathbb{Q}}}&={{\mathbb{Q}}}\otimes V\end{aligned}$$ the corresponding ${{\mathbb{Z}}}$- and ${{\mathbb{Q}}}$-modules. Let $${{{\varphi}_{{\mathbb{Q}}}}}\colon L_{{\mathbb{Q}}}\times L_{{\mathbb{Q}}}\to V_{{\mathbb{Q}}},\quad {{\varphi}_{{\mathbb{Q}}}}(q_1\otimes x_1, q_2\otimes x_2)=q_1q_2\otimes{\varphi}(x_1, x_2)$$ be the induced skew-symmetric bilinear map. Then
1. \[item:G->L\] for every maximal isotropic subgroup $H\subseteq L$, the subspace ${H_{{\mathbb{Q}}}}={{\mathbb{Q}}}\otimes H\subseteq L_{{\mathbb{Q}}}$ is maximal isotropic;
2. \[item:L->G\] for every maximal isotropic subspace ${H_{{\mathbb{Q}}}}\subseteq L_{{\mathbb{Q}}}$, there is a maximal isotropic subgroup $H\subseteq L$ such that ${H_{{\mathbb{Q}}}}={{\mathbb{Q}}}\otimes H$.
In particular, $$\begin{aligned}
\begin{aligned}
{{\cal H}}({{\varphi}})&={{\cal H}}({{{\varphi}_{{\mathbb{Q}}}}}),\\
h({{\varphi}})&=h({{{\varphi}_{{\mathbb{Q}}}}}).
\end{aligned}
\label{eq:hZ=hQ}\end{aligned}$$
\[item:G->L\] Let $H\subseteq L$ be a maximal isotropic subgroup, $H=\langle h_1,\dots,h_n\rangle$. Then ${H_{{\mathbb{Q}}}}={{\mathbb{Q}}}\otimes H\subseteq L_{{\mathbb{Q}}}$ is an isotropic subspace, $\dim{H_{{\mathbb{Q}}}}={\mathop{\mathrm{rk}}\nolimits}H$. Consider $0\ne q\otimes x\in L_{{\mathbb{Q}}}$ such that ${{\varphi}_{{\mathbb{Q}}}}(q\otimes x,{H_{{\mathbb{Q}}}})=0$, i.e., all ${\varphi}(x, h_i)\in V_T$, the torsion subgroup. Then for some $k\ne0$, we have ${\varphi}(kx, h_i)=0$. Since $H$ is maximal, $kx\in H$ and thus $q\otimes x\in H_{{\mathbb{Q}}}$. Therefore, $H_{{\mathbb{Q}}}$ is maximal.
\[item:L->G\] Let $H_{{\mathbb{Q}}}\subseteq L_{{\mathbb{Q}}}$ be a maximal isotropic subspace, ${H_{{\mathbb{Q}}}}=\langle q_1\otimes h_1,\dots,q_n\otimes h_n\rangle$, a basis. Consider $H'=\langle h_1,\dots,h_n\rangle$. Then all ${\varphi}(h_i,h_j)\in V_T$; thus for some $k\ne0$, all ${\varphi}(kh_i, kh_j)=0$. We obtain $H_{{\mathbb{Q}}}={{\mathbb{Q}}}\otimes H''$ for an isotropic subgroup $H''=kH'={\{\,kx\mid x\in H'\,\}}$, $k\ne0$.
Consider a maximal isotropic subgroup $H\supseteq H''$. For any $x\in H$, we have ${\varphi}(x, H'')=0$ and thus ${{\varphi}_{{\mathbb{Q}}}}(1\otimes x,{H_{{\mathbb{Q}}}})=0$. Since $H_{{\mathbb{Q}}}$ is maximal, $1\otimes x\in H_{{\mathbb{Q}}}$. We obtain $H_{{\mathbb{Q}}}={{\mathbb{Q}}}\otimes H$.
Lemma \[lem:h(G)=h(L)\] allows formulating Lemma \[lem:h\_ext-of-scalars\] for fields or ${{\mathbb{Z}}}$:
\[prop:h\_over-diff-rings\] Let $L_R,V_R$ be finitely generated $R$-modules, $R$ being a field or ${{\mathbb{Z}}}$, and ${\varphi}_R\colon L_R\times L_R\to V_R$ be a skew-symmetric bilinear map. Let $R'$ be a field, $R\subseteq R'$, $$\begin{aligned}
L_{R'}&=R'\otimes_R L,\\
V_{R'}&=R'\otimes_R V\end{aligned}$$ modules obtained by extension of scalars, and $
{\varphi}_{R'}\colon L_{R'}\times L_{R'}\to V_{R'}
$ the induced map. Then $$\begin{aligned}
h({\varphi}_R)\le h({\varphi}_{R'}).
\label{eq:h<h}\end{aligned}$$
In particular, in addition to extension of scalars of vector spaces, holds for a group and a corresponding vector space over $F$, ${\mathop{\mathrm{char}}\nolimits}F=0$, since ${{\mathbb{Z}}}\subset{{\mathbb{Q}}}\subseteq F$.
\[sec\_def\_manif\]Isotropy index for manifolds
===============================================
In this section, we introduce maximal isotropic subgroups (subspaces) of the first cohomology group (space) and the isotropy index for manifolds and discuss their geometric meaning and properties.
Definitions
-----------
Let $M$ be a smooth closed orientable $n$-dimensional manifold. Consider the cup product $$\begin{aligned}
{\smile}\colon &H^1(M;R)\times H^1(M;R) \to H^2(M;R),\end{aligned}$$ where $R={{\mathbb{Z}}}$ or $R$ is a field. It is a skew-symmetric bilinear map, and $H^k(M;R)$ are finitely generated $R$-modules; in case of a field, $H^k(M;R)$ are vector spaces.
A submodule $H\subseteq H^1(M;R)$ is called [*isotropic*]{} if it is isotropic under $\smile$ in the sense of Definition \[def:isotropic-space\], i.e., if the restriction of the cup-product to $H\times H$ is zero: ${\smile}|_{H\times H}=0$.
Accordingly, we denote by ${{\cal H}}(M;R)$ the set of ranks of maximal isotropic submodules: $${{\cal H}}(M;R) ={\{\,\dim H\mid H\mbox{\, is a maximal isotropic submodule of } H^1(M;R)\,\}};$$ Proposition \[prop:hS\] below shows that is still the only restriction on ${{\cal H}}(M;R)$, i.e., that almost any set of non-negative integers is ${{\cal H}}(M;R)$ for some manifold $M$.
The [*isotropy index*]{} $$h(M;R)=\max{{\cal H}}(M;R)$$ is the maximum rank of isotropic submodules of $H^1(M;R)$.
Lemma \[lem:h(G)=h(L)\] allows us to work interchangeably with $H^1(M;{{\mathbb{Z}}})$ and $H^1(M;{{\mathbb{Q}}})$:
\[lem:h=h\_Q\] For a smooth closed orientable manifold $M$, there exists a maximal isotropic subgroup $H\subseteq H^1(M;{{\mathbb{Z}}})$, ${\mathop{\mathrm{rk}}\nolimits}H=k$, iff there exists a maximal isotropic subspace $H_Q\subseteq H^1(M;{{\mathbb{Q}}})$, $\dim H_Q=k$, i.e., $$\begin{aligned}
{{\cal H}}(M;{{\mathbb{Z}}})&={{\cal H}}(M;{{\mathbb{Q}}}),\\
h(M;{{\mathbb{Z}}})&=h(M;{{\mathbb{Q}}}).\end{aligned}$$
Geometric meaning of $h(M;{{\mathbb{Z}}})$
------------------------------------------
The notions of ${{\cal H}}(M;{{\mathbb{Z}}})$ and $h(M;{{\mathbb{Z}}})$ have a clear geometric meaning, which can be characterized as follows:
\[def:isotropic-system\] An [*isotropic system*]{} on a manifold $M$, $\dim M\ge2$, is a set of homologically non-intersecting homologically independent smooth closed orientable connected codimension-one submanifolds $X_1,\dots,X_k\subset M$, intersecting transversely: $$\begin{aligned}
[X_i\cap X_j]=0,
\label{eq:[XY]=0}\end{aligned}$$ $i\ne j$; $i,j=1,\dots,k$.
The requirement cannot be simplified to $X_i\cap X_j=\emptyset$, since on some manifolds there exist submanifolds with non-empty, but homologically trivial, intersection:
\[ex:Heisenberg-detailed\] On the Heisenberg 3-nilmanifold, any two homologically independent smooth closed orientable 2-submanifolds have non-empty, but homologically trivial, intersection. This will be shown as Example \[ex:Heisenberg\] below; here we only give a graphical illustration.
The Heisenberg 3-nilmanifold $H^3$ is a $T^2$-bundle over the circle $S^1$, with the monodromy being a Dehn twist $f{{\,:\;}}T^2\to T^2$, defined as the quotient space $$H^3=\frac{[0,1]\times T^2}{(1,x)\sim (0,f(x))},\quad f=\begin{pmatrix} 1 & 0 \\ 1 & 1\end{pmatrix}.$$ For the basis cycles $a$, $c$ of the torus $T^2$, we have $$\begin{aligned}
f_*(a)&=a+c;\\
f_*(c)&=c;\end{aligned}$$ see \[fig:Heisengerg\]([*a*]{}). The cycle $c$ is homologically trivial, being realized by the boundary of a 2-submanifold (torus without a disk) shown in \[fig:Heisengerg\]([*a*]{}). However, for the two submanifolds $T_i=T^2$ shown in \[fig:Heisengerg\]([*b*]{}), we have $[T_1\cap T_2]=c$.
![\[fig:Heisengerg\] The Heisenberg nilmanifold $H^3$, represented as a $T^2$-bundle over the circle $S^1$. The circle is shown as the vertical line $b$; $T^2$ is shown as a horizontal square with the sides $a$ and $c$, the opposite sides of the square being identified. The top is identified with the bottom with the Dehn twist: $a\sim a+c$, $c\sim c$; all four vertical lines are identified. (a) The curve realizing $c$ is the boundary of a 2-submanifold shown as the hatched rectangle, triangle, and another rectangle; thus $c=0$. The triangle forms a disk with two holes, which are glued to the cylinder formed by the two rectangles; the resulting figure is a torus with a disk removed, whose boundary realizes $c$. (b) The two tori $T_1$, $T_2$ intersect by a curve realizing $c$, thus $[T_1\cap T_2]=0$. They cannot be made non-intersecting; see Example \[ex:Heisenberg\].](Fig-Heisenberg-surface-lines-modified){width="100\unitlength"}
(0,0)(100,0) (60,103)[$a$]{} (-8,35)[$b$]{} (87,37)[$c$]{} (25,14) (50,-15)
![\[fig:Heisengerg\] The Heisenberg nilmanifold $H^3$, represented as a $T^2$-bundle over the circle $S^1$. The circle is shown as the vertical line $b$; $T^2$ is shown as a horizontal square with the sides $a$ and $c$, the opposite sides of the square being identified. The top is identified with the bottom with the Dehn twist: $a\sim a+c$, $c\sim c$; all four vertical lines are identified. (a) The curve realizing $c$ is the boundary of a 2-submanifold shown as the hatched rectangle, triangle, and another rectangle; thus $c=0$. The triangle forms a disk with two holes, which are glued to the cylinder formed by the two rectangles; the resulting figure is a torus with a disk removed, whose boundary realizes $c$. (b) The two tori $T_1$, $T_2$ intersect by a curve realizing $c$, thus $[T_1\cap T_2]=0$. They cannot be made non-intersecting; see Example \[ex:Heisenberg\].](Fig-Heisenberg-intersection-lines-modified){width="100\unitlength"}
(0,0)(100,0) (45,63)[$T_1$]{} (103,37)[$T_2$]{} (50,-15)
An algebraic model of $H^3$ can be given as follows: consider the 3-dimensional Heisenberg group over a ring $R$, $$H(R)={\{\,\begin{pmatrix} 1 & x & z \\ 0 & 1 & y \\ 0 & 0 & 1\end{pmatrix}\mid x,y,z\in R\,\}};$$ then the Heisenberg nilmanifold $H^3=H({{\mathbb{R}}})/H({{\mathbb{Z}}})$ is the quotient of the real Heisenberg group by the discrete Heisenberg subgroup. It is a compact orientable connected 3-manifold with Nil geometry.
Definition \[def:isotropic-system\] implies the cardinality of an isotropic system $k\le b_1(M)$, the Betti number; thus each isotropic system is contained in a maximal isotropic system.
The following theorem relies on the fact that homology classes $z\in H_{n-1}(M)$ can be realized by smooth closed orientable connected codimension-one submanifolds.
\[theor:h(M)\_geometry\] Let $M$ be a smooth closed orientable connected manifold, $\dim M\ge2$, and $D\colon H^1(M;{{\mathbb{Z}}})\to H_{n-1}(M)$ be a Poincaré duality map. Then:
1. Let ${\{\,X_i\,\}}$ be a (maximal) isotropic system. Then ${\{\,D^{-1}[X_i]\,\}}$ form a basis of a (maximal) isotropic subgroup $H\subseteq H^1(M;{{\mathbb{Z}}})$.
2. Let ${\{\,x_i\,\}}$ be a basis of a (maximal) isotropic subgroup $H\subseteq H^1(M;{{\mathbb{Z}}})$. Then ${\{\,Dx_i\,\}}$ can be realized by submanifolds $X_i$ that form a (maximal) isotropic system.
In particular,
- ${{\cal H}}(M;{{\mathbb{Z}}})={\{\,k\mid\,X_1,\dots,X_k\subset M\text{ is a maximal isotropic system}\,\}}$;
- the isotropy index $h(M;{{\mathbb{Z}}})$ is the maximum cardinality of an isotropic system of submanifolds of $M$.
Consider an isotropic subgroup $H\subset H^1(M;{{\mathbb{Z}}})$, ${\mathop{\mathrm{rk}}\nolimits}H=k$. Since $H^1(M;{{\mathbb{Z}}})$ is torsion-free, it has a basis, $H=\left\langle u_1,\dots,u_k\right\rangle$. The cup product $${\smile}\colon H^1(M;{{\mathbb{Z}}})\times H^1(M;{{\mathbb{Z}}})\to H^2(M;{{\mathbb{Z}}})$$ is dual to the homology classes intersection map $${\circ}\colon H_{n-1}(M)\times H_{n-1}(M)\to H_{n-2}(M);$$ namely, $$D(u_i\smile u_j)=Du_i\circ Du_j,$$ where $D$ is a Poincaré duality map.
Realize the cycles $Du_i\in H_{n-1}(M)$ by suitable submanifolds $X_i\subset M$, $Du_i=[X_i]$, choosing them to intersect transversely. Then $$\begin{aligned}
\pm[X_i\cap X_j]=[X_i]\circ[X_j]=Du_i\circ Du_j=D(u_i\smile u_j),\end{aligned}$$ where the sign depends on the choice of orientation in $X_i$ and $X_j$. Since $H$ is isotropic, all $u_i\smile u_j=0$; thus $[X_i\cap X_j]=0$ for any $i\ne j$. Since $u_i$ are independent, so are $[X_i]$. If $H$ is maximal then so is this system, because expanding it would, by duality, expand $H$.
Similarly, given a (maximal) system of $k$ such submanifolds $X_i\subset M$, the group $H=\langle D^{-1}[X_i]\rangle\subseteq H^1(M;{{\mathbb{Z}}})$ is a (maximal) isotropic subgroup, ${\mathop{\mathrm{rk}}\nolimits}H=k$.
Examples \[ex:Heisenberg\] and \[ex:KT\] below show that the homological interpretation of the non-intersection requirement \[eq:\[XY\]=0\] is important for Theorem \[theor:h(M)\_geometry\]: some manifolds have fewer non-intersecting submanifolds, $X_i\cap X_j=\emptyset$, with the properties listed in Definition \[def:isotropic-system\], than homologically non-intersecting such submanifolds, $[X_i\cap X_j]=0$. For discussion of systems with $X_i\cap X_j=\emptyset$, see Section \[sec\_corank\].
Properties and examples
-----------------------
Recall that the Betti number $b_k(M;R)={\mathop{\mathrm{rk}}\nolimits}H_k(M;R)$; by definition, $b_k(M)=b_k(M;{{\mathbb{Z}}})$. By the universal coefficient theorem, if $R$ is a field with ${\mathop{\mathrm{char}}\nolimits}R=0$, then $b_k(M;R)=b_k(M)$. Since $H^k(M;R)=H_k(M;R)=0$ for $k>\dim M$, the following statements apply to $S^1$ and a point $*$.
\[lem:h(M)=0\] Let $M$ be a smooth closed orientable manifold; $R={{\mathbb{Z}}}$ or $R$ be a field. Then $$\begin{aligned}
1\le h(M;R)\le b_1(M;R),
\label{eq:1<h<b}\end{aligned}$$ except that $$\begin{aligned}
h(M;R)=0
\label{eq:h=0}\end{aligned}$$ iff any of the following conditions holds:
- $b_1(M;R)=0$, or
- $b_1(M;R)=1$, ${\mathop{\mathrm{char}}\nolimits}R=2$, and the cup product ${\smile}\not\equiv 0$.
Theorem \[theor:h,b\] below states that this lemma gives the only relation between $h(M;R)$ and $b_1(M;R)$.
By definition, we have $h(M;R)\le{\mathop{\mathrm{rk}}\nolimits}H^1(M;R)$. For ${{\mathbb{Z}}}$, by Poincaré duality, $H^1(M;{{\mathbb{Z}}})\cong H_{n-1}(M)$, so ${\mathop{\mathrm{rk}}\nolimits}H^1(M;{{\mathbb{Z}}})=b_{n-1}(M)=b_1(M)$. For a field $F$, $H^1(M;F)\cong H_1(M;F)$, so $\dim H^1(M;F)=\dim H_1(M;F)=b_1(M;F)$. Since the cup product is skew-symmetric, is given by Lemma \[lem:h(L)=0\].
\[ex:RP3\] Consider $M={{\mathbb{R}}}P^3$; it is orientable. Its cohomology ring is $$\begin{aligned}
H^*({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2)\cong\frac{{{\mathbb{Z}}}_2[{\alpha}]}{({\alpha}^4)},\end{aligned}$$ where $|{\alpha}|=1$. Thus each $H^i({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2)$ is a free ${{\mathbb{Z}}}_2$-module with generator ${\alpha}^i$, i.e., $H^i({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2)={{\mathbb{Z}}}_2$. We have $b_i({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2)=1$, and for ${\alpha}\in H^1({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2)$ it holds ${\alpha}\smile{\alpha}\ne0$, i.e., $h({{\mathbb{R}}}P^3;{{{\mathbb{Z}}}_2})=0$.
Obviously, $h(M;R)=b_1(M;R)$ iff ${\smile}\equiv 0$. Since $\ker{\smile}\subseteq H^1(M;R)$ is an isotropic submodule, $h(M;R)\ge\dim\ker{\smile}$. There are, though, better estimates:
\[prop:h-bounds\] Let $M$ be a smooth closed orientable manifold and $k=\dim\ker{\smile}$. For $R={{\mathbb{Z}}}$ or $R$ being a field, with the exception specified below, we have:
1. It holds $$\begin{aligned}
\frac{b_1(M;R)+k\,b_2(M;R)}{b_2(M;R)+1}\le h(M;R)\le\frac{b_1(M;R)\,b_2(M;R)+k}{b_2(M;R)+1};
\label{eq:b+bk<h<bb+k}\end{aligned}$$ in particular, if $b_2(M;R)=1$, then $$\begin{aligned}
h(M;R) = \frac 12(b_1(M;R) + k).
\label{eq:h=1/2}\end{aligned}$$
2. If $\smile$ is surjective, then $$\begin{aligned}
h(M;R)\le k+\frac 12+\sqrt{\left(b_1(M;R)-k-\frac 12\right)^2-2\,b_2(M;R)}.
\label{h=sqrt}\end{aligned}$$
As an exception, if $$\begin{aligned}
\left\{
\begin{array}l
{\mathop{\mathrm{char}}\nolimits}R=2,\\
b_1(M;R)=1,\\
k=0,
\end{array}
\right.
\label{eq:exception}\end{aligned}$$ then $h(M;R)=0$ and of –, only the upper bound in holds.
If $h(M;R)\ne0$, for a field this has been shown in [@Meln3]; for ${{\mathbb{Z}}}$ it follows from Lemma \[lem:h=h\_Q\].
If $h(M;R)=0$, then by Lemma \[lem:h(M)=0\] either $b_1(M;R)=0$ or holds. In the former case, – happen to hold; in , we have $b_2(M;R)=0$. In the latter case, and the lower bound in do not hold, while in we have $b_2(M;R)=1$ and the square root does not exist.
The exception is illustrated by Example \[ex:RP3\]. In fact, and the lower bound in would not need an exception if we took the floor function of the corresponding expressions, which in all other cases except happen to be integer anyway.
For a closed orientable surface of genus $g$, gives $h(M^2_g;R)=g$; for $n$-torus, $n\ge2$, gives $h(T^n;R)=1$. Both cases do not fall under exception , since $b_1(M^2_g)=2g\ne1$ and $b_1(T^n;R)=n\ne1$.
\[ex:H(T)\] Thus, ${{\cal H}}(T^n;R)={\{\,1\,\}}$. Indeed, gives $$h(M;R)=1\quad\text{iff}\quad\cal H(M;R)={\{\,1\,\}}.$$
The following example shows a non-singleton ${{\cal H}}(M;R)$:
\[ex:M22xS1\] ${{\cal H}}(M^2_2\times S^1;{{\mathbb{Z}}})={\{\,1,2\,\}}$. This is seen in \[fig:M22xS1\], but can also be formally proved by Theorem \[theor:dir\_prod\] below.
\[sec\_conn-sum\]Isotropy index of the connected sum of manifolds
=================================================================
For sets, we denote $A+B={\{\,a+b\mid a\in A, b\in B\,\}}$.
\[lem:sum\] Let $L_i$, $V_i$, $i=1,2$, be finite-dimensional vector spaces over a filed $F$ and $\varphi_i\colon L_i\times L_i\to V_i$ be bilinear skew-symmetric maps. Denote $$\begin{aligned}
L&=L_1\oplus L_2,\\
V&=V_1\oplus V_2, \end{aligned}$$ and let $\varphi\colon L\times L\to V$ be a bilinear skew-symmetric map such that $$\begin{aligned}
\varphi|_{L_i\times L_i}&=\varphi_i,\\
\varphi|_{L_1\times L_2}&=0;\end{aligned}$$ i.e., $\varphi$ is defined as component-wise sum of $\varphi_i$: $$\begin{aligned}
\varphi(x,y)=\underbrace{\varphi_1(x_1,y_1)}_{\in V_1}+\underbrace{\varphi_2(x_2,y_2)}_{\in V_2},
\label{phi=phi+phi}\end{aligned}$$ where $x_i,y_i\in L_i$ are projections. Then:
1. A subspace $H\subseteq L$ is maximal isotropic iff $$\begin{aligned}
H=H_1\oplus H_2,
\label{eq:H=H+H}\end{aligned}$$ where $H_i\subseteq L_i$ are maximal isotropic under $\varphi_i$.
2. The set of dimensions of maximal isotropic subspaces $${{\cal H}}({{\varphi}})={{\cal H}}({{\varphi}}_1)+{{\cal H}}({{\varphi}}_2).$$
3. The isotropy index $$h({{\varphi}})=h({{\varphi}}_1)+h({{\varphi}}_2).$$
Note that these conclusions do not necessarily hold for isotropic subspaces that are not maximal. For example, each 1-dimensional subspace $\langle x\rangle$, $x\in L\setminus(L_1\cup L_2)$, is isotropic, but does not hold for it.
By , if $H_i\in L_i$ are isotropic, then $H=H_1\oplus H_2$ is isotropic.
($\Rightarrow$) Let $H\subseteq L$ be a maximal isotropic subspace. Consider $x,y\in H$. By , $\varphi(x_1,y_1)=-\varphi(x_2,y_2) \in V_1\cap V_2=0$, i.e., both projections $p_i(H)\subseteq L_i$ are isotropic. Let $H_i \supseteq p_i(H)$ be maximal isotropic subspaces of $L_i$. Then $H'=H_1\oplus H_2$ is isotropic, and since $H\subseteq H'$ is maximal, $H=H'$.
($\Leftarrow$) Conversely, let $H_i \subseteq L_i$ be maximal isotropic subspaces; then $H=H_1\oplus H_2$ is isotropic. Consider $x=x_1+x_2 \in L\setminus H$, i.e., say, $x_1\notin H_1$. Since $H_1$ is maximal isotropic in $L_1$, there exists $y=y_1\in H_1\subseteq H$ such that $\varphi(x_1,y_1)\ne 0$. By , $\varphi(x,y)=\varphi(x_1,y_1)+0\ne 0$. Thus $H$ is maximal.
\[theor:conn\_sum\] Let $M_1$, $M_2$ be connected closed orientable manifolds with $\dim M_i\ge 2$, and $R$ be a field or $R={{\mathbb{Z}}}$. Then for the connected sum $M_1{\mathrel\#}M_2$:
1. A submodule $H\subseteq H^1(M_1{\mathrel\#}M_2;R)$ is maximal isotropic iff $$H=H_1\oplus H_2,$$ where $H_i\subseteq H^1(M_i;R)$ are maximal isotropic submodules.
2. The set of ranks of maximal isotropic submodules $${{\cal H}}(M_1{\mathrel\#}M_2;R)={{\cal H}}(M_1;R)+{{\cal H}}(M_2;R).$$
3. The isotropy index of the connected sum $$h(M_1 {\mathrel\#}M_2;R)=h(M_1;R)+h(M_2;R).$$
Let $R$ be a field. Denote $L=H^1(M_1{\mathrel\#}M_2;R)$ and $L_i=H^1(M_i;R)$, $i=1,2$. Since $\dim M_i\ge2$, the Mayer-Vietoris sequence gives $L=L_1\oplus L_2$. The additive structure is given by the induced maps of the inclusions; the cup product translates into component-wise product: $$\begin{aligned}
x\smile y=(x_1\smile y_1)+(x_2\smile y_2),\end{aligned}$$ where $x,y\in L$ and $x_i,y_i\in L_i$ are projections. Then, for fields, Lemma \[lem:sum\] gives the result. Now for $R={{\mathbb{Z}}}$, the result follows from Lemma \[lem:h(G)=h(L)\].
\[ex:H(M\^2\_g)\] By Theorem \[theor:conn\_sum\] and given Example \[ex:H(T)\], or by if $g=0$, for a closed orientable surface of genus $g$ it holds $${{\cal H}}(M_g^2;R)={{\cal H}}({\mathrel\#}_{i=1}^g T^2;R)=\sum_{i=1}^g {{\cal H}}(T^2;R)={\{\,g\,\}}.$$
Consider $M=M^2_2\times S^1$ from Example \[ex:M22xS1\] with ${{\cal H}}(M;{{\mathbb{Z}}})={\{\,1,2\,\}}$; see \[fig:M22xS1\]. Then $${{\cal H}}(M{\mathrel\#}M;{{\mathbb{Z}}})={\{\,2,3,4\,\}}$$ and $h(M{\mathrel\#}M;{{\mathbb{Z}}})=4$.
Consider $M=M^2_a\times M^2_b$, surfaces of genus $a$ and $b$, $1\le a\le b$. Theorem \[theor:dir\_prod\] below gives ${{\cal H}}(M;R)={\{\,1,a,b\,\}}.$ Therefore, $${{\cal H}}(M{\mathrel\#}M;R)={\{\,2,a+1,b+1,a+b,2\,a,2\,b\,\}}$$ and $h(M{\mathrel\#}M;R)=2\,b$.
\[sec\_dir-prod\]Isotropy index of the direct product of manifolds
==================================================================
\[lem:xy=yx\] Let $L$ be a vector space and $x,y,u,v\in L$; $x,y\ne 0$. Then $x\otimes v=u\otimes y$ implies $u=ax$, $v=ay$ for some $a$.
Coordinate-wise, we have $$\begin{aligned}
x_iv_j=u_iy_j
\label{eq:xv=uy}\end{aligned}$$ for all $i,j$. For those $i,j$ for which $x_i,y_j\ne0$, this gives $$\frac{u_i}{x_i}=\frac{v_j}{y_j}=a_{ij}.$$ Since $a_{ij}$ does not depend on $i$ or $j$, all $a_{ij}=a$. We obtain $u_i=ax_i$ if $x_i\ne0$ and $v_j=ay_j$ if $y_j\ne0$. If $x_i=0$, gives $0=u_iy_j$ for all $j$, thus $u_i=0$, and similarly $y_j=0$ implies $v_j=0$.
While in Lemma \[lem:sum\] we had $\varphi|_{L_1\times L_2}=0$, now consider ${\mathop{\mathrm{im}}\nolimits}\varphi|_{L_1\times L_2}$ as large as possible:
\[lem:prod\] Let $L_i$, $V_i$, $i=1,2$, be final-dimensional vector spaces over a field $F$ and $\varphi_i\colon L_i\times L_i\to V_i$ be bilinear skew-symmetric maps. Denote $$\begin{aligned}
L&=L_1\oplus L_2,\\
V&=V_1\oplus V_2\oplus V_3,\end{aligned}$$ where $V_3=L_1\otimes L_2$, and let $\varphi\colon L\times L\to V$ be a bilinear skew-symmetric map such that $$\begin{gathered}
\begin{aligned}
\varphi|_{L_i\times L_i}&=\varphi_i,\\
\varphi|_{L_1\times L_2}&=\otimes;
\end{aligned}
\notag
\shortintertext{i.e.,}
\varphi(x,y)=\underbrace{\varphi_1(x_1,y_1)}_{\in V_1}+\underbrace{\varphi_2(x_2,y_2)}_{\in V_2}+\underbrace{x_1\otimes y_2-y_1\otimes x_2}_{\in V_3},
\label{eq:phi=phi+phi+phi}\end{gathered}$$ where $x_i,y_i\in L_i$ are projections. Then:
1. A subspace $H\subseteq L$ is isotropic iff $$\dim H=1 \textforidiot{{\mathop{\mathrm{char}}\nolimits}F=2?} \textrm{ or }H=H_i,$$ where $H_i\subseteq L_i$ is isotropic under $\varphi_i$, for $i=1\textrm{ or }2$.
2. \[item:x2\] The set of dimensions of maximal isotropic subspaces $${{\cal H}}({{\varphi}})={\{\,1\,\}}\cup {{\cal H}}({{\varphi}}_1)\cup {{\cal H}}({{\varphi}}_2)$$ except that ${{\cal H}}({{\varphi}})={{\cal H}}({{\varphi}}_i)$ if $h({\varphi}_j)=0$, i.e., if either
- $L_j=0$ or
- $L_j=F$, ${\mathop{\mathrm{char}}\nolimits}F=2$, and $\varphi_j\not\equiv0$.
3. The isotropy index $$h({{\varphi}})=\max{\{\,h({{\varphi}}_1),h({{\varphi}}_2)\,\}}.$$
Note that in contrast to Lemma \[lem:sum\], the first conclusion does not require $H$ to be maximal.
Let $H$ be an isotropic subspace. We will show that if both projections $p_i(H)\ne 0$, then $\dim H=1$. Consider $x\in H$ such that both projections $x_i\ne 0$. Let $y\in H$. Since $H$ is isotropic and the three components of are independent, we have $$\varphi_1(x_1,y_1)=\varphi_2(x_2,y_2)=x_1\otimes y_2-y_1\otimes x_2=0.$$ By Lemma \[lem:xy=yx\], $y\in\langle x\rangle$. The conditions for $h({\varphi}_j)=0$ in item are given by Lemma \[lem:h(L)=0\].
\[theor:dir\_prod\] Let $M_1$, $M_2$ be connected closed manifolds and $R$ be a field or $R={{\mathbb{Z}}}$. Then for the direct product $M_1 \times M_2$:
1. A submodule $H\subseteq H^1(M_1 \times M_2;R)$ is isotropic iff $${\mathop{\mathrm{rk}}\nolimits}H=1\textrm{ or }H=H_i,$$ where $H_i\subseteq H^1(M_i;R)$ is isotropic for $M_i$, $i=1\text{ or }2$.
2. The set of ranks of maximal isotropic submodules $${{\cal H}}(M_1 \times M_2;R)={\{\,1\,\}}\cup {{\cal H}}(M_1;R)\cup {{\cal H}}(M_2;R)$$ except that ${{\cal H}}(M_1 \times M_2;R)={{\cal H}}(M_i;R)$ if $h(M_j;R)=0$, i.e., if either
- $b_1(M_j;R)=0$, the Betti number, or
- $b_1(M_j;R)=1$, ${\mathop{\mathrm{char}}\nolimits}R=2$, and ${\smile}\not\equiv0$.
3. The isotropy index of the direct product $$h(M_1 \times M_2;R)=\max{\{\,h(M_1;R), h(M_2;R)\,\}}.$$
Let $R$ be a field. Denote $$\begin{aligned}
L_i&=H^1(M_i,R),\quad i=1,2,&L&=H^1(M_1 \times M_2,R),\\
V_i&=H^2(M_i,R),\quad i=1,2,&V&=H^2(M_1 \times M_2,R).\end{aligned}$$ By the Künneth formula, $$\begin{aligned}
L&=L_1\oplus L_2,\\
V&=V_1\oplus V_2\oplus V_3,
\shortintertext{where}
V_3&=L_1\otimes L_2.\end{aligned}$$ By construction, $L_i\smile L_i\subseteq V_i$ for $i=1,2$; $L_1\smile L_2\subseteq V_3$, and holds for the cup-products in $M_1\times M_2$ and $M_i$, respectively. Lemma \[lem:prod\] gives the result for fields and Lemma \[lem:h(G)=h(L)\] for $R={{\mathbb{Z}}}$.
Example \[ex:Heisenberg-detailed\] shows that in Theorem \[theor:dir\_prod\], the direct product cannot be replaced by an arbitrary fiber bundle.
\[ex:h\_R(T\^n)\] By Lemma \[lem:h(M)=0\], $h(S^1;R)=1$, so for a torus $T^n={\mathop\times}_{i=1}^n S^1$, we have $h(T^n;R)=1$.
\[ex:S1\*SN\] By Lemma \[lem:h(M)=0\], $h(S^n;R)=0$, $n\ge 2$, and $h(S^1;R)=1$, so $h(S^n\times S^1;R)=1$.
\[prop:hS\] For any non-empty finite set $S\subset{{\mathbb{Z}}}^*$ of non-negative integers and for $R={{\mathbb{Z}}}$ or $R$ being a field, $S={{\cal H}}(M;R)$ for some smooth closed orientable connected manifold $M$ iff $S={\{\,0\,\}}$ or $0\notin S$.
If $S={\{\,g\,\}}$, then $S={{\cal H}}(M^2_g;R)$, a surface of genus $g$; see Example \[ex:H(M\^2\_g)\]. Let now $S={\{\,s_1,\dots,s_N\,\}}$, $N\ge2$. By the condition, $m=\min S\ge1$. Consider $$\begin{aligned}
2
M_1&=M^2_{s_1-m+1}\times\cdots\times M^2_{s_N-m+1},&\quad&\dim M_1=2N,\\
M_2&=M^2_{m-1}\times S^{2N-2},&&\dim M_2=2N.\end{aligned}$$ By Theorems \[theor:conn\_sum\] and \[theor:dir\_prod\], we obtain ${{\cal H}}(M_1{\mathrel\#}M_2;R)=S$.
Another application of Theorem \[theor:dir\_prod\] can be found in study of the topology of foliations defined by Morse forms. It is known that if the subgroup of $H_{n-1}(M)$ generated by the homology classes of all compact leaves of the foliation is maximal isotropic, then the foliation has no minimal components [@Meln4]. This condition obviously holds true when the foliation has $h(M;{{\mathbb{Z}}})$ homologically independent compact leaves. However, if $M=M_1\times M_2$, in some cases Theorem \[theor:dir\_prod\] allows to conclude that the foliation has no minimal components by considering only one leaf:
\[ex:M22xS1-foliation\] As has been mentioned in Example \[ex:M22xS1\], ${{\cal H}}(M^2_2\times S^1;{{\mathbb{Z}}})={\{\,1,2\,\}}$; see \[fig:M22xS1\]. Even though $h(M^2_2\times S^1,{{\mathbb{Z}}})=2$, if a Morse form foliation has the submanifold $N=M^2_2$ as a leaf, then it has no minimal components. In contrast, nothing can be said about a form that has $T_1=T^2$ as a leaf, because the system ${\{\,T_1\,\}}$ is not maximal.
\[ex:H(M2gxM2g)\] ${{\cal H}}(M^2_a\times M^2_b;{{\mathbb{Z}}})={\{\,1,a,b\,\}}$, $a,b\ge1$. Now consider a cycle $z$ that winds around the $M_1=M^2_a$ and also around the $M_2=M^2_b$, that is, $z=z_1+z_2$, $0\ne z_i\in H^1(M_i,{{\mathbb{Z}}})$. If a Morse form foliation has a leaf dual to $z$, then it has no minimal components.
\[sec\_realization:h,b\]Isotropy index and the first Betti number
=================================================================
By definition of the isotropy index, $$\begin{gathered}
h(M;R)\le b_1(M;R);
\shortintertext{for example:}
\begin{alignedat}2
h(S^1;R)&=1,\quad&b_1(S^1;R)&=1;\\
h(M^2_g;R)&=g,\quad&b_1(M^2_g;R)&=2g;\\
h(T^n;R)&=1,\quad&b_1(T^n;R)&=n.
\end{alignedat}\end{gathered}$$ The only relation between $h(M;R)$, $b_1(M;R)$, and $R$ is given by Lemma \[lem:h(M)=0\]; in particular, any gap between $h(M;R)$ and $b_1(M;R)$ is possible for a given $R$:
\[theor:h,b\] Let $h,b\in{{\mathbb{Z}}}$, and $R$ be a field or $R={{\mathbb{Z}}}$. There exists a connected smooth closed orientable manifold $M$ with $h(M;R)=h$ and $b_1(M;R)=b$ iff any of the following conditions holds:
- $1\le h\le b$, or
- $h=b=0$, or
- $h=0$, $b=1$, and ${\mathop{\mathrm{char}}\nolimits}R=2$.
For $h=b=0$, consider $M=S^n$. For $h=0$ and $b=1$ with ${\mathop{\mathrm{char}}\nolimits}R=2$, consider $M={{\mathbb{R}}}P^3$; see Example \[ex:RP3\]. Let now $1\le h\le b$. Choose $m_i\ge1$ such that $$\begin{aligned}
\sum_{i=1}^{h} m_i=b.\label{sum=b}\end{aligned}$$ For large enough $n$ such that $n-m_i\ge2$ for all $i$, consider an $n$-manifold $$M={\mathop\#}_{i=1}^h\left(T^{m_i}\times S^{n-{m_i}}\right).$$ By Theorem \[theor:dir\_prod\], for each summand $M_i=T^{m_i}\times S^{n-{m_i}}$, we have $h(M_i;R)=1$, while $b_1(M_i;R)=m_i$. Then by Theorem \[theor:conn\_sum\], $$\begin{aligned}
h(M;R)&=\sum_{i=1}^h h(M_i;R)=\sum_{i=1}^h1 =h,\\
b_1(M;R)&=\sum_{i=1}^h b_1(M_i,R)=\sum_{i=1}^h m_i=b.
\qedhere\end{aligned}$$
Note that the construction used in the proof requires $n=\dim M\ge2+\lceil \frac b{h}\rceil$, the ceiling here being the smallest possible value for $\max{\{\,m_i\,\}}$ under . This condition is not restrictive when $h=b$, leading to $n\ge3$; it is not very restrictive when $\frac b2\le h<b$, leading to $n\ge4$, etc.; for $\frac{b}{k+1}\le h<\frac{b}{k}$, $k=1,\dots,b-1$, we need $n\ge k+3$. This requires high dimension when $h\ll b$.
For ${{\mathbb{Z}}}_2$, however, $n=3$ is enough:
For $R={{\mathbb{Z}}}_2$, the manifold in Theorem \[theor:h,b\] can be chosen with any given $\dim M\ge3$.
Let $R={{\mathbb{Z}}}_2$ and $\dim M=3$. For $h=b=0$, consider $M^3=S^3$. Let now $b\ge1$. Consider $$\begin{aligned}
M^3=\left({\mathop\#}_{i=1}^h\left(S^1\times S^2\right)\right){\mathop\#}\left({\mathop\#}_{i=1}^{b-h}{{\mathbb{R}}}P^3\right).
\label{eq:M3}\end{aligned}$$ Example \[ex:S1\*SN\] shows that $h(S^1\times S^2;{{\mathbb{Z}}}_2)=1$, thus Theorem \[theor:conn\_sum\] implies $$\begin{aligned}
4
h(M^3;{{\mathbb{Z}}}_2) &=\sum_{i=1}^h h(S^1\times S^2;{{\mathbb{Z}}}_2)&&{}+\sum_{i=1}^{b-h} h({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2) &&{}=\sum_{i=1}^h1+\sum_{i=1}^{b-h}0 &&{}=h,\\
b_1(M^3;{{\mathbb{Z}}}_2)&=\sum_{i=1}^h b_1(S^1\times S^2;{{\mathbb{Z}}}_2)&&{}+\sum_{i=1}^h b_1({{\mathbb{R}}}P^3;{{\mathbb{Z}}}_2) &&{}=\sum_{i=1}^h1+\sum_{i=1}^{b-h}1 &&{}=b.\end{aligned}$$ This trivially generalizes to $\dim \ge 5$ as $$\begin{aligned}
M^n=M^3\times S^{n-3}
\label{eq:M3timesS}\end{aligned}$$
Let now $\dim M=4$. For $1\le h<b$, we use with one summand less in , namely, $$M^3=\left({\mathop\#}_{i=1}^h\left(S^1\times S^2\right)\right){\mathop\#}\left({\mathop\#}_{i=1}^{b-h-1}{{\mathbb{R}}}P^3\right).$$ For $h=b$, consider $
M^4={\mathop\#}_{i=1}^h(S^1\times S^3)
$.
Finally, for $h=0$, $b=1$, consider an Enriques surface $X$. Indeed,[^1] $X=K3/\sigma$, where $\sigma$ is an orientation-preserving fixed point-free involution; note that a $K3$ surface is simply connected. Then $H^1(X;{{\mathbb{Z}}}_2)={\mathop{\mathrm{Hom}}\nolimits}(\pi_1(X),{{\mathbb{Z}}}_2)={{\mathbb{Z}}}_2$; thus $b_1(X;{{\mathbb{Z}}}_2)=1$. For $0\ne x\in H^1(X;{{\mathbb{Z}}})$, $x\smile x$ is a reduction (mod 2) of $\beta x$, where $\beta{{\,:\;}}H^1(X;{{\mathbb{Z}}}_2)\to H^2(X;{{\mathbb{Z}}})$ is the Bockstein homomorphism. Suppose $x\smile x=0$, i.e., for some $y\in H^2(X;{{\mathbb{Z}}})$ we have $\beta x=2y$ with $\beta x\ne 0$ because $H^1(X;{{\mathbb{Z}}})=0$. Since $\beta x$ is $2$-torsion, we obtain that $0\ne\pi^*y\in H^2(K3;{{\mathbb{Z}}})$ is $4$-torsion, where $\pi$ is the quotient map, while the latter group is torsion-free. Thus $h(X;{{\mathbb{Z}}})=0$.
\[sec\_corank\]Isotropy index and the co-rank of the fundamental group
======================================================================
In this section, we give a lower bound on $h(M;R)$ stronger than $1$ from .
The [*co-rank of the fundamental group*]{} of a smooth closed connected manifold $M$ is the maximum rank of a free quotient group of $\pi_1(M)$; we denote it by ${{b_1'(M)}}$.
While $h(M;{{\mathbb{Z}}})$ is the maximum number of homologically non-intersecting submanifolds $[X_i\cap X_j]=0$ (Theorem \[theor:h(M)\_geometry\]), ${{b_1'(M)}}$ strengthens the condition to $X_i\cap X_j=\emptyset$:
\[theor:b’(M)\_geometry\] The co-rank of the fundamental group ${{b_1'(M)}}$ is the maximum number of non-intersecting homologically independent smooth closed orientable connected codimension-one submanifolds $X_i\subset M$: $$X_i\cap X_j=\emptyset,$$ $i\ne j$, $i,j=1,\dots,{{b_1'(M)}}$.
The condition for $X_i$ to be smooth closed orientable is important for deduce that ${{b_1'(M)}}\le h(M)$. Jaco’s [@Jaco72 Theorem 2.1] is formulated for a compact $M$, possibly with $\partial M$; $X_i$ are also compact and $\partial X_i\subseteq\partial M$. Since our $M$ is closed, then $X_i$ are also closed. These $X_i$ are two-sided; since $M$ is orientable, so are $X_i$. These $X_i$ are not necessary smooth, but any homology class can be realized by a smooth $X_i$.
Accordingly, properties of ${{b_1'(M)}}$ closely resemble those of $h(M;{{\mathbb{Z}}})$. Similarly to –, it holds [@Gelb18]: $$\begin{gathered}
{{b_1'(M)}}=0\textrm{\quad iff\quad}b_1(M)=0
\label{eq:b'=0-iff-b1=0}
\shortintertext{and otherwise}
1\le {{b_1'(M)}}\le b_1(M);
\label{eq:1<b'<b}\end{gathered}$$ in particular, $h(M;{{\mathbb{Z}}})=0$ iff ${{b_1'(M)}}=0$. Exactly as in Theorems \[theor:conn\_sum\] and \[theor:dir\_prod\], for the connected sum, $\dim M_i\ge 2$, except for non-orientable surfaces, and the direct product it holds $$\begin{aligned}
2b_1'(M_1{\mathrel\#}M_2)&=b_1'(M_1)+b_1'(M_2),&\quad&\text{\cite{Harvey}}\\
b_1'(M_1\times M_2)&=\max{\{\,b_1'(M_1), b_1'(M_2)\,\}}.&&\text{\cite{Gelb18}}\end{aligned}$$
\[ex:b’=h\] Non-surprisingly, for many manifolds ${{b_1'(M)}}=h(M;R)$:
- For the closed orientable surface, $b_1'(M^2_g)=g$ [@Leininger] and $h(M^2_g;R)=g$ [@Meln3]; see Example \[ex:H(M\^2\_g)\].
- For $n$-torus, $b_1'(T^n)=1$ [@Gelb17] and $h(T^n;R)=1$ [@Meln3]; see Example \[ex:h\_R(T\^n)\].
- For manifolds with quasi-Kähler and 1-formal fundamental group, for example, for compact Kähler manifolds, ${{b_1'(M)}}=h(M;{{\mathbb{C}}})$ [@Dimca-Pa-Su].
- For $M={\mathop\#}_{i=1}^h\left(T^{m_i}\times S^{n-{m_i}}\right)$ from Theorem \[theor:h,b\], it holds ${{b_1'(M)}}=h(M;R)$.
A non-trivial theorem from [@Gelb18] implies that – represent the only relation between ${{b_1'(M)}}$ and $b_1(M)$ for any given $\dim M$. The last item in Example \[ex:b’=h\] shows that the construction from Theorem \[theor:h,b\] gives an elementary proof of this fact for large enough $\dim M$:
Let $b',b\in{{\mathbb{Z}}}$. There exists a connected smooth closed orientable manifold $M$ with ${{b_1'(M)}}=b'$ and $b_1(M)=b$ iff either $$\begin{aligned}
2
b'=b=0,&\quad\text{see~\eqref{eq:b'=0-iff-b1=0}, or}\\
1\le b'\le b,&\quad\text{see~\eqref{eq:1<b'<b}.}\end{aligned}$$
Comparing Theorems \[theor:h(M)\_geometry\] and \[theor:b’(M)\_geometry\] gives $$\begin{aligned}
{{b_1'(M)}}\le h(M;{{\mathbb{Z}}});
\label{b'<h(M;Z)}\end{aligned}$$ together with this gives a geometric proof of lower and upper bounds on the isotropy index $h(M;{{\mathbb{Z}}})$, which have been obtained indirectly in [@Gelb10]. We extend this to fields of characteristic zero:
\[prop:b’<h<b1\] Let $R={{\mathbb{Z}}}$ or $R$ be a field, ${\mathop{\mathrm{char}}\nolimits}R=0$. For the co-rank of the fundamental group ${{b_1'(M)}}$, the isotropy index $h(M;R)$, and the first Betti number $b_1(M)$ it holds $$\begin{aligned}
{{b_1'(M)}}\le h(M;R)\le b_1(M).
\label{h(M;R)<b1(M)}\end{aligned}$$
By Proposition \[prop:h\_over-diff-rings\], for a field $F$ with ${\mathop{\mathrm{char}}\nolimits}F=0$, we have $h(M;{{\mathbb{Z}}})\le h(M;F)$. Equations and complete the proof: $${{b_1'(M)}}\le h(M;{{\mathbb{Z}}})\le h(M;F)\le b_1(M).
\qedhere$$
Both bounds in are exact (see Example \[ex:b’=h\] and Theorem \[theor:h,b\]); in particular, as we have shown, in many cases ${{b_1'(M)}}$ is a very strong lower bound for $h(M)$. However, both inequalities can also be strict:
\[ex:Heisenberg\] Consider the Heisenberg nilmanifold $H^3$. Its fundamental group $\pi_1(H^3)$ is nilpotent, so $b'_1(H^3)=1$. Since $H^1(H^3,{{\mathbb{Z}}})={{\mathbb{Z}}}^2$ with zero cup-product [@Lambe-Pr], we have $$1=b'_1(H^3)<h(H^3;{{\mathbb{Z}}})=b_1(H^3)=2.$$
\[ex:KT\] The Kodaira–Thurston nilmanifold $M=H^3\times S^1$ gives an example of $${{b_1'(M)}}<h(M;{{\mathbb{Z}}})<b_1(M).$$ Indeed, the fundamental group $\pi_1(M)$ is nilpotent, so ${{b_1'(M)}}=1$; by Theorem \[theor:dir\_prod\] and given Example \[ex:Heisenberg\], $h(M;{{\mathbb{Z}}})=2$; and, obviously, $b_1(M)=3$.
[10]{}
[Bab-Laz-conf]{} E. M. Babalic C. I. Lazaroiu Foliated backgrounds for [M]{}-theory compactifications [(II)]{} Proceedings of the Joint Meeting on Quantum Field Theory and Nonlinear Dynamics, Sinaia, Romania 2014
[Bab-Laz]{} E. M. Babalic C. I. Lazaroiu Singular foliations for [M]{}-theory compactification Journal of High Energy Physics 03 2015 116
[Bauer]{} I. Bauer Irrational pencils on non-compact algebraic manifolds Int. J. Math. 8 4 1997 441–450
[Catanese]{} F. Catanese Fibered surfaces, varieties isogeneous to a product and related moduli spaces Amer. J. Math. 122 2000 1–44
[Dimca]{} A. Dimca On the isotropic subspace theorems Bull. Math. Soc. Sci. Math. Roumanie 51(99) 4 2008 307–324
[Dimca-Pa-Su]{} A. Dimca, S. Papadima A. Suciu Quasi-[K]{}ähler groups, 3-manifold groups, and formality Mathematische Zeitschrift 268 1–2 2011 169–186
[Dimca-Su]{} A. Dimca A. Suciu Which 3-manifold groups are [K]{}ähler groups? J. European Math. Soc. 11 3 2009 521–528
[Gelb17]{} I. Gelbukh Co-rank and [Betti]{} number of a group Czech. Math. J. 2015 65 2 565–567
[Gelb10]{} I. Gelbukh Number of minimal components and homologically independent compact leaves for a [M]{}orse form foliation Stud. Sci. Math. Hung. 46 4 2009 547–557
[Gelb09]{} I. Gelbukh On the structure of a [M]{}orse form foliation Czech. Math. J. 59 1 2009 207–220
[Gelb08]{} I. Gelbukh Presence of minimal components in a [M]{}orse form foliation Differ. Geom. Appl. 22 2005 189–198
[Gelb13]{} I. Gelbukh Structure of a [M]{}orse form foliation on a closed surface in terms of genus Differ. Geom. Appl. 29 4 2011 473–492
[Gelb18]{} I. Gelbukh The co-rank of the fundamental group: The direct product, the first [B]{}etti number, and the topology of foliations Mathematica Slovaca 2016
[Gelb13a]{} I. Gelbukh The number of minimal components and homologically independent compact leaves of a weakly generic [M]{}orse form on a closed surface Rocky Mt. J. Math. 43 5 2013 1537–1552
[Harvey]{} S. Harvey On the cut number of a 3-manifold Geom. Topol. 6 2002 409–424
[Jaco72]{} W. Jaco Geometric realizations for free quotients J. Austral. Math. Soc. 14 1972 411–418
[Lambe-Pr]{} L. A. Lambe S. B. Priddy Cohomology of nilmanifolds and torsion-free, nilpotent groups Transactions of the American Mathematical Society 273 1 1982 39–55
[Leininger]{} C. J. Leininger A. W. Reid The co-rank conjecture for 3-manifold groups Algebraic and Geometric Topology 2 2002 37–50
[Meln4]{} I. A. Mel’nikova A test for compactness of a foliation Math. Notes 58 6 1995 1302–1305
[Meln2]{} I. A. Mel’nikova A test for non-compactness of the foliation of a [M]{}orse form Russ. Math. Surveys 50 2 1995 444–445
[Meln3]{} I. A. Mel’nikova Maximal isotropic subspaces of skew-symmetric bilinear mapping Mosc. Univ. Math. Bull. 54 4 1999 1–3
[^1]: Example contributed by a colleague who preferred not to be named.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Explainable models in Artificial Intelligence are often employed to ensure transparency and accountability of AI systems. The fidelity of the explanations are dependent upon the algorithms used as well as on the fidelity of the data. Many real world datasets have missing values that can greatly influence explanation fidelity. The standard way to deal with such scenarios is imputation. This can, however, lead to situations where the imputed values may correspond to a setting which refer to counterfactuals. Acting on explanations from AI models with imputed values may lead to unsafe outcomes. In this paper, we explore different settings where AI models with imputation can be problematic and describe ways to address such scenarios.'
author:
- '\'
title: The Challenge of Imputation in Explainable Artificial Intelligence Models
---
Introduction
============
Even though the field of Artificial Intelligence is more than sixty years old, it is only in the last decade or so that AI systems are being increasingly interwoven into the fabric of the socio-technical apparatus of the society and are thus having a massive impact on society. This increasing incorporation of AI has led to increased calls for accountability and regulation of AI systems [@gunning2017explainable]. Model explanations are considered to be one of the most important ways to provide accountability of AI systems. The model explanations, however, can only be as good as the data on which the algorithms are based. This is where the issue of missing and imputed data becomes pivotal for model explanations. In some domains like healthcare, almost all datasets have missing values [@cismondi2013missing]. As many applications of AI in healthcare are patient-oriented, decisions that are informed by AI and ML models can potentially have significant clinical consequences. Additionally, the requirements for a responsible and robust AI solution in critical domains like healthcare, criminal justice system etc. require additional considerations that affect safety and compliance with regulations [@ahmad2018interpretable].
One of the main reasons why imputation is used in AI and machine learning models is that it allows the use of all available data for model building instead of restricting oneself to a subset of completed records, likely achieve via deletion. The net effect is that models built using imputed data often result in better predictive performance. Consequently, imputation may be necessary if performance is more important than explainability. Even outside of data missingness, researchers acknowledge that there is a trade-off between explainable AI models and their predictive performance [@weller2017challenges] *i.e.,* in general the more explainable a model is the less predictive power it will have and vice versa. Thus, Deep Learning models, which are epitome of black box models, are relatively opaque but models like linear regression models are more straightforward to interpret. The choice of the underlying algorithm for model explanation is often dependent upon the risk involved [@ahmad2018interpretable]. For example, in a problem predicting the likely need for palliative care, the physician must know why the prediction is being made since the prediction may changes the plan of care for a patient. In this scenario predictive performance could be sacrificed for intelligibility of the prediction. On the other hand, if the application is to predict the upcoming patient census in an emergency department the ED charge nurse may be less concerned with model explainability.
Concerns regarding the safety of AI systems become paramount when we consider that incorrect imputation can lead to explanations that may not have fidelity with the underlying phenomenon and thus may lead to potentially harmful actions, especially in decision support systems. Consequently, questions regarding imputation are of paramount importance from a safety and even regulatory perspective in high risk fields like healthcare, criminal justice system, and elsewhere. In this paper we will explore multiple scenarios that can arise when dealing with imputation in model explanations in AI. To illustrate the problems with imputation, we consider use cases in the healthcare domain using data from a hospital system in the mid-Western United States consisting of data from 76 thousand patients, similar but not identical to the dataset used in [@ahmad2018death].
Imputation with Unsafe Outcomes
===============================
It is important to note that explainability is just one mechanism to ensure safety of AI systems and to mitigate against unwarranted and unforeseen outcomes. Optimizing for explinability via post-hoc models while simultaneously optimizing for performance via imputation may lead to unsafe outcomes as we demonstrate with examples from healthcare and the criminal justice system.
Use Case: Imputation and Patient Safety
---------------------------------------
One of the main dangers for using imputed values for model explanations is that the explanations that are produced may in fact have factors which are absurd or do not make domain sense. To illustrate, consider the following example based off of the problem of predicting the length of stay of a patient at a hospital at the time of admission. Suppose that a black box model is used (e.g., Extreme Gradient Boosting) as the underlying predictive model. One way to extract explanations from this models is to use post-hoc models like LIME [@ribeiro2016should] or Shapley Values [@shapley1953value]. To illustrate such a case, consider the output for model explanation in Table \[tab:lime\] for predicting length of stay in a hospital using the LIME model. The Table gives the factor and its associated value, as well as the relative importance of the factor for the explanation and whether it was imputed or not. The third most important factor, the albumin level which refers to a lab test, is an imputed value.
Suppose the end user for this system is a clinician who is trying to understand why the predictions are being made. To get the full context she looks at her patient’s lab results (*e.g.,* a list of the most recent lab tests and results for the patient.) She will observe a missing value for Albumin but when she looks at the explanation for the predictions and sees the lab value for a lab which was not actually performed. The problem with this scenario is that it could cast doubt on the entire machine learning solution in the mind of the user since the explanation for the prediction is at odds with the patient history. Consider the alternative ways in which this scenario can play out *e.g.,* if the explanation for the prediction is that the patient has a low readmission risk due to a low Troponin. Troponin refers to a group of proteins found in skeletal and heart muscle fibers that are responsible for regulating muscular contraction [@ebashi1971troponin]. However if Troponin has not been evaluated, the physician may overlook that fact and assume that an acute coronary event has already been ruled out. However if the acute coronary event is indeed the real reason for the patient is at risk then it is may lead to a catastrophic outcome.
Additionally imputing from a biased population (hospitalized patients, for example) does not give one a true representation of the value in a patient without pathology *e.g.,* Troponin is only measured in people with an expected heart attack or other acute cardiac issue. Therefore, the distribution of values for Troponin does not represent that of the non-diseased population. If a 20 year old is in the Emergency Department for abdominal pain from appendicitis, one should not impute a Troponin. What this example illustrates is that in some contexts there is additional context which may not be captured by data and imputing without domain knowledge can be potentially hazardous.
**Feature** **Value** **Importance** **Imputed**
---------------------- ----------- ---------------- -------------
Readmits last year 2 0.67 0
Creatinine Level 1.6 0.54 0
Albumin Level 2.4 0.36 1
Avg. Cigarettes/week 28 0.18 0
Systolic BP 90 0.17 1
: LOS Related Explanation Factor for a patient[]{data-label="tab:lime"}
Use Case: Imputation in Criminal Justice System
-----------------------------------------------
Another use case where imputation can lead to adverse consequences is in the criminal justice system. If the prediction task is to predict if the person is likely to re-offend or not, then imputing values may lead to scenarios where the person is deemed low risk because variables related to the person’s risk profile are imputed based on people who are similar to him in the larger population. The problem with this approach using imputation is that if the sample is biased against minority populations then imputation may always leads to imputing negative tendencies and thus higher risk scores for the minority population. This may happen even if demographic features are not used in model building [@berk2018fairness] and because such factors are not used they will also not show up as explanation factors. The end user will only see that the person is a high risk because of certain, possibly violent, tendencies or past history. If the imputed nature of such explanations is not highlighted then this is likely to lead to faulty recommendations and even wrongfully long sentences.
Alternates to Imputation in Model Explanations
==============================================
Since model performance is a requirement in most machine learning applications and explanation fidelity is needed, it is important to consider alternatives to imputation. One way to address the problem of missing values in some domains is to use indicator variables because the presence and absence of certain variable is still useful information. In Electronic Health Records (EHR) datasets majority of the labs and vitals are missing for most patients because testing for all labs is unnecessary and thus the labs were never ordered. Most machine learning predictive models infer the missing values via imputation. If a result exists for a lab then that implies that the ordering clinician deemed it necessary. Therefore, the fact that the lab was ordered is itself useful information. We propose that this information can be incorporated in machine learning models in at least two ways. In the first scheme, indicator variables for features like labs and vitals can be used. Instead of the original values for the variables we record their presence and absence as follows: $$\begin{aligned}
f(x)=
\begin{cases}
0, &\textit{if}\ x=null \\
1, &\textit{otherwise}
\end{cases}
\end{aligned}$$ The alternate method is to use central values *i.e.,* average, median, truncated average etc. or central values by cohort for variables like labs and vitals as follows: $$f(x)=
\begin{cases}
norm(x), ~\textit{if}\ x=null \\
x, ~\textit{otherwise}
\end{cases}$$
It should be noted that even though one can use imputed variables instead of the corresponding variable with the actual and missing values, in most cases the expected performance of a model based on the indicator variables would be less as compared to a model with imputed values. Consider Decision Trees where the split of the decision tree branch needs to be decided for a variable in integer space *e.g.,* age. A popular scheme to do this is to consider the midpoints between actual data points and determine which splits give the best information gain or other information theoretic criteria.
Assuming that most variable spaces will be non-binary spaces, the split that one gets from the indicator variables will be different since the loss of entropy given by the equation would be different. In other words the models which are created from the imputed variables vs. the indicator variables would be different. To illustrate this, we build multiple prediction models on the mid-Western EHR data described above. The results of prediction for the models are given in Table \[tab:results\]. The model with no missing values refers to a model that was built with only using variables with no missing values. We note that even for the model which used variables with missing values, we did not use a variable where more than fifty percent of the values were missing. Since the space of such variables is much smaller as compared to the space considered for all the variables, the performance of the models is quite mediocre *i.e.,* an MAE of 4.09 days. The results for the model with a sophisticated imputation technique like MICE are quite good with an MAE of 2.08 days and that for the model with indicator variables is 2.46 days.
Lastly, we also tried model building with average values and the results are better than what we get for indicator variables model but with the caveat that these values do not correspond to anything that was measured but rather the values are inferred for normalcy. While the indicator variables models is not as good as the imputation model, it may still be sufficiently good to be used in a production setting. The added advantage of using such a model would be that the model explanations would actually correspond to how the data is *i.e.,* indicate the presence or absence of a variable.
**Model** **MAE** **MSE**
--------------------------------- --------- ---------
Model with no Missing Variables 4.09 213.07
Model with MICE Imputation 2.08 13.29
Model with Indicator Variables 2.46 18.05
Model with Averages Imputed 2.40 17.80
: Results Summary for LOS Models in days[]{data-label="tab:results"}
Explanability vs. Types of Data Missingness
===========================================
All imputation methods add bias to the data which in turn affects the fidelity of the model built based on the imputed data. This implies that model fidelity is dependent upon the nature of the missing data and the imputation technique. Data can be, Missing at random, not missing at random and completely missing at random [@little2019statistical]. Consider how each of these missingness types affect model explanations: If the data is missing completing at random (MCAR) then even though imputation adds bias to the data, the introduced bias would be local *i.e.,* at the instance level and not global since the data is not systematically missing. On the other hand, if the data is missing at random (MAR) then it means that the missing data is at least partially dependent upon other variables in the data *e.g.,* in a survey people who are in the service sector may be less likely to report their income, clinicians may not order certain labs to be tested if they think that the patient is less likely to have certain conditions. Imputing in such cases, add bias to the data at the global level.
Imputation for data that is missing at random can be done by employing domain knowledge or some other known aspect of data *e.g.,* expected values for populations with normal characteristics may be imputed for some laboratory tests if the patients do not have certain conditions. Lastly, if the data is not missing at random (NAMR) then sophisticated statistical methods like MICE, 3D-MICE are often used used for imputation. MICE or multiple imputation by chained equations [@azur2011multiple] is a method where each missing variable is imputed by a separate model but it maintains consistency between imputations by means of passive imputation. It should be noted that however, all imputation methods introduce biases and studies have shown that imputed values can be wildly off from real values [@luo20173d]. If the values are off then the explanations will be incorrect. Thus, explanations for AI models are less likely to be off in case of the data missingness is because of MAR or MCAR but imputation for NMAR data could occasionally lead to explanations that may not correspond to how the model actually works. We address these and other concerns in the section on operationalization of explanations.
Implementations of tree based models like XGBoost handle missing values by having a default direction for nodes with missing values in the current instance set [@chen2016xgboost]. The direction of the missing data is thus used as a proxy of data missingness. Such implementations of tree based algorithms minimize error loss in the training phase. However this also means that if the distribution of missing values in the test set is different from that of the training set then the performance of the predictive models will suffer and consequently the quality of explanations generated by the models would also degrade.
Operationalizing Explanations with Imputation
=============================================
An important axis of accountability of Artificial Intelligence systems is Fairness. There are multiple notions of what constitutes fairness in machine learning [@srivastava2019mathematical], to illustrate the effect of imputation on fairness and explainability we focus on group level notions of fairness. A system that is fair need not be explainable if the underlying algorithm can ensure that the various groups of interest are being scored in a fair manner even when the model is blackbox. Requiring the model to be explainable and interpretable may require it impute values about groups which could render the model unfair since imputed values can add bias against certain protected groups *e.g.,* ethnicity, race, gender etc. The choice of the trade-off should be made with respect to the use case and the application domain.
It has been noted that generating and operationalizing explanations is not strictly a machine learning problem [@weller2017challenges]. The same machine learning model may be used to generate disparate explanations if the use case or the end users are different [@ahmad2018interpretable]. The risk profile of the end user determines how explanations should be operationalized. If the use of explainable models is an absolute requirement in a setting where there are a large number of missing values then appropriate disclaims regarding data missing and imputation should be given to the end user. Even with such messaging it is still possible that automation bias can creep in and lead to unsafe recommendations and outcomes [@mullainathan2017does]. Additionally, imputation of missing values can lead to models which are unfair and thus require special attention [@martinez2019fairness].
Conclusion
==========
In this paper we have highlighted a number of issues related to extracting explanations from AI and machine learning models that use imputation of missing data. If such issues are not surfaced to the end user then acting upon explanations from such models can lead to dire consequence in terms of human safety and well being. The case of incorrect explanations in the healthcare domain is especially acute. At the minimum use of explanations at the local instance level should be accompanied by appropriate disclaimers and end users should be properly educated about the potential hazards of incorrect explanations. The other alternative is to take an engineering approach to explanations and assume that assurance of AI systems takes precedence over explainability. Just as one’s choice of algorithm for prediction problems needs to take into account the risk profile of the prediction, the same should also apply for imputation.
[10]{}
Muhammad A Ahmad, Carly Eckert, Greg McKelvey, Kiyana Zolfagar, Anam Zahid, and Ankur Teredesai. Death vs. data science: Predicting end of life. In [*Thirty-Second AAAI Conference on Artificial Intelligence*]{}, 2018.
Muhammad Aurangzeb Ahmad, Carly Eckert, and Ankur Teredesai. Interpretable machine learning in healthcare. In [*Proceedings of the 2018 ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics*]{}, pages 559–560. ACM, 2018.
Melissa J Azur, Elizabeth A Stuart, Constantine Frangakis, and Philip J Leaf. Multiple imputation by chained equations: what is it and how does it work? , 20(1):40–49, 2011.
Richard Berk, Hoda Heidari, Shahin Jabbari, Michael Kearns, and Aaron Roth. Fairness in criminal justice risk assessments: The state of the art. , page 0049124118782533, 2018.
Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In [*Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining*]{}, pages 785–794. ACM, 2016.
Federico Cismondi, Andr[é]{} S Fialho, Susana M Vieira, Shane R Reti, Jo[ã]{}O MC Sousa, and Stan N Finkelstein. Missing data in medical databases: Impute, delete or classify? , 58(1):63–72, 2013.
Setsuro Ebashi, Takeyuki WAKABAYASHI, and Fumiko EBASHI. Troponin and its components. , 69(2):441–445, 1971.
David Gunning. Explainable artificial intelligence (xai). , 2, 2017.
Roderick JA Little and Donald B Rubin. , volume 793. John Wiley & Sons, 2019.
Yuan Luo, Peter Szolovits, Anand S Dighe, and Jason M Baron. 3d-mice: integration of cross-sectional and longitudinal imputation for multi-analyte longitudinal clinical data. , 25(6):645–653, 2017.
Fernando Mart[í]{}nez-Plumed, C[è]{}sar Ferri, David Nieves, and Jos[é]{} Hern[á]{}ndez-Orallo. Fairness and missing values. , 2019.
Sendhil Mullainathan and Ziad Obermeyer. Does machine learning automate moral hazard and error? , 107(5):476–80, 2017.
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In [*Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining*]{}, pages 1135–1144. ACM, 2016.
Lloyd S Shapley. A value for n-person games. , 2(28):307–317, 1953.
Megha Srivastava, Hoda Heidari, and Andreas Krause. Mathematical notions vs. human perception of fairness: A descriptive approach to fairness for machine learning. , 2019.
Adrian Weller. Challenges for transparency. , 2017.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Early detection of melanoma is difficult for the human eye but a crucial step towards reducing its death rate. Computerized detection of these melanoma and other skin lesions is necessary. The central research question in this paper is *“How to segment skin lesion images using a neural network with low available data?”*. This question is divided into three sub questions regarding best performing network structure, training data and training method. First theory associated with these questions is discussed. Literature states that U-net CNN structures have excellent performances on the segmentation task, more training data increases network performance and utilizing transfer learning enables networks to generalize to new data better.
To validate these findings in the literature two experiments are conducted. The first experiment trains a network on data sets of different size. The second experiment proposes twelve network structures and trains them on the same data set. The experimental results support the findings in the literature. The FCN16 and FCN32 networks perform best in the accuracy, intersection over union and mean BF1 Score metric. Concluding from these results the skin lesion segmentation network is a fully convolutional structure with a skip architecture and an encoder depth of either one or two. Weights of this network should be initialized using transfer learning from the pre trained VGG16 network. Training data should be cropped to reduce complexity and augmented during training to reduce the likelihood of overfitting.
author:
- |
Beril Sirmacek, Max Kivits\
Robotics and Mechatronics\
University of Twente\
Enschede, The Netherlands\
`[email protected]`\
bibliography:
- 'template.bib'
title: Semantic Segmentation of Skin Lesions using a Small Data Set
---
Introduction {#ch:introduction}
============
A skin lesion is defined as damage or abnormality found in the skin, for instance chronic skin diseases like psoriasis or melanoma, a type of skin cancer. For melanoma early detection is crucial as it can be used to avoid most deaths caused by this type of cancer, and most skin cancers in general [@MelanomaFacts][@Geller2015FocusDeaths]. When detected in an early stage a simple excision is a curative treatment and the 5-year relative survival rate is 97%. This drops to less than 50% for stage III melanoma, decreasing exponentially with each subsequent stage [@SurvivalStage].
Visually early stage melanomas are difficult to distinguish from their benign counterparts. This leads to many missed or false biopsies, causing great personal and financial harm. The importance of early and accurate melanoma detection is obvious. Experts with specialized equipment can very accurately diagnose even very early stage melanomas. However, this equipment is very expensive and not enough experts are available. To solve this problem digital skin imaging techniques have been developed, increasing the efficiency and accuracy of diagnosis and reducing the need for expert personnel.
In this paper such a digital skin imaging technique is presented. A neural network is designed to semantically segment images of skin lesions, labeling each pixel as either ’Lesion’ or ’Skin’. A data set has been generated at the University of Twente (UT). This set contains 41 images taken by a regular smartphone camera. These images display different body parts containing skin lesion area. Machine learning networks require large amounts of data to train. State of the art networks train for weeks on ImageNet, a database containing over 14 million images [@HeDeepRecognition] [@Russakovsky2015ImageNetChallenge]. Compared to ImageNet the size of the data set provided by the University is exceptionally small.
Machine learning is a rapidly growing discipline. The number of fields machine learning can be successfully applied in is only limited by the amount of available data in those fields. In an attempt to find solutions to this problem and construct a functioning network for the UT, maximizing performance with a small data set is the central theme of this paper. The research question is: *“How to segment skin lesion images using a neural network with low available data?”*
In order to better understand the problem and focus research multiple sub questions are formulated:
1. How does training data impact specific and general segmentation performance?
2. How does network structure impact specific and general segmentation performance?
3. How does training method impact specific and general segmentation performance?
In order to give insight into these sub questions two experiments are conduced. First a network is trained on different data sets. To find the best performing network their segmentation performances are compared. The second experiment compares the segmentation performance of different network structures trained on the same data set. To drawn conclusions their performances are compared. The second experiment also yields observations regarding the third sub question. Some of the network stuctures trained in this experiment have weights partially transfered from the pre trained VGG16 network. Conclusions on the impact of training methods are drawn from the comparison of these networks to randomly initialized networks.
The second section gives background information regarding segmentation and neural network theory, discussing the optimal structure for the segmentation task. Overfitting and strategies to prevent it from occuring are discussed. The difference between randomly initializing network weights and transfering them from pre-trained networks is explained. Strategies to deal with a sparse data set are given. In the final part of this section the different metrics used to compare network performance are explained.
The third section discusses the experimental setup of the conducted experiments. The network structures used are given. Information of the metrics by which performance will be measured is also provided. Two publicly available data sets are used in this paper. All used data sets are listed below and specified in more detail in the third section \[ch:method\].
- 41 Skin lesion images provided by the University of Twente
- 126 Skin lesion images made publicly available by E. Nasr et al. [@Nasr-EsfahaniDenseSegmentation]
- 10 Skin lesion images publicly available in the ISIC Archive [@ISDISISICArchive].
Section four lists all the results obtained in the experiments done in this paper. Discussion of these results can be found in section five. In this section, observations on the experimental results are made and conclusions regarding the research sub questions are drawn.
Section six summarizes the observations made in section five and concludes the paper by answering the main research question.
Background {#ch:theory}
==========
Semantics are the relation between words and symbols to their respective meanings. In computer vision, semantic segmentation is the practice of labeling each pixel of an image in order to obtain regions within that image that share the same high level features. For instance mapping each pixel of a driving scene image to either road, cars or pedestrian. Image segmentation is a fundamental task in computer vision because it enables machines to extract contextual information and make decisions based on it.
Traditionally semantic segmentation is done by means of pre processing methods like thresholding, clustering or edge detection [@Celebi2009LesionImages.] [@Sumithra2015SegmentationDiagnosis]. An extremely valued property of segmentation algorithms is their robustness to input variability. The aforementioned “handcrafted” methods usually struggle greatly in this area and require manual attention when input variations occur.
Contemporary segmentation methods instead leverage deep learning techniques which are more resilient and effective. Steady increase in processing power and availability of labeled data have caused an increase in interest in the machine learning field. Deep learning models based on neural networks have outperformed traditional methods in the annual ImageNet Large Scale Visual Recognition Challenge (ILSVRC) since 2012 [@Russakovsky2015ImageNetChallenge]. Today machine learning is used extensively in many fields such as computer vision, marketing and drug discovery [@Ching2018OpportunitiesMedicine].
Neural Network Layers {#sc:networkLayers}
---------------------
Artificial Neural networks (ANNs or simply “networks”) are composed of connected layers of computational nodes called neurons that work together to optimize the network output. The first network layer is called the input layer. It takes the network input, which are usually a vector or matrix. The following layers are called hidden layers and the final layer is the output layer.
Connections between neurons are weighted. Each layer performs an operation on the output of the previous layer. A network input changes the neuron activations in the input layer. Neuron activations then propagate through the network layers until the output layer is reached. An optimization algorithm compares the network output with the desired output and determines the optimal weights for the neuron connections. This optimization process is called network training and is responsible for an ANNs’ ability to ’learn’.
A lot of different neural network structures exist. A common type of computer vision network is the classification network. This network tries to classify an input image by calculating the probabilities of all the classes it is trained on. A general classification network is given in figure \[fig:basicClassNet\], this particular network is designed to classify digits from the MNIST data set. Most state of the art performances in computer vision challenges are achieved using convolutional neural networks(CNNs). These are a special type of network that use convolutional layers. Although naming convention suggests otherwise, these layers use 2D cross-correlation operations to extract image features from their input. This operation is defined as:
$$Y_{i,j} = \sum _{\delta_{i},\delta_{j},k}^{} X_{i+\delta_{i},j+\delta_{j},k} \cdot K_{\delta_{i},\delta_{j},k}$$
For output $ Y_{i,j} $ at position $i,j$, with input $ X_{i,j}$, filter kernel $K$, input dimension $k$ and x,y filter length $\delta{i}$ and $\delta{j}$ respectively. A filter or kernel is moved over the entire input vector, producing a receptive field. The output of the layer is a dot product of the receptive field and the kernel. A visual representation of this operation is given in figure \[fig:basicConv\]. The step size of this operation is defined as its stride. Before network training kernel size and stride are defined. Generally the kernel is randomly initialized.
![Commonly used classification network structure, taken from [@PeemenLNCSPlatform][]{data-label="fig:basicClassNet"}](images/BasicSegnet.jpg){width="\textwidth"}
![Visualization of 3x3 Cross-correlation Operation []{data-label="fig:basicConv"}](images/basicConv.JPG){width="325pt"}
Another commonly used layer in CNNs is the batch normalization layer. This layer normalizes the output of previous layers by subtracting their batch mean and dividing by their variance, as seen in figure \[fig:batchNorm\]. This is useful to reduce the influence of extremely high neuron activation values on the rest of the network. This enables networks to use a higher learning rate without the risk of overfitting and allows networks to reach a more optimal trained state. Only normalizing a layer can change the information it is representing, to compensate for this the total transform function of this layer has to be able to represent the identity transform [@Ioffe2015BatchShift]. In order to do this a learnable shift parameter $\beta$ and a scale parameter $\gamma$ are added to the batch normalization transform.
![Batch Normalizing Transform, applied to activation $x$ over a mini-batch, taken from [@Ioffe2015BatchShift][]{data-label="fig:batchNorm"}](images/batchNorm.jpg){width="275pt"}
Activation layers are usually implemented after normalization layers. These layers transform their input by applying an activation function to them. Inspired by biological neurons, activation functions aim to mimic the threshold potential during neural excitement. These functions are usually non-linear. An example is the Sigmoid activation function, figure \[fig:Sigmoid\]. This function bounds the input to either 0 or 1 at very small or high activation levels respectively. This function was heavily used but recently the rectifier linear unit or ReLu, figure \[fig:ReLu\], is gaining popularity. Compared to the Sigmoid, ReLu is computationally more efficient. It also solves the Sigmoids’ vanishing gradient problem. This is a problem that occurs during training of the network where the Sigmoid function causes the optimization algorithm to stall.
[200pt]{} ![Activation Functions[]{data-label="fig:ActivationFunc"}](images/Sigmoid.png "fig:"){width="180pt"}
[200pt]{} ![Activation Functions[]{data-label="fig:ActivationFunc"}](images/ReLu.png "fig:"){width="180pt"}
Another extremely common layer in CNNs is the pooling layer. This layer reduces spatial dimensions by downsampling or ’pooling’ its input and producing a single output from it. The Max Pooling layer is the most common, illustrated in figure \[fig:MaxPool\]. Pooling servers two purposes. It reduces computational difficulty by reducing the number of parameters in the network. Pooling also aids in object detection by generalizing high resolution data to lower resolution information [@AndyThomas2017ConvolutionalLearning], illustrated in figure \[fig:PoolDemo\].
![Visualization of a 2x2 Max Pooling operation []{data-label="fig:MaxPool"}](images/MaxPool.png){width="275pt"}
![Visualization of Generalization when combining Convolutional and Max Pooling layers [@AndyThomas2017ConvolutionalLearning][]{data-label="fig:PoolDemo"}](images/PoolingDemo.jpg){width="300pt"}
Overfitting & Underfitting {#sc:overfitting}
--------------------------
Overfitting is defined as: “The production of an analysis which corresponds too closely or exactly to a particular set of data, and may therefore fail to fit additional data or predict future observations reliably”, from the Oxford dictionary [@OverfittingDictionaries]. Underfitting occurs when a statistical model assumes a relationship between data points that is too simple. Figure \[fig:overfit\] shows a visualization of over- and underfitting a polynomial on a small data set. MSE is the mean squared error on the validation set [@UnderfittingOverfitting].
![Under- and Overfitting a polynomial curve[]{data-label="fig:overfit"}](images/overfit.png){width="420pt"}
Overfitting in neural networks manifests as poor performance on the test set compared to the training data set. This signifies a loss of generalization. Using a small amount of training data significantly increases likelihood of overfitting. Increasing network width has been shown to reduce the effects of overfitting [@Zagoruyko2016WideNetworks]. The generalization effect caused by the combination of convolution and pooling layers discussed in section \[sc:networkLayers\] also combats overfitting.
Regularization or *Weight Decay* is an effective method to reduce overfitting [@Schmidhuber2014DeepOverview]. Many different regularization techniques exist. Most work by adding a penalty term to the loss function of a network. Two common and similar techniques are L1 and L2 regularization. L2 or *Ridge Regression* adds a squared magnitude term of the weights vector. L1 or *Lasso Regression* adds only an absolute value term of the weights vector. Compared to L1, loss functions with a L2 term are more influential to statistical outliers.In practice L1 and L2 are commonly used when training on large amounts of data.
Early stopping is an often implemented regularization technique. Up to a point the training process improves the networks performance on data outside the training and test sets. Past this point any improvements made on the training and testing set comes at the expense of network generalization performance. The aim of early stopping is to stop network training at this turning point by providing certain conditions. An often used early stopping condition is requiring improvement in validation accuracy within the last X validation performance tests.
Dropout is another effective regularization technique that aims at reducing overfitting by preventing complex co-adaptations on training data. Networks with dropout ignore a ratio of neurons activations during training proportionate to its dropout rate. Batch normalization is often favored over dropout in convolutional architectures. [@Ioffe2015BatchShift].
Network Structure
-----------------
A network structure hat performs very well on segmentation tasks is the deconvolutional or U-net structure [@Ronneberger2015U-Net:Segmentation]. This structure is build up from a combination of repetitive sections. One section usually contains a combination of convolution, batch normalization and activation layers followed by a pooling layer. An example section is given in figure \[fig:section\]. U-net structures combine these sections to form an encoding and decoding structure, as illustrated in figure \[fig:segNetStruct\]. Encoding is the descending part of the structure. This aims at extracting low level information like textures by reducing resolution using convolution and pooling operations. Decoding is the ascending part of the structure. It acts as the transpose of an encoding layer, increasing the resolution until reaching the input image size. In the decoding structure all convolution layers are replaced by deconvolution layers that perform a transpose convolution operation illustrated in fig \[fig:transConv\]. All Pooling layers are replaced by layers performing upsampling operations. Nearest neighbour and bilinear interpolation are the two most common upsampling algorithms, illustrated in figure \[fig:upsampling\].
![Example of multiple layer segmentation network structure section[]{data-label="fig:section"}](images/block.png){width="300pt"}
![Visualization of transpose convolution operation, taken from [@Dumoulin2016ALearning][]{data-label="fig:transConv"}](images/transConv.png){width="300pt"}
![Visualization of upsampling algorithms[]{data-label="fig:upsampling"}](images/upsampling.png){width="300pt"}
U-net structures connect these encoding and decoding parts by means of a ’skip architecture’, visible as grey arrows in figure \[fig:segNetStruct\]. The intuitive reasoning behind implementing these connections is to combine semantic information from deep layer feature maps with environmental information from more shallow layers. This forces the network to look at environmental information in hopes of creating a network that is able to process contextual information. The total decrease of resolution within a network is defined as network depth. Every network layer also has a width, defined as the total amount of neurons within that layer.
![High level U-net segmentation network structure, taken from [@Ronneberger2015U-Net:Segmentation][]{data-label="fig:segNetStruct"}](images/segNetStruct.png){width="300pt"}
Although U-net structures generally perform very well on segmentation tasks it gives no further specific details like network depth, convolution size or network structure section composition. These specifications are usually called network parameters or *Hyper-parameters*. Optimal network parameters depend on the task they are required to perform. Determining the optimal network depth is subject of much debate, with opinion split between favoring either network depth or width. Increase of either has been proven to yield increased network performance [@Simonyan2015VERYRECOGNITION] [@Zagoruyko2016WideNetworks]. Choosing network depth and width is always a consideration between performance and computational efficiency. An optimum between these two attributes is generally found by experimentation.
Image input resolution is another parameters to consider. Higher input resolutions allow the network to detect features more accurately. This can cause a very significant increase in network performance. A major drawback of increasing input resolution is the major increase in computational cost. An increase input resolution causes a proportional increase in neurons in the input and subsequent layers. This greatly increases the number of parameters of the network and the computational cost of the training process. Another parameter that is not directly defined by the U-net structure is the kernel size in convolution layers. Kernel size determines the resolution at which a convolution layer searches for features. Kernel size can range from 1x1 to the entire input resolution. 1x1 kernels look for features in singular pixels, disregarding any information contained in surrounding pixels. Kernels spanning the entire input range take all pixels into account when looking for features, likely skipping important details. In recent years researchers have started to favor 3x3 kernels over bigger kernels [@KrizhevskyImageNetNetworks] [@Simonyan2015VERYRECOGNITION] for image processing tasks.
In practice the optimal kernel size is found by experimentation. This extends to many Hyper-parameters. To my knowledge there exists no generic algorithm to *a priori* determine the optimal network structure for a given problem. Change in a network parameter heavily influences other network parameters. There are however a few general parameters optimization techniques. A commonly used strategy is starting with the simplest possible network and using experimental validation to determine the optimal level of network complexity. Experiments are often done with the dropout parameter discussed in section \[sc:overfitting\]. Experiments on network depth and width are also common practice. The metric used in these experiments is the ratio of network performance to computational cost, with more optimal networks scoring higher.
Training Methods
----------------
The training process is responsible for a networks’ performance. Network training aims to minimize a loss function associated with the network. This loss function depends on neuron weights and other learnable parameters. This function is generally composed of an error and a regularization term. The error term evaluates the performance of a network by comparing the output to the training data set. The regularization term prevents overfitting by forcing a network to favor less complex solutions over more complex ones.
Figure \[fig:lossSimple\] contains a hypothetical loss function $f(w)$. This loss function only depends on two weights: $W_{1}$ and $W_{2}$. The training process aims at reaching the global minimum of the loss function, in this case the point $W^{*}$. At any point A the first and second derivatives of the function can be calculated. They can respectively be written as the gradient vector $\nabla _{i}\textup{f(w)}$ and Hessian matrix $H_{i,j}\textup{f(w)}$:
$$\nabla _{i}f(w)= df/dw_{i}\qquad (i = 1,...,n)$$
$$H_{i,j}f(w) = d^2f/dw_{i} \cdot dw_{j} \qquad (i,j = 1,...,n)$$
Most practical networks have a lot more learnable parameters and contain many non-linearities, complicating the loss function. Figure \[fig:lossHard\] serves to illustrate this, displaying the loss landscape of the very deep ResNet-110 network [@HeDeepRecognition]. Due to the complex nature of loss functions and the non-linearities within them it is not possible to find closed training algorithms to obtain the minima. Instead weights are often initialized randomly and the network is trained using iterative methods. Finding the global minima is still extremely difficult because of the many local minima present in more complex loss functions. Often training stops when a relatively small local minima is reached.
Gradient or stochastic gradient descent (SGD) is the most commonly used iterative algorithm. Each iteration is formulated as equation \[eq:gradientDescend\], starting at point $W_{0}$. Each iteration the gradient vector is calculated and a step is taken in the reverse direction of the gradient $g$. Since the gradient vector gives information of the direction of quickest ascend of a function, moving in the reverse direction would yield the quickest descend. The size of the step taken every iteration is based on the training rate $\eta$ which usually decreases in size as training progresses. Gradient descend is a first order algorithm as it only requires information of the gradient vector, reducing computational cost. This difference causes gradient descend to be favored when training large networks.
A common addition to gradient descent is momentum. SGD with momentum updates the weights of a network it is optimizing based on a moving average of past iterations. Instead of updating weights every iteration. This addition almost always increases performance.
$$\label{eq:gradientDescend}
W_{i+1} = W_{i} - g_{i} \cdot \eta_{i},\qquad i=0,1,...$$
[200pt]{} ![Loss functions[]{data-label="fig:ActivationFunc"}](images/lossSimple.png "fig:"){width="180pt"}
[200pt]{} ![Loss functions[]{data-label="fig:ActivationFunc"}](images/lossHard.png "fig:"){width="180pt"}
Another training algorithm is the Newton’s method. This is a numerical analysis method can be used to find 0 crossings of functions. In network training it functions as a second order algorithm, requiring information of both the gradient vector and the hessian matrix. Each iteration follows equation \[eq:newton\], with $H_{i}$ the Hessian of the loss function $f(w_{i})$. The main advantage of second order methods is a significant increase in convergence speed. A disadvantage of these methods is their information requirement of the Hessian, increasing computational cost. There exits other algorithms like the Conjugate gradient and Quasi-Newton method that aim at reducing this computational cost whilst still making accurate steps.
$$\label{eq:newton}
W_{i+1} = W_{i} - (H_{i}^{-1} \cdot g_{i}) \cdot \eta_{i}, \qquad i=0,1,...$$
In CNNs that process images, the first few convolution layers usually develop similar kernels during training. These kernels are called Gabor filters and appear to be similar to the filter response of human cortical cells [@Marcelja1980MathematicalCells]. This similarity between image processing networks suggests the feasibility of a training method called transfer learning. Evidence shows that copying weights from a pre trained network, instead of randomly initializing them, is beneficial to network performance. Even when the pre trained network is designed for very different input data and only a few layers are transferred [@YosinskiHowNetworks].
Data manipulation
-----------------
Training a network on a bigger data set is an effective way to reduce overfitting and increase its generalization performance. Whether there is an upper limit to this effect or not is a topic of discussion. Papers written by A. Halvey and M. Banko et al. claim data set size is the main limiting factor of statistical model effectiveness [@2009TheData] [@BankoScalingDisambiguation]. Z. Xiangxin et al. have a contrasting opinion and conjecture that “the greatest gains in detection performance will continue to derive from improved representations and learning algorithms that can make efficient use of large datasets.”[@ZhuDoData].
In any case it remains beneficial to acquire more training data. Qualitative and labeled data is seldom publicly available. Another method of acquiring more training data is using data augmentation. Existing training data can be scaled, rotated, translated or otherwise transformed to produce new data. Training a network on augmented training data increases its generalization performance by forcing it to learn from contextual information.
Another data augmentation technique is adding different kinds of noise to training data. This almost always increases network performance and sometimes even causes a network to converge faster[@LuoDeepNoise].
A third strategy is to generate more data using *Generative Adversarial Networks* (GANs). GANs are a machine learning model used to generate images from noise. GANs can be trained to generate images similar to the data they are trained on. Transfer learning can be applied to a pre trained GAN, configuring it to generate useful training images. This is a very effective method of increasing training data set size and increase network performance, according to L. Sixt et al.[@SixtRENDERGAN:DATA].
Pre processing
--------------
Using filters to enhance network training data is another effective data augmentation technique. The goal of this technique is to increase visibility of textures and other features, reducing the difficulty of training a network to identify those features.
Performance Metrics
-------------------
Segmentation performance of networks can be measured using different metrics. One of these metrics is the confusion matrix, displayed in figure \[fig:TPFN\]. Another useful metric is label, mean and weighted mean accuracy. Equation \[eq:labacc\] shows how label accuracy is calculated, with P total number of actual positives. Mean accuracy is the average accuracy of all labels. Weighted accuracy is the weighted accuracy of all labels.
![A Confusion Matrix[]{data-label="fig:TPFN"}](images/TPFN.png){width="200pt"}
$$\textup{ACC} = \frac{\textup{TP}}{\textup{P}}
\label{eq:labacc}$$
Intersection over Union (IoU) is a commonly used metric. It is the ration of the intersection of the actual and predicted label to their union. IoU is visualized in figure \[fig:IoU\]. When looking at a network with a large accuracy its results can still look quite far from optimal. This deceiving effect of the accuracy metric is why IoU is often used when comparing segmentation networks.
![Visualization of the Intersection over Union metric, from [@IntersectionPyImageSearch][]{data-label="fig:IoU"}](images/IoU.png){width="200pt"}
A third metric is the F1 score. It uses the notion of precision and recall. Precision is a measure of the number of true positives to the number of total perceived positives. It is a measure of the amount of false positives a system generates. The precision metric increases in importance when the consequences of false positives are very negative. Recall is the number of true positives divided by the number of true positives and false negatives. A low recall indicates a large number of false negatives.
Precision (PPV) and recall (TPR) are calculated as: $$\textup{PPV} = \frac{\textup{TP}}{\textup{TP+FP}}
\label{PPV}$$ $$\textup{TPR} = \frac{\textup{TP}}{\textup{TP+FN}}
\label{TPR}$$
F1 score is the harmonic mean of precision and recall. This gives information of the balance a system strikes between both metrics. $$\textup{F1} = 2 \cdot \frac{\textup{PPV} \cdot \textup{TPR}}{\textup{PPV+TPR}}
\label{PPV}$$
Methods {#ch:method}
=======
In order to answer the research question and sub-questions Two experiments are conducted. The Training Data experiment serves to examine the impact of training data on specific and general segmentation performance. The impact of network structure and training method is examined in the Network Structure experiment. Figure \[fig:method\] shows a block diagram of the experimental setup used in both experiments.
![Block Diagram of Experimental Setup[]{data-label="fig:method"}](images/method.png){width="420pt"}
All preprocessing, network training, analysis and post-processing are done using the Matlab R2018b software [@MathWorksMathWorksMATLAB]. All experiments are done on a desktop computer, specifications are listed in table \[tab:specs\]
**Component** **Specification**
--------------- -----------------------
CPU i5 6600k @3.5GHz
RAM 16GB @1200MHz
GPU GTX 970 4GB Vram
OS Windows 10 pro 64-bit
: Specifications of the hardware used in this experiment[]{data-label="tab:specs"}
Data Sets {#sc:dataset}
---------
The definition of a skin lesion is very broad and no additional information is given regarding the type of lesions the network should be optimized for. Because of this, combined with the scarcity of properly diagnosed skin lesion data sets, the network is trained to detect general skin lesions. This allows the use of publicly available skin lesion data sets, increasing the amount of training data available.
The data used in these experiments is gathered from three different sources. All images are RGB. he DFCN data set is made publicly available by the E. Nasr et al. [@Nasr-EsfahaniDenseSegmentation]. The ISIC data set is taken from the ISIC Archive under public CC0 license. The UT image set is built at the University of Twente. Labels for this image set are made using the MatLab [Image Labeler](https://nl.mathworks.com/help/vision/ref/imagelabeler-app.html). Data set information can be found below in table \[tab:dataset\].
**Data set** **Author** **Used Images** **Image Size** **Reference** **Link**
-------------- ---------------- ----------------- ------------------------------ ----------------------------------- -----------------------------------------------------
UT Lab 41 2560 x 1920 $\times$ $\times$
DFCN E. Nasr et al. 126 600 x 400 [@Nasr-EsfahaniDenseSegmentation] [DFCN Github](https://github.com/ebrahimnasr/DFCN/)
ISIC ISDIS & IDS 10 \[1022-2592\] x \[767-1944\] [@ISDISISICArchive] [ISIC Archive](https://www.isic-archive.com)
: Data set Specifications[]{data-label="tab:dataset"}
The Dermnet, DermnetNZ, Dermquest and the Brazilian Dermatology Atlas Databases were contacted without reply. The ISIC database [@ISDISISICArchive] contains vast amount of images. However due to time and computational power constraints only a small subset is used.
In order to reduce computational cost all data is resized to 360 x 480 prior to training. Labels are grayscale files. Each label pixel is assigned a labelPixelID of 0,1 or 2 representing “Background”, “Skin” or “Lesion” respectively. Compared to the other dermatological data sets, our smartphone images from the UT data set contain significantly more background. Training the network to be able to separate background from useful input would significantly increase training difficulty. To combat this during network training all image pixels that are classed as “Background” in the corresponding label are disregarded. An example image label pair and their overlay is given in figure \[fig:imglab\].
![Overlay of a Label and corresponding Image[]{data-label="fig:imglab"}](images/overlay.png){width="420pt"}
During network training images are augmented in order to reduce overfitting. Augmentation details can be found in table \[tab:augment\]. The “Skin” label is much more common compared to the “Lesion” label. This causes the network to favor optimizing for the “Skin” label over the “Lesion” label which reduces segmentation performance. In order to adjust for this imbalance class weighting is used in the final network layer. Class weights are calculated by dividing class pixels by total pixels averaged over all labels.
**Augmenter** **Range**
------------------------ ----------------------
Random X, Y Reflection True
Random X Translation \[-100 100\]px
Random Rotation \[-30 30\]$^{\circ}$
Random X, Y Scaling \[0.75,1.5\]
: Data Augmentation Settings[]{data-label="tab:augment"}
Training Data Experiment
------------------------
This experiment serves to answer the first subquestion:
1. How does training data impact specific and general segmentation performance?
Figure \[fig:SGN3\] shows a plot of the network structure used in this experiment. This SGN3 network is a basic U-net structure with an encoder depth of 3 with 2 convolutional layers per depth level. It is trained from scratch thrice on three different data sets. All other parameters are kept the same. 70% of the data set is used to train the network, the remaining 30% is used as test set. After training specific performance metrics are measured on the test set. The ISIC data set is not used for training. Instead it serves as a hold-out set. After training network performance is tested on this set, which serves as an indicator of network generalization performance.
![Plot of SGN3 Network Structure used in the first experiment[]{data-label="fig:SGN3"}](images/lgraphSGN3_Structure.png){width="420pt"}
Table \[tab:trainingdata\] contains data set specifications of the three sets used in this experiment. The data in the *cropped* Training Data experiment contains cropped versions sampled from the original UT data set. Each original image in the UT set is sampled 10 times. Each image contains a crop of different, or a combination of, interesting regions.
This data is further enhanced for the *augmented* Training Data experiment. Histogram equalization, median filtering and edge enhancement pre processing filters are applied to each training images. Five types of noise are applied ten times to a copy of each image for a total of 50 additional images per original image. Table \[tab:noisespec\] contains specifications on the type and strength of noise. $P(x,y,c)$ is the pixel intensity at width $x$, height $y$ and channel $c$. $|P|$ it the average pixel intensity of the image.
**Training Data Experiment** **Data Sets** **Augmentation** **Total Images**
------------------------------ --------------- ---------------------------- ------------------
*original* UT None 41
*cropped* UT + DFCN Cropping 536
*augmented* UT + DFCN Cropping, Filtering, Noise 27336
: Specifications of training data used in the first experiment[]{data-label="tab:trainingdata"}
**Noise Type** **Times applied** **Strength**
--------------------------- ------------------- --------------------------------------------------------------------------------------------------------------------
Gaussian 10 $\sigma^2=0.01, \mu = 0$
Gaussian - Local Variance 10 $\sigma^2=0.01 \cdot \frac{P(h,w,c)}{|P|}, \mu = 0$
Poisson 10 [Poisson algorithm](https://nl.mathworks.com/help/images/ref/imnoise.html#mw_226e1fb2-f53a-4e49-9bb1-6b167fc2eac1)
Speckle 10 Multiplicative with $\sigma^2=0.04, \mu = 0$
Salt & pepper 10 Noise density $\rho= 0.05$
: Specifications of Noise Types in the *augmented* Training Data experiment[]{data-label="tab:noisespec"}
Training options for this experiment are listed in table \[tab:opttrain\]. Stochastic gradient descend with momentum 0.9 is used as the optimization algorithm. A mini-batch size of 1 is very small but necessary to run the script on a lower end computer. Every epoch all images in the training set are shuffled to prevent adaptations of the network to the order in which images are presented. All scripts are going to be shared on the github repository of the first author.
**Training option** **Setting**
--------------------- -------------
Algorithm SGDM
Momentum 0.9
Learning rate 0.003
L2 Regularization 0.0005
Maximum Epoch 100
Mini-batch Size 1
Validation Data Test Set
Validation Patience 10 Epochs
Shuffling Every Epoch
: Training options used in the Training Data experiment[]{data-label="tab:opttrain"}
Network structure Experiment
----------------------------
This experiment attempts to answer the second and third subquestions:
1. How does network structure impact specific and general segmentation performance
2. How does training method impact specific and general segmentation performance?
In this experiment multiple different network structures are trained. Only the structure differs, all other parameters are kept constant. Almost all training options are the same as those used in the Training Data experiment. To reduce computation time the number of maximum epochs have been reduced to 50. To improve network performance results the early stopping condition has been relaxed to 25 validation iterations. All networks are based on the U-net encoder decoder structure. Figures \[fig:SGN1\] to \[fig:SGNVGG16\] show plots of the networks trained in this experiment.
Networks SGN1-6 are straightforward U-net implementations. SGN1 has an encoder depth of 1, SGN6 has a depth of 6. Each encoder section contains two convolution layers with batch normalization, ReLu activation layers and a max pooling layer. Each decoder section receives information from the decoding section below and its corresponding encoding section. The weights of all SGN networks are randomly initialized.
VGG16 and VGG19 are similar networks based on the 16 layer deep VGG16 classification network [@Simonyan2015VERYRECOGNITION]. VGG16 has an encoder depth of 5. Its first two encoding sections have two convolution layers, the last three sections have three convolution layers. VGG19 is based on the pre trained VGG19 network. It is slightly deeper than the VGG16 network. The weights of both VGG16 and VGG19 are initialized using weights of their original counterparts, their decoder weights are initialized randomly.
Introduced by J. Long et al. [@LongFullySegmentation], the FCN network structures are similar to the SGN networks. Each decoder section of the FCN structures take information of the decoder section below and a pooled combination of the decoder section below and the input of the encoding section on the same depth. The encoder depth of FCN32, FC16 and FC8 is 1, 2 and 3 respectively. Compared to the SGN networks the skip connections in the FCN networks pass information earlier. The Weights of the FCN8, 16 and 32 networks are initialized using VGG16 weights.
SGNVGG16 has exactly the same structure as VGG16. It has the same encoder depth of 5 with the first two layers containing two convolution layers per section and the other sections containing three. The only difference are the weights which are all randomly initialized.
All networks are trained on the same data set used in the *cropped* Training Data experiment, see table \[tab:trainingdata\]. This data set contains a total of 536 images.
[200pt]{} ![Plots of SGN1-6 Networks used in the Network Structure experiment[]{data-label="fig:SGNplot"}](images/lgraphSGN1_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of SGN1-6 Networks used in the Network Structure experiment[]{data-label="fig:SGNplot"}](images/lgraphSGN2_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of SGN1-6 Networks used in the Network Structure experiment[]{data-label="fig:SGNplot"}](images/lgraphSGN3_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of SGN1-6 Networks used in the Network Structure experiment[]{data-label="fig:SGNplot"}](images/lgraphSGN4_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of SGN1-6 Networks used in the Network Structure experiment[]{data-label="fig:SGNplot"}](images/lgraphSGN5_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of SGN1-6 Networks used in the Network Structure experiment[]{data-label="fig:SGNplot"}](images/lgraphSGN6_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of network structures used in the Network Structure experiment[]{data-label="fig:ActivationFunc"}](images/lgraphVGG16_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of network structures used in the Network Structure experiment[]{data-label="fig:ActivationFunc"}](images/lgraphVGG19_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of network structures used in the Network Structure experiment[]{data-label="fig:ActivationFunc"}](images/lgraphFCN8_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of network structures used in the Network Structure experiment[]{data-label="fig:ActivationFunc"}](images/lgraphFCN16_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of network structures used in the Network Structure experiment[]{data-label="fig:ActivationFunc"}](images/lgraphFCN32_Structure.png "fig:"){width="200pt"}
[200pt]{} ![Plots of network structures used in the Network Structure experiment[]{data-label="fig:ActivationFunc"}](images/lgraphSGNVGG16_Structure.png "fig:"){width="200pt"}
Performance Metrics
-------------------
Performance of the networks in both experiments is tested on the test and hold-out set. The training data set in the Training Data experiment varies between each of the tested networks. Since the test set is 30% of total training data this means the test set also varies. Networks in the Network Structure experiment all use the same training data and test sets. The hold-out set consists of 10 images from the ISIC database and remains constant throughout all experiments.
Performance on the test set gauges the networks’ specific performance. Performance on the hold-out set shows the networks capability to generalize to new data, or generalization performance. Performance metrics used are the confusion matrix and the label specific accuracy, intersection over union (IoU) and Mean Boundary F1 Score which is the F1 score calculated at label boundaries. This provides information of a networks segmentation performance at lesion boundary regions. This is very useful when aiming to apply post processing to the networks output. Training time is also taken into consideration.
Results and Discussion {#ch:discussion}
======================
How does training data impact segmentation performance?
-------------------------------------------------------
Due to computational constraints the performance of the SGN3 network on the Augmented set can not be tested. The SNG3 network trained on the *original* data is referred to as the *original* network. The SNG3 network trained on the *cropped* data set is referred to as the *cropped* network. Several observations are made. Compared to the *original* network the skin label accuracy on the test set of the *cropped* network is 6.65% better, but the lesion accuracy is 14.42% worse. On the contrary the skin label IoU of both networks are similar but the lesion IoU of the *cropped* network is 82.49% better. The *original* network has better Mean BF1 scores on both the skin and lesion label with 11.53% and 18.02% respectively.
On the ISIC set these differences are less pronounced. On this set the *cropped* network has 3.61% and 2.21% better skin and lesion accuracy. The skin and lesion IoU of the *cropped* network is 3.86% and 8.13% better. The *original* network has better Mean BF1 performance with 2.14% and 3.18% increase over the *cropped* network for skin and lesions. Figure \[fig:ogex\] and \[fig:cropex\] show a comparison of network performance and segmentation examples of the test set.
Figure \[datacomp\] shows a comparison of the Accuracy and IoU averaged over both labels of the *original* and *cropped* network. The *cropped* network outperforms the *original* network in every IoU metric. Accuracy of the *cropped* network is 0.028 worse on the test but 0.019 better on the ISIC set. The *cropped* network generalizes better to new data.
A skin lesion segmentation model needs high precision to function in a clinical setting because of the serious consequences and cost of false positives. With this in mind the *cropped* network arguably outperforms the *original* network, despite slightly lower test set accuracy. The *original* network correctly assigns more pixels but the *cropped* network more often assigns groups of pixels correctly. The *cropped* network also appears more robust to non-uniform skin textures and variations in image lighting. These observations make the *cropped* network more suitable to identify lesion area.
The Mean BF1 scores of the networks do not fully support this conclusion. The *original* network scores better on the test set and ISIC set for both labels. However this difference is small. Averaged over both labels the difference in mean BF1 score is 0.0069 and 0.0072 on the test and ISIC set respectively.
It is important to emphasize the difference in size of the test sets. The test sets consist of 30% of total training data. The test set for the *original* network contains 12 images, the test set for the *cropped* network contains 161. Because of this conclusions drawn from these observations have arguable relevance.
From the experimental results can be concluded that: (1) More training data is related to better segmentation performance. (2) More training data is related to better network generalization.
![Mean label Accuracy and IoU of *original* and *cropped* network[]{data-label="datacomp"}](images/DATACOMP.png){width="420pt"}
[200pt]{}
[200pt]{}
![Test set accuracy and segmentation examples on the *cropped* data set[]{data-label="fig:cropacc"}](images/MeanOGCROP.png){width="420pt"}
How does network structure impact segmentation performance?
-----------------------------------------------------------
When comparing all performance metrics, the FCN32 and FCN16 perform best. FCN32 has a mean label accuracy, IoU and BF1 score of 0.962, 0.810 and 0.551 respectively. FCN16 has 0.961, 0.810 and 0.563 for the same metrics, both on the test set. FCN32 slightly outperforms FCN16 in the accuracy metric. FCN16 has a slightly higher BF1 Score. On the ISIC set FCN32 achieves an mean label accuracy, IoU and BF1 score of 0.759, 0.658 and 0.548 respectively. For each of these metrics FC16 achieves 0.820, 0.701 and 0.509. FCN16 outperforms FCN32 with 8.08% and 6.49% in mean label accuracy and IoU. FCN32 has a 7.68% increased BF1 score compared to FCN16. FCN16 appears to segment more coarsely compared to FCN32. This is likely a result of the difference in network structure. FC16 has an encoder depth of 2, FCN32 has a depth of 1. This supports the intuitive belief that networks with larger encoder depths look for finer grained structures. This is likely the reason FCN32 has a better mean BF1 score but worse accuracy and IoU performance. BF1 score is more heavily influenced by falsely labeled pixels in the label boundary region. Compared to FCN32, FCN16 has coarser segmentation behavior which would result in more erroneously labeled boundary pixels.
Surprisingly FCN8 achieves very poor scores on almost all performance metrics, scoring much lower than the FCN16 and FCN32 networks. FCN8 has a similar architecture to FCN16 and 32 but has one added encoding section. All networks have the same amount of convolution layers. Compared to other metrics, FCN8 scores remarkably well on lesion label accuracy. On the Test set FCN8 scores 0.707 and 0.746 on the skin and lesion accuracy metric. Interestingly the lesion accuracy is higher than the skin accuracy. Only the very shallow SGN1 network has the same result. On the ISIC set these scores decrease to 0.308 and 0.861. The skin label accuracy decreases dramatically with 0.399. However lesion accuracy increases with 0.114. This shows that the significant decrease of performance of the FCN8 network might be explained by class weighting. For the FCN8 structure the magnitude class weighting could have been to high causing the network to heavily favor optimizing for lesion accuracy, decreasing its overall performance.
On average SGN structures perform much worse compared to FCN structures across all metrics, especially when comparing network generalization performance on the ISIC set. This decrease in performance could be explained by the difference in insertion position of the skip architectures. SGN structures use skip architectures that combine information from the decoder section below with the output of the encoder section on the same depth. Skip architectures in FCN structures also use information from the decoder section below but combine it with information from a pooling layer. This pooling layer takes information from the start of the encoder section on the same depth and combines it information of the decoder section below. This combination of less processed information pooled with more heavily processed information are likely the reason FCN structures perform better.
Training time also heavily differs between the SGN and FCN structures. FCN8, 16 and 32 have trained for 4:50, 3:46 and 3:49 respectively. Disregarding the shallow SGN1 and SGN2 networks, FCN networks train on average 48:20 longer compared to SGN networks. Training of the FCN networks also behaved much more erratic, with frequent changes in loss and accuracy during training orders of magnitude bigger compared to the SGN networks. This is likely the reason for the increased average training time since all networks stopped training at some point by achieving the early stopping condition. Training of the FCN networks was more erratic and therefore took longer to achieve the stopping condition. This property of the FCN networks can also indicate a higher variation when revising the experiment.
The VGG structures also perform well. VGG16 has a mean label accuracy, IoU and BF1 Score of 0.936, 0.775 and 0.539. VGG19 scores 0.850, 0.756 and 0.520 in these categories. Interestingly the VGG16 network outperforms its deeper VGG19 counterpart. When looking at individual label accuracy VGG16 scores 0.957 and 0.914 for the skin and lesion label respectively. VGG19 achieves the highest skin label accuracy of 0.993 but a more modest 0.701 for lesion accuracy. This effect is more pronounced on the ISIC set. Label specific accuracy for VGG16 are 0.906 and 0.644 for skin and lesion respectively. VGG19 scores a 0.948 and 0.326 skin and lesion accuracy. This is likely due to the difference in size of the skin and lesion labels. For larger networks the class weighting has not dissuaded the network enough to fit for both classes. Overfitting could also have occured but is less likely due to the excellent skin accuracy of both networks on the ISIC set. These networks likely require either more regularization during training to achieve maximum performance.
Another interesting observation is the effect of encoder depth on the performance of the SGN networks, shown in figure \[fig:encode\]. IoU, Mean BF1 score and skin accuracy on the test set all have upward trends when increasing encoder depth, with maximum performance arguably achieved by the SGN5 network. Interestingly the lesion accuracy has a downward trend with SGN4 performing much worse than SGN1 to 3. Comparing these observations to the performance of the SGN networks on the ISIC set in figure \[fig:encodeISIC\] shows encoder depth is positively related to network performance. Performance of SGN on the ISIC set is much worse compared to SGN4 and 6. This shows SGN5 is slightly overfitting on the test set. The same downward trend in lesion accuracy exists on the ISIC set. This indicates that the training algorithm judges the best weight changes each iteration to be those that reflect a more positive change in skin accuracy in more shallow networks. A deeper network is better able to optimize for lesion accuracy as well.
![Test set performance of SGN1-6[]{data-label="fig:encode"}](images/encoderdepth.png){width="420pt"}
![ISIC set performance of SGN1-6[]{data-label="fig:encodeISIC"}](images/encoderISIC.png){width="420pt"}
Figure \[fig:comp1\] and \[fig:comp2\] show a comparison between the performance of SGN5 and SGNVGG16 on the test and ISIC set. On the test set the SGN5 network performs better compared to the SGNVGG16 network. However the SGNVGG16 outperforms the SGN5 network when tested on the ISIC set. This indicates the SGN5 network is overfitting on the training data and has worse generalization performance than SGNVGG16. The SGN5 and SGNVGG16 networks are very similar. Both have an encoder depth of five. SGN5 has two convolutional layers per encoder section. SGNVGG16 is modeled after the pre trained VGG16 network and has two convolutional layers in the first two layers and three convolutional layers in the last three layers. To explain whether the difference in performance of these networks is due to this difference in structure or to the random nature of network training requires more investigation.
![Comparison of performance of SGN5 and SGNVGG16 on the test set[]{data-label="fig:comp1"}](images/SGNVGGhold.png){width="420pt"}
![Comparison of performance of SGN5 and SGNVGG16 on the ISIC set[]{data-label="fig:comp2"}](images/SGNVGGtest.png){width="420pt"}
From these observations the following conclusions can be drawn: (3) Increasing the number of convolution layers per encoding section is more effective than increasing network encoder depth. (4) When using skip architectures, combining information from a deep layer with relatively unprocessed information from a higher layer is more effective than combining it with more heavily processed information. (5) Training FCN network structures using early stopping regularization increases training time. (6) For SGN structures increased encoding depth is related to increased generalization performance.
How does training method impact segmentation performance?
---------------------------------------------------------
Training a network that uses weights of a pre trained network is called transfer learning. Training a network that uses randomly initialized weights is referred to as training from scratch.
Figure \[fig:methcomptest\] and \[fig:methcomphold\] show a comparison of test and ISIC set performance metrics of the VGG16 and SGNVGG16 networks. The only difference between these networks is the initialization method of their weights, their structures are identical. SGNVGG16 weights are initialized randomly, VGG16 weights are initialized using weights of the well-known VGG16 network. Since the original VGG16 network is a classification network only the encoder weights of the version used in this experiment are transferred. VGG16 decoder weights are initialized randomly.
On the test set, the VGG16 network outperforms the SGNVG16 network in every performance metric except for skin label accuracy. Compared to VGG16 the skin label accuracy of the SGNVGG16 is 0.068 better. When testing on the ISIC set VGG16 performs much better by every metric. The biggest difference is the performance of the networks when labeling lesion pixels. When comparing the ISIC set metrics to the test set metrics VGG16 has a 20.73%, 18.49% and 23.72% reduction in mean accuracy, IoU and BF1 Score respectively. SGNVGG16 has a 37.07%, 46.80% and 35.49% reduction in mean accuracy, IoU and BF1 Score. Figure \[fig:traindecr\] shows this difference. From this can be concluded the VGG16 is better able to generalize to new data. This supports the notion in the literature that suggests transfer learning increases network generalization[@YosinskiHowNetworks].
Another observation is the difference in training time when comparing VGG16 to SGNVGG16. VGG16 took 2:54:29 to train, SGNVGG16 took 3:44:47 using the same training options. This is a significant reduction of 50 minutes and 18 seconds.
The FCN networks also have weights based on the pre trained VGG16 network. As with the VGG16 network used in this experiment only the encoder weights are transferred and decoder weights are randomly initialized. The VGG19 network also uses encoder weights transferred from its pre trained counterpart. All FCN and VGG networks have above average performance metrics. The extent of the influence transfer learning has on performance is difficult to identify. However it is safe to say transfer learning has aided the performance of these networks.
From these observations can be concluded: (7) when comparing transfer learning to learning from scratch, transfer learning yields better specific and general performance even when weights are only partially transferred. (8) The increase of performance by utilizing transfer learning is greater on general performance than specific performance. (9) Utilizing transfer learning significantly reduces training time.
![Test set Performance metrics Comparison of VGG16 and SGNVGG16 networks[]{data-label="fig:methcomptest"}](images/SGNVGG16comptest.png){width="420pt"}
![ISIC set Performance metrics Comparison of VGG16 and SGNVGG16 networks[]{data-label="fig:methcomphold"}](images/SGNVGG16comphold.png){width="420pt"}
![Comparison of the percentage decrease in label-averaged performance metrics between the test and ISIC set[]{data-label="fig:traindecr"}](images/traindecr.png){width="420pt"}
Shortcomings and Clinical Implementation
----------------------------------------
Training neural networks is a stochastic process. Training the same network keeping all hyperparameters constant yields different results every time. In order to validate the conclusions drawn from this experiment insight into the variance of this experiment is necessary. Especially the FCN structures have very erratic loss functions while training, possibly indicating a high variance. Valuable future research would be investigating the variance of the experiments by repeating them multiple times. This is necessary to validate the conclusions drawn in this paper.
Other future research could include experimenting with different training options. Experimental results provided in this paper are obtained using very similar training options. Exploring the impact of training options on network structure is necessary to validate the best performing network structures. SGN structures could possibly outperform FCN structures if given more training time and a more relaxed early stopping condition. More research should be conducted into the underlying reasons of the excellent performance of the FCN networks. This could be further comparison between SGN, FCN and similar network structures. Or research into the significant difference in training behavior of the SGN and FCN structures.
Testing networks on the Augmented sets could unfortunately not be done due to computational constraints. Experimenting with the Augmented set and further research into the impact of training data is necessary to validate the conclusions drawn from the Training Data experiment. Further research could include comparing the performance of different structures trained on a large number of different data sets. To more accurately discern the consequence on network performance of different pre processing filters, a similar structure should be trained on differently pre processed data sets. Researching the consequences addition of noise yields on network performance can also yield valuable insight.
In order for this system to be implemented in a clinical environment it needs additional work. The accuracy of the segmentation achieved in this paper is not enough to pass clinical performance tests and should be improved. This improvement can come from changes in the machine learning structure or from added post processing. This system could be employed to do a rough segmentation of image material and use other segmentation methods like active contouring to achieve a high final segmentation performance.
In order to accurately track a patients lesion over time a method needs to be devised to present the system with an image taken at the same distance from the skin. If the system does not recognize scale the resulting tracking could be very inaccurate due to changes in distance between measurement images. This could be done using any distance measuring tool, for instance a laser that measures distance attached to the camera. Another method to solve this would be to implement a subsystem that relates skin lesion size to other visual clues like freckles, tattoos or other irregularities.
Other functionality that is valuable to add for use in a clinical setting is a severity classifier. This classifier would give insight into the severity of a lesion or tumor by means of shape and texture analysis. Generally non-uniform shapes and textures are related with more active and severe tumors. To further increase accuracy of the severity classification thickness and relative redness analysis functionality can be added. Lesion thickness is difficult to measure and therefore difficult to monitor, making it a useful addition to a clinical lesion model. In order to add this functionality some sort of 3D data should be gathered.
As the first experiment suggests, maximum performance of the networks in this paper can only be achieved when a large set of training is available. If qualitatively labeled data remains sparse. research into different machine learning models should be conducted. Less complex methods like support vector machines can yield better performance on smaller data sets. These methods also reduce the computational cost of network training and inference time. A decrease in training time can aid in development and research. A lower inference time is a valuable property in clinical settings, where real time processing of images is necessary.
When the lack of labeled data turns out to be the most limiting factor of further improvement, research into unsupervised learning methods can be invaluable. Unsupervised learning methods do not require labeled data. Auto encoders are an example of these type of networks. These networks deconstruct training data into a smaller dimensional representation often called code. It then attempts to reconstruct the original data from the code layer and used back propagation methods to improve this process. Because these networks do not require training data to be labeled finding appropriate training data is much simpler. Another potential method of dealing with sparse training data is the usage of Generative Adversarial Networks to generate new training data. Using GANs for training data generation should be researched further.
Conclusion {#ch-conclusion}
==========
To perform semantic segmentation on skin lesion pictures multiple U-net structures are constructed and tested. The FCN32 and FCN16 structures perform best by label accuracy, IoU and Mean BF1 metrics. All networks using weights partially initialized by the VGG16 network perform above average, except for the FCN8 network. The first experiment yields some observations regarding the first research sub question. The SGN3 network trained on a cropped set generalizes better to new data. The impact of training data on specific and general segmentation performance is: (1) More training data is related to better segmentation performance. (2) More training data is related to better network generalization.
The impact of network structure on specific and general segmentation performance is researched by means of the second experiment. Observations made are: (3) Increasing the number of convolution layers per encoding section is more effective than increasing network encoder depth. (4) When using skip architectures, combining information from a deep layer with relatively unprocessed information from a higher layer is more effective than combining it with more heavily processed information. (5) Training FCN network structures using early stopping regularization increases training time. (6) For SGN structures increased encoding depth is related to increased generalization performance.
The second experiment also yields observations regarding the research sub question ’How does training method impact specific and general segmentation performance?’ From these observations can be concluded: (7) when comparing transfer learning to learning from scratch, transfer learning yields better specific and general performance even when weights are only partially transferred. (8) The increase of performance by utilizing transfer learning is greater on general performance than specific performance. (9) Utilizing transfer learning significantly reduces training time.
Combining these conclusions yields insight into the research question “*How to segment skin lesion images using a neural network with low available data?*” Of the methods tested in this paper the best approach is using a FCN32/16 network structure with weights transferred from the pre-trained VGG16 network, trained on a representative data set. This set should be pre processed using cropping to increase the information the network can draw from it during training, especially when the set is small.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'John M. Brewer, Matthew Giguere, & Debra A. Fischer'
bibliography:
- 'ms.bib'
title: 'CHIRON TOOLS: Integrated Target Submission, Scheduling and Observing Systems for a High Resolution Fiber Fed Spectrograph'
---
Introduction
============
CHIRON is a fiber fed spectrometer with spectral resolution up to $R = 137\,000$ designed for extra-solar planet searches [@2010SPIE.7735E.149S; @2012SPIE.8446E..0BS; @Tokovinin:2013:submitted] at the 1.5-meter telescope at Cerro Tololo Inter-America Observatory (CTIO). After initial commissioning, we found that significant effort was required to schedule the queue-based instrument for multiple users with different time sensitive requests. We also found that errors in creating and executing the nightly observing scripts resulted in time losses of up to an hour per night and $\sim 10$% loss of data. It became evident that in addition to the planned automated data reduction pipeline, a new software interface was needed to submit targets, to generate nightly observing scripts, and to execute and log observations in order to take full advantage of the new instrument. We had significantly under-estimated the full scope of work required to commission the spectrometer; the software effort required to make this a facility instrument nearly doubled the effort required to build the instrument.
Historically, astronomers traveled to the telescope to carry out their observing programs. However, that model has changed in the past decade to more economical models where telescope operators execute observing programs or the data are acquired robotically. The 1.5-meter telescope at CTIO is part of the Small to Moderate Aperture Research Telescopes (SMARTS) consortium [@2010SPIE.7737E..31S] that is partly supported by the National Optical Astronomical Observatories (NOAO). Users are either awarded time through the NOAO proposal process or they purchase telescope time from SMARTS on an hourly basis. The data are generally obtained by telescope operators. During the commissioning of CHIRON, shared risk time was opened to the community and there was a large influx of users with time allocations between 5 and 100 hours over the semester. Initially, PIs e-mailed target lists to the scheduler at Yale and nightly scripts were assembled and sent as PDF files to the telescope operator at CTIO. The operator was responsible for taking all calibrations and logging any issues with observing. While still working out kinks in the instrument control system (ICS), missed observations or calibrations were frequent as were simple typos both in creating the observing script and in entering script information into the ICS and telescope control system (TCS).
We developed a software management system with the goal of improving the efficiency and scientific productivity of the 1.5-meter CHIRON queue. Our specific objectives were to reduce the time needed to build nightly scripts (with targets from multiple users and with different time sensitive scheduling requirements) and to reduce the errors in executing the scripts. The full list of goals and requirements of the software can be found in §\[sec:goals\_reqs\]. Rapid deployment of this management system was also critical and this resulted in some software design choices that were narrowly tailored to our particular instrument and queue setup. However, the design pattern can be easily adapted for other facility instruments and we offer our code, Tools for Observing, Operating, Logging and Scheduling (TOOLS), to anyone interested in adapting it for their use.
Goals and Requirements {#sec:goals_reqs}
======================
The project goal was to improve the efficiency and scientific return of CHIRON. The following requirements were set to automate management of the CHIRON queue.
- *Rapid Deployment* The project had to be completed as quickly as possible to reduce the data loss and to handle scheduling allocations for the upcoming semester.
- *Accurate & Complete Target Information* Proper scheduling of a CHIRON observation requires specification of the target name, coordinates, brightness, slit choice, binning, calibrations (including additional observations of calibration targets), iodine cell position, number of exposures, and desired SNR and/or exposure time. E-mailed information often contained typos or incomplete information which resulted in repeated correspondence or incorrect or missed observations.
- *Simplified process for creating Nightly Observing Scripts* As the number of users increased, it began taking up to two hours to schedule and format the observing scripts for each night.
- *Reduction in Observing Errors* The data taking GUI contained several fields that the telescope operator needed to fill for each observation. This was a time-consuming process with many failure modes. Observations were unusable if the slit, binning, I2 cell, or exposure time were not correctly set, coordinates were not entered correctly or if a standard comparison star or the nightly calibrations were skipped. If a typo was made in the object or program name, up to an hour was spent the next day trying to match observations to individual investigator programs. If coordinates were incorrectly typed into the TCS, dark time was lost tracking down this error or the wrong star was observed.
- *Detailed Observing Log* Weather and early system problems also contributed to loss of observations. Communication regarding successful observations and the reasons for missed observations was a high priority and we needed a system that would not unduly burden or distract the telescope operator.
- *Secure Target Information* With the possibility of competing projects, it was deemed important that target information be treated as proprietary for individual programs. Observers are only able to access information about their own programs with CHIRON TOOLS.
- *Automated Data Reduction and Backup* The automated reduction pipeline was integrated with our scheduling queue to save time and eliminate human handling errors. As observations are archived and spectra extracted, the queue system is automatically updated to let scientists know that they can retrieve their raw and reduced data.
- *Use Existing Hardware* For both cost savings and rapid turnaround, we built the software system integrated with the existing MacOS based servers and RAID system at Yale and the Mac and Linux machines used for observing at CTIO. To ensure that the applications would be easily portable, we kept largely to standard open source tools available on all Unix based systems.
- *Intuitive Graphical User Interface* An intuitive GUI was a high priority to reduce errors and training time for the tools.
- *Robust to Internet Outages* We needed a flexible automated system which would also tolerate interruptions in internet connectivity.
Application Design
==================
Hardware and Software {#sec:hardware_software}
---------------------
The Yale Exoplanets web server is a Mac Pro running Mac OS X Server. At CTIO, there is an iMac running the standard version of Mac OS X. The Instrument Control System and Telescope Control Systems are Linux and VxWorks machines. Communication between Yale and CTIO all happens between the two Mac systems. A diagram of the system can be seen in Figure \[fig:app\_layout\]
The software used to construct the web applications is a combination of PHP, SQL, HTML, CSS, Javascript, XML, JSON, and C-Shell and run on a combination of Apache2 and MySQL servers. This is a common setup for web applications and is known as LAMP Linux or MAMP on MacOS where AMP refers to Apache, MySQL, PHP. Bi-directional replication of a portion of the database between Yale and CTIO is handled by means of SSH Tunnels which use the Mac OS X launchd system to dynamically connect when needed and when a connection is available. The xinetd system, prevalent on Linux, could also have been used but launched was a flexible, easy and pre-installed option on our MacOS systems.
Observer Web App
----------------
The first stage of development focused on a web application which allowed users with allocations for a given semester to log in and upload their target information along with specifications for the observations. With many new principal investigators (PIs) per semester, it is important to automate account setup as much as possible. After SMARTS time or the NOAO Time Allocation Committee (TAC) has awarded time to each observer, we use a PHP script to insert new users and add proposal and allocation information into the database. Since all PIs are required to provide an e-mail address with their proposal we allow them to activate their account by providing their name and proposal ID and a link is e-mailed to them to allow them to set a user name and password. The PIs receive a brief tutorial that explains how to fill out the target (package) requests when they are allocated CHIRON time.
Once logged in, the PI can create one or more ’plans’ per semester to organize their target lists. Adding targets is accomplished with a context sensitive form which ensures that only options pertaining to the particular type of observation and slit type are available. Targets can be grouped together as ’packages’ with calibrations or standard star observations and the total time (including approximate targeting and readout) is deducted from their allocation dynamically. Target submission has a cutoff date each semester after which the PI can no longer add or edit targets.
Once scheduling is complete, the PI can monitor the observing plan to see the status of their observations and collect their data. When targets (or packages) are added to a nightly observing script, an annotation is displayed on the PI observing plan page to show the date when the observation will occur. The annotation is automatically updated when any action occurs. If the target has been observed, a link to the raw and reduced data appears. If the target is skipped because of bad weather or mechanical failure, an auto-generated explanation appears. Importantly, all annotations are automatically saved to the PI observing plan as soon as the telescope operator selects a note from a pop-up menu on the observing script. In addition to the annotations, users are notified via e-mail automatically once the spectra are extracted and wavelength calibrated and are ready to be downloaded.
Nightly Observing Scripts
-------------------------
The previous spectrograph on the CTIO 1.5m only had a handful of users per semester. Nightly observing scripts were assembled in an Excel spreadsheet and exported as a PDF for the Telescope Operator. The large increase in the number of users per semester, and per night, required a better system.
When a scheduler creates a new Nightly Observing Script (NOS) for a given night, by default they are shown all targets which have not yet been observed, sorted in RA order. Notes attached by the investigator to both their plan and individual targets are available to the scheduler for reference. The scheduler can click on targets to add them to the script and they are immediately displayed at the top of the page along with any calibrations attached to the target. The script displays the approximate start time and duration of the observations based on the requested integration time, number of exposures, and acquisition and readout times. An additional column in the script shows the scheduler if the target is visible at that point in the night and whether it is rising or setting. They can then choose to re-organize the script by changing the sequence number of any line items.
Once a script is complete, the script is sent to the Interactive Observing Script (IOS) application on the computers in Chile. This process involves copying the necessary information for the telescope operator to records in the database which are replicated to the database at CTIO when there is an Internet connection. If the script needs to be updated at any point, it can be re-sent with each submission resulting in a separately versioned script.
Writing good scheduling software is complex, time consuming, and still usually requires human intervention. The NOS tool has allowed us to eliminate most of the scheduling time without spending a lot on software development. Since both target and script information are stored in the database, it will be possible to add automated scheduling software at a later date and still retain the NOS interface for human tweaking of automatically generated schedules.
Interactive Observing Script {#sec:IOS}
----------------------------
We developed quality control pages to monitor the nightly observations and found that on average, 40% of the observations had an error in the FITS header information. In minor cases, such as a typo in the object name, we could correct the FITS headers. However, correcting FITS headers for the observing queue was outside the scope of work that we could commit for CHIRON. For more significant errors, such as an incorrect slit setting, binning, coordinates, I2 cell or exposure time, the data might not even be useable by the PI. Therefore, we developed an interactive observing script (IOS) for use by the telescope operator. This web-based app simplified observing for the telescope operator by sending target coordinates to the TCS and auto-filling all fields in the data-taking GUI.
We defined an interface and a set of requirements for the IOS web application as well as an application programming interface (API) for a piece of middleware, the Instrument and Telescope Connector which could take command line requests and return XML responses. This allowed the web application developer (Methanie Binder) to develop to the API without knowledge of how the telescope or instrument control software worked. It also allowed later additions such as integration with the exposure meter to take place solely within the middleware without having to call back the web developer.
When a nightly observing script is completed (typically at Yale by the CHIRON queue manager), targets are locked from further changes in the PI observing plan and the NOS is copied into a smaller table structure to reduce the data flow for replication at CTIO. The interactive observing script is auto-generated from the NOS for the telescope operator and includes the formatted list of targets and calibrations with active links to any extended notes provided by the PI when scheduling a target package. Occasional Target of Opportunity (ToO) requests for a different instrument can also be added to the script.
Each line in the IOS has buttons allowing the operator to send or skip the observation or calibration. Selecting a line item sends all relevant information to the middleware which in turn updates the data taker and (if it is a target) the telescope coordinates. Once a line item has been sent, it is possible to ’resend’ in the event of an error. A pop-up menu is included for each line of the IOS so that the telescope operator can quickly tag the reason (weather, instrument failure, other) for skipping particular observations. If the internet connection is active, any information recorded by the telescope operator is automatically replicated back to the master CHIRON database and appears immediately on the PI’s observing plan. If the internet connection is down, the information will be replicated as soon as the connection returns.
Integration with Instrument and Telescope Controls
--------------------------------------------------
As discussed in §\[sec:IOS\], the IOS web application does not attempt to talk directly with the instrument control system (ICS) or the telescope control system (TCS), but instead communicates via a command-line tool which returns XML responses. This segregation avoids the need for a web programmer to make updates when changes are made to the controller software or to the machines that host the ICS and TCS.
Both the ICS and TCS live on separate machines but do not currently have network sockets listening for commands. Instead, tasks such as sending coordinates to the telescope or updating settings in the data taker must run through command line tools on the respective machines. The middleware shell script handles the network communication and encapsulates the syntax for both the ICS and TCS tools. The API consists of a single command, send\_object, with arguments consisting of all information required for both systems and whether the object is a science target or a calibration.
The middleware is launched by the web app and executes remote commands on the ICS and TCS computers as appropriate. Status information is returned to standard out which is read by the web app and returned to the browser. The script continues to monitor the progress of the setting changes every second and writes log information to a file, which can be reviewed later. A status window is launched so that the telescope operator can see when all of the changes are complete. Though the middleware is capable of monitoring the exposures, it currently only returns status information to the web application about the success of setting the ICS and TCS items.
Database Design and Replication {#sec:db_design}
-------------------------------
We designed a relational database to manage all of the information for targets, scripts, logging, and access to the applications. The database can be viewed as separate logical units (Figure \[fig:db\_layout\]). With the exception of the *Script Execution* unit, it was designed as a single normalized[@Codd:1970:RMD:362384.362685] set of tables. The structure and fields are generic to any queue based astronomical observing with time allocation blocks assigned by semester. Information specific to the instrument is contained within the *object* and *decker* (slit and fiber options) tables and the related *nos\_objects* table.
Due to sporadic outages and slow-downs of the Internet connection to the 1.5 meter telescope at CTIO and to avoid impacting the mountain network during observing times, we attempted to limit the traffic needed to link the database on the mountain to the one at Yale. The nightly observing script tables (*nos* and *nos\_objects*) duplicate all information for scheduled targets contained in the main scheduling tables (*scripts* and *script\_objs*) as well as necessary information from linked tables (such as *objects* and *deckers*) in order to reduce the amount of information replicated to CTIO. The simplified structure has the side-benefit of obscuring proprietary observer information at the telescope and includes date-stamps and version information for reference.
Replication of the *Site Access* and *Script Execution* table groups along with the *users* table is accomplished using the built-in MySQL replication over secure SSH tunnels as described in §\[sec:hardware\_software\]. Most of the tables are read-only at CTIO which eased the replication setup. Only the results from the nightly script (*nos\_results* table) can be updated in both locations; this is the only table participating in bi-directional replication. The others are all synced from Yale to CTIO and generally have very low update frequencies.
During an observing run, the IOS application used by the telescope operator reads and writes to the local database on the iMac at CTIO. Log entries are saved to the *nos\_results* table which is replicated on-demand when an Internet connection is present and typically consists of less than a few hundred bytes of data per observation.
Reduction Pipeline
------------------
To take advantage of the lower demand on intercontinental bandwidth at night, data are rsynced from CTIO to Yale on an hourly basis. This allows processing, analysis and distribution to take place the next morning. Data are then processed in IDL using a modified version of the *REDUCE* package created by Piskunov & Valenti . Both the raw and extracted and wavelength calibrated reduced files are in the FITS standard . Files are packaged and compressed into tarfiles and their locations are entered into the CHIRON MySQL database at the end of the pipeline. When users log in to their CHIRON accounts, hyperlinks to the combined raw and reduced tar files appear in each target row, allowing users to download their data.
Conclusions
===========
Reduced Time Requirements
-------------------------
One of the primary reasons for spending the time to develop this software was to reduce the amount of time we were spending on the mundane tasks of observation planning and distributing the data. Initially, construction of nightly observing scripts required up to two hours per night including time spent communicating with observers. Schedulers now see a list of unobserved targets and can click to add them to and re-order them in the script which shows target elevation and time available. The PI portion of the application has already ensured that all necessary data is in the database. After full implementation of the software, script preparation has been reduced to $\sim$15 minutes per night. The system has been easy to maintain, requiring little intervention other than operating system updates.
We have no data on how much time PIs spent exporting their observing lists before the software upgrade. However, we know that $\sim$10% of the data was unusable and repeat observations were required. There have been a few requests for features which would make target submission smoother such as the ability to duplicate observations or entire plans and the ability to upload basic target information (name, RA, DEC, V mag) in bulk. These are sensible upgrades but there is no funding for this effort.
Improved Science Output
-----------------------
The second reason for developing this software was to improve the efficiency and science return of the instrument. Before implementing this software, up to 40% of the observations had some type of error in the FITS header; roughly 10% of data was unusable because of incorrect settings manually entered into the data-taking GUI or because calibration targets were missed; manually typing instrument control settings in the GUI and target coordinates for the TCS took several minutes for every target or up to an hour per night. The new software system has eliminated the instrument setup errors (any errors that exist are due to incorrect target requests by the PI) and there are no longer any lost observations due to missed calibrations. Thus, we estimate that we have recovered nearly two hours of observing time every night.
Lessons learned
---------------
It is worth restating the obvious: in developing a facility instrument, the task of organizing observations, data reduction pipelines and distributing reduced data is a significant effort. We estimate that the time to develop the scheduling software (including the PI user interface, the NOS, the middleware, the IOS) totaled to 6 month FTE effort. This was all an effort that we did not plan for when proposing to build CHIRON. The algorithms provided in this paper can help in planning the scope of software integration.
Since the previous spectrograph was thought to have adequate software, the need for significantly revised software was not recognized until commissioning the instrument. In order to fully exploit the capabilities of a new instrument it is critical to evaluate the entire process from proposal through delivery of reduced data. Identifying and planning for efficiency bottlenecks in the process can be critical to the success of the instrument and should be included in the scope of work for the hardware.
Summary
=======
While commissioning the CHIRON spectrograph at CTIO, challenges with the target collection, nightly planning, and data distribution were identified which led to the development of this suite of software applications. Overall instrument performance was greatly improved by a development effort which was relatively modest in comparison to time spent on the instrument design and construction itself. However, analysis of the new use cases of the instrument prior to commissioning would have further reduced development time and made for smoother commissioning.
We share our experience in building a facility-class software interface and offer our software as a template on which to build similar target submission, scheduling, and observing systems for facility instruments. Due to the focus on rapid deployment, parts of our software are necessarily tied closely to CHIRON and the specifics of the Yale and CTIO systems. The software is available upon request, as-is, but could be adapted to other instruments and settings. To facilitate this, we have carefully segregated instrument specific items in the database and used middleware to abstract the communication of object information to the instrument.
Thanks are due to Marco Bonati for making the data taker remotely accessible and Andrei Tokovinin for helpful discussions and feedback. We thank Michele Beleu for keeping the project on schedule, and Methanie Binder for rapid turnaround on the web design. We also gratefully acknowledge the support of NASA grant NNX12AC01G and NSF grant AST1109727.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We prove the existence of a strong coupling expansion for a classical $\lambda\phi^4$ field theory in agreement with the duality principle in perturbation theory put forward in \[M.Frasca, Phys. Rev. A [**58**]{}, 3439 (1998)\]. The leading order solution is a snoidal wave taking the place of plane waves of the free theory. We compute the first order correction and show that higher order terms do renormalize the leading order solution.'
author:
- Marco Frasca
title: 'Dual perturbation expansion for a classical $\lambda\phi^4$ field theory'
---
Introduction
============
Quantum field theory is generally formulated to grant a proper weak perturbation theory starting from a free field. This approach has been proved largely successful as relevant techniques to cope with infinities and other anomalies following this approach has been devised [@wei1]. Besides, renormalization group methods have permitted to extract non-perturbative results form these theories [@wei2; @zj].
Notwithstanding the long established tradition of a weak perturbation theory, the need for analysis in the strong coupling limit has become increasingly relevant since seventies when the discovery of quantum chromodynamics made the situation very difficult to manage. Bender and colleagues devised an approach for a strong coupling quantum field theory [@bend1; @bend2; @bend3] but this method proved difficult to apply as all the relevant quantities depend on a cut-off present in the perturbation series and difficulties in resumming these series and taking the corresponding limit to zero arose.
Since the pioneering work of Bender and Wu [@bw1; @bw2] a proper framework to test strong coupling theories has been the anharmonic oscillator giving also precious informations about the large order behavior of weak perturbation expansions. Recently, Kleinert showed as the variational method for generating strong coupling expansions from a given weak perturbation series, initially devised for the anharmonic oscillator, produced striking results when applied to a $\lambda\phi^4$ theory to compute critical exponents without renormalization group[@kle1; @kle2; @kle3].
In this paper we follow a different aim. We would like to know explicitly the perturbation solutions of the field theory as also attempted in the Bender and colleagues works. But our approach will be somewhat different as it will rely on the recent devised principle of duality in perturbation theory [@fra1; @fra2]. This principle is based on the symmetry observation that the choice of a perturbation term in a differential equation is arbitrary and one links the series obtained interchanging a term with another. Dual series having the development parameter inverted are then obtained. So, from a given equation dual perturbation series can be generated.
We will see that the leading order solution is given by a homogeneous equation and is represented by a snoidal wave. This leading order solution is renormalized by higher order terms. The model we consider will be taken with the mass term always positive. We also limit our considerations to the two dimensional case even if generalization to higher dimensions is straightforward. The leading order solution will need a cut-off, that is the space volume should be limited, as also happens at the free field theory. Anyhow, classical solutions of the $\lambda\phi^4$ theory display a lot of interesting features otherwise not obtainable and this should also grant the proper working of a perturbation method whose proof of existence is the main aim of this work.
Quantum field theory in this same limit, $\lambda\rightarrow\infty$, also exists and we proved this in ref.[@fraprd]. We give here a brief description of this theory to prove the fully consistency and generality of our approach. The most important point to be emphasized is that this method yields the behavior of a quantum field theory in the infrared regime with the relative mass spectrum. The paper is structured in the following way. In sec.\[sec2\] we present the duality principle in perturbation theory. In sec.\[sec3\] the dual series is computed for a $\lambda\phi^4$ field theory showing how the leading order solution gets renormalized by higher order terms and the first order correction is computed. In sec.\[sec4\], we describe the quantum field theory in the strong coupling limit as obtained in [@fraprd]. Finally, in sec.\[sec5\] conclusions are given.
Duality in perturbation theory and adiabatic approximation\[sec2\]
==================================================================
In this section we will show how a duality principle holds in perturbation theory showing how to derive a strong coupling expansion with the leading order ruled by an adiabatic dynamics in order to study the evolution of a physical system. This section is similar to the one in [@fra3] but we need it in order to make this paper self-contained.
We consider the following perturbation problem $$\label{eq:eq1}
\partial_t u = L(u) + \lambda V(u)$$ being $\lambda$ an arbitrary ordering parameter: As is well known an expansion parameter is obtained by the computation of the series itself. The standard approach assume the limit $\lambda\rightarrow 0$ and putting $$u = u_0 + \lambda u_1 +\ldots$$ one gets the equations for the series $$\begin{aligned}
\partial_t u_0 &=& L(u_0) \\ \nonumber
\partial_t u_1 &=& L'(u_0)u_1 + V(u_0) \\ \nonumber
&\vdots&\end{aligned}$$ where a derivative with respect to the ordering parameter is indicated by a prime. We recognize here a conventional small perturbation theory as it should be. But the ordering parameter is just a conventional matter and so one may ask what does it mean to consider $L(u)$ as a perturbation instead with respect to the same parameter. Indeed one formally could write the set of equations $$\begin{aligned}
\label{eq:set}
\partial_t v_0 &=& V(v_0) \\ \nonumber
\partial_t v_1 &=& V'(v_0)v_1 + L(v_0) \\ \nonumber
&\vdots&\end{aligned}$$ where we have interchanged $L(u)$ and $V(u)$ and renamed the solution as $v$. The question to be answered is what is the expansion parameter now and what derivative the prime means. To answer this question we rescale the time variable as $\tau = \lambda t$ into eq.(\[eq:eq1\]) obtaining the equation $$\label{eq:eq2}
\lambda\partial_{\tau} u = L(u) + \lambda V(u)$$ and let us introduce the small parameter $\epsilon=\frac{1}{\lambda}$. It easy to see that applying again the small perturbation theory to the parameter $\epsilon\rightarrow 0$ we get the set of equations (\[eq:set\]) but now the time is scaled as $t/\epsilon$, that is, at the leading order the development parameter of the series will enter into the scale of the time evolution producing a proper slowing down ruled by the equation $$\label{eq:lead}
\epsilon\partial_t v_0 = V(v_0)$$ that we can recognize as an equation for adiabatic evolution that in the proper limit $\epsilon\rightarrow 0$ will give the static solution $V(u_0)=0$. We never assume this latter solution but rather we will study the evolution of eq.(\[eq:lead\]). Finally, the proof is complete as we have obtained a dual series $$u = v_0 + \frac{1}{\lambda} v_1 +\ldots$$ by simply interchanging the terms for doing perturbation theory. This is a strong coupling expansion holding in the limit $\lambda\rightarrow\infty$ dual to the small perturbation theory $\lambda\rightarrow 0$ we started with and having an adiabatic equation at the leading order.
It is interesting to note that, for a partial differential equation, we can be forced into a homogeneous equation because, generally, if we require also a scaling into space variables we gain no knowledge at all on the evolution of a physical system. On the other side, requiring a scaling on the space variables and not on the time variable will wash away any evolution of the system. So, on most physical systems a strong perturbation means also a homogeneous solution but this is not a general rule. As an example one should consider fluid dynamics where two regimes dual each other can be found depending if it is the Eulerian or the Navier-Stokes term to prevail.
Dual perturbation series for a classical $\lambda\phi^4$ field theory\[sec3\]
=============================================================================
We consider a field with the following Hamiltonian $$H = \int d^{D-1}x\left[\frac{1}{2}\pi^2+\frac{1}{2}(\partial_x\phi)^2+V(\phi)\right]$$ being $D$ the spacetime dimensionality and $V(\phi)=\frac{1}{2}\phi^2+\frac{\lambda}{4}\phi^4$. For our aim we will assume $\phi$ real but the extension to a higher number of fields will have no problems. In this way the Hamilton equations become $$\begin{aligned}
\partial_t\phi &=& \pi \\ \nonumber
\partial_t\pi &=& \partial_x^2\phi -\phi -\lambda\phi^3.\end{aligned}$$ We can limit our analysis to the $D=2$ case being straightforward the extension to higher dimensions. Small perturbation theory is easily obtained by taking $\lambda$ as a small parameter giving the following set of equations $$\begin{aligned}
\partial_t\phi_0 &=& \pi_0 \\ \nonumber
\partial_t\phi_1 &=& \pi_1 \\ \nonumber
\partial_t\phi_2 &=& \pi_2 \\ \nonumber
&\vdots& \\ \nonumber
\partial_t\pi_0 &=& \partial_x^2\phi_0 -\phi_0 \\ \nonumber
\partial_t\pi_1 &=& \partial_x^2\phi_1 -\phi_1 -\phi_0^3 \\ \nonumber
\partial_t\pi_2 &=& \partial_x^2\phi_2 -\phi_2 -3\phi_0^2\phi_1 \\ \nonumber
&\vdots& \end{aligned}$$ where it easily seen that the free theory, $\Box\phi_0 + \phi_0 = 0$, is the leading order solution. Our aim is to derive a dual perturbation series to this one and, at the same time, to derive the dual leading order solution to the standard plane waves of the free theory.
In order to reach our aim, following the principle of duality in perturbation theory as described in sec.\[sec2\], we put $$\begin{aligned}
\tau &=& \sqrt{\lambda}t \\ \nonumber
\pi &=& \sqrt{\lambda}\left(\pi_0 + \frac{1}{\lambda}\pi_1 + \frac{1}{\lambda^2}\pi_2 + \ldots\right) \\ \nonumber
\phi &=& \phi_0 + \frac{1}{\lambda}\phi_1 + \frac{1}{\lambda^2}\phi_2 + \ldots.\end{aligned}$$ This choice is similar to the one applied in general relativity [@fra3]. It is straightforward to obtain the following not trivial set of equations $$\begin{aligned}
\partial_{\tau}\phi_0 &=& \pi_0 \\ \nonumber
\partial_{\tau}\phi_1 &=& \pi_1 \\ \nonumber
\partial_{\tau}\phi_2 &=& \pi_2 \\ \nonumber
&\vdots& \\ \nonumber
\partial_{\tau}\pi_0 &=& -\phi_0^3 \\ \nonumber
\partial_{\tau}\pi_1 &=& \partial_x^2\phi_0-\phi_0-3\phi_0^2\phi_1 \\ \nonumber
\partial_{\tau}\pi_2 &=& \partial_x^2\phi_1-\phi_1-3\phi_0\phi_1^2-3\phi_0^2\phi_2 \\ \nonumber
&\vdots&\end{aligned}$$ whose solution proves the existence of a dual perturbation series for the classical $\lambda\phi^4$ theory. At a first glance we notice that the leading order is ruled by a homogeneous equation in space as already pointed out it could happen in sec.\[sec2\]. Secondly, we notice that the perturbation is now the operator $\partial_x^2-1$. Finally, we have a different time scale that makes possible to apply the adiabatic approximation that for this set of equations reduce just to a WKB in time due to the fact that we are considering the dual limit $\lambda\rightarrow\infty$ [@fra1]. These points clarify the differences between our approach and the one devised by Bender and colleagues [@bend1; @bend2; @bend3].
Now we can build the dual leading order solution to the free theory. The leading order is ruled by the homogeneous equation $$\partial_{\tau}^2\phi_0 + \phi_0^3 = 0$$ that has the solution $$\phi_0 = A{\rm sn}\left[\pm\frac{A}{\sqrt{2}}(\tau+B),i\right]$$ being $\rm sn$ the snoidal Jacobi function with modulus $k^2=-1$, and $A$ and $B$ two integration constants that can be taken dependent on the space coordinates. Imposing Lorentz invariance we can rewrite the above solution in the final form $$\phi_0 = A{\rm sn}\left[\pm\frac{A}{\sqrt{2}}(\tau+kx),i\right]$$ representing a wave with a different period in space and time. In order to normalize this solution we consider the energy functional given by $$E = \int dx\left[\frac{1}{2}(\partial_t\phi)^2+\frac{1}{2}(\partial_x\phi)^2+V(\phi)\right]$$ that at this order would be given by $$E_0 = \lambda\int dx\left[\frac{1}{2}(\partial_{\tau}\phi_0)^2+\phi_0^3\right]$$ that is unbounded. This means that, as also happens to the free theory, we have to normalize on a box. So, we take a line with a length $L$ and integrate as $$E_0 = \lambda\int^{\frac{L}{2}}_{-\frac{L}{2}}dx
\left[\frac{1}{2}(\partial_{\tau}\phi_0)^2+\phi_0^3\right]$$ that is $$E_0=\frac{\lambda}{4}A^4L.$$ In turn this means that the values of $k$ are quantized as $$\label{eq:kn}
k_n = n\frac{L_0}{L}$$ being $L_0=\frac{4\sqrt{2}K(i)}{A}$ an intrinsic length of the system, $n$ an integer and $K(i)$ the elliptic integral [@gr] $$K(k)=\int_0^{\frac{\pi}{2}}d\theta\frac{1}{\sqrt{1-k^2\sin^2\theta}}=
F\left(\frac{\pi}{2},k\right)$$ computed for $k^2=-1$. In the same way we can compute the period in time as $$\label{eq:T}
T_0=\frac{4\sqrt{2}K(i)}{A}=L_0$$ that, as should be expected on the basis of the slowing down in time in a strongly perturbed system, means $T=T_0/\sqrt{\lambda}$. So, we can rewrite our solution as $$\phi_0 = A{\rm sn}\left[\pm 4K(i)\left(\frac{\tau}{T_0}+n\frac{x}{L}\right),i\right]$$ representing the dual solutions to the plane waves of the free theory for the weak coupling perturbation theory.
The equation to solve at the first order is given by $$\partial_{\tau}^2\phi_1 = -k_n^2\phi_0-\phi_0-3\phi_0^2\phi_1.$$ To solve this equation we firstly note that the solution can be cast in the form $$\phi_1 = \tilde\phi_1-\frac{k_n^2}{2}\phi_0$$ leaving us the equation to solve $$\partial_{\tau}^2\tilde\phi_1 = -\phi_0-3\phi_0^2\tilde\phi_1.$$ Further we notice that due to the adiabatic time scale $\tau$ we do not need to solve exactly this equation. Rather a WKB solution is enough giving $$\tilde\phi_1 = -\frac{1}{2\sqrt{3}}\frac{1}{\sqrt{\phi_0(\tau,x)}}
\int_0^{\tau}d\tau'\frac{\sqrt{\phi_0(\tau',x)}}
{\cos\left[\sqrt{3}\int_0^{\tau'}d\tau''\phi_0(\tau'',x)\right]}
\sin\left[\sqrt{3}\int_0^{\tau}d\tilde\tau\phi_0(\tilde\tau,x)\right].$$ The main result we obtain is that the leading order solution gets a correction from the first order. This situation is the same at the higher orders and we can compute the following renormalized leading order solution $$\phi_{0R}(\tau,x) = \left(1-\frac{k_n^2}{2\lambda}
-\frac{k_n^4}{8\lambda^2}
+\frac{5k_n^6}{16\lambda^3}+\ldots\right)\phi_0(\tau,x)$$ and the field is renormalized also at the classical level. The renormalization constant does depend on the cut-off $L$ but in a way that the limit $L\rightarrow\infty$ means the continuous limit for $k_n$.
We can compute the ground state energy with the renormalized leading order solution. This can be accomplished by taking into account the integrals $$\label{eq:i1}
\int_{-\frac{L}{2}}^{\frac{L}{2}}dx{\rm sn}^2\left[\frac{A}{\sqrt{2}}(\tau+k_nx),i\right]=L$$ and $$\label{eq:i2}
\int_{-\frac{L}{2}}^{\frac{L}{2}}dx{\rm sn}^4\left[\frac{A}{\sqrt{2}}(\tau+k_nx),i\right]=\frac{L}{3}$$ giving rise to the expression for the ground state energy $$E_g = Z_E\frac{\lambda}{4}LA^4$$ being $$Z_E = 1+\frac{2}{A^2\lambda}\left(1-\frac{k_n^2A^2}{3}\right)
-\frac{2k_n^2}{A^2\lambda^2}\left(1+\frac{k_n^2A^2}{6}\right)+\frac{k_n^6}{\lambda^3}
+O\left(\frac{1}{\lambda^4}\right)$$ where we can see that the leading term of the energy is renormalized by $Z_E=Z_E(\lambda,A,k_n)$. As it should be expected, due to the extensive properties of energy, we have a direct proportionality with the cut-off $L$. We also note the correct scaling of energy with $\lambda$. Finally we note that the integrals (\[eq:i1\]) and (\[eq:i2\]) could not be as easy to compute in a higher dimensional theory.
Strongly coupled quantum field theory\[sec4\]
=============================================
Quantum field theory in the strong coupling limit, describing the behaviour in the infrared regime of a $\lambda\phi^4$ theory, has been obtained in [@fraprd]. The starting point is the classical theory obtained sec.\[sec3\] where is noticed that the dual perturbation theory can be obtained considering as a perturbation the term $\partial^2_x-1$. Then, we know that a spatial homogeneous equation rules the theory at the leading order but we also proved numerically in [@fraprd] that in the limit $\lambda\rightarrow\infty$ the Green function method does hold. Putting all this together gives the following generating functional $$Z[j]=\exp\left[\frac{i}{2}\int d^Dy_1d^Dy_2\frac{\delta}{\delta j(y_1)}(-\nabla^2+1)\delta^D(y_1-y_2)
\frac{\delta}{\delta j(y_2)}\right]Z_0[j]$$ being $$Z_0[j]=\exp\left[\frac{i}{2}\int d^Dx_1d^Dx_2j(x_1)\Delta(x_1-x_2)j(x_2)\right]$$ with the Feynman propagator $$\Delta(x_2-x_1)=\delta^{D-1}(x_2-x_1)[G(t_2-t_1)+G(t_1-t_2)]$$ being $$\label{eq:gf}
G(t)=\theta(t)\left(\frac{2}{\lambda}\right)^{\frac{1}{4}}
{\rm sn}\left[\left(\frac{\lambda}{2}\right)^{\frac{1}{4}}t,i\right].$$ In the frequency domain the Feynman propagator is given by $$\Delta(\omega)=\sum_{n=0}^\infty\frac{B_n}{\omega^2-\omega_n^2+i\epsilon}$$ being $$B_n=(2n+1)\frac{\pi^2}{K^2(i)}\frac{(-1)^{n+1}e^{-(n+\frac{1}{2})\pi}}{1+e^{-(2n+1)\pi}}.$$ and the mass spectrum of the theory given by $$\omega_n = \left(n+\frac{1}{2}\right)\frac{\pi}{K(i)}\left(\frac{\lambda}{2}\right)^{\frac{1}{4}}$$ with a mass gap $\delta_0=\frac{\pi}{2K(i)}\left(\frac{\lambda}{2}\right)^{\frac{1}{4}}$. A proportionality appears between the mass gap and the inverse of the period $T$ given in eq.(\[eq:T\]). The quantum theory just fixes the classical amplitude $A$ and the equation for mass spectrum resembles the one that can be obtained by a WKB formula.
Application to Yang-Mills theory has been also given obtaining the relative mass spectrum and mass gap [@fraym]. This result can be obtained by mapping the infrared perturbation theory for the $\lambda\phi^4$ theory described above to the dual perturbation theory for a Yang-Mills theory. For SU(N) one has $\omega_n=\left(n+\frac{1}{2}\right)\frac{\pi}{K(i)}\left(\frac{g^2N}{2}\right)^{\frac{1}{4}}$ with a mass gap given by $\delta_0 = \frac{\pi}{2K(i)}\left(\frac{g^2N}{2}\right)^{\frac{1}{4}}$, being $g$ the Yang-Mills coupling constant, that are adimensional being normalized to a constant that is the integration constant of a Yang-Mills theory. Anyhow we point out that in higher dimensions the most relevant aspect should be the coupling constant renormalization that can make the method not so straightforward to apply.
Conclusions\[sec5\]
===================
We were able to show the existence of a dual perturbation solution to the $D=2$ classical $\lambda\phi^4$ theory and we proved that the dual leading order solutions are snoidal waves. We also have given explicitly the first order correction and the renormalization of the leading order solution producing a properly corrected energy.
Besides, in [@fraprd] we proved the existence of a infrared quantum field theory obtained with the same approach applied to the classical theory. This proves the fully consistency and the generality of this approach in solving partial differential equations in any situation. A couple of considerations are in order. Till now the ability to get explicit solutions to field theory in the strong coupling regime was much limited against the very fine ability to get high precision values for the relevant observables. This trend may change with the extension of the above approach to quantum field theory, as already done in general relativity, opening the way to face physical problems that today are manageable only by numerical computation.
[99]{} S. Weinberg, [*The Quantum Theory of Fields*]{}, Vol. I (Cambridge University Press, 1995). S. Weinberg, [*The Quantum Theory of Fields*]{}, Vol. II (Cambridge University Press, 1996) J. Zinn-Justin, [*Quantum Field Theory and Critical Phenomena*]{}, (Clarendon Press, 1996). C. M. Bender, F. Cooper, G. S. Guralnik and D. H. Sharp, Phys. Rev. D [**19**]{}, 1865 (1979). C. M. Bender, F. Cooper, G. S. Guralnik, H. Moreno, R. Roskies and D. H. Sharp, Phys. Rev. Lett. [**45**]{}, 501 (1980). C. M. Bender, F. Cooper, R. Kenway and L. M. Simmons, Jr., Phys. Rev. D [**24**]{}, 2693 (1981). C. M. Bender and T. T. Wu, Phys. Rev. [**184**]{}, 1231 (1969). C. M. Bender and T. T. Wu, Phys. Rev. Lett. [**27**]{}, 461 (1971). H. Kleinert, Phys. Rev. D [**57**]{}, 2264 (1998). H. Kleinert, Phys. Rev. D [**58**]{}, 107702 (1998). H. Kleinert, Phys. Rev. D [**60**]{}, 085001 (1999). M. Frasca, Phys. Rev. A [**58**]{}, 3439 (1998). M. Frasca, Phys. Rev. A [**60**]{}, 573 (1999). M. Frasca, Phys. Rev. D [**73**]{}, 027701 (2006); Erratum: ibid., 049902 (2006) \[hep-th/0511068\]. M. Frasca, hep-th/0508246, to appear in International Journal of Modern Physics D. I. S. Gradshteyn, I. M. Ryzhik, [*Table of Integrals, Series, and Products*]{}, (Academic Press, 2000). M. Frasca, hep-th/0511173.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using Fedosov theory of deformation quantization of endomorphism bundle we construct several models of pure geometric, deformed vacuum gravity, corresponding to arbitrary symplectic noncommutativity tensor. Deformations of Einstein-Hilbert and Palatini actions are investigated. Coordinate covariant field equations are derived up to the second order of the deformation parameter. For some models they are solved and explicit corrections to an arbitrary Ricci-flat metric are pointed out. The relation to the theory of Seiberg-Witten map is also studied and the correspondence to the spacetime noncommutativity described by Fedosov $*$-product of functions is explained.'
author:
- 'Micha[ł]{} Dobrski'
title: |
On some models of geometric\
noncommutative general relativity
---
Introduction
============
The present paper is dedicated to a study of some possible global and geometric models of relativity on noncommutative spacetimes within the framework of Fedosov deformation quantization of endomorphism bundle. The motivation for such investigation originates in the conviction that whatever “noncommutative gravity” would be, it should preserve the basic symmetry of the classical theory – the full diffeomorphism invariance. Presented analysis aims at showing that Fedosov quantization of endomorphism bundle can serve as a tool for building geometric field theories on noncommutative spacetimes.
The general strategy we are going to adopt can be summarized in the following steps.
1. Take some symplectic manifold and an action on it which leads to the general relativity.
2. Rewrite the action by representing Lagrangian as a product of endomorphisms of some bundle.
3. Replace the product of endomorphisms by Fedosov $*$-product of endomorphisms.
4. Replace the integral by Fedosov trace functional.
5. Do the variations to obtain field equations.
6. Observe that steps *3* and *4* together with results of [@dobrski2] induce that the theory is locally equivalent to the theory with Seiberg-Witten map applied on endomorphisms.
There is vast literature concerning construction of noncommutative gravity by means of Moyal product and Seiberg-Witten map. Hence, one can point out series of works [@calmet1; @calmet2; @mukherjee] based on combination of infinitesimal $so(3,1)$ gauging with infinitesimal coordinate transformations, preserving (at first order of deformation) constant deformation parameter $\theta^{ij}$. Another approach is given by [@chamseddine1; @chamseddine2; @cardella; @chaichian; @mukherjee0A], where $SO(4,1)$ (or $U(2,2)$) symmetry is investigated. In such setting gauge potential carries information about both tetrad field and the usual $SO(3,1)$-connection. The standard gravity is recovered by the procedure of contraction of the gauge group. There are also investigations based on some variants of $SL(2,\mathbb{C})$ symmetry [@chamseddine3; @garciacompean]. The common feature of all of these approaches are vanishing first order corrections to the field equations. On the other hand, the common issue is the lack of diffeomorphism invariance[^1]. More general types of noncommutativity were also studied – eg. Lie algebraic one [@mukherjee2] or given by Kontsevich theory [@miaozhangsl2c; @miaozhangu22]. In [@marculescu] theories based on Moyal product and Seiberg-Witten map were geometrized. The resulting structure is invariant under passive diffeomorphisms, but at the price of nonassociativity of the corresponding $*$-product. One should also mention some other approaches to noncommutative gravity related somehow to $*$-products and Seiberg-Witten map. These are [@aschieri1; @aschieri2; @aschieri3; @aschieri4; @meyer; @compean2], where the method of Lie algebra twisting has been used to represent deformation of diffeomorphism symmetry. One of the remarkable results of [@aschieri3] is the construction of an action which is geometric (i.e. described by globally defined 4-form) being simultaneously invariant under deformed diffeomorphism symmetry. Finally there are investigations which are strictly related to some particular models emerging in the context of the string theory, e.g. [@steinacker1; @steinacker2].
The paper is organized as follows. In section 2 a brief overview of results of Fedosov construction is presented, and also some further conventions are fixed. In sections 3 and 4 deformations of Einstein-Hilbert and Palatini actions are investigated. Fifth section is devoted to analysis of interrelation between presented models and the theory of Seiberg-Witten map. We also clarify, how obtained results are related to the noncommutativity of the spacetime described by Fedosov $*$-product of functions. Finally some concluding remarks (section 6) are given.
Preliminaries
=============
Fedosov construction
--------------------
The main tool used in this paper is Fedosov construction of deformation quantization of endomorphism bundle formulated in [@fedosov]. We are not going to concern technical or “internal” details of this theory (which are interesting and beautiful on their own) but rather to make use of some of its particular results. Interested reader may find short exposition of Fedosov construction in its simplest, suitable for present pourposes form in [@dobrski2]. Further geometric and algebraic interpretations are provided by [@emmrwein; @farkas]. Some other analysis and examples can be found in [@tosiek; @tosiek2; @tosiek3]. Thus, we limit ourselves to the very brief, notation-fixing description of Fedosov $*$-product.
The starting point is given by Fedosov manifold $({\mathcal{M}},\omega,{\partial^{S}})$ i.e. $2n$-dimensional symplectic manifold $({\mathcal{M}},\omega)$ with some fixed symplectic (torsionless and preserving $\omega$) connection ${\partial^{S}}$ [@gelfand; @bielgutt]. The corresponding Poisson tensor (given by the inverse of $\omega_{ij}$) is going to be denoted as $\Lambda^{ij}$. These data generate[^2] global, geometric and associative deformation of product of functions on ${\mathcal{M}}$. Its explicit form can be computed up to arbitrary power of deformation parameter $h$ (which has nothing to do with Planck constant in our context) by means of Fedosov’s recursive techniques.
For the vector bundle ${\mathcal{E}}$ over ${\mathcal{M}}$, equipped with a connection ${\partial^{{\mathcal{E}}}}$ one can construct global, geometric and associative deformation of product of sections of ${\operatorname{End}({\mathcal{E}})}$. Locally it can be understood as a deformation of product of matrices. Denoting by $\partial={\partial^{S}}\otimes 1+ 1\otimes {\partial^{{\mathcal{E}}}}$ the connection in $T{\mathcal{M}}\otimes{\mathcal{E}}$ (and by the same symbol its natural extension to any other tensor product of $T{\mathcal{M}}$, $T^*{\mathcal{M}}$, ${\mathcal{E}}$ and ${\mathcal{E}}^*$) one may calculate that for arbitrary two sections $F,G \in C^{\infty}({\operatorname{End}({\mathcal{E}})})[[h]]$ the Fedosov $*$-product is given up to $h^2$ by the expression $$\begin{gathered}
\label{fedosov_endstar}
F*G=FG-\frac{{\mathrm{i}}h}{2} \Lambda^{a b} \partial_a F \partial_b G+\\
-\frac{h^2}{8}\Lambda^{ab}\Lambda^{cd}\Big(\{\partial_b F,{R^{{\mathcal{E}}}}_{ac}\}\partial_d G + \partial_b F \{{R^{{\mathcal{E}}}}_{ac},\partial_d G\} +\partial_{(a} \partial_{c)} F \partial_{(b} \partial_{d)} G\Big) + O(h^3)\end{gathered}$$ where ${R^{{\mathcal{E}}}}_{ab}=\frac{\partial}{\partial x^a}{\Gamma^{{\mathcal{E}}}}_b-\frac{\partial}{\partial x^b}{\Gamma^{{\mathcal{E}}}}_a+[{\Gamma^{{\mathcal{E}}}}_a,{\Gamma^{{\mathcal{E}}}}_b]$ (for ${\partial^{{\mathcal{E}}}}_i=\frac{\partial}{\partial x^i}+{\Gamma^{{\mathcal{E}}}}_i$) is the curvature of ${\partial^{{\mathcal{E}}}}$, and $\{\cdot\,,\cdot\}$ stands for the anticommutator. It is clear that in above formula usual product of endomorphisms (noncommutative from the beginning) has been used. For the special case of flat ${\partial^{{\mathcal{E}}}}$ and the local frame with ${\Gamma^{{\mathcal{E}}}}\equiv 0$, the Fedosov $*$-product of endomorphisms becomes product of matrices with commutative multiplication of entries replaced by noncommutative Fedosov product of functions. Such product of matrices is going to be denoted as $*_S$. (The same symbol will be used for the Fedosov product of functions). If additionally ${\partial^{S}}$ is flat and we work in local Darboux coordinates for which coefficients of ${\partial^{S}}$ vanish, then Fedosov product of functions becomes Moyal product $*_T$. Thus, in such special case, we are dealing with multiplication used in [@seibwitt] for the description of deformed gauge transformations.
The object which needs some more attention is Fedosov trace functional ([@fedosov] section 5.6). Given some Fedosov product $*$ one is able to construct trace functional $\operatorname{tr}_*$ taking values in $\mathbb{C}[[h]]$ and acting on compactly supported sections belonging to $C^{\infty}({\operatorname{End}({\mathcal{E}})})[[h]]$, with the property $$\label{tracecommutation}
\operatorname{tr}_*(F * G)=\operatorname{tr}_*(G * F).$$ If one requires additionally, that for arbitrary (global or local) isomorphism $M$ between $*$-algebras with products $*_1$ and $*_2$ (i.e. for $M$ fulfilling $M(F*_1G)=M(F)*_2M(G)$) the relation $$\label{traceinvar}
\operatorname{tr}_{*_1}(F)=\operatorname{tr}_{*_2}(M(F))$$ holds, then it follows that the trace functional is unique up to a constant normalizing factor. The proof of this fact relies on the observation that for the Moyal product $*_T$ the trace is given by $$\label{tracemoyal}
\operatorname{tr}_{*_T}(F)= \mathrm{const} \int_{\mathbb{R}^{2n}} \operatorname{Tr}(F) \frac{\omega^n}{n!}$$ where $\operatorname{Tr}$ stands for the trace of a matrix, and on possibility of representing $\operatorname{tr}_*$ in terms of traces on Moyal algebras by a partition of unity $\{\rho_i\}$ and a compatible set of local isomorphisms $\{M_i\}$ between $*$ and the Moyal product. It turns out that $\operatorname{tr}_*$ is independent of particular choice of $\{\rho_i\}$ and $\{A_i\}$. Unlike convention of [@fedosov], we fix the normalizing constant to be equal $1$. Construction presented in [@fedosov] enables calculation of explicit form of $\operatorname{tr}_*$. Up to $h^2$ it reads[^3] $$\label{fedosov_trace}
\operatorname{tr}_* (F)=\int_{{\mathcal{M}}} \operatorname{Tr}\Bigg(F + \frac{{\mathrm{i}}h }{2}\Lambda^{a b}{R^{{\mathcal{E}}}}_{a b} F
+h^2 \bigg(-\frac{3}{8} \Lambda^{[a b} \Lambda^{c d]}{R^{{\mathcal{E}}}}_{a b} {R^{{\mathcal{E}}}}_{c d} +s_2 1\bigg)F + O(h^3) \Bigg) \frac{\omega^n}{n!}$$ where $1$ is the identity endomorphism and the scalar[^4] $$\label{s2_def}
s_2=\frac{1}{64} \Lambda^{[a b} \Lambda^{c d]}\tensor{{{\accentset{S}{R}}}}{^k_{lab}}\tensor{{{\accentset{S}{R}}}}{^l_{kcd}} + \frac{1}{48}\Lambda^{ab}\Lambda^{cd}{\partial^{S}}_e {\partial^{S}}_a \tensor{{{\accentset{S}{R}}}}{^e_{bcd}}$$ has been introduced for sake of simplicity of further notations. In above formula $\tensor{{{\accentset{S}{R}}}}{^i_{jab}}$ stands for the curvature tensor of ${\partial^{S}}$. It is useful to write down explicit form of $\operatorname{tr}_*(F*G)$. Substitution of (\[fedosov\_endstar\]) into (\[fedosov\_trace\]) after some manipulations yields[^5] $$\begin{gathered}
\operatorname{tr}_* (F*G)
=\int_{{\mathcal{M}}} \operatorname{Tr}\Bigg(FG+\frac{{\mathrm{i}}h}{4}\Lambda^{a b}{R^{{\mathcal{E}}}}_{a b} \{F,G\}
+h^2 \Bigg[ s_2 FG
+\frac{1}{8}\Lambda^{a b}\Lambda^{c d}\bigg({R^{{\mathcal{E}}}}_{a b}[\partial_c F, \partial_d G]+\\
- \partial_{(a} \partial_{c)} F \partial_{(b} \partial_{d)} G
-\frac{3}{2} {R^{{\mathcal{E}}}}_{[a b} {R^{{\mathcal{E}}}}_{c d]}\{F,G\} \bigg)
\Bigg] + O(h^3) \Bigg) \frac{\omega^n}{n!}\end{gathered}$$
Some further conventions
------------------------
The important problem related to the programme presented in the introduction is the incompatibility of the volume forms. In (\[fedosov\_trace\]) the symplectic volume form ${\operatorname{vol}_{S}}=\frac{\omega^n}{n!}$ must be used, and in general relativity the metric one ${\operatorname{vol}_{M}}=\sqrt{-g} {\mathrm{d}}x^1\wedge\dots\wedge {\mathrm{d}}x^{2n}$ more or less explicitly appears. Since the two must be proportional one can write ${\operatorname{vol}_{M}}=v {\operatorname{vol}_{S}}$ defining the function $v:{\mathcal{M}}\to \mathbb{R}$ this way. The above mentioned incompatibility should be handled somehow, and in what follows two possible approaches are investigated. First, one can simply rescale one of the endomorphisms by multiplying them by $v$. Thus, let us fix the convention that $\breve{F}=vF$. The other option is given by introducing endomorphism $V=v1$ which multiplies endomorphism under the action. Both methods are completely equivalent at the undeformed level, but become different after deformation.
Let us also point out the following issue concerning the tangent bundle $T{\mathcal{M}}$. In presented models it appears in two distinct roles. First as a “component” of bundle ${\mathcal{E}}$, and then as an object which carries information about the symplectic structure and the covariant derivations producing quantization formalism. This distinction becomes important when applying covariant derivation to tensors involving indices from both copies of $T{\mathcal{M}}$. The $\partial$ connection acts in such case by means of ${\partial^{{\mathcal{E}}}}$ (which is going to be chosen as a metric connection) and the symplectic connection ${\partial^{S}}$ respectively. Thus, one needs some way of “marking” indices which should be differentiated by ${\partial^{{\mathcal{E}}}}$ and we are going to put prime at them (e.g. $\tensor{{R^{{\mathcal{E}}}}{}}{^{a'}_{b'lm}}$). The ambiguity may be also postponed by using index-free notation for endomorphisms, and this approach is also used. Finally, the primes are omitted in the field equations, as they are no longer needed and may tend to obscure the result.
Finally, let us mention that all indices in subsequent sections are manipulated by means of corresponding metric tensors. (With exception of relations (\[eh\_wab\_seh1a\])–(\[eh\_wab\_seh2b\]), (\[sol-g2-seh1a\]) and (\[sol-g2-seh1b\]) where the undeformed part of metric is used. This is also recalled within the text). These metric tensors are $g_{ab}$ and also $\eta_{AB}$ for the case of deformation of Palatini action. To avoid ambiguities (or even inconsistencies), we abandon convention of using symplectic form for raising or lowering indices. All formulae taken from Fedosov theory are rewritten in such manner, that they do not involve manipulation of indices by means of symplectic form.
Einstein-Hilbert action
=======================
Now, let us analyze some possible applications of Fedosov theory in the general relativity on noncommutative spacetime. We are going to proceed using programme sketched in the introduction and to assume, that symplectic form $\omega$ and compatible symplectic connection ${\partial^{S}}$ are fixed. First, let us focus on the Einstein-Hilbert action. Thus, there is a metric $g_{ab}$ with determinant $g$, its torsionless Levi-Civita connection $\nabla$, Riemann curvature tensor $\tensor{R}{^{a'}_{b'cd}}$ (also used with all indices primed[^6] $\tensor{R}{^{a'}_{b'c'd'}}$), Ricci tensor $R_{a'b'}=\tensor{R}{^{c'}_{a'c'b'}}$ and Ricci scalar $R$. Field equations are going to be derived by the variation of the metric.
Let us introduce the notation $\tensor{{\underline{R}}}{^{a'}_{b'}}=\tensor{R}{^{a'}_{b'}}$ and $\tensor{{\underline{\underline{R}}}}{^{a'b'}_{c'd'}}=\tensor{R}{^{a'b'}_{c'd'}}$. (This becomes convenient when distinguishing between endomorphisms ${\underline{\underline{R}}}$, ${\underline{R}}$ and the scalar $R$). Also, let $$\begin{split}
Y^{ijk'l'}=&\Lambda^{[ij} \Lambda^{ab]}\tensor{R}{^{k'l'}_{ab}}\\
X^{i'j'k'l'}=&\Lambda^{[a b} \Lambda^{c d]} {\tensor{R}{^{i'j'}_{ab}}}{\tensor{R}{^{k'l'}_{cd}}} =\tensor{R}{^{i'j'}_{ab}}\tensor{Y}{^{abk'l'}}\\
Z=&\frac{1}{\sqrt{-g}}\Lambda^{ij}\Lambda^{kl}{\partial^{S}}_i {\partial^{S}}_k {\partial^{S}}_j {\partial^{S}}_l \sqrt{-g}
\end{split}$$
Deformed actions and field equations
------------------------------------
### ${\underline{\breve{R}}}$ as an endomorphism of $T{\mathcal{M}}$
The Einstein-Hilbert action can be quickly rewritten as $$\mathcal{S}_{EH_{1A}}=\int_{{\mathcal{M}}}\operatorname{Tr}{\underline{\breve{R}}}\,\frac{\omega^n}{n!}$$ Thus, we are going to treat rescaled Ricci tensor $\tensor{{\underline{\breve{R}}}}{^{a'}_{b'}}=v\tensor{{\underline{R}}}{^{a'}_{b'}}$ as an endomorphism of ${\mathcal{E}}=T{\mathcal{M}}$. In order to define $*$-product of endomorphisms one needs some connection in ${\mathcal{E}}$. Let us fix ${\partial^{{\mathcal{E}}}}=\nabla$ and consequently ${R^{{\mathcal{E}}}}$ is given by the Riemann tensor. The corresponding $*$-product is going to be denoted by $*_{EH_1}$. Under these assumptions the deformed action is given by $$\begin{split}
\widehat{\mathcal{S}}_{EH_{1A}}=\operatorname{tr}_{*_{EH_1}}({\underline{\breve{R}}})=&\int_{{\mathcal{M}}} \operatorname{Tr}\Bigg({\underline{\breve{R}}}+h^2 \bigg(-\frac{3}{8} \Lambda^{[a b} \Lambda^{c d]}{R^{{\mathcal{E}}}}_{a b} {R^{{\mathcal{E}}}}_{c d} +s_2\bigg){\underline{\breve{R}}}+ O(h^3) \Bigg) \frac{\omega^n}{n!}\\
=&\int_{{\mathcal{M}}} \bigg(R
-\frac{3}{8}h^2 \tensor{X}{^{k'}_{l'}^{l'}_{m'}}\tensor{R}{^{m'}_{k'}}
+h^2 s_2 R + O(h^3) \bigg) {\operatorname{vol}_{M}}\end{split}$$ Variation of the metric yields the following field equations $$\begin{split}
&R^{ab}-\frac{1}{2}g^{ab}R+h^2\Bigg[\frac{3}{8}\bigg(
-{\tensor{R}{^{(a}_k}}{\tensor{X}{_l^{b)}^k^l}}
+\frac{1}{2}{\tensor{R}{^k_l}}{\tensor{X}{^l_m^m_k}}g^{ab}+
\nabla_k \nabla^{(a} {\tensor{X}{^{b)}_l^{lk}}}
-\frac{1}{2}\nabla_l \nabla^l {\tensor{X}{^a_k^{kb}}}\\
&-\frac{1}{2}g^{ab}\nabla_k \nabla_l {\tensor{X}{^k_m^{ml}}}
-2\nabla_k \nabla^l\left({\tensor{R}{^{(a}_m}}{\tensor{Y}{_l^{b)mk}}}\right)
+2\nabla_k \nabla_l\left({\tensor{R}{^{km}}}{\tensor{Y}{^{l(a}_m^{b)}}} \right)
\bigg)
-\frac{1}{2}g^{ab}Rs_2\\
&+R^{ab}s_2
+g^{ab}\nabla_l \nabla^l s_2-\nabla^a \nabla^b s_2
\Bigg]+O(h^3)=0
\end{split}$$
### ${\underline{R}}$ and $V$ as endomorphisms of $T{\mathcal{M}}$ {#RVendTM}
Now, keeping unmodified $*$-product structure given by $*_{EH_1}$, we are going to investigate another possibility of forcing correct volume form at $h=0$. The Einstein-Hilbert action written as $$\mathcal{S}_{EH_{1B}}=\int_{{\mathcal{M}}}\operatorname{Tr}{\underline{R}}V\,\frac{\omega^n}{n!}$$ may be deformed into[^7] $$\begin{split}
&\widehat{\mathcal{S}}_{EH_{1B}}=\operatorname{tr}_{*_{EH_1}}({\underline{R}}*_{EH_1} V)\\
&=\int_{{\mathcal{M}}} \operatorname{Tr}\Bigg({\underline{R}}V
+h^2 \bigg(-\frac{1}{8}\Lambda^{a b}\Lambda^{c d}\Big(
\partial_{(a} \partial_{c)} {\underline{R}}\partial_{(b} \partial_{d)} V
+3 {R^{{\mathcal{E}}}}_{[a b} {R^{{\mathcal{E}}}}_{c d]}{\underline{R}}V \Big)+s_2 {\underline{R}}V
\bigg) + O(h^3) \Bigg) \frac{\omega^n}{n!}\\
&=\int_{{\mathcal{M}}} \bigg(R
-\frac{3}{8}h^2 \tensor{X}{^{k'}_{l'}^{l'}_{m'}}\tensor{R}{^{m'}_{k'}}
-\frac{1}{8}h^2 \Lambda^{a b}\Lambda^{c d} {\partial^{S}}_{b}{\partial^{S}}_{d}{\partial^{S}}_{a}{\partial^{S}}_{c} R
+h^2 s_2 R + O(h^3) \bigg) {\operatorname{vol}_{M}}\end{split}$$ Then, the field equations become $$\begin{split}
&R^{ab}-\frac{1}{2}g^{ab}R+h^2\Bigg[\frac{3}{8}\bigg(
-{\tensor{R}{^{(a}_k}}{\tensor{X}{_l^{b)}^k^l}}
+\frac{1}{2}{\tensor{R}{^k_l}}{\tensor{X}{^l_m^m_k}}g^{ab}+
\nabla_k \nabla^{(a} {\tensor{X}{^{b)}_l^{lk}}}
-\frac{1}{2}\nabla_l \nabla^l {\tensor{X}{^a_k^{kb}}}\\
&-\frac{1}{2}g^{ab}\nabla_k \nabla_l {\tensor{X}{^k_m^{ml}}}
-2\nabla_k \nabla^l\left({\tensor{R}{^{(a}_m}}{\tensor{Y}{_l^{b)mk}}}\right)
+2\nabla_k \nabla_l\left({\tensor{R}{^{km}}}{\tensor{Y}{^{l(a}_m^{b)}}} \right)
\bigg)
+\frac{1}{8}\bigg(
-{\tensor{R}{^{ab}}}Z\\
&+\nabla^a \nabla^b Z-g^{ab}\nabla_l\nabla^l Z
+\frac{1}{2}g^{ab}\Lambda^{jk}\Lambda^{lm} {\partial^{S}}_{k}{\partial^{S}}_{m}{\partial^{S}}_{j}{\partial^{S}}_{l}R
\bigg)
-\frac{1}{2}g^{ab}Rs_2+R^{ab}s_2 \\
&+g^{ab}\nabla_l \nabla^l s_2-\nabla^a \nabla^b s_2
\Bigg]+O(h^3)=0
\end{split}$$
### ${\underline{\underline{\breve{R}}}}$ as an endomorphism of $T{\mathcal{M}}\otimes T{\mathcal{M}}$
This time, we start with the action $$\mathcal{S}_{EH_{2A}}=\int_{{\mathcal{M}}}\operatorname{Tr}{\underline{\underline{\breve{R}}}}\,\frac{\omega^n}{n!}$$ Here, the rescaled Riemann tensor is treated as an endomorphism of ${\mathcal{E}}=T{\mathcal{M}}\otimes T{\mathcal{M}}$ whose action on $l \in T{\mathcal{M}}\otimes T{\mathcal{M}}$ yields $({\underline{\underline{\breve{R}}}}l)^{a'b'} = v\tensor{R}{^{a'b'}_{c'd'}}l^{c'd'}$. As a connection in ${\mathcal{E}}$ we take ${\partial^{{\mathcal{E}}}}=\nabla \otimes 1+1 \otimes \nabla$. Its curvature is given by ${R^{{\mathcal{E}}}}_{ab}=R^{\nabla}_{ab}\otimes1+1\otimes R^{\nabla}_{ab}$, with $R^{\nabla}_{ab}$ being the curvature of $\nabla$ treated as an endomorphism of $T{\mathcal{M}}$. Let $*_{EH_2}$ be the corresponding $*$-product. Thus $$\begin{split}
\widehat{\mathcal{S}}_{EH_{2A}}=&\operatorname{tr}_{*_{EH_2}}({\underline{\underline{\breve{R}}}})=\int_{{\mathcal{M}}} \operatorname{Tr}\Bigg({\underline{\underline{\breve{R}}}}+h^2 \bigg(-\frac{3}{8} \Lambda^{[a b} \Lambda^{c d]}{R^{{\mathcal{E}}}}_{a b} {R^{{\mathcal{E}}}}_{c d} +s_2\bigg){\underline{\underline{\breve{R}}}}+ O(h^3) \Bigg) \frac{\omega^n}{n!}\\
=&\int_{{\mathcal{M}}} \bigg(R
-\frac{3}{4}h^2 \Big(\tensor{X}{^{k'}_{l'}^{l'}_{m'}}\tensor{R}{^{m'}_{k'}}
+ \tensor{X}{^{k'}_{l'}^{m'}_{p'}}\tensor{R}{^{l'p'}_{k'm'}}\Big) +h^2 s_2 R + O(h^3) \bigg) {\operatorname{vol}_{M}}\end{split}$$ yielding $$\begin{split}
&R^{ab}-\frac{1}{2}g^{ab}R+h^2\Bigg[\frac{3}{4}\bigg(
-{\tensor{R}{^{(a}_k}}{\tensor{X}{_l^{b)}^k^l}}
+\frac{1}{2}{\tensor{R}{^k_l}}{\tensor{X}{^l_m^m_k}}g^{ab}+
\nabla_k \nabla^{(a} {\tensor{X}{^{b)}_l^{lk}}}
-\frac{1}{2}\nabla_l \nabla^l {\tensor{X}{^a_k^{kb}}}\\
&-\frac{1}{2}g^{ab}\nabla_k \nabla_l {\tensor{X}{^k_m^{ml}}}
-2\nabla_k \nabla^l\left({\tensor{R}{^{(a}_m}}{\tensor{Y}{_l^{b)mk}}}\right)
+2\nabla_k \nabla_l\left({\tensor{R}{^{km}}}{\tensor{Y}{^{l(a}_m^{b)}}} \right)
+{\tensor{R}{_{km}^{l(a}}}{\tensor{X}{^{b)m}^k_l}}\\
&+\frac{1}{2}\nabla_k \nabla_l {\tensor{X}{^{k(ab)l}}}
+2 \nabla_k \nabla_l\left( {\tensor{R}{^{mjk(a}}}{\tensor{Y}{^{b)l}_{mj}}} \right)
+\frac{1}{2}{\tensor{R}{^{lm}_{jk}}}{\tensor{X}{^j_l^k_m}}g^{ab}
\bigg)
-\frac{1}{2}g^{ab}Rs_2+R^{ab}s_2 \\
&+g^{ab}\nabla_l \nabla^l s_2-\nabla^a \nabla^b s_2
\Bigg]+O(h^3)=0
\end{split}$$
### ${\underline{\underline{R}}}$ and $V$ as endomorphisms of $T{\mathcal{M}}\otimes T{\mathcal{M}}$
Analogously to section \[RVendTM\], one may keep the product $*_{EH_2}$ unchanged, but rewrite the action using $V$ $$\mathcal{S}_{EH_{2B}}=\int_{{\mathcal{M}}}\operatorname{Tr}{\underline{\underline{R}}}V\,\frac{\omega^n}{n!}$$ After the deformation the action takes form $$\begin{split}
&\widehat{\mathcal{S}}_{EH_{2B}}=\operatorname{tr}_{*_{EH_2}}({\underline{\underline{R}}}*_{EH_2} V)\\
&=\int_{{\mathcal{M}}} \operatorname{Tr}\Bigg({\underline{\underline{R}}}V
+h^2 \bigg(-\frac{1}{8}\Lambda^{a b}\Lambda^{c d}\Big(
\partial_{(a} \partial_{c)} {\underline{\underline{R}}}\partial_{(b} \partial_{d)} V
+3 {R^{{\mathcal{E}}}}_{[a b} {R^{{\mathcal{E}}}}_{c d]}{\underline{\underline{R}}}V \Big)+s_2 {\underline{\underline{R}}}V
\bigg) + O(h^3) \Bigg) \frac{\omega^n}{n!}\\
&=\int_{{\mathcal{M}}} \bigg(R
-\frac{3}{4}h^2 \Big(\tensor{X}{^{k'}_{l'}^{l'}_{m'}}\tensor{R}{^{m'}_{k'}}
+ \tensor{X}{^{k'}_{l'}^{m'}_{p'}}\tensor{R}{^{l'p'}_{k'm'}}\Big)
-\frac{1}{8}h^2 \Lambda^{a b}\Lambda^{c d} {\partial^{S}}_{b}{\partial^{S}}_{d}{\partial^{S}}_{a}{\partial^{S}}_{c} R+h^2 s_2 R\\
&\quad+ O(h^3)\bigg) {\operatorname{vol}_{M}}\end{split}$$ The field equations are given by $$\begin{split}
&R^{ab}-\frac{1}{2}g^{ab}R+h^2\Bigg[\frac{3}{4}\bigg(
-{\tensor{R}{^{(a}_k}}{\tensor{X}{_l^{b)}^k^l}}
+\frac{1}{2}{\tensor{R}{^k_l}}{\tensor{X}{^l_m^m_k}}g^{ab}+
\nabla_k \nabla^{(a} {\tensor{X}{^{b)}_l^{lk}}}
-\frac{1}{2}\nabla_l \nabla^l {\tensor{X}{^a_k^{kb}}}\\
&-\frac{1}{2}g^{ab}\nabla_k \nabla_l {\tensor{X}{^k_m^{ml}}}
-2\nabla_k \nabla^l\left({\tensor{R}{^{(a}_m}}{\tensor{Y}{_l^{b)mk}}}\right)
+2\nabla_k \nabla_l\left({\tensor{R}{^{km}}}{\tensor{Y}{^{l(a}_m^{b)}}} \right)
+{\tensor{R}{_{km}^{l(a}}}{\tensor{X}{^{b)m}^k_l}}\\
&+\frac{1}{2}\nabla_k \nabla_l {\tensor{X}{^{k(ab)l}}}
+2 \nabla_k \nabla_l\left( {\tensor{R}{^{mjk(a}}}{\tensor{Y}{^{b)l}_{mj}}} \right)
+\frac{1}{2}{\tensor{R}{^{lm}_{jk}}}{\tensor{X}{^j_l^k_m}}g^{ab}
\bigg)
+\frac{1}{8}\bigg(
-{\tensor{R}{^{ab}}}Z\\
&+\nabla^a \nabla^b Z-g^{ab}\nabla_l\nabla^l Z
+\frac{1}{2}g^{ab}\Lambda^{jk}\Lambda^{lm} {\partial^{S}}_{k}{\partial^{S}}_{m}{\partial^{S}}_{j}{\partial^{S}}_{l}R
\bigg)
-\frac{1}{2}g^{ab}Rs_2+R^{ab}s_2 \\
&+g^{ab}\nabla_l \nabla^l s_2-\nabla^a \nabla^b s_2
\Bigg]+O(h^3)=0
\end{split}$$
Structure of deformed theories
------------------------------
Let us briefly comment formulae obtained in the previous subsection. In all cases the $h^1$ terms in deformed actions have vanished due to $\tensor{R}{^k_{lab}}\tensor{R}{^l_k}=0$. Also, in all deformed Lagrangians one is dealing with $h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_s=h^2 s_2 R$ term, originating in the part of the trace formula (\[fedosov\_trace\]) generated by the curvature of symplectic connection. ${\accentset{\mathit{(2)}}{\mathcal{L}}}_s$ gives rise to the field equations by the expression $$-\frac{1}{2}g^{ab}Rs_2+R^{ab}s_2 +g^{ab}\nabla_l \nabla^l s_2-\nabla^a \nabla^b s_2,$$ and describes interaction of the metric $g_{ab}$ with symplectic connection, which defines $s_2$.
Now, let us write Lagrangians (with respect to ${\operatorname{vol}_{M}}$) as $$\begin{aligned}
\widehat{\mathcal{L}}_{EH_{1A}}&=\mathcal{L}_{EH}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\breve{R}}}}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_s +O(h^3),\\
\widehat{\mathcal{L}}_{EH_{1B}}&=\mathcal{L}_{EH}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{R}}*_{EH_1} V}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_s +O(h^3),\\
\widehat{\mathcal{L}}_{EH_{2A}}&=\mathcal{L}_{EH}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{\breve{R}}}}}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_s +O(h^3),\\
\widehat{\mathcal{L}}_{EH_{2B}}&=\mathcal{L}_{EH}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{R}}}*_{EH_1} V}+h^2{\accentset{\mathit{(2)}}{\mathcal{L}}}_s +O(h^3),\end{aligned}$$ with $\mathcal{L}_{EH}=R$. Hence, ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\breve{R}}}}$, ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{R}}*_{EH_1} V}$, ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{\breve{R}}}}}$ and ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{R}}}*_{EH_1} V}$ represent terms produced by the part of the trace generated by ${R^{{\mathcal{E}}}}$. It follows that choosing ${\mathcal{E}}=T{\mathcal{M}}$ and ${\underline{\breve{R}}}$ as the endomorphism yields ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\breve{R}}}}=-\frac{3}{8} \tensor{X}{^{k'}_{l'}^{l'}_{m'}}\tensor{R}{^{m'}_{k'}}$, while taking ${\mathcal{E}}=T{\mathcal{M}}\otimes T{\mathcal{M}}$ and ${\underline{\underline{\breve{R}}}}$ produces ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{\breve{R}}}}}=-\frac{3}{4} \Big(\tensor{X}{^{k'}_{l'}^{l'}_{m'}}\tensor{R}{^{m'}_{k'}}
+ \tensor{X}{^{k'}_{l'}^{m'}_{p'}}\tensor{R}{^{l'p'}_{k'm'}}\Big)$. Thus one can write the relation ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{\breve{R}}}}}=2{\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\breve{R}}}}-\frac{3}{4}\tensor{X}{^{k'}_{l'}^{m'}_{p'}}\tensor{R}{^{l'p'}_{k'm'}}$. Switching from endomorphism rescaling to multiplication by $V$ influences the deformed Lagrangians by ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{R}}*_{EH_1} V}={\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\breve{R}}}}-\frac{1}{8}h^2 \Lambda^{a b}\Lambda^{c d} {\partial^{S}}_{b}{\partial^{S}}_{d}{\partial^{S}}_{a}{\partial^{S}}_{c} R$ , and ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{R}}}*_{EH_2} V}={\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\underline{\breve{R}}}}}-\frac{1}{8}h^2 \Lambda^{a b}\Lambda^{c d} {\partial^{S}}_{b}{\partial^{S}}_{d}{\partial^{S}}_{a}{\partial^{S}}_{c} R$. The term ${\accentset{\mathit{(2)}}{\mathcal{L}}}_{{\underline{\breve{R}}}}$ contributes to the field equations by $$\begin{gathered}
\frac{3}{8}\bigg(
-{\tensor{R}{^{(a}_k}}{\tensor{X}{_l^{b)}^k^l}}
+\frac{1}{2}{\tensor{R}{^k_l}}{\tensor{X}{^l_m^m_k}}g^{ab}+
\nabla_k \nabla^{(a} {\tensor{X}{^{b)}_l^{lk}}}
-\frac{1}{2}\nabla_l \nabla^l {\tensor{X}{^a_k^{kb}}}\\
-\frac{1}{2}g^{ab}\nabla_k \nabla_l {\tensor{X}{^k_m^{ml}}}
-2\nabla_k \nabla^l\left({\tensor{R}{^{(a}_m}}{\tensor{Y}{_l^{b)mk}}}\right)
+2\nabla_k \nabla_l\left({\tensor{R}{^{km}}}{\tensor{Y}{^{l(a}_m^{b)}}} \right)
\bigg).\end{gathered}$$ Analogously, from $-\frac{3}{4}\tensor{X}{^{k'}_{l'}^{m'}_{p'}}\tensor{R}{^{l'p'}_{k'm'}}$ one obtains $$\frac{3}{4}\bigg({\tensor{R}{_{km}^{l(a}}}{\tensor{X}{^{b)m}^k_l}}
+\frac{1}{2}\nabla_k \nabla_l {\tensor{X}{^{k(ab)l}}}
+2 \nabla_k \nabla_l\left( {\tensor{R}{^{mjk(a}}}{\tensor{Y}{^{b)l}_{mj}}} \right)
+\frac{1}{2}{\tensor{R}{^{lm}_{jk}}}{\tensor{X}{^j_l^k_m}}g^{ab}
\bigg).$$ Finally, the expression $-\frac{1}{8}h^2 \Lambda^{a b}\Lambda^{c d} {\partial^{S}}_{b}{\partial^{S}}_{d}{\partial^{S}}_{a}{\partial^{S}}_{c} R$ is responsible for $$\frac{1}{8}\bigg(
-{\tensor{R}{^{ab}}}Z
+\nabla^a \nabla^b Z-g^{ab}\nabla_l\nabla^l Z
+\frac{1}{2}g^{ab}\Lambda^{jk}\Lambda^{lm} {\partial^{S}}_{k}{\partial^{S}}_{m}{\partial^{S}}_{j}{\partial^{S}}_{l}R
\bigg),$$ being the second source of terms involving symplectic connection.
Corrections to the metric
-------------------------
In all considered cases field equations are of the form $G_{ab}=W_{ab}+O(h^3)$, where $G_{ab}=R_{ab}-\frac{1}{2}R g_{ab}$ is the Einstein tensor, and the term $W_{ab}$ is of $h^2$ order i.e. $W_{ab}=h^2 {\accentset{\mathit{(2)}}{W}}_{ab}+O(h^3)$. Let us investigate how $W_{ab}$ influences the metric. For this pourpose one can rewrite $g_{ab}$ as a formal power series with respect to $h$ $$g_{ab}={\accentset{\mathit{(0)}}{g}}_{ab}+h{\accentset{\mathit{(1)}}{g}}_{ab}+h^2 \,{\accentset{\mathit{(2)}}{g}}_{ab}+\dots$$ Coefficients of Levi-Civita connection corresponding to $g_{ab}$ can be written as $${\tensor{\Gamma}{^a_{bc}}}={\tensor{{\accentset{\mathit{(0)}}{\Gamma}}}{^a_{bc}}}+h{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{bc}}}+h^2\,{\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}}+\dots$$ One can quite easily calculate that
$$\begin{aligned}
\label{eh_gamma_zero}
{\tensor{{\accentset{\mathit{(0)}}{\Gamma}}}{^a_{bc}}} &=& \frac{1}{2}{\accentset{\mathit{(0)}}{g}}^{\;ak}\left( \frac{\partial {\accentset{\mathit{(0)}}{g}}_{kb}}{\partial x^c }+ \frac{\partial {\accentset{\mathit{(0)}}{g}}_{kc}}{\partial x^b } -
\frac{\partial {\accentset{\mathit{(0)}}{g}}_{bc}}{\partial x^k }\right),\\
\label{eh_gamma_one}
{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{bc}}} &=& \frac{1}{2}{\accentset{\mathit{(0)}}{g}}^{\;ak}\left({\accentset{\mathit{(0)}}{\nabla}}_{c} {\accentset{\mathit{(1)}}{g}}_{kb} + {\accentset{\mathit{(0)}}{\nabla}}_{b} {\accentset{\mathit{(1)}}{g}}_{kc} - {\accentset{\mathit{(0)}}{\nabla}}_{k} {\accentset{\mathit{(1)}}{g}}_{bc}\right),\\
{\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}} &=& \frac{1}{2}{\accentset{\mathit{(0)}}{g}}^{\;ak}\left({\accentset{\mathit{(0)}}{\nabla}}_{c} {\accentset{\mathit{(2)}}{g}}_{kb} + {\accentset{\mathit{(0)}}{\nabla}}_{b} {\accentset{\mathit{(2)}}{g}}_{kc} - {\accentset{\mathit{(0)}}{\nabla}}_{k} {\accentset{\mathit{(2)}}{g}}_{bc}\right)
- {\accentset{\mathit{(0)}}{g}}^{\;ak\;} {\accentset{\mathit{(1)}}{g}}_{kl} {\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^{\,l}_{bc}}},\end{aligned}$$
where ${\accentset{\mathit{(0)}}{\nabla}}$ denotes Levi-Civita connection of metric ${\accentset{\mathit{(0)}}{g}}_{ab}$. Observe that ${\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{bc}}}$ and ${\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}}$ are tensorial objects. Hence, for the Riemann tensor $${\tensor{R}{^a_{bcd}}}={\tensor{{\accentset{\mathit{(0)}}{R}}}{^a_{bcd}}}+h{\tensor{{\accentset{\mathit{(1)}}{R}}}{^a_{bcd}}}+h^2{\tensor{{\accentset{\mathit{(2)}}{R}}}{^a_{bcd}}}+\dots$$ one obtains
\[eh\_riemann\_corr\] $$\begin{aligned}
{\tensor{{\accentset{\mathit{(1)}}{R}}}{^a_{bcd}}} &=& 2{\accentset{\mathit{(0)}}{\nabla}}_{[c} {\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{d]b}}},\\
{\tensor{{\accentset{\mathit{(2)}}{R}}}{^a_{bcd}}} &=& 2{\accentset{\mathit{(0)}}{\nabla}}_{[c} {\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{d]b}}} + 2{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{k[c}}}{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^k_{d]b}}},\end{aligned}$$
and ${\tensor{{\accentset{\mathit{(0)}}{R}}}{^a_{bcd}}}$ is the Riemann tensor of metric ${\accentset{\mathit{(0)}}{g}}_{ab}$. Substituting above relations into field equations and analyzing terms at $h^0$ and $h^1$ one calculates that
\[eh\_g\_corr\] $$\begin{gathered}
\label{eh_g0_corr}
{\tensor{{\accentset{\mathit{(0)}}{R}}}{_{ab}}} = 0,\\
\label{eh_g1_corr}
{\accentset{\mathit{(0)}}{g}}^{\;kl}\left({\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(1)}}{g}}_{bl} + {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(1)}}{g}}_{al} - {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(1)}}{g}}_{ab} - {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(1)}}{g}}_{kl} \right)=0,\end{gathered}$$ where ${\tensor{{\accentset{\mathit{(0)}}{R}}}{_{ab}}}$ is zeroth order term in power series expansion of $R_{ab}$ and also Ricci tensor of ${\accentset{\mathit{(0)}}{g}}_{ab}$. For $h^2$ the following relation can be derived $$\begin{gathered}
\label{eh_g2_corr}
{\accentset{\mathit{(0)}}{g}}^{\;kl}\left({\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(2)}}{g}}_{bl} + {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(2)}}{g}}_{al} - {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(2)}}{g}}_{ab} - {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(2)}}{g}}_{kl}
\right)=\\
=2{\accentset{\mathit{(2)}}{W}}_{ab}- \frac{1}{n-1}{\accentset{\mathit{(0)}}{g}}_{ab} {\accentset{\mathit{(2)}}{W}}
-4{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^k_{l[k}}}{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^l_{b]a}}}
+4\,{\accentset{\mathit{(0)}}{g}}^{\;rk} {\accentset{\mathit{(0)}}{\nabla}}_{[r} \left(\,{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^l_{b]a}}} {\accentset{\mathit{(1)}}{g}}_{kl} \right),\end{gathered}$$ where ${\accentset{\mathit{(2)}}{W}}={\accentset{\mathit{(0)}}{g}}^{\;rs} {\accentset{\mathit{(2)}}{W}}_{rs}$. The term ${\accentset{\mathit{(2)}}{W}}_{ab}$ is given by the following formulae[^8]:
- for $\widehat{\mathcal{S}}_{EH_{1A}}$ $$\label{eh_wab_seh1a}
\begin{split}
{\accentset{\mathit{(2)}}{W}}_{ab}=&-\frac{3}{8}\Bigg(
{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_{(a} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{b)}_l^{lk}}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l {\tensor{{\accentset{\mathit{(0)}}{X}}}{_a_k^k_b}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\tensor{{\accentset{\mathit{(0)}}{X}}}{^k_m^{ml}}}
\Bigg)\\
&-{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l s_2+{\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b s_2,
\end{split}$$
- for $\widehat{\mathcal{S}}_{EH_{1B}}$ $$\begin{split}
{\accentset{\mathit{(2)}}{W}}_{ab}=&-\frac{3}{8}\Bigg(
{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_{(a} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{b)}_l^{lk}}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l {\tensor{{\accentset{\mathit{(0)}}{X}}}{_a_k^k_b}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\tensor{{\accentset{\mathit{(0)}}{X}}}{^k_m^{ml}}}
\Bigg)\\
&-\frac{1}{8}\Bigg(
{\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(0)}}{Z}}-{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_l{\accentset{\mathit{(0)}}{\nabla}}^l {\accentset{\mathit{(0)}}{Z}}
\Bigg)
-{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l s_2+{\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b s_2,
\end{split}$$
- for $\widehat{\mathcal{S}}_{EH_{2A}}$ $$\begin{split}
{\accentset{\mathit{(2)}}{W}}_{ab}=&-\frac{3}{4}\Bigg(
{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_{(a} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{b)}_l^{lk}}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l {\tensor{{\accentset{\mathit{(0)}}{X}}}{_a_k^k_b}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\tensor{{\accentset{\mathit{(0)}}{X}}}{^k_m^{ml}}}
+{\tensor{{\accentset{\mathit{(0)}}{R}}}{_{km}^l_{(a}}}{\tensor{{\accentset{\mathit{(0)}}{X}}}{_{b)}^m^k_l}}\\
&+\frac{1}{2}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\tensor{{\accentset{\mathit{(0)}}{X}}}{^{k}_{(ab)}^l}}
+2 {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l\left( {\tensor{{\accentset{\mathit{(0)}}{R}}}{^{\,mjk}_{(a}}}{\tensor{{\accentset{\mathit{(0)}}{Y}}}{_{b)}^l_{mj}}} \right)
+\frac{1}{2}{\tensor{{\accentset{\mathit{(0)}}{R}}}{^{\,lm}_{jk}}}{\tensor{{\accentset{\mathit{(0)}}{X}}}{^j_l^k_m}}{\accentset{\mathit{(0)}}{g}}_{ab}
\Bigg)\\
&
-{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l s_2+{\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b s_2,
\end{split}$$
- for $\widehat{\mathcal{S}}_{EH_{2B}}$ $$\label{eh_wab_seh2b}
\begin{split}
{\accentset{\mathit{(2)}}{W}}_{ab}=&-\frac{3}{4}\Bigg(
{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_{(a} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{b)}_l^{lk}}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l {\tensor{{\accentset{\mathit{(0)}}{X}}}{_a_k^k_b}}
-\frac{1}{2}{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\tensor{{\accentset{\mathit{(0)}}{X}}}{^k_m^{ml}}}
+{\tensor{{\accentset{\mathit{(0)}}{R}}}{_{km}^l_{(a}}}{\tensor{{\accentset{\mathit{(0)}}{X}}}{_{b)}^m^k_l}}\\
&+\frac{1}{2}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\tensor{{\accentset{\mathit{(0)}}{X}}}{^{k}_{(ab)}^l}}
+2 {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l\left( {\tensor{{\accentset{\mathit{(0)}}{R}}}{^{\,mjk}_{(a}}}{\tensor{{\accentset{\mathit{(0)}}{Y}}}{_{b)}^l_{mj}}} \right)
+\frac{1}{2}{\tensor{{\accentset{\mathit{(0)}}{R}}}{^{\,lm}_{jk}}}{\tensor{{\accentset{\mathit{(0)}}{X}}}{^j_l^k_m}}{\accentset{\mathit{(0)}}{g}}_{ab}
\Bigg)\\
&-\frac{1}{8}\Bigg(
{\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(0)}}{Z}}-{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_l{\accentset{\mathit{(0)}}{\nabla}}^l {\accentset{\mathit{(0)}}{Z}}
\Bigg)
-{\accentset{\mathit{(0)}}{g}}_{ab}{\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(0)}}{\nabla}}^l s_2+{\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b s_2,
\end{split}$$
where ${\tensor{{\accentset{\mathit{(0)}}{X}}}{^{ijkl}}}$, ${\tensor{{\accentset{\mathit{(0)}}{Y}}}{^{ijkl}}}$, ${\accentset{\mathit{(0)}}{Z}}$ are zeroth order terms in power series expansion of $X_{ijkl}$, $Y_{ijkl}$ and $Z$, which can be expressed by means of ${\tensor{{\accentset{\mathit{(0)}}{R}}}{^i_{jkl}}}$ and ${\accentset{\mathit{(0)}}{g}}_{ab}$.
Thus, in all cases equations which describe deformed metric are of the same structure. At $h^0$ one is dealing with arbitrary Ricci-flat metric ${\accentset{\mathit{(0)}}{g}}_{ab}$. The $h^1$ correction ${\accentset{\mathit{(1)}}{g}}_{ab}$ can be understood as a classical[^9] (undeformed) first order perturbation of ${\accentset{\mathit{(0)}}{g}}_{ab}$, governed by the linear homogeneous equations (\[eh\_g1\_corr\]). Noncommutativity appears for the first time at $h^2$. Correction ${\accentset{\mathit{(2)}}{g}}_{ab}$ is given by linear, inhomogeneous equations (\[eh\_g2\_corr\]). The homogeneous part of (\[eh\_g2\_corr\]) is expressed by the same linear operator as that of (\[eh\_g1\_corr\]). The inhomogeneous part consists of two groups of terms – the one describing interaction with first order perturbation ${\accentset{\mathit{(1)}}{g}}_{ab}$, and the other given by $W_{ab}$, with purely noncommutative origin. Discarding first order classical perturbation by putting ${\accentset{\mathit{(1)}}{g}}_{ab}=0$, we are able to point out special solution of (\[eh\_g2\_corr\]) for actions $\widehat{\mathcal{S}}_{EH_{1A}}$ and $\widehat{\mathcal{S}}_{EH_{1B}}$. It reads $$\label{sol-g2-seh1a}
{\accentset{\mathit{(2)}}{g}}_{ab}=-\frac{3}{8} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{ak}^k_b}} - \frac{1}{n-1}\left(s_2 -\frac{3}{16} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{mk}^{km}}} \right){\accentset{\mathit{(0)}}{g}}_{ab}$$ for $\widehat{\mathcal{S}}_{EH_{1A}}$, and $$\label{sol-g2-seh1b}
{\accentset{\mathit{(2)}}{g}}_{ab}=-\frac{3}{8} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{ak}^k_b}} - \frac{1}{n-1}\left(s_2 -\frac{1}{8}{\accentset{\mathit{(0)}}{Z}} -\frac{3}{16} {\tensor{{\accentset{\mathit{(0)}}{X}}}{_{mk}^{km}}} \right){\accentset{\mathit{(0)}}{g}}_{ab}$$ for $\widehat{\mathcal{S}}_{EH_{1B}}$. (Here, like in (\[eh\_wab\_seh1a\]) – (\[eh\_wab\_seh2b\]), indices at ${\tensor{{\accentset{\mathit{(0)}}{X}}}{}}$ are manipulated by ${\accentset{\mathit{(0)}}{g}}_{ab}$). Let us observe that the difference between above solutions and arbitrary other solution of (\[eh\_g2\_corr\]), with ${\accentset{\mathit{(1)}}{g}}_{ab}=0$, must be a solution of homogeneous variant of (\[eh\_g2\_corr\]). Thus, such a difference may be interpreted as a classical perturbation of metric ${\accentset{\mathit{(0)}}{g}}_{ab}$. For this reason one can regard (\[sol-g2-seh1a\]) and (\[sol-g2-seh1b\]) as the solutions carrying full information about considered noncommutativity at $h^2$.
Palatini action
===============
Now, let us switch to the Palatini formalism with the connection and the tetrad field as separate dynamical variables. Thus, one is dealing with the vector bundle ${\mathcal{L}}$ for which $SO(3,1)$ transformations preserve the canonical form of the Lorentzian metric $\eta_{AB}$. The bundle ${\mathcal{L}}$ is equipped with some metric-compatible connection $\partial^{{\mathcal{L}}}$ . Its local coefficients are denoted as $\tensor{\undertilde{\Gamma}}{^{AB}_i}$ and are antisymmetric in $^{AB}$. The corresponding curvature is given by ${\tensor{\undertilde{R}}{^A_{Bij}}}$. The bundle ${\mathcal{E}}$ is taken to be ${\mathcal{L}}\otimes T{\mathcal{M}}$. The tetrad field ${\tensor{\theta}{^A_{b'}}}$ induces the metric $g_{a'b'}={\tensor{\theta}{^A_{a'}}}\eta_{AB}{\tensor{\theta}{^B_{b'}}}$ and the metric connection $\nabla$ in $T{\mathcal{M}}$ (not necessarily torsionless). Local coefficients of $\nabla$ can be computed from the expression $\tensor{\Gamma}{^{i'}_{j'k}}
={\tensor{\theta}{_A^{i'}}}\tensor{\undertilde{\Gamma}}{^A_B_k}{\tensor{\theta}{^B_{j'}}}+{\tensor{\theta}{_A^{i'}}}\frac{\partial}{\partial x^k}{\tensor{\theta}{^A_{j'}}}$. The curvature tensors are related by $\tensor{R}{^{i'}_{j'kl}}={\tensor{\theta}{_A^{i'}}}{\tensor{\undertilde{R}}{^A_{Bkl}}}{\tensor{\theta}{^B_{j'}}}$. As a connection in ${\mathcal{E}}$ we choose ${\partial^{{\mathcal{E}}}}=\partial^{{\mathcal{L}}} \otimes 1+1 \otimes \nabla$. These data encode the $*$-product $*_P$.
We are going to make use of the following two endomorphisms of ${\mathcal{L}}\otimes T{\mathcal{M}}$: ${\tensor{\undertilde{\breve{R}}}{}}$, i.e. rescaled[^10] by $v$ version of ${\tensor{\undertilde{R}}{^A_B^{a'}_{b'}}}$ (defined by the curvature of $\partial^{{\mathcal{L}}}$ and the tetrad which raises index ${}^{a'}$), and ${\tensor{\Theta}{}}$ given by ${\tensor{\Theta}{^A_B^{a'}_{b'}}}={\tensor{\theta}{^{Aa'}}}{\tensor{\theta}{_{Bb'}}}$. As a starting point one may take the following version of Palatini action $$\mathcal{S}_{P}=\int_{{\mathcal{M}}}\operatorname{Tr}{\tensor{\undertilde{\breve{R}}}{}}\Theta\,\frac{\omega^n}{n!}$$ The deformation procedure yields particularly simple expression due to $\partial_i {\tensor{\Theta}{}}=0$, $\operatorname{Tr}({R^{{\mathcal{E}}}}_{a b} \{{\tensor{\undertilde{R}}{}},{\tensor{\Theta}{}}\})=0$ and $\operatorname{Tr}({R^{{\mathcal{E}}}}_{a b}{R^{{\mathcal{E}}}}_{c d} \{{\tensor{\undertilde{R}}{}},{\tensor{\Theta}{}}\})=0$. $$\label{pal_def_act_h2}
\widehat{\mathcal{S}}_{P}=\operatorname{tr}_{*_P}({\tensor{\undertilde{\breve{R}}}{}}*{\tensor{\Theta}{}})=\int_{{\mathcal{M}}}\operatorname{Tr}\Big({\tensor{\undertilde{R}}{}}{\tensor{\Theta}{}} +h^2 s_2{\tensor{\undertilde{R}}{}}{\tensor{\Theta}{}} +O(h^3)\Big)\operatorname{vol}_M$$ The variation with respect to $\delta{\tensor{\theta}{}}$ leads to the equations
$$(1+h^2 s_2)\left(R_{ab}-\frac{1}{2}g_{ab}R\right)+O(h^3)=0$$
clearly equivalent (up to $h^2$) to the condition $R_{ab}=0$. The variation of the connection field $\delta\undertilde{\Gamma}$ produces [^11] $$\label{palatinitorsion}
(1+h^2 s_2){\tensor{Q}{^a_{bc}}}=\frac{h^2}{n-1} \tensor{\delta}{^a_{[b}} \frac{\partial s_2}{\partial x^{c]}}+O(h^3)
$$
where ${\tensor{Q}{^a_{bc}}}=\tensor{\Gamma}{^a_{cb}}-\tensor{\Gamma}{^a_{bc}}$ is the torsion tensor of the connection $\nabla$. Thus, one obtains the theory with vanishing Ricci tensor and nonvanishing torsion generated by the scalar $s_2$. A quick calculation shows that the trace-free part of ${\tensor{Q}{^c_{ab}}}$ is equal to zero. Equation (\[palatinitorsion\]) means that for $${\tensor{Q}{^a_{bc}}}={\tensor{{\accentset{\mathit{(0)}}{Q}}}{^a_{bc}}}+h{\tensor{{\accentset{\mathit{(1)}}{Q}}}{^a_{bc}}}+h^2\,{\tensor{{\accentset{\mathit{(2)}}{Q}}}{^a_{bc}}}+\dots$$ one has ${\tensor{{\accentset{\mathit{(0)}}{Q}}}{^a_{bc}}}={\tensor{{\accentset{\mathit{(1)}}{Q}}}{^a_{bc}}}=0$ and $${\tensor{{\accentset{\mathit{(2)}}{Q}}}{^a_{bc}}}=\frac{1}{n-1} \tensor{\delta}{^a_{[b}} \frac{\partial s_2}{\partial x^{c]}}.$$ Connection coefficients for $\nabla$ are given by $${\tensor{\Gamma}{^a_{bc}}}=\frac{1}{2} g^{ak} \left(\frac{\partial g_{bk}}{\partial x^c}+
\frac{\partial g_{ck}}{\partial x^b}-\frac{\partial g_{bc}}{\partial x^j}+
{\tensor{Q}{_b_k_c}}+{\tensor{Q}{_c_k_b}}-{\tensor{Q}{_k_b_c}}
\right).$$ Hence, ${\tensor{{\accentset{\mathit{(0)}}{\Gamma}}}{^a_{bc}}}$ and ${\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{bc}}}$ are still expressed by relations (\[eh\_gamma\_zero\]) and (\[eh\_gamma\_one\]). For ${\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}}$ one computes that
$$\label{pal_gam2_corr}
{\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}} = \frac{1}{2}{\accentset{\mathit{(0)}}{g}}^{\;ak}\left({\accentset{\mathit{(0)}}{\nabla}}_{c} {\accentset{\mathit{(2)}}{g}}_{kb} + {\accentset{\mathit{(0)}}{\nabla}}_{b} {\accentset{\mathit{(2)}}{g}}_{kc} - {\accentset{\mathit{(0)}}{\nabla}}_{k} {\accentset{\mathit{(2)}}{g}}_{bc}\right)
- {\accentset{\mathit{(0)}}{g}}^{\;ak\;} {\accentset{\mathit{(1)}}{g}}_{kl} {\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^{\,l}_{bc}}}
+\frac{1}{2(n-1)}\left(
\tensor{\delta}{^a_c} \frac{\partial s_2}{\partial x^b }-
{\accentset{\mathit{(0)}}{g}}_{bc} {\accentset{\mathit{(0)}}{g}}^{\;ak} \frac{\partial s_2}{\partial x^k }
\right).$$
Corrections to Riemann tensor are again given by (\[eh\_riemann\_corr\]). Substituting them to $R_{ab}=0$ we obtain that for $h^0$ and $h^1$ relations (\[eh\_g0\_corr\]) and (\[eh\_g1\_corr\]) remain valid. However, equations for ${\accentset{\mathit{(2)}}{g}}_{ab}$ take the following form $$\begin{gathered}
\label{pal_g2_corr}
{\accentset{\mathit{(0)}}{g}}^{\;kl}\left({\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(2)}}{g}}_{bl} + {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(2)}}{g}}_{al} - {\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l {\accentset{\mathit{(2)}}{g}}_{ab} - {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b {\accentset{\mathit{(2)}}{g}}_{kl}
\right)=\\
=2 {\accentset{\mathit{(0)}}{\nabla}}_a {\accentset{\mathit{(0)}}{\nabla}}_b s_2 + \frac{1}{n-1} \, {\accentset{\mathit{(0)}}{g}}_{ab} {\accentset{\mathit{(0)}}{g}}^{\;kl}{\accentset{\mathit{(0)}}{\nabla}}_k {\accentset{\mathit{(0)}}{\nabla}}_l s_2
-4{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^k_{l[k}}}{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^l_{b]a}}}
+4\,{\accentset{\mathit{(0)}}{g}}^{\;rk} {\accentset{\mathit{(0)}}{\nabla}}_{[r} \left(\,{\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^l_{b]a}}} {\accentset{\mathit{(1)}}{g}}_{kl}\right).\end{gathered}$$
Like in the case of Einstein-Hilbert action, one can easily guess special solution of (\[pal\_g2\_corr\]) by requiring that ${\accentset{\mathit{(1)}}{g}}_{ab}=0$, i.e. that classical first order perturbation vanish. It reads $$\label{pal_spec_sol}
{\accentset{\mathit{(2)}}{g}}_{ab}=- \frac{1}{n-1}s_2{\accentset{\mathit{(0)}}{g}}_{ab}$$ For such case the correction ${\tensor{{\accentset{\mathit{(1)}}{\Gamma}}}{^a_{bc}}}$ is equal to zero, and ${\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}}$ is given by $$\label{pal_spec_sol_conn}
{\tensor{{\accentset{\mathit{(2)}}{\Gamma}}}{^a_{bc}}}=- \frac{1}{2(n-1)}\tensor{\delta}{^a_b} \frac{\partial s_2}{\partial x^c}.$$ Repeating arguments of the previous section, one can point out that arbitrary other solution of (\[pal\_g2\_corr\]), with ${\accentset{\mathit{(1)}}{g}}_{ab}=0$, differs from (\[pal\_spec\_sol\]) by a classical perturbation of ${\accentset{\mathit{(0)}}{g}}_{ab}$. Let us observe that since deformation of the action given by (\[pal\_def\_act\_h2\]) depends up to $h^2$ solely on $s_2$, then corrections (\[pal\_spec\_sol\]), (\[pal\_spec\_sol\_conn\]) are also expressible in terms of $s_2$. Now, $s_2$ is related to the curvature of ${\partial^{S}}$ by the formula (\[pal\_def\_act\_h2\]). In particular this means, that for the Moyal case of flat ${\partial^{S}}$, one is dealing with undeformed theory even at $h^2$.
Relation to the theory of Seiberg-Witten map
============================================
Let us explain how proposed models can be understood in terms of the theory of Seiberg-Witten map. This becomes quite straightforward when one combines results of [@dobrski2] with the property (\[traceinvar\]). Indeed, what [@dobrski2] states is that Seiberg-Witten map is an local isomorphism of $*$-product algebras, while the relation (\[traceinvar\]) says that the trace functional is invariant on such isomorphisms.
More precisely, suppose that one prescribes to each frame $e$ in ${\mathcal{E}}$ $*$-product isomorphism $M_{\langle e \rangle}$ which transforms the initial global product $*$ to the local one $*_S$. (Recall that $*_S$ is nothing but matrix multiplication with commutative product of entries replaced by Fedosov $*$-product of functions). Thus $$M_{\langle e \rangle}(F_{\langle e \rangle}*G_{\langle e \rangle})=M_{\langle e \rangle}(F_{\langle e \rangle})*_S M_{\langle e \rangle}(G_{\langle e \rangle})$$ where $F_{\langle e \rangle}$, $G_{\langle e \rangle}$ are matrices representing endomorphisms $F$ and $G$ in the frame $e$. It turns out ([@dobrski2] section 3.1) that if we switch to different frame $\widetilde{e}=eg^{-1}$ then $M_{\langle e \rangle}$ and $M_{\langle \widetilde{e} \rangle}$ are related by $$\label{seibwitt_covar}
M_{\langle \widetilde{e} \rangle}(F_{\langle \widetilde{e} \rangle})=
\widehat{g}_{{\langle e \rangle}}(g,{\Gamma^{{\mathcal{E}}}}) *_S
M_{\langle e \rangle}(F_{\langle e \rangle}) *_S
\widehat{g}_{{\langle e \rangle}}^{-1}(g,{\Gamma^{{\mathcal{E}}}})$$ with $\widehat{g}_{{\langle e \rangle}}(g,{\Gamma^{{\mathcal{E}}}})=g+O(h)$ dependent both on $g$ and connection one-forms ${\Gamma^{{\mathcal{E}}}}_i$ in the frame $e$ and their derivatives. Moreover, if we combine two gauge transformations, then $\widehat{g}$ fulfills “consistency conditions” (compare [@jurco0; @schupp]) given by $$\label{triv_g_consistency}
\widehat{g}_{{\langle e \rangle}}(g'g,{\Gamma^{{\mathcal{E}}}})=\widehat{g}_{{\langle \tilde{e} \rangle}}(g',g{\Gamma^{{\mathcal{E}}}}g^{-1}+g {\mathrm{d}}g^{-1})*_S \widehat{g}_{{\langle e \rangle}}(g,{\Gamma^{{\mathcal{E}}}}).$$ Thus, $M$ and $\widehat{g}$ behave exactly like Seiberg-Witten map [@seibwitt]. Indeed, if $M$ is set up with Fedosov’s techniques of generating $*$-product isomorphisms, then one can compute[^12] $M$ and $\widehat{g}$, and for the case of $*_S$ given by Moyal product $*_T$ obtain results which are well-known expressions for Seiberg-Witten map ([@dobrski2] section 4).
We are going to rewrite investigated actions in terms of Seiberg-Witten map. Let $M(F)=\widehat{F}$, as it is justified by relations (\[triv\_g\_consistency\]) and (\[seibwitt\_covar\]). Also, let us separately distinguish Moyal case of $*_S=*_T$, for which ${\partial^{S}}$ is flat, one works in Darboux coordinates with coefficients of ${\partial^{S}}$ equal to zero, and the trace functional $\operatorname{tr}_{*_T}$ is given by the integral (\[tracemoyal\]). Suppose that supports of endomorphisms under consideration are small enough to be covered by a single frame in ${\mathcal{E}}$, and – in Moyal case – by a single Darboux coordinates. Then, due to property (\[traceinvar\]), actions considered in this paper can be locally rewritten as follows.
[r| l |l]{} & arbitrary $*_S$ & $*_S=*_T$\
------------------------------------------------------------------------
$\widehat{\mathcal{S}}_{EH_{1A}}=$ & $\displaystyle \operatorname{tr}_{*_S}(\,\widehat{{\underline{\breve{R}}}\,})$ & $\displaystyle \int_{\mathbb{R}^{2n}}\operatorname{Tr}(\,\widehat{{\underline{\breve{R}}}\,}) \; {\mathrm{d}}^{2n}x$\
------------------------------------------------------------------------
$\widehat{\mathcal{S}}_{EH_{1B}}=$ & $\displaystyle \operatorname{tr}_{*_S}(\widehat{{\underline{R}}\,} *_S \widehat{V})$ & $\displaystyle \int_{\mathbb{R}^{2n}}\operatorname{Tr}(\widehat{{\underline{R}}\,} *_T \widehat{V}) \; {\mathrm{d}}^{2n}x$\
------------------------------------------------------------------------
$\widehat{\mathcal{S}}_{EH_{2A}}=$ & $\displaystyle \operatorname{tr}_{*_S}(\,\widehat{{\underline{\underline{\breve{R}}}}\,})$ & $\displaystyle \int_{\mathbb{R}^{2n}}\operatorname{Tr}(\,\widehat{{\underline{\underline{\breve{R}}}}\,}) \; {\mathrm{d}}^{2n}x$\
------------------------------------------------------------------------
$\widehat{\mathcal{S}}_{EH_{2B}}=$ & $\displaystyle \operatorname{tr}_{*_S}(\widehat{{\underline{\underline{R}}}\,} *_S \widehat{V})$ & $\displaystyle \int_{\mathbb{R}^{2n}}\operatorname{Tr}(\widehat{{\underline{\underline{R}}}\,} *_T \widehat{V}) \;{\mathrm{d}}^{2n}x$\
------------------------------------------------------------------------
$\widehat{\mathcal{S}}_{P}=$ & $\displaystyle \operatorname{tr}_{*_S}(\widehat{{\tensor{\undertilde{\breve{R}}}{}}\,} *_S \widehat{\Theta})$ & $\displaystyle \int_{\mathbb{R}^{2n}}\operatorname{Tr}(\widehat{{\tensor{\undertilde{\breve{R}}}{}}\,} *_T \widehat{\Theta})\; {\mathrm{d}}^{2n}x$
Let us observe that such setting clarifies how considered models are related to the spacetime noncommutativity described by $*_S$. Indeed, due to (\[tracecommutation\]), above mentioned local versions of action functionals are invariant with respect to gauge transformations (\[seibwitt\_covar\]) realized by means of $*_S$. Thus, one is able to reasonably claim that models considered in this paper correspond to noncommutativity of spacetime generated by Fedosov product of functions $*_S$.
Discussion
==========
We have obtained number of nonequivalent geometric deformations of vacuum Einstein relativity. They have been analyzed at $h^2$ order, starting from the action functional, through field equations, up to corrections to the metric which have been explicitly given for the case of $\widehat{\mathcal{S}}_{EH_{1A}}$, $\widehat{\mathcal{S}}_{EH_{1B}}$ and $\widehat{\mathcal{S}}_{P}$. Using results of [@dobrski2], we have pointed out the relation between proposed models, the theory of Seiberg-Witten map, and the noncommutativity of the spacetime described by Fedosov $*$-product generated by symplectic form $\omega$ and symplectic connection ${\partial^{S}}$.
The construction scheme we have adopted, relies on the geometric deformation of product of endomorphisms, but it does not include deformations of other geometric data like connection, tensor product, exterior algebra of forms or contraction operator. (Approaches aiming at modifying various structures of classical geometry in the deformation quantization framework certainly exist. These are e.g. [@zumino; @chaichian2; @vassilevich]). The advantage of our approach consists in immediate interpretation in terms of Seiberg-Witten map. On the other hand, the price is that the noncommutativity does not appear as a fundamental structure modifying all the geometry, but rather may seem to be a kind of “extra interaction” entering to action functionals via the procedure described in the introduction.
The multiplicity of models arises as a consequence of ambiguity in translating traditional action functionals to the language of traces of endomorphisms of some bundle. From the gauge simplicity point of view, actions $\widehat{\mathcal{S}}_{EH_{1A}}$ and $\widehat{\mathcal{S}}_{EH_{1B}}$ seem to be most straightforward as they correspond to the natural $GL(2n,\mathbb{R})$ gauging. On the other hand, action $\widehat{\mathcal{S}}_{P}$ produces especially simple expressions for deformed field equations and for corrections to the metric.
The remarkable problem related to presented models concerns incompatibility of the volume forms – metric and symplectic ones. Both proposed solutions (rescaling one of the endomorphisms and multiplication by $V$) seem to be a bit unnatural. One can suspect that this problem is related to fixing symplectic structure as a nondynamical background. Notice however, that Fedosov construction provides natural framework for the variation of the symplectic data. Moreover, it could turn out that some refinements to the Fedosov theory should be made, to put the metric into the internal structure of the deformation quantization procedure. Such considerations are hoped to be covered in author’s subsequent work.
Let us briefly discuss diffeomorphism invariance of proposed models. Clearly they are diffeomorphism invariant in the passive sense, since all actions, field equations and derived corrections to the metrics are given in either explicitly global or coordinate covariant manner. However, they are not invariant under active diffeomorphisms. Again, this issue originates in fixing symplectic data as a nondynamical background. Such observation is a further argument for considering dynamics of $\omega$ and ${\partial^{S}}$ as a natural next step within Fedosov formalism.
Due to the symmetries of the Riemann tensor, in all considered cases imaginary terms at $h^1$ have vanished. It must be stressed however, that we have no clear evidence that the same stays true for other odd powers of $h$. Thus, some further analysis of the reality of proposed actions should be performed. This suggests deeper investigation of the structure of the trace functional, which seems to be rather hard task (but not hopless, as it can inferred from Fedosov’s results [@fedosov_on_the_trace; @fedosov_atiyah_bott_patodi] on relating $\operatorname{tr}_*(1)$ to integrals of characteristic classes of $T{\mathcal{M}}$ and ${\operatorname{End}({\mathcal{E}})}$). On the other hand, construction of some appropriate involution operator in the Fedosov algebra may be useful and it is also matter of author’s further interest.
Finally one could be interested, how the present work is related to the well known existence of *closed* $*$-products (compare e.g. [@felder]). First of all, the existence of such products has been investigated for functions but not for endomorphisms (to the best of author’s knowledge). Moreover, if one is going to treat Seiberg-Witten map in more or less fundamental manner, then nontriviality of the trace is what should be expected. Indeed, as it was argumented in the previous section, the nontrivial trace could be interpreted as the object carrying information about the globalization of Seiberg-Witten map.
Acknowledgments {#acknowledgments .unnumbered}
===============
I am very grateful to professor Maciej Przanowski for his thorough interest in my work and many helpful remarks. I am also indebted to professors Piotr Kosiński and Anatol Odzijewicz for a number of valuable comments.
[99]{} M. Dobrski, Int. J. Geom. Meth. Mod. Phys. **8**, 411 (2011), `arXiv:0904.4409` X. Calmet and A. Kobakhidze, Phys. Rev. D **72**, 045010 (2005) X. Calmet and A. Kobakhidze, Phys. Rev. D **74**, 047702 (2006) P. Mukherjee and A. Saha, Phys. Rev. D **74**, 027702 (2006) A.H. Chamseddine, Phys. Lett. B **504**, 33 (2001) A.H. Chamseddine, J.Math.Phys. **44**, 2534 (2003) M. A. Cardella and D. Zanon, Class. Quant. Grav. **20**, L95-L104 (2003) M. Chaichian, A. Tureanu and G. Zet, Phys. Lett. B **660**, 573 (2008) P. Mukherjee and A. Saha, Phys. Rev. D **77**, 064014 (2008) A.H. Chamseddine, Phys. Rev. D **69**, 024015 (2004) H. Garcia-Compean, O. Obregon, C. Ramirez and M. Sabido, Phys. Rev. D **68**, 044015 (2003) C. Rovelli and M. Gaul in *Towards Quantum Gravity*, proceedings of the XXXV International Winter School on Theoretical Physics, Polanica, Poland, edited by J. Kowalski-Glikman (Springer, Berlin/Heidelberg, 2000) C. Heinicke in *Compendium of Quantum Physics* edited by D. Greenberger, K. Hentschel and F. Weinert (Springer, Berlin/Heidelberg, 2009) R. Banerjee, P. Mukherjee and S. Samanta, Phys Rev. D **75**, 125020 (2007) Y.-G. Miao and S.-J. Zhang, Phys Rev. D. **82**, 084017 (2010) Y.-G. Miao, Z. Xue and S.-J. Zhang, Phys Rev. D. **83**, 024023 (2011) S. Marculescu and F. Ruiz Ruiz, Phys. Rev. D **79**, 025004 (2009) P. Aschieri, C. Blohmann, M. Dimitrijević, F. Meyer, P. Schupp and J. Wess, Class. Quant. Grav. **22**, 3511 (2005) P. Aschieri, M. Dimitrijević, F. Meyer and J. Wess, Class. Quant. Grav. **23**, 1883 (2006) P. Aschieri and L. Castellani, JHEP **0906**, 086 (2009) P. Aschieri and L. Castellani, J. Geom. Phys. **60**, 375 (2010) F. Meyer, `hep-th/0510188` S. Estrada-Jiménez, H. García-Compeán, O. Obregón, and C. Ramírez, Phys. Rev. D **78**, 124008 (2008) H. Steinacker, JHEP **0712**, 049 (2007) H. Steinacker, Nucl. Phys. B **810**, 1 (2009) B. V. Fedosov, *Deformation quantization and index theory* (Akademie Verlag, Berlin, 1996) C. Emmrich and A. Weinstein, in *Lie Theory and Geometry. In Honor of Bertram Kostant*, (Birkhäuser 1994), `hep-th/9311094` D. R. Farkas, Lett. Math. Phys. **51**, 161 (2000) M. Gadella, M. A. del Olmo and J. Tosiek, J. Geom. Phys. **55**, 316 (2005) M. Gadella, M. A. del Olmo and J. Tosiek, Annals Phys. **307**, 272 (2003) J. Tosiek, `arXiv:0907.4911` I. Gelfand, V. Retakh and M. Shubin, Adv. Math. **136**, 104 (1998) P. Bieliavsky, M. Cahen, S. Gutt, J. Rawnsley and L. Schwachhöfer, Int. J. Geom. Meth. Mod. Phys. **3**, 375 (2006) N. Seiberg and E. Witten, JHEP **9909**, 032 (1999) `http://www.exact.es` B. Jurčo, L. Möller S. Schraml, P. Schupp and J. Wess, Eur. Phys. J. C **21**, 383 (2001) P. Schupp, Fortschr. Phys. **54**, 165 (2006) C. Misner, K. Thorne and J. A. Wheeler, *Gravitation*, (W.H. Freeman and Company, San Francisco, 1973) B.V. Fedosov, in *Deformation quantization: proceedings of the meeting of theoretical physicists and mathematicians, Strasbourg, May 31-June 2, 2001*, (de Gruyter, Berlin 2002), `MPIM2001-75` B. V. Fedosov, Comm. Math. Phys. **209**, 691 (2000) S. McCurdy and B. Zumino, AIP Conf. Proc. **1200**, 204 (2010), `arXiv:0910.0459` M. Chaichian, M. Oksanen, A. Tureanu and G. Zet, Phys. Rev. D **81**, 085026 (2010) D. V. Vassilevich, Class. Quantum Grav. **27** 095020 (2010) G. Felder and B. Shoikhet, Lett. Math. Phys. **53**, 75 (2000)
[^1]: We are going to distinguish *passive* and *active* diffeomorphism invariance (compare e.g. [@rovelli; @heinicke]). Here we mean noninvariance in both above senses.
[^2]: Precisely, one has also to fix curvature and normalizing condition for Abelian connection generating Fedosov $*$-product [@fedosov]. Within this paper, standard normalization $\mu \equiv 0$ and curvature $\Omega=-\omega$ are used.
[^3]: The computation leading to (\[fedosov\_trace\]) is quite laborious as one has to deal with connection coefficients which in final step massively cancel and the remaining terms can be grouped to yield tensorial expressions. Large parts of this work has been performed with the significant use of xAct tensor manipulation package [@xact].
[^4]: Index $_2$ corresponds to the presence $s_2$ at $h^2$. Such defined scalar is a symplectic part of what is called *trace density* in [@fedosov]. With similar conventions $s_1=0$.
[^5]: The formula has been rearranged to explicitly exhibit symmetry $\operatorname{tr}_*(F*G)=\operatorname{tr}_*(G*F)$. For the term at $h$ this can be done quickly using integration by parts and definition of ${R^{{\mathcal{E}}}}$. For terms at $h^2$ one can proceed in a following manner. *1)* Take what appears at $h^2$ after simple substitution of (\[fedosov\_endstar\]) into (\[fedosov\_trace\]). Let it call $h^2Q(F,G)$. *2)* Rewrite it as $h^2/2(Q(F,G)+Q(G,F))+h^2/2(Q(F,G)-Q(G,F))$. Drop the antisymmetric part. *3)* Check that discarded terms are indeed equal to zero (as it should be, by the construction). When integrated by parts, the terms with single covariant derivative of ${R^{{\mathcal{E}}}}$ vanish in virtue of Bianchi identity, while the ones with double $\partial$ can be replaced by ${R^{{\mathcal{E}}}}$ and in turn, sum up with remaining terms to give 0. Such calculation can be treated as an additional verification of the formula $(\ref{fedosov_trace})$.
[^6]: As we use exactly the same frame (e.g. coordinate one) for both primed and unprimed indices we can consistently define primed tensors from unprimed ones and vice versa. The prime is used only as a marking for covariant derivation $\partial$.
[^7]: The term with $\partial_{(a} \partial_{c)} {\underline{R}}\partial_{(b} \partial_{d)} V$ is integrated by parts twice, then the covariant derivatives are commuted with the trace and the torsionless property of ${\partial^{S}}$ is used to get rid of symmetrizations.
[^8]: In equations (\[eh\_wab\_seh1a\]–\[eh\_wab\_seh2b\]) indices are manipulated by means of metric ${\accentset{\mathit{(0)}}{g}}_{ab}$.
[^9]: Arguments leading to (\[eh\_g1\_corr\]) are essentially identical to standard calculations concerning small perturbations of classical vacuum relativity, e.g. in *shortwave formalism* ([@misthrnwh] §35.13).
[^10]: The function $v$ modifying ${\tensor{\undertilde{R}}{}}$ is taken with respect to the metric $g_{ab}$ induced by the tetrad. Obviously $\operatorname{vol}_M$ and the volume form given by the determinant of $\theta$ coincide in such case.
[^11]: The variation gives $$w(\tensor{\Gamma}{^c_{ab}}-\tensor{\Gamma}{^c_{ba}})=\\
\tensor{\delta}{^c_a}\left(\frac{\partial w}{\partial x^b} - \tensor{\Gamma}{^d_{bd}}w \right)
-
\tensor{\delta}{^c_b}\left(\frac{\partial w}{\partial x^a} - \tensor{\Gamma}{^d_{ad}}w \right)$$ with the tensor density $w=\sqrt{-g}(1+h^2s_2)$. Contraction of this relation enables expressing $\tensor{\Gamma}{^d_{ad}}$ in terms of $\tensor{\Gamma}{^d_{da}}$, leading in turn to (\[palatinitorsion\]).
[^12]: Fedosov construction enables *computation* of Seiberg-Witten map, up to arbitrary order in $h$, by its recursive techniques. This situation is rather different from the usual framework, where the Seiberg-Witten equations must be *solved*.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'There are some approaches, either based on General Relativity (GR) or modified gravity, that use galaxy rotation curves to derive the matter density of the corresponding galaxy, and this procedure would either indicate a partial or a complete elimination of dark matter in galaxies. Here we review these approaches, clarify the difficulties on this inverted procedure, present a method for evaluating them, and use it to test two specific approaches that are based on GR: the Cooperstock-Tieu (CT) and the Balasin-Grumiller (BG) approaches. Using this new method, we find that neither of the tested approaches can satisfactorily fit the observational data without dark matter. The CT approach results can be significantly improved if some dark matter is considered, while for the BG approach no usual dark matter halo can improve its results.'
author:
- |
Álefe O. F. de Almeida, Oliver F. Piattella and Davi C. Rodrigues[^1]\
Departamento de Física, Universidade Federal do Espírito Santo, Av. Fernando Ferrari 514, Vitória, ES, 29075-910 Brazil
bibliography:
- 'bibdavi2014B.bib'
title: A method for evaluating models that use galaxy rotation curves to derive the density profiles
---
\[firstpage\]
gravitation, dark matter, galaxies: spiral, galaxies: kinematics and dynamics
Introduction
============
Besides the unknown nature of dark matter, the standard model of cosmology ($\Lambda$CDM) is also facing difficulties [e.g., @2009MNRAS.397.1169D; @deBlok:2009sp; @2011AJ....142...24O; @BoylanKolchin:2011de; @BoylanKolchin:2011dk; @Weinberg:2013aya; @Pawlowski:2015qta]. There is hope that these issues may be solvable within the $\Lambda$CDM model [e.g., @Governato:2012fa; @DelPopolo:2014yta; @Onorbe:2015ija], but the solutions depend on baryonic physics details with which is difficult to deal semi-analytically or through simulations. On the other hand, the answer may be related with more issues than the baryonic physics alone, and may depend on the nature of dark matter [@Moore:1994yx; @Colin:2000dn; @Hu:2000ke; @Zavala:2009ms; @Foot:2014uba], on refinements on the gravitational side [@Capozziello:2012ie; @Lora:2013fla; @Rodrigues:2014xka], or perhaps on both.
Galaxy rotation curves (RCs) constitute one of the most clear and useful test on the existence of either dark matter or non-Newtonian gravity in galaxies at low redshift. The determination of the dark matter profile in a galaxy is based on the following schematic procedure [see e.g., @Sofue:2000jx; @Courteau:2013cjm]: the observed light is converted into mass densities for the stellar and the gaseous parts. For the stellar part, the conversion depends on the stellar mass-to-light ratio, which depends on the dominant stelar population. From these mass densities, one derives the corresponding Newtonian potentials, and therefore their individual contributions to the RC. These contributions are typically far from being sufficient to reproduce the observed RC, the difference being attributed to dark matter. In order to compute the dark matter contribution, the usual procedure is to assume a dark matter halo profile that depends on some free parameters which are fitted to the observed RC.
From Newtonian gravity without assumptions on the matter distribution, it is not possible to infer the mass density of a disk galaxy from the observational RC alone, even if it is assumed that all the matter is in a thin axisymmetric disk [@0691084459]. What can be done is to compare the RC generated by a given mass density profile, with some free parameters, to the observed RC.
On the other hand, some non-Newtonian proposals [@Cooperstock:2006dt; @CoimbraAraujo:2007zz; @Dey:2014gka; @2015arXiv150807491M] use the inverse procedure: the observed RC is used as the input from which the mass distribution is derived. For some galaxies, the RC fits of these theories seem satisfactory, but these publications lack a detailed investigation with respect to the baryonic matter data inferred from observations. Models that use this inverse route have not been yet properly tested and confronted with results from other approaches, it is the purpose of this new method to be able to properly test and compare them. Also, there are other nontrivial approaches that have never been tested with RC data, and this inverse procedure may prove useful to test them [e.g., @Vogt:2007zza; @Balasin:2006cg; @Rahaman:2008dw; @Vieira:2013zba].
To address the latter issue, we propose here the effective Newtonian RC method. To exemplify it, two relativistic approaches are selected, the one of the Refs. [@Cooperstock:2006dt; @Cooperstock:2007sc; @Carrick:2011ac], which we label CT, and the one of Ref. [@Balasin:2006cg], which we label BG. Both of them use GR in 4D spacetime as the source for gravitational dynamics. Details, merits and criticisms on these approaches are presented in the next sections. This is the first time that the BG approach is studied with realistic galaxy data.
The next section reviews the CT and the BG approaches, sec. \[sec3\] presents the effective Newtonian method in generality and its two applications, sec. \[sec4\] shows the results, and in sec. \[sec5\] we present our conclusions and discussions.
A brief review on two relativistic approaches {#sec2}
=============================================
General considerations
----------------------
In order to determine the distribution of dark matter in galaxies, the standard approach is to use Newtonian gravity. The motivation for doing so comes usually from the following: $i$) the assumption that GR is the gravitational theory to be considered; $ii$) that galaxies seem to be stationary systems whose Newtonian potential is small (typically about $\sim 10^{-6}-10^{-8}$, in $c = 1$ units), and $iii$) that the typical speeds are at most about a few hundred km/s (i.e., $\lesssim 10^{-3}$, using $c = 1$ units). These small numbers [*suggest*]{} that GR corrections to Newtonian dynamics, namely on the rotation curve (RC), are smaller than 1$\%$, and therefore significantly smaller than the typical uncertainties associated with the astrophysical data from galaxies. Some authors agree with the assumption of using GR in galaxies (item $i$), but found that those small numbers in the items $ii$ and $iii$ may lead to significative consequences on dark matter distribution in galaxies and corrections to the RCs larger than 10$\%$ [@Cooperstock:2006dt; @Carrick:2011ac; @2015arXiv150807491M; @Balasin:2006cg; @RamosCaro:2012rz].
The CT approach [@Cooperstock:2006ti; @Cooperstock:2006dt; @Cooperstock:2007sc; @Carrick:2011ac; @2015arXiv150807491M] has received a number of criticisms on the theoretical basis, which the authors claim to have answered [@Carrick:2011ac]. These criticisms focus on whether the CT approach is indeed fully embedded in GR with a single kind of matter given by a disk of dust. Apart from the theoretical issues, the authors in Refs. [@Cooperstock:2006dt; @Carrick:2011ac] also do a strong claim, which is that their approach is capable of reproducing the internal dynamics of about 10 commonly studied galaxies without the need for dark matter, or with only a small amount of it. It appears that, apart from the present work, there is only a single work that criticises the CT phenomenological consequences [@Fuchs:2006pr]. Their criticism considers a single galaxy, the Milky Way, and not its rotation curve, but only the velocity dispersion of stars outside the galactic plane or in the Sun neighbourhood. The latter is a valid criticism, but too specific and more prone to observational systematical errors that could erroneously invalidate the model.
The BG approach [@Balasin:2006cg] is actually a bifurcation of the CT approach that avoids certain issues on the CT solution at the galactic plane ($z = 0$) [@Vogt:2005va; @Balasin:2006cg]. It is claimed that this approach cannot remove dark matter in galaxies, but it can significantly reduce its total amount [@Balasin:2006cg].
The Cooperstock & Tieu (CT) approach
------------------------------------
The CT approach [@Cooperstock:2006dt; @Cooperstock:2006ti; @Cooperstock:2007sc; @Carrick:2011ac; @2015arXiv150807491M] starts from the assumption that all the relevant matter in a galaxy can be modelled by an axisymmetric stationary dust fluid, and that the spacetime metric can be written as, using $c = 1$ units and standard conventions of the cylindrical coordinates $(r, \phi, z)$, $$\label{metricact}
ds^2 = -e^{w}(dt-Nd\phi)^2 + e^{-w}r^2d\phi^2 + e^{\nu-w}(dr^2+u \, dz^2),$$ where $w,N,\nu,u$ are functions that only depend on the coordinates $r,z$. The above line element is not necessarily the most convenient to work on the dynamics of galaxies, but it is the most general with the desired symmetry [e.g., see Chap. 7 of @0226870332]. Since all the considered matter is dust, it is possible to select coordinates to reduce the number of functions that the metric depends on [@Balasin:2006cg], such that the line element becomes $$\label{metricact}
ds^2 = -(dt-Nd\phi)^2 + r^2d\phi^2 + e^{\nu}(dr^2 + dz^2).$$
By either performing a suitable coordinate change [@Carrick:2011ac] or an ADM splitting to unveil the lapse function and the shift vector [@Balasin:2006cg], an asymptotic observer at rest with respect to the galaxy center would perceive space rotation with a velocity profile given by $$\label{velocity}
V= \frac{N}{r}\,.$$
The GR field equations impose limits on the form of $N$, in particular from the field equations, $$\label{eq1first}
N_{rr}+N_{zz}-\frac{N_{r}}{r}=0\:.$$ For $z\geq 0$, the following is a valid expression for $V$ [@Cooperstock:2006dt], $$\label{velocidadect}
V_{\mbox{\tiny CT}} (r, z\geq 0)= - \sum_{n}D_{n}e^{-k_{n} z }J_{1}\left ( k_{n}r \right )\:,$$ where $J_\alpha$ is a Bessel function of the first kind, and $k_n$ and $D_n$ are arbitrary constants.
The main point of the CT approach is to present a non-Newtonian solution for galaxies that can fit the data without dark matter or a small amount of it, and this non-Newtonian solution needs not to be the most general case. To this end, @Cooperstock:2006dt have found that they could achieve interesting results by setting the constants $k_n$ to be the $n$-th root of the Bessel function $J_0(k_n \, r_{\mbox{\scriptsize max}})$, where $r_{\mbox{\scriptsize max}}$ is the radius of the farthest observed circular velocity data of a given galaxy. Therefore, at $z=0$, it is not a surprise that eq. (\[velocidadect\]) can fit very well the rotation curve of galaxies, it is just a kind of Fourier-Bessel series, which can actually fit almost any curve defined in the interval $(0,r_{\mbox{\scriptsize max}})$. The nontrivial part of eq. (\[velocidadect\]) is the $z$ dependence. Hence, it should be clear that the fact that this approach can match very well the observational rotation curve data at $z=0$ is irrelevant for the phenomenology, it is a triviality, since it can fit any curve. What is not trivial is whether the corresponding inferred mass distribution matches the observed baryonic density.
The extension of eq. (\[velocidadect\]) by using $ V(r,z) = V(r,-z)$ is a problematic one and was criticised in particular by @Vogt:2005va [@Balasin:2006cg] [see however @Carrick:2011ac].
The connection between the velocity profile and the matter distribution is derived from the following GR equation, $$\label{eq2}
\frac{N_{r}^2 + N_{z}^2}{r^2}= 8\pi G \rho \, e^\nu \approx 8\pi G \rho \:.$$
To apply this approach to galaxy RC’s, from the RC data one uses eq. (\[velocidadect\]) to derive the $D_n$ coefficients. The fit of the curve (\[velocidadect\]) to the observational RC can be as good as one wants, the higher precision one demands, the larger is the number of $D_n$’s to be fitted. Typically this approach uses about 10 coefficients [@Cooperstock:2006dt] (this is just a matter of convention and was found to be suitable to a certain number of galaxies). Hence, the RC fit alone of this approach is physically irrelevant. The physically important consequences are derived from the matter distribution $\rho$ that is inferred from the RC from eq. (\[eq2\]).
The high number of free parameters that the CT approach uses, when confronted to phenomenological results [e.g., @Salucci:2007tm], seem to indicate alone that either this model is unrealistic or that there must exist some way to reduce its number of free parameters. Indeed, the rule to only pick the first $n$ roots of the Bessel function is arbitrary. It may exist a rule to select three particular roots of the Bessel function depending on certain galaxy parameters (e.g., the disc scale length), and hence three $D_n$ constants only, which would lead to reasonable galaxy RC fits. Hence, although it is inconvenient that no such rule for selecting the best three $D_n$ parameters is known, the argument on the number of free parameters is not sufficient to discredit this approach.
The Balasin & Grumiller (BG) approach
-------------------------------------
The BG approach is a bifurcation of the CT one. It starts from the same line element (\[metricact\]) and the same energy momentum tensor, but their solution respects the reflection symmetry about the $z=0$ plane, contrary to eq. (\[velocidadect\]).
The exact GR field equations derived from the line element (\[metricact\]) and the energy momentum tensor $T_\mu^\nu = \rho U_\mu U^\nu$ read
$$\begin{aligned}
2r\nu_{r}+ N_{r}^2-N_{z}^2 &=&0\;, \label{eq3} \\
r\nu_{z}+ N_{r}N_{z}&=&0\;, \label{eq4} \\
\nu_{rr}+\nu_{zz}+\frac{1}{2r^2}(N_{r}^2+N_{z}^2)&=&0\;, \label{eq5} \\
N_{rr}+N_{zz}-\frac{N_{r}}{r}&=&0\:, \label{eq1} \\
\frac{N_{r}^2 + N_{z}^2}{r^2}&=& 8\pi G \rho \, e^{\nu}\:. \label{densitybg}\end{aligned}$$
@Balasin:2006cg present the following solution for eq.(\[eq1\]), $$\label{solucaobalasin1}
N(r,z)= A_0 + \int_{0}^{\infty}\cos(\lambda z)(r\lambda)A(\lambda)K_{1}(\lambda r)d\lambda\:,$$ where $A(\lambda)$ is a “sufficiently regular” arbitrary function, $A_0$ is a constant and $K_1$ is a modified Bessel function of the second kind. Since the relation between $V$ and $N$ is given by eq. (\[velocity\]), it should be clear that selecting $A(\lambda)$ to fit the observed RC should not be seen as the physical output of this approach, but the physical input. It is shown that a suitable choice of the $A(\lambda)$ function can lead to the following velocity profile, $$\begin{aligned}
\label{velocidadebg}
V_{\mbox{\tiny BG}}(r,z) =&& \mbox{\hspace*{-0.7cm}}\frac{(R-r_{0})V_{0}}r + \frac{V_{0}}{2r }\sum_{\pm}\left(\sqrt{(z\pm r_{0})^2 + r^2}- \right. \nonumber \\
&& \mbox{\hspace*{-0.7cm}} \left. - \sqrt{(z\pm R)^2+r^2}\right),
\end{aligned}$$ with, $|z|< r_0$. Hence, at $z=0$, $$\label{velocidadebg}
V_{\mbox{\tiny BG}}(r,0)= \frac{V_0}r \left ( R-r_{0} + \sqrt{r_{0}^2 + r^2}-\sqrt{ R^2+r^2} \right).$$
This profile includes three stages, first the linear increase (for $r \lesssim r_0$), then the constant velocity $V \sim V_0$ regime (for $ r_0 \lesssim r \lesssim R$), and a $1/r$ decrease for $r \gg R$. In practice, for many galaxies, the parameter $R$ cannot be accurately derived from the observational RC, since the transition to a decreasing RC cannot be seen up to the last RC data.
For circular velocities much lower than the speed of light, and assuming that at $r<r_0$ this approach should coincide with Newtonian gravity, Ref. [@Balasin:2006cg] shows that $\nu$ is (close to) a constant. Finally, by comparing the differences between Newtonian gravity and their GR approach at the plateau part of the RC, it is argued that this GR approach may significantly reduce the need of dark matter (the estimated differences being about $30 \%$ of the total matter).
In the following, to simplify the problem in this first step, we will consider both the CT and the BG approaches without dark matter.
The effective Newtonian rotation curve method {#sec3}
=============================================
General considerations
----------------------
The purpose of this method is to properly and feasibly evaluate models in which the observational RC is used as the model input, while the mass density profile is derived from the latter. The essential feature is to circumvent the use of the commonly unknown matter density error bars by a proper, model dependent, transposition of the observational RC error bars to an effective Newtonian RC. In the end, the method provides an effective RC with error bars that should be fitted with the usual Newtonian procedures.
For the majority of the works on galaxy RC data, and for diverse reasons, there is no profile stating the values of the baryonic density at each radius and its corresponding uncertainty. Therefore, if a model can derive a baryonic density profile by certain means, it is not obvious how to compare it with the expected baryonic profile from the observations.
The proposed method uses that the relevant uncertainties are already in part encoded in the RC error bars. In Refs. [@2008AJ....136.2648D; @Gentile:2004tb], like in many others, changes in the redshift data at the same galaxy radius are the main contribution to the RC error bars at that radius. Therefore, the RC error bars contain information on the violation of perfect axial symmetry, and hence they include information on the maximal confidence one can have on any axially symmetric model.
The method that is here proposed depends on the realisation of two minimization procedures. The first one is to derive the model parameters that best fit the observed RC (this fit does not depend on neither the baryonic or the dark matter densities). The second minimization is used to derive the baryonic (and dark matter) parameters, and to yield the relevant quantities to evaluate the goodness of the fit. If one is suspicious on the value of a baryonic model parameter, say the disc scale length, or wants to consider it within a certain range, one can do the last fit with different values of that parameter (without the need of redoing the first fit). The method can be briefly summarised in the following steps:\
- The model circular velocity at $z=0$, which is designated by $V(r, p_i)$, where $p_i$ represent the model parameters, is fitted to the observed RC. This RC is described by the table whose $k$-th line reads ($r_k, V_{\mbox{\tiny Obs}, k}, \delta V_{\mbox{\tiny Obs}, k}$), where $r_k$ is the radius of the galaxy whose corresponding circular velocity is $V_{\mbox{\tiny Obs},k}$ with a 1$\sigma$ error bar given by $\delta V_{\mbox{\tiny Obs},k}$. The fit determines the best fit parameters $\bar p_i$ and the corresponding error bars $\delta p_i$.
- From $V(r, \bar p_i \pm \delta p_i)$ one can (numerically) determine the corresponding mass density profile as a function of the model parameters, $\rho(r, z, \bar p_i \pm \delta p_i)$, for instance from eqs. (\[velocity\], \[eq2\]).
- From $\rho(r, z, \bar p_i \pm \delta p_i)$ one can determine the effective Newtonian circular velocity at $z=0$. The latter is written as $V_{\mbox{\scriptsize eN}}(r, \bar p_i \pm \delta p_i)$ and it is defined as being the circular velocity derived from Newtonian gravity for the matter density $\rho(r,z, \bar p_i \pm \delta p_i)$.
- The effective Newtonian RC data with error bars is built from $V_{\mbox{\scriptsize eN}}$. These data can be expressed as a table whose $k$-th line is given by ($r_k, \bar V_{\mbox{\scriptsize eN},k}, \delta V_{\mbox{\scriptsize eN}, k}$), where $r_k$ assumes the same values of the original data on the observational RC, $\bar V_{\mbox{\scriptsize eN},k} = V_{\mbox{\scriptsize eN}}(r_k, \bar p_i)$ and $\delta V_{\mbox{\scriptsize eN},k}$ is an approximation for the corresponding $1\sigma$ error bar, which is detailed afterwards.
- The astrophysical expectation on the gas and stellar densities, together possibly with a given dark matter profile, are used to derive the Newtonian circular velocity $V_{\mbox{\scriptsize N}}$, which will depend on baryonic parameters (like the mass-to-light ratios) and possibly on dark matter parameters as well.
- If the gravitation theory being considered is compatible with both the observational RC and the matter content assumed for the galaxy, then $V_{\mbox{\scriptsize eN}}$ and $V_{\mbox{\scriptsize N}}$ should be mutually compatible. Hence, one fits $V_{\mbox{\scriptsize N}}$ to the effective Newtonian RC data, thus deriving the baryonic (and dark matter) parameters, and deriving the quantities $\chi^2$ and $\chi^2_{\mbox{\scriptsize red}}$. The latter are the quantities that have physical information on the quality of the fit and that can be compared to other approaches.
The fit procedure step by step {#subsec:fitprocedure}
------------------------------
Here we describe in detail the procedures associated to the proposed method in four steps:
1. [*The derivation of $\bar p_i$ and $\rho$.*]{} The minimisation of certain $\chi^2$ is used to compute the best fit parameters for $V(r,p_i)$ in regard to the observed RC data. The corresponding $\chi^2$ quantity is $$\label{chisquare}
\chi^2_p =\sum_{k=1}^{N}\left(\frac{V(r_{k},p_i)-V_{\mbox{\scriptsize Obs},k}}{\delta V_{\mbox{\scriptsize Obs},k}}\right)^2\;.$$ The subscript $p$ is a reminder that the purpose of $\chi^2_p$ is to derive the model parameters $p_i$ from the observational RC data (this does not constitute the main model results). The values of $p_i$ that minimise $\chi^2_p$ are denoted by $\bar p_i$, and $N$ is the total number of observational data points of the circular velocity $V_{\mbox{\scriptsize Obs}}$. From the knowledge of $V(r,\bar p_i)$ it is straightforward to evaluate the matter density $\rho(r,z,\bar p_i)$. It is also possible to evaluate $\rho$ for all the values of $p_i$ inside the range given by the error bars $\delta p_i$, hence one can derive $\rho(r,z,\bar p_i \pm \delta p_i)$.
2. [*The derivation of $V_{\mbox{\scriptsize eN}}$.*]{} The effective Newtonian circular velocity can be derived by solving the Poisson equation, $\nabla^ 2\Phi(r,z,\bar p_i)=4\pi G \rho(r,z,\bar p_i)$, and using $V^{2}_{\mbox{\scriptsize eN}}(r,\bar p_i) =r\partial_r \Phi(r,\bar p_i)$. In particular, the expression for $V_{\mbox{\scriptsize eN}}$ can be directly evaluated from [@0691084459] $$\begin{aligned}
&& \mbox{\hspace*{-0.7cm}}V^2_{\mbox{\scriptsize eN}}(r,p_i) = r \, \partial_r \Phi(r,z=0,p_i) \, , \nonumber \\[.1in]
&& \mbox{\hspace*{-0.7cm}}= -G r \partial_r \int_{-\pi}^\pi d\varphi' \int_{-\infty}^\infty dz' \int_0^\infty dr' \times \nonumber \\
&& \;\;\;\;\;\;\;\;\; \times \frac{\rho(r', z', p_i)}{\sqrt{r^2 + {r'}^2 + {z'}^2 - 2 r r' \cos(\varphi')}} \, r' \, , \nonumber \\[.1in]
&& \mbox{\hspace*{-0.7cm}}= - 2 G r \int_0^\infty dz' \int_0^\infty dr' \times \label{VeNgeneral} \\
&& \;\;\;\;\;\;\;\;\; \times \rho(r',z',p_i) \, \partial_r \left( \frac{4 K\left( \frac{4 r r'}{(r + r')^2 + {z'}^2} \right ) }{\sqrt{(r + {r'})^2 + {z'}^2}}\right) \, r' \, . \nonumber\end{aligned}$$ In the above, $K$ is the complete elliptic integral defined by $$K(x) = F(\pi/2,x) = \int_0^{\pi/2} d\theta (1 - x \sin^2(\theta))^{-1/2}.$$ From the above, it is possible to derive $V_{\mbox{\scriptsize eN}}$ for all the values of $p_i$ inside their 1$\sigma$ uncertainties, that is, one can find $V_{\mbox{\scriptsize eN}}(r, \bar p_i \pm \delta p_i)$.
3. [*The effective Newtonian RC.*]{} The purpose of this step is to generate the relevant data, with error bars, that should be used in the next and final fitting procedure. The observational RC is described by the data $(r_k, V_{\mbox{\tiny Obs},k}, \delta V_{\mbox{\tiny Obs},k})$, where $k$ runs from 1 to $N$. The effective Newtonian RC data are given by $(r_k, \bar V_{\mbox{\tiny eN},k}, \delta V_{\mbox{\tiny eN},k})$. In order to avoid the introduction of any bias towards any radii, the same radial values $r_k$ used for the observational RC also appear for the effective Newtonian RC. The quantity $\bar V_{\mbox{\tiny eN},k}$ is simply $ V_{\mbox{\tiny eN}}(r_k,\bar p_i)$ and $\delta V_{\mbox{\tiny eN},k}$ is its corresponding $1\sigma$ error bar. A straightforward procedure to derive the latter goes as follows: firstly one finds $V_{\mbox{\tiny max},k}$ and $V_{\mbox{\tiny min},k}$, which are respectively the maximum and the minimum of $V_{\mbox{\scriptsize eN}}(r_k, p_i)$, with fixed $r_k$, such that $\chi^2(p_i) \leq \chi^2_{\mbox{\tiny min}} + \Delta \chi^2$, where $\Delta \chi^2$ is the constant associated to a 1$\sigma$ uncertainty considering the total number of the model parameters ($p_i$). This guarantees that $V_{\mbox{\tiny max},k}$ is the maximum value achievable for $V_{\mbox{\scriptsize eN},k}$ inside the 1$\sigma$ confidence region. Ideally one should compute the full probability density function (PDF), but depending on the model it may be either exactly valid or be a reasonable approximation to assume a Gaussian distribution, If the error bars are not exactly symmetric (but are not far from being symmetric), the $1\sigma$ uncertainty $\delta V_{\mbox{\scriptsize eN},k}$ is set as the maximum between $V_{\mbox{\tiny max},k} - \bar V_{\mbox{\scriptsize eN},k}$ and $\bar V_{\mbox{\scriptsize eN},k} - V_{\mbox{\tiny min},k}$. It should be noted that $\Delta \chi^2$ increases with the number of model parameters $p_i$, and hence in general the larger is the number of parameters $p_i$, the larger will be the uncertainties $\delta V_{\mbox{\scriptsize eN},k}$.
4. [*The derivation of the baryonic and dark matter parameters.*]{} Since all the galaxy matter is composed by either baryonic or dark matter, the total Newtonian circular velocity can be expressed by $$\label{velocitybar}
V^2_{\mbox{\scriptsize N}}=V^2_{\mbox{\scriptsize disk}} + V^2_{\mbox{\scriptsize bulge}}+ V^2_{\mbox{\scriptsize gas}} + V^2_{\mbox{\scriptsize dark matter}}\;.$$ To be clear, $V^2_{\mbox{\scriptsize N}}$ is directly derived from certain matter densities as given above, while $V^2_{\mbox{\scriptsize eN}}$ is derived from the observational RC and from the use of the chosen non-Newtonian gravitation.
For concreteness, here it is considered that the stellar mass-to-light ratios of the bulge and the disk ($\Upsilon_{\mbox{\scriptsize*B}}$ and $\Upsilon_{\mbox{\scriptsize*D}}$) are the only baryonic parameters that are not sufficiently constrained by the observations and need to be fitted. The dark matter contribution will not be considered at the moment, that is, $V^2_{\mbox{\scriptsize dark matter}}=0$. In conclusion, for the assumptions above, $V^2_{\mbox{\scriptsize N}}=V^2_{\mbox{\scriptsize N}} (r, \Upsilon_{\mbox{\scriptsize*B}}, \Upsilon_{\mbox{\scriptsize*D}})$.
If the gravitation theory being considered is compatible with the observational RC and the matter content assumed for the galaxy, then $V_{\mbox{\scriptsize eN}}$ and $V_{\mbox{\scriptsize N}}$ should be mutually compatible. Since $V_{\mbox{\scriptsize N}}$ depends on free parameters, one should evaluate a second and last $\chi^2$ minimization, whose quantity to be minimized reads, $$\label{chisquareeN}
\chi^2 =\sum_{k=1}^{N}\left(\frac{V_{\mbox{\scriptsize N}}(r_{k},\Upsilon_{*\mbox{\scriptsize D}},\Upsilon_{*\mbox{\scriptsize B}})-\bar V_{\mbox{\scriptsize eN},k}}{\delta V_{\mbox{\scriptsize eN},k}}\right)^2\;.$$ It is this last $\chi^2$, and the reduced chi-square computed from it ($\chi^2_{\mbox{\scriptsize red}}$), the quantities that should be used to compare different approaches, not $\chi^2_p$.
Application to the CT approach
------------------------------
To apply the effective Newtonian method to the CT approach, we follow the steps detailed in Sec. \[subsec:fitprocedure\].\
1. [*The derivation of $\bar p_i$ and $\rho$*]{}. The results for $\bar p_i$ and its corresponding error bars can be seen in Table \[tab:parametrospi\]. The $p_i$ parameters for this approach correspond to the $D_n$ parameters in eq. (\[velocidadect\]).
For all the six galaxies of this sample we followed the procedure of @Cooperstock:2006dt of adopting 10 parameters to be fit. For this first fit, the CT approach with 10 parameters could easily fit the observational RC. This can be seen from the values of $\chi^2_{\mbox{\scriptsize p,red}}$ in Table \[tab:tabelaresultadoct\]. Although 10 parameters is more than the usual number of parameters used to fit galaxies, this quantity depends on the chosen profile. For the CT approach, which uses eq. (\[velocidadect\]), less then five parameters only leads to good fits (i.e., $\chi^2_{\mbox{\scriptsize p,red}} \sim 1$) for a few galaxies, typically those whose RC slowly and smoothly increases and hence do not need the high frequency terms of the expansion (\[velocidadect\]). There are examples in which 10 parameters are not sufficient [@2015arXiv150807491M].
The derivation of $\rho$ from the fitted circular velocity $V$, at the region with observational RC data, comes from the combination of eqs. (\[velocity\], \[velocidadect\], \[eq2\]).
In general, for the evaluation of $V_{{\mbox{\scriptsize eN}}}$, it is necessary to consider an extension of $\rho$ beyond the farthest observational RC data, whose radius is $r_{{{\mbox{\scriptsize max}}}}$. Namely, the larger is the density beyond $r_{{\mbox{\scriptsize max}}}$, the smaller becomes $V_{{\mbox{\scriptsize eN}}}$ close to $r_{{\mbox{\scriptsize max}}}$ [this is a known Newtonian effect in axisymmetric systems @0691084459]. In principle, one can extend $\rho$ beyond $r_{{\mbox{\scriptsize max}}}$ by simply extending the circular velocity curve towards larger $r$ and using eq. (\[velocidadect\]). But, as explained in detail by @Cooperstock:2006ti, there is no need to use the same $D_n$ and the same $k_n$ beyond $r_{{\mbox{\scriptsize max}}}$, and physically reasonable extensions usually require different values for the latter parameters. From the phenomenological perspective, for sure the baryonic mass density of galaxies must drop at larger radius. As a phenomenologically simple and viable approximation for the total baryonic matter beyond the last observed RC data, we adopt $$\rho(r \geq r_{\mbox{\scriptsize max}}, z) = e^{({r_{\mbox{\scriptsize max}} - r})/r_{\mbox{\scriptsize d}}} \rho(r_{\mbox{\scriptsize max}}, z). \label{eq:rhoCTextension}$$ This extension is specially natural for the case of a disk galaxy with negligible gas content, since it is just an extension of a Freeman disk [@1970ApJ...160..811F]. The gas density usually decays slower than the stellar component, hence for galaxies with significative amount of gas, the above approximation will cease to be a good one at some radius. Nonetheless, the impact of such deviations on $V_{{\mbox{\scriptsize eN}}}$ is insignificant, since only the density beyond but close to $r_{{\mbox{\scriptsize max}}}$ should contribute significantly to $V_{{\mbox{\scriptsize eN}}}$. Moreover, due to the exponential decrease, and the small density at $r_{{\mbox{\scriptsize max}}}$, changes on $r_{{\mbox{\scriptsize d}}}$ by a factor of two have small or negligible impact on $V_{{\mbox{\scriptsize eN}}}$.\
2. [*The derivation of $V_{\mbox{\scriptsize eN}}$*]{}. Since, with the extension above, $\rho$ is known in the complete space, deriving $V_{\mbox{\scriptsize eN}}$ reduces to computing the integral (\[VeNgeneral\]). A technical difficulty can be promptly spotted, and it comes from the large number of parameters that $V_{{\mbox{\scriptsize eN}}}(r,D_n)$ depends on. This difficulty will have consequences to the next step. On the other hand, it is computationally easy to derive the effective Newtonian circular velocity with the best fit $D_n$ parameters, which is written as $\bar V_{{\mbox{\scriptsize eN}}}(r)$.\
3. [*The effective Newtonian RC data.*]{} As detailed in the previous section, these data can be expressed through a table given by $(r_k, \bar V_{{\mbox{\scriptsize eN,k}}}, \delta V_{{\mbox{\scriptsize eN,k}}} )$, hence at this step one should compute the error bars $\delta V_{{\mbox{\scriptsize eN,k}}}$. To this end, it is necessary to perform both a minimization and a maximization of $V_{{\mbox{\scriptsize eN}}}(r,D_n)$ with the constraint $\chi^2(p_i) \leq \chi^2_{{\mbox{\scriptsize min}}} + \Delta \chi^2$ at each radius $r_k$. Thus, for each observational RC data point, and for each one of the six galaxies, one should derive constrained maximizations and minimizaitons with 10 free parameters.
For the particular case of the CT approach, it is not essencial to compute $\delta V_{{\mbox{\scriptsize eN,k}}}$ to conclude that this model (without a large amount of dark matter) cannot describe the astrophysical data of galaxies, since in spite of the error bars values, the $V_{{\mbox{\scriptsize eN}}}$ and $V_{{\mbox{\scriptsize N}}}$ are systematically incompatible. Moreover, it is not computationally easy to evaluate $\delta V_{{\mbox{\scriptsize eN,k}}}$ for the CT approach with ten $D_n$ parameters.
To evaluate the CT approach with fewer than 10 parameters is helpful as an illustration and to serve as a basis for an estimation of $\delta V{{\mbox{\scriptsize eN}}}$ when the 10 parameters are considered. The galaxy ESO 116-G12 was selected to be analysed with the CT approach and with only three $D_n$ parameters. The results are in Fig. \[resultadosct3\] and Table \[tab:tabelaresultadoct3\]. The derived values of $\delta V_{{\mbox{\scriptsize eN}}}$ ranges from 0.6 km/s to 4.5 km/s. With the exception of the first point, all the others have $\delta V_{{\mbox{\scriptsize eN}}}$ larger than 1 km/s. The mean $\delta V_{{\mbox{\scriptsize eN}}}$ is about 3 km/s.
{width="90.00000%"}
------------- -- ----------------------------------- ---------------------------------------- ---------- ------------------------------------ -- ----------- ------------------------------------ -- --
Galaxy
$\chi^2_{\mbox{\scriptsize{p}}} $ $\chi^{2}_{\mbox{\scriptsize{p,red}}}$ $\chi^2$ $\chi^2_{\mbox{\scriptsize{red}}}$ $\chi^2 $ $\chi^2_{\mbox{\scriptsize{red}}}$
ESO 116-G12 31.60 2.63 1341.35 95.81 31.15 2.60
------------- -- ----------------------------------- ---------------------------------------- ---------- ------------------------------------ -- ----------- ------------------------------------ -- --
4. [*The derivation of the baryonic parameters.*]{} From the fit of $V_{{\mbox{\scriptsize N}}}$ to the effective Newtonian RC, one derives $\chi^2$, $\chi^2_{{\mbox{\scriptsize red}}}$, $\Upsilon_{* {{\mbox{\scriptsize B}}}}$ and $\Upsilon_{*{{\mbox{\scriptsize D}}}}$. This is a straightforward procedure, and the results are presented and commented in the next section.
Application to the BG approach
------------------------------
It is easier to apply the effective Newtonian method to the BG than to the CT one for some reasons. The numerical integrals are faster to compute, the model always use 3 parameters $(p_i)$ instead of 10, and the extension of $\rho$ beyond the last observed RC data, $r_{{\mbox{\scriptsize max}}}$, is already included as part of this model.\
1. [*The derivation of $\bar p_i$ and $\rho$*]{}. The results for $\bar p_i$ and its corresponding error bars are displayed in Table \[tab:parametrospi\]. The $p_i$ parameters for this approach correspond to the three parameters in eq. (\[velocidadebg\]), i.e., $R$, $r_0$ and $V_0$. This first fit that fixes the $p_i$ parameters yields values for $\chi^2_{{{\mbox{\scriptsize p,red}}}}$ in the range from $0.5$ to $1.6$, thus indicating that the velocity profile of this approach is reasonable for describing the RC of galaxies.
The derivation of $\rho$ from the fitted circular velocity $V$ comes from the combination of eqs. (\[velocity\], \[densitybg\], \[velocidadebg\]). The extension of $\rho$ beyond $r_{{\mbox{\scriptsize max}}}$ is direct in this BG approach, and essentially it depends on a single parameter ($R$). For some galaxies the value of this parameter can be constrained to lie within some kpc’s, but for others, specially those whose RC is monotonously increasing up to $r_{{\mbox{\scriptsize max}}}$, there is no maximum for $R$.\
2. [*The derivation of $V_{{\mbox{\scriptsize eN}}}$.*]{} The effective Newtonian circular velocity $V_{{\mbox{\scriptsize eN}}}$ is directly computed from eq. (\[VeNgeneral\]).\
3. [*The effective Newtonian RC data.*]{} At this step one should compute the error bars $\delta V_{{\mbox{\scriptsize eN,k}}}$. To this end, it is necessary to perform both a minimisation and a maximisation of $V_{{\mbox{\scriptsize eN}}}(r,r_0, V_0,R)$ with the constraint $\chi^2(r_0,V_0,R) \leq \chi^2_{{\mbox{\scriptsize min}}} + \Delta \chi^2$ at each radius $r_k$. Thus, for each observational RC data point, and for each one of the six galaxies, one should derive constrained maximisations and minimisations with 3 free parameters. The derived error bars were either symmetric or close to symmetric, and they were all symmetrized taking the largest value. This was done for all the six galaxies that are in this work evaluated.
4. [*The derivation of the baryonic parameters.*]{} From the fit of $V_{{\mbox{\scriptsize N}}}$ to the effective Newtonian RC, one derives $\chi^2$, $\chi^2_{{\mbox{\scriptsize red}}}$, $\Upsilon_{* {{\mbox{\scriptsize B}}}}$ and $\Upsilon_{*{{\mbox{\scriptsize D}}}}$. This is a straightforward procedure, and the results are presented and commented in the next section.
Results {#sec4}
=======
The results of the fit procedures for CT and BG are in the Tables \[tab:parametrospi\], \[tab:tabelaresultadoct\] and \[tab:tabelaresultadobg\], and the RC plots are shown in Fig. \[resultados1\].
[lrrrrrrrrrrrr]{}\
parameters& & DDO 154 & & ESO 116-G12 & & ESO 287-G13 & & NGC 2403 2D& & NGC 2841& & NGC 3198 1D\
$D_{1} \left(\mbox{km/s}\right)$& &$303\pm 10$& &$962\pm30$& &$3455\pm 87$& & $1864\pm 13$& &$1.155\pm0.017$ & &$4407\pm 54$\
$D_{2}$ (km/s)& &$6.3^{+5.7}_{-5.8}$& &$34\pm22$& &$200^{+61}_{-69}$& & $111.2\pm8.3$& &$1303^{+91}_{-90}$ & &$415\pm31$\
$D_{3}$ (km/s)& &$10.4\pm4.4$& &$27\pm13$& & $191^{+47}_{-46}$ & &$103.5\pm6.3$& & $928^{+78}_{-76}$& &$280\pm23$\
$D_{4}$ (km/s)& &$-1.1\pm3.5$& &$-4\pm14$& &$22\pm 44$ & &$26.5\pm5.5$& & $252\pm55$& &$55\pm21$\
$D_{5}$ (km/s)& &$4.8\pm3.0$& &$9.7^{+9.6}_{-9.7}$& &$60^{+49}_{-43}$ & &$13.4\pm5.0$& & $293\pm58$& &$21.9\pm5.9$\
$D_{6}$ (km/s)& &$-2.1\pm2.8$& &$1\pm11$& &$12\pm23$ & &$11.4\pm4.5$& & $101^{+46}_{-47}$& &$5\pm16$\
$D_{7}$ (km/s)& &$1.4^{+2.7}_{-2.6}$& &$2.1^{+8.0}_{-7.9}$& &$1\pm43$ & &$7.1\pm4.1$& & $112^{+46}_{-47}$& &$26\pm14$\
$D_{8}$ (km/s)& &$-0.3^{+2.3}_{-2.4}$& &$-1.4\pm8.4$& &$3^{+16}_{-15}$& &$7.5\pm3.8$& & $17^{+40}_{-41}$& &$9\pm14$\
$D_{9}$ (km/s)& &$0.7\pm2.0$& &$0.6\pm7.5$& &$16^{+29}_{-28}$& &$8.7\pm3.5$& & $33^{+39}_{-38}$& &$11\pm13$\
$D_{10}$ (km/s)& &$0.2\pm1.5$& &$1.9\pm6.5$& &$-3^{+18}_{-16}$& &$7.9\pm3.2$& & $60^{+28}_{-27}$& &$-0.1\pm11$\
\[.2in\]\
parameters & & DDO 154 & & ESO 116-G12 & &ESO 287-G13 & & NGC 2403 2D& & NGC 2841& & NGC 3198 1D\
\
$R \;\left(\mbox{kpc}\right)$& &$2.1^{+\infty}_{-2.1 }\times 10^7$& & $63^{+\infty}_{-40}$& &$6.7^{+\infty}_{-6.7}\times 10^{7}$& & $4.37^{+\infty}_{-0.10}\times 10^7$& &$109^{+13}_{-11}$ & &$78.9^{+12}_{-9.3}$\
$r_0\;\left(\mbox{kpc}\right)$& &$1.18^{+0.13}_{-0.12}$& &$1.79^{+0.56}_{-0.38}$& &$1.308^{+0.095}_{-0.090}$& & $0.706^{+0.034}_{-0.033}$& &$0.31^{+0.14}_{-0.13}$ & &$2.01^{+0.19}_{-0.18}$\
$V_{0}\;\left(\mbox{km/s}\right)$& &$58.3^{+5.1}_{-1.8}$& &$146^{+40}_{-19}$& & $191.9^{+6.2}_{-2.3}$ & &$141.14^{+0.77}_{-0.76}$& & $345.4^{+7.1}_{-6.8}$& &$197.3^{+6.8}_{-6.3}$\
------------- -- ---------------------------------- ---------------------------------------- ---------- ------------------------------------ -- ---------------------------------- ---------------------------------------- ---------- ------------------------------------ -- ----------- --------------------------------------
Galaxy
$\chi^2_{\mbox{\scriptsize{p}}}$ $\chi^{2}_{\mbox{\scriptsize{p,red}}}$ $\chi^2$ $\chi^2_{\mbox{\scriptsize{red}}}$ $\chi^2_{\mbox{\scriptsize{p}}}$ $\chi^{2}_{\mbox{\scriptsize{p,red}}}$ $\chi^2$ $\chi^2_{\mbox{\scriptsize{red}}}$ $\chi^2 $ $\chi^{2}_{\mbox{\scriptsize{red}}}$
DDO 154 5.93 0.12 103.56 1.73 26.46 0.45 53.29 0.89 50.42 0.87
ESO 116-G12 8.84 1.77 200.75 14.34 63.38 1.22 12.36 0.88 31.15 2.60
ESO 287-G13 13.96 0.87 358.70 14.35 37.68 1.63 2278.12 91.12 36.33 1.58
NGC 2403 2D 239.29 0.86 9200.02 32.17 275.66 0.96 17802.40 62.25 155.59 0.55
NGC 2841 58.18 0.44 23468.80 168.84 75.79 0.55 147.06 1.06 26.52 0.19
NGC 3198 1D 22.14 0.26 2918.39 31.38 59.54 0.65 3733.66 40.14 115.67 1.27
------------- -- ---------------------------------- ---------------------------------------- ---------- ------------------------------------ -- ---------------------------------- ---------------------------------------- ---------- ------------------------------------ -- ----------- --------------------------------------
The CT approach with 10 parameters needs at least about $ 10^5$ times more computational time than the BG approach, hence some approximation for $\delta V_{{\mbox{\scriptsize eN}}}$ was necessary. All the error bars on this approach were taken to be the same with a common value of 4.5 km/s, based on the maximum error value of the case with 3 parameters presented in Fig. \[resultadosct3\]. The latter is plausible since additional parameters on this approach only add Bessel functions of higher frequency, and hence the error bars derived with ten free parameters are not expected to be much larger then the three parameters ones.
The plots in Fig. (\[resultados1\]) clearly show that, for all the galaxies modelled with the CT approach, the form of the effective Newtonian RC (the grey squares in the plots) systematically does not match the form of the Newtonian circular velocity $V_{{\mbox{\scriptsize N}}}$. For all these six galaxies, the best fit Newtonian circular velocity $V_{{\mbox{\scriptsize N}}}$ is too high for small radii, and becomes too low at large radii. The latter behaviour is a clear indication that adding a dark matter halo would significantly improve the fit. By adding dark matter, the effective Newtonian RC is the same, but $V_{{\mbox{\scriptsize N}}}$ changes by the addition of the new component whose most significative contribution to the RC appears at large radii.
For the BG approach, there is no evidence of the same CT systematics. However, there is a less significative tendency with the opposite behaviour, that is, the $V_{{\mbox{\scriptsize N}}}$ curve is too high at large radii. Hence, no significative improvement on the fits are expected if some dark matter profile is considered (at least considering the usual dark matter profiles whose density profile decreases much slower than the baryonic density).
The values of $\chi^2_{{\mbox{\scriptsize p}}}$ in Table \[tab:tabelaresultadoct\] are significantly lower for the CT approach than for the BG one. This is expected, since the first has more free parameters to fit the observational RC. A reduced $\chi^2$ analysis indicates that the BG approach fits better the observational RC, in the sense that its $\chi^2_{{\mbox{\scriptsize red,p}}}$ values are closer to 1.
A good fit related to $\chi^2_{{\mbox{\scriptsize p}}}$ is just a minimum requirement for the proposed model to work, it is not sufficient to show that the model is a good one. For the CT and the BG approaches, the $\chi^2$ values are associated to the effective Newtonian RC fit. It is the fit related to $\chi^2$ that is the physically meaningful fit.
------------- -- ------------------------------------- ------------------------------------- -- ------------------------------------- ------------------------------------- -- ------------------------------------- ------------------------------------- -- --------------------------------------------------- ---------------------------------------------------
Galaxy
$\Upsilon_{\mbox{\scriptsize{*D}}}$ $\Upsilon_{\mbox{\scriptsize{*B}}}$ $\Upsilon_{\mbox{\scriptsize{*D}}}$ $\Upsilon_{\mbox{\scriptsize{*B}}}$ $\Upsilon_{\mbox{\scriptsize{*D}}}$ $\Upsilon_{\mbox{\scriptsize{*B}}}$ $\langle\Upsilon_{\mbox{\scriptsize{*D}}}\rangle$ $\langle\Upsilon_{\mbox{\scriptsize{*B}}}\rangle$
DDO 154 4.18 - 3.24 - 1.25 - 0.2-0.6 -
ESO 116-G12 0.80 - 0.55 - 0.05 - 0.5-1.8 -
ESO 287-G13 1.16 - 0.63 - 1.69 - 0.5-1.8 -
NGC 2403 2D 2.39 0.00 0.23 2.17 0.32 0.63 0.2-0.8 0.3-1.2
NGC 2841 1.66 0.00 0.005 0.24 0.72 1.28 0.4-1.5 0.4-1.7
NGC 3198 1D 1.07 - 0.41 - 0.51 - 0.4-1.6 -
------------- -- ------------------------------------- ------------------------------------- -- ------------------------------------- ------------------------------------- -- ------------------------------------- ------------------------------------- -- --------------------------------------------------- ---------------------------------------------------
Table \[tab:tabelaresultadobg\] shows the stellar mass-to-light ratios. The expected ranges for ESO 116-G12 and ESO 287-G13 are the same stated by @Gentile:2004tb. The other galaxies expectations come from @2008AJ....136.2648D. We considered a factor two of uncertainty to generate the stated ranges in this table [@Bell:2000jt; @Meidt:2014mqa], hence the lower bound is found by dividing the expected value from @2008AJ....136.2648D by two, and the upper bound by multiplying it by two. The CT approach has a tendency towards higher $\Upsilon_*$ values, while the BG one tends towards low $\Upsilon_*$ values. This indicates that, by adding a dark matter halo to these approaches, the CT one may benefit from it, achieving better agreement with the expected $\Upsilon_*$ values, but the BG approach cannot improve and may worsen the $\Upsilon_*$ concordance if the presence of dark matter is considered.
Conclusions and Discussion {#sec5}
==========================
There are some models that use the observational rotation curve (RC) data of galaxies to derive the matter density, while the usual route is the opposite one. This class of models have appeared in the context of pure General Relativity (GR) in four dimensional spacetime [e.g. @Cooperstock:2006dt; @Balasin:2006cg], in higher dimensions [e.g., @CoimbraAraujo:2007zz], or in GR extensions [e.g., @Dey:2014gka]. Not always the corresponding papers have clearly stated that they were doing this inversion, and sometimes a simple and successful fit to the observational RC was claimed as an evidence that the proposal can model the internal dynamics of galaxies, even without dark matter. To simply mimic the form of some galaxy RCs is not sufficient (this is one of the general criticisms of @Salucci:2005ip).
Here we propose a new method to properly evaluate the application of these approaches when confronted to observational galaxy data. The method only relies on data that can be commonly found in publications on galaxy RCs. Namely, it depends on the observational RC data, the stellar density profile and the gaseous density profile.
The method consists of converting the observational RC into a dataset of an effective RC that should be fitted using standard Newtonian gravity procedures. The latter dataset is called the effective Newtonian RC. This conversion is model dependent. The method can be used to consider non-Newtonian gravity models with or without some dark matter.
The method was here applied to two approaches based on GR, the CT [@Cooperstock:2006dt] and the BG [@Balasin:2006cg] approaches. Merits and issues related to the theoretical basis of these approaches were briefly reviewed in sec. \[sec2\]. Our focus was here on testing their phenomenological results, in particular since some publications related to the CT approach [e.g., @Cooperstock:2006dt; @Carrick:2011ac] stated that this approach is capable of achieving good results on the RC fitting of diverse galaxies without dark matter (or with a small amount of it). The BG approach was here confronted with the astrophysical data of galaxies for the first time.
The application of the effective Newtonian RC method has shown that both the approaches have strong problems fitting galaxy RCs without dark matter (the selected sample favours the BG approach over the CT one). The method also indicates that if dark matter is considered, the BG approach cannot improve its results significantly, but the CT approach can.
For an example, we consider the case of the BG approach applied to the galaxy ESO 287-G13, Fig. \[resultados1\]. The BG RC could match nicely the observational RC, but is effective Newtonian RC at large radii is smaller than the contribution from the gas alone, hence this model and the data are not compatible (this can also be seen from its large $\chi^2_{{\mbox{\scriptsize red}}}$ value at Table \[tab:tabelaresultadoct\]). Moreover, by adding dark matter to the analyses, the problem is not alleviated, but it increases. By adding any dark matter halo whose most significative RC contribution is at large radii, the Newtonian curve $V_{{\mbox{\scriptsize N}}}$ (the dashed cyan curve in the plots) will become higher at large radii. This is the opposite to what happens for the same galaxy with the CT approach. For the latter, the $V_{{\mbox{\scriptsize N}}}$ curve is below the effective Newtonian RC at large radii.
Beyond the two GR approaches tested here, we expect that the evaluation of other models could benefit from the method that was here introduced.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Fabio Iocco for critically reading and commenting on a previous version of this paper; W.J.G. de Blok and G. Gentile for providing relevant data on galaxies that are used in this paper; Fred Cooperstock and Steven Tieu for clarifying an issue we had with their galaxy fitting procedure and Alberto Saa for relevant discussions on the subject. AOFA thanks CAPES (Brazil) for support. OFP and DCR thank CNPq (Brazil) and FAPES (Brazil) for partial support.
\[lastpage\]
[^1]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Lie algebroids are by no means natural as an infinitesimal counterpart of groupoids. In this paper we propose a functorial construction called *Nishimura algebroids* for an infinitesimal counterpart of groupoids. Nishimura algebroids, intended for differential geometry, are of the same vein as Lawvere’s functorial notion of *algebraic theory* and Ehresmann’s functorial notion of theory called *sketches*. We study *totally intransitive* Nishimura algebroids in detail. Finally we show that Nishimura algebroids naturally give rise to Lie algebroids.'
author:
- |
Hirokazu Nishimura\
Institute of Mathematics, University of Tsukuba\
Tsukuba, Ibaraki, 305-8571, Japan
title: A Functorial Approach to the Infinitesimal Theory of Groupoids
---
\[s0\]Introduction
==================
Many mathematicians innocently believe that *infinitesimalization* is no other than *linearization*. We contend that linearization is only a tiny portion of infinitesimalization. It is true that Lie algebras are the linearization of Lie groups, but it is by no means true that Lie algebras are the infinitesimalization of Lie groups. The fortunate success of the theory of Lie algebras together with their correspondence with Lie groups unfortunately enhanced their wrong conviction and blurred what are to be really the infinitesimalization of groups and, more generally, groupoids.
In this paper we propose, after the manners of Lawvere’s functorial construction of *algebraic theory* and Ehresmann’s functorial notion of theory called *sketches,* a functorial construction of *Nishimura algebroids* for the infinitesimalization of groupoids. After giving some preliminaries and fixing notation in the coming section, we will introduce our main notion of Nishimura algebroid in 6 steps. Then we will study totally intransitve Nishimura algebroids, in which the main result is that the linear part of any totally intransitve Nishimura algebroid is a Lie algebra bundle. As our final investigation we will show that Nishimura algebroids naturally give rise to Lie algebroids.
\[s1\]Preliminaries
===================
Synthetic Differential Geometry
-------------------------------
Our standard reference on synthetic differential geometry is Lavendhomme [@l1]. In synthetic differential geometry we generally work within a good topos. If the reader is willing to know how to get such a topos, he or she is referred to Kock [@k1] or Moerdijk and Reyes [@mr]. We denote by $\mathbb{R}$ the internal set of real numbers, which is endowed with a cornucopia of nilpotent infinitesimals persuant to the general Kock-Lawvere axiom. The internal category $\mathbf{Inf}$ of *infinitesimal spaces* comes contravariantly from the external category of Weil algebras over the set of real numbers by taking $\mathrm{Spec}_{\mathbb{R}}$. In particular, the infinitesimal space corresponding to the set of real numbers as a Weil algebra is denoted by $1$.We should note that every infinitesimal space $\mathcal{D}$ has a distinguished point, namely, $0_{\mathcal{D}}$ (often written simply $0$), and every morphism in $\mathbf{Inf}$ preserves distinguished points. An arbitrarily chosen microlinear space $M$ shall be fixed throughout the rest of this paper.
\[p1\]Groupoids
---------------
Our standard reference on groupoids is [@ma]. Let $\mathcal{D}$ be an object in $\mathbf{Inf}$. Given $m\in M$ and a groupoid $G$ over $M$ with its object inclusion map $\mathrm{id}:M\rightarrow G$ and its source and target projections $\alpha,\beta:G\rightarrow M$, we denote by $\mathcal{A}_{m}^{\mathcal{D}}G$ the totality of mappings $\gamma:\mathcal{D}\rightarrow
G$ with $\gamma(0_{\mathcal{D}})=\mathrm{id}_{m}$ and $(\alpha\circ
\gamma)(d)=m$ for any $d\in\mathcal{D}$. We denote by $\mathcal{A}^{\mathcal{D}}G$ the set-theoretic union of $\mathcal{A}_{m}^{\mathcal{D}}G$’s for all $m\in M$. The canonical projection $\pi:\mathcal{A}^{\mathcal{D}}G\rightarrow M$ is defined as is expected. The *anchor* $\mathbf{a}_{G}^{\mathcal{D}}:\mathcal{A}^{\mathcal{D}}G\rightarrow M^{\mathcal{D}}$ is defined to be simply $$\mathbf{a}_{G}^{\mathcal{D}}(\gamma)=\beta\circ\gamma$$ for any $\gamma\in\mathcal{A}^{\mathcal{D}}G$, where $M^{\mathcal{D}}$ is the space of mappings of $\mathcal{D}$ into $M$. We note that if the groupoid $G$ is the pair groupoid $M\times M$, then $\mathcal{A}^{\mathcal{D}}(M\times M)$ can canonically be identified with $M^{\mathcal{D}}$. We write $\mathbf{I}G$ for the inner subgroupoid of $G$, for which the reader is referred to p.14 of [@ma].
Simplicial Spaces
-----------------
The notion of *simplicial space* was discussed by Nishimura [@n11] and [@n13], where simplicial spaces were called *simplicial objects* in the former paper, while they were called *simplicial infinitesimal spaces* in the latter paper. *Simplicial spaces* are spaces of the form $$D^{m}\{\mathcal{S}\}=\{(d_{1},...,d_{m})\in D^{m}|d_{i_{1}}...d_{i_{k}}=0\text{ for any }(i_{1},...,i_{k})\in\mathcal{S}\}\text{,}$$ where $\mathcal{S}$ is a finite set of sequences $(i_{1},...,i_{k})$ of natural numbers with $1\leq i_{1}<...<i_{k}\leq m$. By way of example, we have $D(2)=D^{2}\{(1,2)\}$ and $D(3)=D^{3}\{(1,2),(1,3),(2,3)\}$. Given two simplicial spaces $D^{m}\{\mathcal{S}\}$ and $D^{n}\{\mathcal{T}\}$, we define another simplicial space $D^{m}\{\mathcal{S}\}\oplus D^{n}\{\mathcal{T}\}$ to be $$\begin{aligned}
& D^{m}\{\mathcal{S}\}\oplus D^{n}\{\mathcal{T}\}\\
& =\{(d_{1},...,d_{m},e_{1},...,e_{n})\in D^{m+n}|d_{i_{1}}...d_{i_{k}}=0\text{ for any }(i_{1},...,i_{k})\in\mathcal{S}\text{, }\\
e_{j_{1}}...e_{j_{l}} & =0\text{ for any }(j_{1},...,j_{l})\in
\mathcal{T}\text{, }d_{i}e_{j}=0\text{ for any }1\leq i\leq m\text{ and }1\leq
j\leq n\}\end{aligned}$$ We denote by $\mathbf{Simp}$ the full subcategory of $\mathbf{Inf}$ whose objects are all simplicial spaces. Obviously the category $\mathbf{Simp}$ is closed under direct products. The category $\mathbf{Simp}$ has finite coproducts. In particular, it has the initial object $\mathbf{1}$, which is also the terminal object.
Nishimura Algebroids
====================
Let $M$ be a microlinear space. We will introduce our main notion of *Nishimura algebroid* *over* $M$ step by step, so that the text is divided into six subsections.
*Nishimura Algebroids*$_{1}$
----------------------------
A *Nishimura algebroid*$_{1}$ *over* $M$ is simply a contravariant functor $\mathcal{A}$ from the category $\mathbf{Simp}$ of simplicial spaces to the category $\mathbf{MLS}_{M}$ of microlinear spaces over $M$ mapping finite coproducts in $\mathbf{Simp}$ to finite products in $\mathbf{MLS}_{M}$.
Given a simplicial space $\mathcal{D}$ in $\mathbf{Simp}$, we will usually write $\pi:\mathcal{A}^{\mathcal{D}}\rightarrow M$ for $\mathcal{A}(\mathcal{D})$. In particular, we will often write $\mathcal{A}^{n}$ in place of $\mathcal{A}^{D^{n}}$. We will simply write $\pi$ for the projection to $M$ in preference to such a more detailed notation as $\pi_{\mathcal{A},\mathcal{D}}$, which should not cause any possible confusion. Given $m\in M$, we write $\mathcal{A}_{m}^{\mathcal{D}}$ for $\{x\in\mathcal{A}^{\mathcal{D}}\mid\pi(x)=m\}$. Given a morphism $f:\mathcal{D\rightarrow D}^{\prime}$ in $\mathbf{Simp}$, we will usually write $\mathcal{A}^{f}:\mathcal{A}^{\mathcal{D}^{\prime}}\rightarrow\mathcal{A}^{\mathcal{D}}$ for $\mathcal{A}(f)$. Given $m\in M$, there is a unique element in $\mathcal{A}_{m}^{\mathbf{1}}$, which we denote by $\mathbf{0}_{m}^{\mathbf{1}}$. Given an object $\mathcal{D}$ in $\mathbf{Simp}$, we define $\mathbf{0}_{m}^{\mathcal{D}}\in\mathcal{A}_{m}^{\mathcal{D}}$ to be $$\mathbf{0}_{m}^{\mathcal{D}}=\mathcal{A}^{\mathcal{D\rightarrow}\mathbf{1}}(\mathbf{0}_{m}^{\mathbf{1}})$$
By assigning the spac $M^{\mathcal{D}}$ of mappings from $\mathcal{D}$ into $M$ to each object $\mathcal{D}$ in $\mathbf{Simp}$ and assigning $M^{f}:M^{\mathcal{D}^{\prime}}\rightarrow M^{\mathcal{D}}$ to each morphism $f:\mathcal{D\rightarrow D}^{\prime}$ in $\mathbf{Simp}$, we have a *Nishimura algebroid*$_{1}$ *over* $M$ to be called the standard *Nishimura algebroid*$_{1}$ *over* $M$ and to be denoted by $\mathcal{S}_{M}$ or more simply by $\mathcal{S}$.
Let $G$ be a groupoid over $M$. By assigning $\mathcal{A}^{\mathcal{D}}G$ to each object $\mathcal{D}$ in $\mathbf{Simp}$ and assigning $\mathcal{A}^{f}G:\mathcal{A}^{\mathcal{D}^{\prime}}G\rightarrow\mathcal{A}^{\mathcal{D}}G$ to each morphism $f:\mathcal{D\rightarrow D}^{\prime}$ in $\mathbf{Simp}$, we have a *Nishimura algebroid*$_{1}$ *over* $M$ to be denoted by $\mathcal{A}G$.
Each $\sigma\in\mathfrak{S}_{n}$ induces a morphism $\sigma:D^{n}\rightarrow
D^{n}$ in $\mathbf{Simp}$ such that $$\sigma(d_{1},...,d_{n})=(d_{\sigma(1)},...,d_{\sigma(n)})$$ for any $(d_{1},d_{2})\in D^{2}$. Given $x\in\mathcal{A}^{n}$, we will often write $^{\sigma}x$ for $\mathcal{A}^{\sigma}(x)$. It is easy to see that $$^{\tau\sigma}x=^{\tau}(^{\sigma}x)$$ for any $x\in\mathcal{A}^{n}$ and any $\sigma,\tau\in\mathfrak{S}_{n}$.
Given $x\in\mathcal{A}^{n}$ and $a\in\mathbb{R}$, we define $a\underset
{i}{\cdot}x$ ($1\leq i\leq n$) to be $$a\underset{i}{\cdot}x=\mathcal{A}^{((d_{1},,,,,d_{n})\in D^{n}\mapsto
(d_{1},...,d_{i-1},ad_{i},d_{i+1},...,d_{n})\in D^{n})}(x)$$
*Nishimura Algebroids*$_{2}$
----------------------------
A Nishimura algebroid$_{1}$ $A$ over $M$ is called a *Nishimura algebroid*$_{2}$ *over* $M$ if the application of $\mathcal{A}$ to any quasi-colimit diagram in $\mathbf{Simp}$ results in a limit diagram.
The notion of *Nishimura algebroid*$_{2}$ *over* $M$ can be regarded as a partial algebrization of microlinearity.
The standard Nishimura algebroid$_{1}$ $\mathcal{S}_{M}$ over $M$ is a Nishimura algebroid$_{2}$ over $M$. This follows simply from our assumption that $M$ is a microlinear space.
Let $G$ be a groupoid over $M$. Then the Nishimura algebroid$_{1}$ $\mathcal{A}G$ over $M$ is a Nishimura algebroid$_{2}$ over $M$. This follows simply from our assumption that $M$ and $G$ are microlinear spaces.
Let $\mathcal{A}$ be a Nishimura algebroid$_{2}$ over $M$. Let $m\in M$ with $x,y\in\mathcal{A}_{m}^{1}$. By using the quasi-colimit diagram (1) of small objects referred to in Proposition 6 (§2.2) of Lavendhomme [@l1], there exists a unique $z\in\mathcal{A}^{D\oplus D}$ with $\mathcal{A}^{(d\in
D\longmapsto(d,0)\in D\oplus D)}(z)=x$ and $\mathcal{A}^{(d\in D\longmapsto
(0,d)\in D\oplus D)}(z)=y$. We define $x+y$ to be $\mathcal{A}^{(d\in
D\longmapsto(d,d)\in D\oplus D)}(z)$. Given $a\in\mathbb{R}$, we define $ax$ to be $\mathcal{A}^{(d\in D\mapsto ad\in D)}(x)\in\mathcal{A}_{m}^{1}$. With these operations we have
\[tn2.1\]Given a Nishimura algebroid$_{2}$ $\mathcal{A}$ over $M$, $\mathcal{A}_{m}^{1}$ is an $\mathbb{R}$-module for any $m\in M$.
The proof is essentially a familiar proof that $\mathcal{S}_{m}^{1}$ is an $\mathbb{R}$-module, for which the reader is referred, e.g., to Lavendhomme [@l1], §3.1, Proposition 1. What we should do is only to reformulate the familiar proof genuinely in terms of diagrams. The details can safely be left to the reader.
Let $\mathcal{A}$ be a Nishimura algebroid$_{2}$ over $M$ with $m\in M$.
Let $x,y\in\mathcal{A}_{m}^{2}$ with $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(x)\nonumber\\
& =\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y)
\label{n2.3}$$ By using the quasi-colimit diagram of small objects at page 92 of Lavendhomme [@l1], we are sure that there exists a unique $z\in\mathcal{A}_{m}^{D^{2}\oplus D}$ with $$\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(z)=x \label{n2.1}$$ and $$\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(z)=y \label{n2.2}$$ We define $y\overset{\cdot}{-}x\in\mathcal{A}_{m}^{1}$ to be $\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}(z)$.
\[tn2.2\]Let $x,y\in\mathcal{A}^{2}$ abide by (\[n2.3\]). Then we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(y)\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(x)\\
& =y\overset{\cdot}{-}x\end{aligned}$$
Let $z\in\mathcal{A}_{m}^{D^{2}\oplus D}$ obedient to (\[n2.1\]) and (\[n2.2\]). Then we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{2},d_{1},d_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1},0)\in D^{2}\oplus
D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(x)\end{aligned}$$ while we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(d_{2},d_{1},d_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1},d_{1}d_{2})\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(y)\end{aligned}$$ Therefore we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(y)\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(x)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{2},d_{1},d_{3})\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}(z)\\
& =y\overset{\cdot}{-}x\end{aligned}$$ This completes the proof.
Let $x,y\in\mathcal{A}^{2}$ abide by (\[n2.3\]). Then we have $$x\overset{\cdot}{-}y=-(y\overset{\cdot}{-}x)$$
Let $z\in\mathcal{A}^{D^{2}\oplus D}$ abide by the conditions (\[n2.1\]) and (\[n2.2\]). Let $u\in\mathcal{A}^{D^{2}\oplus D}$ be $$u=\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3})\in D^{2}\oplus D)}(z)$$ Then we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(z)\\
& =y\end{aligned}$$ while we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(z)\\
& =x\end{aligned}$$ Therefore we have $$\begin{aligned}
& x\overset{\cdot}{-}y\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}(u)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,-d)\in D^{2}\oplus D)}(z)\\
& =-(y\overset{\cdot}{-}x)\end{aligned}$$ This completes the proof.
Let $x,y\in\mathcal{A}^{2}$ abide by (\[n2.3\]) with $a\in\mathbb{R}$. Then we have $$a\underset{i}{\cdot}y\overset{\cdot}{-}a\underset{i}{\cdot}x=a(y\overset
{\cdot}{-}x)\text{ \ \ \ \ \ \ (}i=1,2\text{)}$$
Here we deal only with the case $i=1$, leaving the other case to the reader. Let $z\in\mathcal{A}^{D^{2}\oplus D}$ abide by the conditions (\[n2.1\]) and (\[n2.2\]). Let $u\in\mathcal{A}^{D^{2}\oplus D}$ be $$u=\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(ad_{1},d_{2},ad_{3})\in D^{2}\oplus D)}(z)$$ Then we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(ad_{1},d_{2},ad_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(ad_{1},d_{2},0)\in D^{2}\oplus
D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(ad_{1},d_{2})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(z)\\
& =a\underset{1}{\cdot}x\end{aligned}$$ while we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(ad_{1},d_{2},ad_{3})\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(ad_{1},d_{2},ad_{1}d_{2})\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(ad_{1},d_{2})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(z)\\
& =a\underset{1}{\cdot}y\end{aligned}$$ Therefore we have $$\begin{aligned}
& a\underset{i}{\cdot}y\overset{\cdot}{-}a\underset{i}{\cdot}x\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}(u)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(ad_{1},d_{2},ad_{3})\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,ad)\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{(d\in D\mapsto ad\in D)}\circ\mathcal{A}^{(d\in
D\mapsto(0,0,d)\in D^{2}\oplus D)}(z)\\
& =a(y\overset{\cdot}{-}x)\end{aligned}$$ This completes the proof.
The following diagram is a quasi-colimit diagram: $$\begin{array}
[c]{cccccccc}
& & D^{2} & & \overset{\mathbf{i}}{\leftarrow} & D\oplus D & & \\
& \overset{\mathbf{i}}{\nearrow} & & \overset{\mathbf{\varphi}_{1}}{\searrow}
& & & \overset{\mathbf{i}}{\searrow} & \\
D\oplus D & & & & D^{2}\oplus D\oplus D & \overset{\mathbf{\varphi}_{3}}{\longleftarrow} & & D^{2}\\
& \underset{\mathbf{i}}{\searrow} & & \underset{\mathbf{\varphi}_{2}}{\nearrow} & & & \underset{\mathbf{i}}{\nearrow} & \\
& & D^{2} & & \underset{\mathbf{i}}{\leftarrow} & D\oplus D & &
\end{array}$$ where $\mathbf{i}:D\oplus D\rightarrow D^{2}$ is the canonical injection, and $\mathbf{\varphi}_{1},\mathbf{\varphi}_{2},\mathbf{\varphi}_{3}:D^{2}\rightarrow D^{2}\oplus D\oplus D$ are defined to be $$\begin{aligned}
\mathbf{\varphi}_{1}(d_{1},d_{2}) & =(d_{1},d_{2},0,0)\\
\mathbf{\varphi}_{2}(d_{1},d_{2}) & =(d_{1},d_{2},d_{1}d_{2},0)\\
\mathbf{\varphi}_{3}(d_{1},d_{2}) & =(d_{1},d_{2},0,d_{1}d_{2})\end{aligned}$$
Let $x,y,z\in\mathcal{A}^{2}$ with $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(z)\end{aligned}$$ Then we have $$(y\overset{\cdot}{-}x)+(z\overset{\cdot}{-}y)+(x\overset{\cdot}{-}z)=\mathbf{0}$$
Let $u\in\mathcal{A}^{D^{2}\oplus D\oplus D}$ be the unique one such that $$\begin{aligned}
x & =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0,0)\in
D^{2}\oplus D\oplus D)}(u)\\
y & =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2},0)\in D^{2}\oplus D\oplus D)}(u)\\
z & =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0,d_{1}d_{2})\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ The unique existence of such $u\in\mathcal{A}^{D^{2}\oplus D\oplus D}$ is guaranteed by the above lemma. Since we have $$\begin{aligned}
& x\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0,0)\in
D^{2}\oplus D\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{3},0)\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ and $$\begin{aligned}
& y\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2},0)\in
D^{2}\oplus D\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(d_{1},d_{2},d_{3},0)\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ we have $$\begin{aligned}
& y\overset{\cdot}{-}x\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{3},0)\in
D^{2}\oplus D\oplus D)}(u)\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d,0)\in D^{2}\oplus D\oplus D)}(u)
\label{n2.10}$$ Since we have $$\begin{aligned}
& y\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2},0)\in
D^{2}\oplus D\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3},d_{3})\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ and $$\begin{aligned}
& z\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0,d_{1}d_{2})\in
D^{2}\oplus D\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3},d_{3})\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ we have $$\begin{aligned}
& z\overset{\cdot}{-}y\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},d_{1}d_{2}-d_{3},d_{3})\in D^{2}\oplus D\oplus D)}(u)\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,-d,d)\in D^{2}\oplus D\oplus D)}(u)
\label{n2.11}$$ Since we have $$\begin{aligned}
& z\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0,d_{1}d_{2})\in
D^{2}\oplus D\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},0,d_{1}d_{2}-d_{3})\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ and $$\begin{aligned}
& x\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0,0)\in
D^{2}\oplus D\oplus D)}(u)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(d_{1},d_{2},0,d_{1}d_{2}-d_{3})\in D^{2}\oplus D\oplus D)}(u)\end{aligned}$$ we have $$\begin{aligned}
& x\overset{\cdot}{-}z\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2},0,d_{1}d_{2}-d_{3})\in D^{2}\oplus D\oplus D)}(u)\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,0,-d)\in D^{2}\oplus D\oplus D)}(u)
\label{n2.12}$$ Since we have $$\begin{aligned}
& y\overset{\cdot}{-}x\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d,0)\in D^{2}\oplus D\oplus D)}(u)\text{
\ \ [by (\ref{n2.10})]}\\
& =\mathcal{A}^{(d\in D\mapsto(d,0)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(0,0,d_{1}-d_{2},d_{2})\in D^{2}\oplus
D\oplus D)}(u)\end{aligned}$$ and $$\begin{aligned}
& z\overset{\cdot}{-}y\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,-d,d)\in D^{2}\oplus D\oplus D)}(u)\text{
\ \ \ [by (\ref{n2.11})]}\\
& =\mathcal{A}^{(d\in D\mapsto(0,d)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(0,0,d_{1}-d_{2},d_{2})\in D^{2}\oplus
D\oplus D)}(u)\end{aligned}$$ we have $$\begin{aligned}
& (y\overset{\cdot}{-}x)+(z\overset{\cdot}{-}y)\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(0,0,d_{1}-d_{2},d_{2})\in D^{2}\oplus
D\oplus D)}(u)\nonumber\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,0,d)\in D^{2}\oplus D\oplus D)}(u)
\label{n2.13}$$ Since we have $$\begin{aligned}
& (y\overset{\cdot}{-}x)+(z\overset{\cdot}{-}y)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,0,d)\in D^{2}\oplus D\oplus D)}(u)\text{
\ \ \ [by (\ref{n2.13})]}\\
& =\mathcal{A}^{(d\in D\mapsto(d,0)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(0,0,0,d_{1}-d_{2})\in D^{2}\oplus D\oplus
D)}(u)\end{aligned}$$ and $$\begin{aligned}
& x\overset{\cdot}{-}z\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,0,-d)\in D^{2}\oplus D\oplus D)}(u)\text{
\ \ \ \ [by (\ref{n2.12})]}\\
& =\mathcal{A}^{(d\in D\mapsto(0,d)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(0,0,0,d_{1}-d_{2})\in D^{2}\oplus D\oplus
D)}(u)\end{aligned}$$ we have $$\begin{aligned}
& \{(y\overset{\cdot}{-}x)+(z\overset{\cdot}{-}y)\}+(x\overset{\cdot}{-}z)\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(0,0,0,d_{1}-d_{2})\in D^{2}\oplus D\oplus
D)}(u)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,0,0)\in D^{2}\oplus D\oplus D)}(u)\\
& =\mathbf{0}$$ This completes the proof.
Let $x,y\in\mathcal{A}_{m}^{3}$ with $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2},d_{3})\in D\times(D\oplus D)\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(x)\nonumber\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D\times(D\oplus D)\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(y) \label{n2.4}$$ . By using the first quasi-colimit diagram of small objects in Lemma 2.1 of Nishimura [@n11], we are sure that there exists a unique $z\in
\mathcal{A}_{m}^{D^{4}\{(2,4),(3,4)\}}$ with $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{2},d_{3},0)\in
D^{4}\{(2,4),(3,4)\})}(z)=x$$ and $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{2},d_{3},d_{2}d_{3})\in D^{4}\{(2,4),(3,4)\})}(z)=y$$ We define $y\underset{1}{\overset{\cdot}{-}}x\in\mathcal{A}_{m}^{2}$ to be $\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},0,0,d_{2})\in
D^{4}\{(2,4),(3,4)\})}(z)$.
Let $x,y\in\mathcal{A}_{m}^{3}$ with $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\{(1,3)\}\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(x)\nonumber\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\{(1,3)\}\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(y) \label{n2.5}$$ By using the second quasi-colimit diagram of small objects in Lemma 2.1 of Nishimura [@n11], we are sure that there exists a unique $z\in
\mathcal{A}_{m}^{D^{4}\{(1,4),(3,4)\}}$ with $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{2},d_{3},0)\in
D^{4}\{(1,4),(3,4)\})}(z)=x$$ and $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{2},d_{3},d_{1}d_{3})\in D^{4}\{(1,4),(3,4)\})}(z)=y$$ We define $y\underset{2}{\overset{\cdot}{-}}x\in\mathcal{A}_{m}^{2}$ to be $\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(0,d_{1},0,d_{2})\in
D^{4}\{(1,4),(3,4)\})}(z)$.
Let $x,y\in\mathcal{A}_{m}^{3}$ with $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2},d_{3})\in(D\oplus D)\times D\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(x)\label{n2.6}\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in(D\oplus D)\times D\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(y)\nonumber\end{aligned}$$ By using the third quasi-colimit diagram of small objects in Lemma 2.1 of Nishimura [@n11], we are sure that there exists a unique $z\in
\mathcal{A}_{m}^{D^{4}\{(1,4),(2,4)\}}$ with $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{2},d_{3},0)\in
D^{4}\{(1,4),(2,4)\})}(z)=x$$ and $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{2},d_{3},d_{1}d_{2})\in D^{4}\{(1,4),(2,4)\})}(z)=y$$ We define $y\underset{3}{\overset{\cdot}{-}}x\in\mathcal{A}_{m}^{2}$ to be $\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(0,0,d_{1},d_{2})\in
D^{4}\{(1,4),(2,4)\})}(z)$.
\[tn2.3\]Let $x,y\in\mathcal{A}_{m}^{3}$.
1. If they satisfy (\[n2.4\]), then we have $$\begin{aligned}
& y\underset{1}{\overset{\cdot}{-}}x\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{2},d_{1},d_{3})\in
D^{3})}(y)\underset{2}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{2},d_{1},d_{3})\in D^{3})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{3},d_{2},d_{1})\in
D^{3})}(y)\underset{3}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{3},d_{2},d_{1})\in D^{3})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in
D^{3})}(y)\underset{1}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{1},d_{3},d_{2})\in D^{3})}(x)\end{aligned}$$
2. If they satisfy (\[n2.5\]), then we have $$\begin{aligned}
& y\underset{2}{\overset{\cdot}{-}}x\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{2},d_{1},d_{3})\in
D^{3})}(y)\underset{1}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{2},d_{1},d_{3})\in D^{3})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{3},d_{2},d_{1})\in
D^{3})}(y)\underset{2}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{3},d_{2},d_{1})\in D^{3})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in
D^{3})}(y)\underset{3}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{1},d_{3},d_{2})\in D^{3})}(x)\end{aligned}$$
3. If they satisfy (\[n2.6\]), then we have $$\begin{aligned}
& y\underset{3}{\overset{\cdot}{-}}x\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{2},d_{1},d_{3})\in
D^{3})}(y)\underset{3}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{2},d_{1},d_{3})\in D^{3})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{3},d_{2},d_{1})\in
D^{3})}(y)\underset{1}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{3},d_{2},d_{1})\in D^{3})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in
D^{3})}(y)\underset{2}{\overset{\cdot}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{1},d_{3},d_{2})\in D^{3})}(x)\end{aligned}$$
The proof is similar to that in Proposition \[tn2.2\]. The details can safely be left to the reader.
Now we have
\[tn2.4\]The four strong differences $\overset{\cdot}{-}$, $\underset
{1}{\overset{\cdot}{-}}$, $\underset{2}{\overset{\cdot}{-}}$ and $\underset
{3}{\overset{\cdot}{-}}$ satisfy the general Jacobi identity. I.e., given $x_{123},x_{132},x_{213},x_{231},x_{312},x_{321}\in\mathcal{A}^{3}$, as long as the following three expressions are well defined, they sum up only to vanish: $$\begin{aligned}
& (x_{123}\overset{\cdot}{\underset{1}{-}}x_{132})\overset{\cdot}{-}(x_{231}\overset{\cdot}{\underset{1}{-}}x_{321})\\
& (x_{231}\overset{\cdot}{\underset{2}{-}}x_{213})\overset{\cdot}{-}(x_{312}\overset{\cdot}{\underset{2}{-}}x_{132})\\
& (x_{312}\overset{\cdot}{\underset{3}{-}}x_{321})\overset{\cdot}{-}(x_{123}\overset{\cdot}{\underset{3}{-}}x_{213})\end{aligned}$$
The theorem was already proved in case of the standard Nishimura algebroid $\mathcal{S}_{M}$ in Nishimura’s [@n12], §3. What we should do is only to reformulate the above proof genuinely in terms of diagrams. The details can safely be left to the reader.
*Nishimura algebroids*$_{3}$
----------------------------
A Nishimura algebroid$_{2}$ $\mathcal{A}$ over $M$ is called a *Nishimura algebroid*$_{3}$ *over* $M$ providing that it is endowed with a natural transformation $\mathbf{a}$ from $\mathcal{A}$ to the standard Nishimura algebroid$_{2}$ $\mathcal{S}_{M}$ to be called the anchor natural transformation.
The standard Nishimura algebroid$_{2}$ $\mathcal{S}_{M}$ over $M$ is canonically a Nishimura algebroid$_{3}$ over $M$ endowed with the identity natural transformation of $\mathcal{S}_{M}$.
Let $G$ be a groupoid over $M$. Then the Nishimura algebroid$_{2}$ $\mathcal{A}G$ over $M$ is a Nishimura algebroid$_{3}$ over $M$ endowed with the anchor natural transformation assigning $\mathbf{a}_{G}^{\mathcal{D}}:\mathcal{A}^{\mathcal{D}}\rightarrow M^{\mathcal{D}}$ to each object $\mathcal{D}$ in $\mathbf{Simp}$.
*Nishimura Algebroids*$_{4}$
----------------------------
We denote by $\otimes_{\mathcal{A}}$, or more simply by $\otimes$, the contravariant functor which assigns $\mathcal{D}_{1}\otimes\mathcal{D}_{2}=\{(\zeta,x)\in(\mathcal{A}^{\mathcal{D}_{2}})^{\mathcal{D}_{1}}\times\mathcal{A}^{\mathcal{D}_{1}}\mid\mathbf{a}(x)=\pi^{\mathcal{D}_{1}}(\zeta)\}$ to each object $(\mathcal{D}_{1},\mathcal{D}_{2})$ in $\mathbf{Simp}\times\mathbf{Simp}$ and which assigns $f\otimes g=(\zeta
\in(\mathcal{A}^{\mathcal{D}_{2}})^{\mathcal{D}_{1}}\mapsto\mathcal{A}^{f}\circ\zeta\circ\mathcal{A}^{g}\in(\mathcal{A}^{\mathcal{D}_{2}^{\prime}})^{\mathcal{D}_{1}^{\prime}},\mathcal{A}^{g}):\mathcal{D}_{1}\otimes
\mathcal{D}_{2}\rightarrow\mathcal{D}_{1}^{\prime}\otimes\mathcal{D}_{2}^{\prime}$ to each morphism $(f,g):(\mathcal{D}_{1}^{\prime},\mathcal{D}_{2}^{\prime})\rightarrow(\mathcal{D}_{1},\mathcal{D}_{2})$ in $\mathbf{Simp}\times\mathbf{Simp}$, where $(\mathcal{A}^{\mathcal{D}_{2}})^{\mathcal{D}_{1}}$ denotes the space of mappings from the infinitesimal space $\mathcal{D}_{1}$ to $\mathcal{A}^{\mathcal{D}_{2}}$, and $\pi
^{\mathcal{D}_{1}}(\zeta)$ assigns $\pi(\zeta(d))$ to each $d\in
\mathcal{D}_{1}$. We denote by $\widetilde{\otimes}_{\mathcal{A}}$, or more simply by $\widetilde{\otimes\text{,}}$ the contravariant functor which assigns $\mathcal{D}_{1}\widetilde{\otimes}\mathcal{D}_{2}=\mathcal{A}^{\mathcal{D}_{1}\times\mathcal{D}_{2}}$ to each object $(\mathcal{D}_{1},\mathcal{D}_{2})$ in $\mathbf{Simp}\times\mathbf{Simp}$ and which assigns $f\widetilde{\otimes}g=\mathcal{A}^{f\times g}:\mathcal{A}^{\mathcal{D}_{1}\times\mathcal{D}_{2}}\rightarrow\mathcal{A}^{\mathcal{D}_{1}^{\prime
}\times\mathcal{D}_{2}^{\prime}}$ to each morphism $(f,g):(\mathcal{D}_{1}^{\prime},\mathcal{D}_{2}^{\prime})\rightarrow(\mathcal{D}_{1},\mathcal{D}_{2})$ in $\mathbf{Simp}\times\mathbf{Simp}$.
\[d2.4\]A Nishimura algebroid$_{3}$ $\mathcal{A}$ over $M$ is called a *Nishimura algebroid*$_{4}$ *over* $M$ providing that it is endowed with a natural isomorphism $\ast_{\mathcal{A}}$ (denoted more simply $\ast$ unless there is possible confusion) from the contravariant functor $\otimes$ to the contravariant functor $\widetilde{\otimes}$ abiding by the following conditions:
1. For any $(\zeta,x)\in\mathcal{D}_{1}\otimes\mathcal{D}_{2}$ with $(\mathcal{D}_{1},\mathcal{D}_{2})$ in $\mathbf{Simp}\times\mathbf{Simp}$, we have $$\pi(\zeta\ast x)=\pi(x)$$ and $$\mathbf{a}(\zeta\ast x)=\mathbf{a}^{\mathcal{D}_{1}}(\zeta)$$ where $\mathbf{a}^{\mathcal{D}_{1}}(\zeta)$ assigns $\mathbf{a}(\zeta
(d_{1}))(d_{2})$ to each $(d_{1},d_{2})\in\mathcal{D}_{1}\times\mathcal{D}_{2}$.
2. Let $\mathbf{i}_{j}:\mathcal{D}_{j}\rightarrow\mathcal{D}_{1}\times\mathcal{D}_{2}$ be the canonical injection with $\mathbf{p}_{j}:\mathcal{D}_{1}\times\mathcal{D}_{2}\rightarrow\mathcal{D}_{j}$ the canonical projection ($j=1,2$). Then we have $$\mathcal{A}^{\mathbf{i}_{1}}(\zeta\ast x)=x$$ and $$\mathcal{A}^{\mathbf{i}_{2}}(\zeta\ast x)=\zeta(0_{\mathcal{D}_{2}})$$ for any $(\zeta,x)\in\mathcal{D}_{1}\otimes\mathcal{D}_{2}$, while we have $$\mathcal{A}^{\mathbf{p}_{1}}(y)=(d\in\mathcal{D}_{1}\mapsto\mathbf{0}_{(\mathbf{a}y)(d)}^{\mathcal{D}_{2}})\ast y$$ for any $y\in\mathcal{A}^{\mathcal{D}_{1}}$ and $$\mathcal{A}^{\mathbf{p}_{2}}(z)=(d\in\mathcal{D}_{1}\mapsto z)\ast
\mathbf{0}_{\pi(z)}^{\mathcal{D}_{1}}$$ for any $z\in\mathcal{A}^{\mathcal{D}_{2}}$.
3. Let $f\in\mathbb{R}^{\mathcal{D}}$. For any $(\zeta,x)\in(\mathcal{D}_{1}\times...\times\mathcal{D}_{n})\otimes\mathcal{D}$, we have $$\begin{aligned}
& \mathcal{A}^{((d,d_{1},...,d_{n})\in\mathcal{D\times D}_{1}\times
...\times\mathcal{D}_{n}\mapsto(d,d_{1},...,d_{i-1},f(d)d_{i},d_{i+1},...,d_{n})\in\mathcal{D\times D}_{1}\times...\times\mathcal{D}_{n})}(\zeta\ast x)\\
& =\{d\in\mathcal{D}\mapsto\mathcal{A}^{((d_{1},...,d_{n})\in\mathcal{D}_{1}\times...\times\mathcal{D}_{n}\mapsto(d_{1},...,d_{i-1},f(d)d_{i},d_{i+1},...,d_{n})\in\mathcal{D}_{1}\times...\times\mathcal{D}_{n})}\zeta(d)\in\mathcal{A}^{\mathcal{D}_{1}\times...\times\mathcal{D}_{n}}\}\ast
x\text{ \ \ \ }\\
\text{(}1 & \leq i\leq n\text{)}$$
4. For any $x\in\mathcal{A}^{\mathcal{D}_{1}}$, any $\zeta_{1}\in(\mathcal{A}^{\mathcal{D}_{2}})^{\mathcal{D}_{1}}$ and any $\zeta_{2}\in(\mathcal{A}^{\mathcal{D}_{3}})^{\mathcal{D}_{1}\times\mathcal{D}_{2}}$ with $\mathbf{a}(x)=\pi^{\mathcal{D}_{1}}(\zeta_{1})$ and $\mathbf{a}^{\mathcal{D}_{1}}(\zeta_{1})=\pi^{\mathcal{D}_{1}\times\mathcal{D}_{2}}(\zeta_{2})$, we have $$\zeta_{2}\ast(\zeta_{1}\ast x)=(\zeta_{2}\ast^{D}\zeta_{1})\ast x$$ where $\zeta_{2}\ast^{\mathcal{D}_{1}}\zeta_{1}\in(\mathcal{A}^{\mathcal{D}_{1}\times\mathcal{D}_{2}})^{\mathcal{D}_{1}}$ is defined to be $$(\zeta_{2}\ast^{\mathcal{D}_{1}}\zeta_{1})(d)=\zeta_{2}(d,\cdot)\ast\zeta
_{1}(d)$$ for any $d\in\mathcal{D}_{1}$.
\[t2.4.1\]What we require in our definition of Nishimura algebroid$_{4}$ over $M$ is that while multiplication seen in groupoids is no longer in view in Nishimura algebroids, the remnants of multiplication and its associativity are to be still in view. Multiplication seems completely lost in the traditional definition of Lie algebroid.
The standard Nishimura algebroid$_{3}$ $\mathcal{S}_{M}$ over $M$ is canonically a Nishimura algebroid$_{4}$ over $M$ provided that $\zeta
\ast_{\mathcal{S}_{M}}x\in\mathcal{S}_{M}^{\mathcal{D}_{1}\times
\mathcal{D}_{2}}$ is defined to be $$(d_{1},d_{2})\in\mathcal{D}_{1}\times\mathcal{D}_{2}\mapsto\zeta(d_{1})(d_{2})\in M$$
Let $G$ be a groupoid over $M$. The Nishimura algebroid$_{3}$ $\mathcal{A}G$ over $M$ is a Nishimura algebroid$_{4}$ over $M$ provided that $\zeta
\ast_{\mathcal{A}G}x\in(\mathcal{A}G)^{\mathcal{D}_{1}\times\mathcal{D}_{2}}$ is defined to be $$(d_{1},d_{2})\in\mathcal{D}_{1}\times\mathcal{D}_{2}\mapsto\zeta(d_{1})(d_{2})x(d_{1})\in G$$
Now we give some results holding for any Nishimura algebroid$_{4}$ $\mathcal{A}$ over $M$.
\[t2.4.2\]There is a bijective correspondence between the mappings $\Phi:D\rightarrow\mathcal{A}_{m}^{1}$ and the elements $x\in\mathcal{A}_{m}^{2} $ with $\mathcal{A}^{(d\in D\mapsto(d,0)\in D^{2})}(x)=\mathbf{0}_{m}^{D}$.
This follows simply from the first condition in the definition of Nishimura algebroid$_{4}$ over $M$, which claims that the assignment of $\Phi
\ast\mathbf{0}_{m}^{D}\in\mathcal{A}_{m}^{2}$ to each mapping $\Phi
:D\rightarrow\mathcal{A}_{m}^{1}$ gives such a bijective correspondence.
It is easy to see that
\[l2.1\]Let $\mathbf{p}_{1}:\mathcal{D}_{1}\times\mathcal{D}_{2}\rightarrow\mathcal{D}_{1}$ be the canonical projection as in the second condition of Definition \[d2.4\]. Then we have $$\mathcal{A}^{\mathbf{p}_{1}}(\mathbf{0}_{m}^{\mathcal{D}_{1}})=\mathbf{0}_{m}^{\mathcal{D}_{1}\times\mathcal{D}_{2}}$$
As an easy consequence of the above proposition, we have
\[t2.4.3\]Given a Nishimura algebroid$_{4}$ $\mathcal{A}$ over $M$ with $m\in M$, the $\mathbb{R}$-module $\mathcal{A}_{m}^{1}$ is Euclidean.
We have already proved that $\mathcal{A}_{m}^{1}$ is naturally an $\mathbb{R}
$-module. Let $\varphi:D\rightarrow\mathcal{A}_{m}^{1}$ be a mapping. We will consider another mapping $\Phi:D\rightarrow\mathcal{A}_{m}^{1}$ defined to be $$\Phi(d)=\varphi(d)-\varphi(0)$$ for any $d\in D$. Let us consider $x=\Phi\ast\mathbf{0}_{m}^{D}\in
\mathcal{A}_{m}^{2}$. We have $\mathcal{A}^{(d\in D\mapsto(d,0))}(x)=\mathbf{0}_{m}^{D}$, while it is easy to see that $\mathcal{A}^{(d\in
D\mapsto(0,d))}(x)=\Phi(0)=\mathbf{0}_{m}^{D}$. Therefore there is a unique $y\in\mathcal{A}_{m}^{1}$ with $\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto
d_{1}d_{2}\in D)}(y)=x$. Let us consider $\mathcal{A}^{((d_{1},d_{2})\in
D^{2}\mapsto d_{2}\in D)}(y)=y\ast\mathbf{0}_{m}^{D}\in\mathcal{A}_{m}^{2}$. Then it is easy to see that $$\begin{aligned}
(d & \in D\longmapsto dy)\ast\mathbf{0}_{m}^{D}\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{1}d_{2})\in
D^{2})}(\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto d_{2}\in D)}(y))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto d_{1}d_{2}\in D)}(y)\\
& =x\end{aligned}$$ Therefore we have $\Phi\ast\mathbf{0}_{m}^{D}=(d\in D\longmapsto
dy)\ast\mathbf{0}_{m}^{D}$, which implies that $$\varphi(d)-\varphi(0)=dy$$ for any $d\in D$. To see the uniqueness of such $y\in\mathcal{A}_{m}^{1}$, let us suppose that some $z\in\mathcal{A}_{m}^{1}$ satisfies $$dz=\mathbf{0}_{m}^{D}$$ for any $d\in D$. Since $z\ast\mathbf{0}_{m}^{D}=\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto d_{2}\in D)}(z)$, we have $$\begin{aligned}
& (d\in D\rightarrow\mathbf{0}_{m}^{D})\ast\mathbf{0}_{m}^{D}\\
& =(d\in D\longmapsto dz)\ast\mathbf{0}_{m}^{D}\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{1}d_{2})\in
D^{2})}(\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto d_{2}\in D)}(z))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto d_{1}d_{2}\in D)}(z)\end{aligned}$$ Since $(d\in D\rightarrow\mathbf{0}_{m}^{D})\ast\mathbf{0}_{m}^{D}=\mathbf{0}_{m}^{D^{2}}$ by Lemma \[l2.1\] and the second condition of Definition \[d2.4\], the desired uniqueness follows from Proposition 1 (§2.2) of Lavendhomme [@l1].
Now we will discuss the relationship between $\ast$ and strong differences.
\[t2.6\]
1. For any $\zeta_{1},\zeta_{2}\in(\mathcal{A}^{2})^{D}$ and any $x\in\mathcal{A}^{1}$ with $$\mathbf{a}(x)=\pi^{D}(\zeta_{1})=\pi^{D}(\zeta_{2})$$ and $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(\zeta_{1}(d))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(\zeta_{2}(d))\end{aligned}$$ for any $d\in D$, we have $$(\zeta_{2}\overset{\cdot}{-}\zeta_{1})\ast x=\zeta_{2}\ast x\underset
{1}{\overset{\cdot}{-}}\zeta_{1}\ast x$$ where $\zeta_{2}\overset{\cdot}{-}\zeta_{1}\in(\mathcal{A}^{1})^{D}$ is defined to be $$(\zeta_{2}\overset{\cdot}{-}\zeta_{1})(d)=\zeta_{2}(d)\overset{\cdot}{-}\zeta_{1}(d)$$ for any $d\in D$.
2. For any $x,y\in\mathcal{A}^{2}$ and any $\zeta\in(\mathcal{A}^{1})^{D^{2}\oplus D}$ with $$\mathbf{a}(x)=(d_{1},d_{2})\in D^{2}\mapsto\pi(\zeta(d_{1},d_{2},0))$$ $$\mathbf{a}(y)=(d_{1},d_{2})\in D^{2}\mapsto\pi(x(d_{1},d_{2},d_{1}d_{2}))$$ and $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(x)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y)\end{aligned}$$ we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}\\
(\{\zeta\circ(d & \in D\mapsto(0,0,d)\in D^{2}\oplus D)\}\ast(y\overset
{\cdot}{-}x))\\
& =\{\zeta\circ((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)\}\ast y\underset{3}{\overset{\cdot}{-}}\\
\{\zeta\circ((d_{1},d_{2}) & \in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)\}\ast x\end{aligned}$$
It suffices to note that given an object $\mathcal{D}$ in $\mathbf{Simp}$, the contravariant functor $\widetilde{\otimes}\mathcal{D}$ (resp. $\mathcal{D}\widetilde{\otimes}$) and therefore the functor $\otimes\mathcal{D}$ (resp. $\mathcal{D}\otimes$) map every quasi-colimit diagram of small objects in $\mathbf{Simp}$ to a limit diagram. Therefore the proof is merely a reformulation of Proposition 2.6 of Nishimura [@n11]. The details can safely be left to the reader.
*Nishimura Algebroids*$_{5}$
----------------------------
A Nishimura algebroid$_{4}$ $\mathcal{A}$ over $M$ is called a *Nishimura algebroid*$_{5}$ *over* $M$ providing that the anchor natural transformation $\mathbf{a}$ from $\mathcal{A}$ to the standard Nishimura algebroid$_{4}$ $\mathcal{S}_{M}$ is a homomorphism of Nishimura algebroids$_{4}$ over $M$. In other words, a Nishimura algebroid$_{4}$ $\mathcal{A}$ over $M$ is a *Nishimura algebroid*$_{5}$ *over* $M$ providing that for any $(\zeta,x)\in\mathcal{D}_{1}\otimes_{\mathcal{A}}\mathcal{D}_{2}$ with $(\mathcal{D}_{1},\mathcal{D}_{2})$ in $\mathbf{Simp}\times\mathbf{Simp}$, we have $$\mathbf{a}(\zeta\ast_{\mathcal{A}}x)=\mathbf{a}^{\mathcal{D}_{1}}(\zeta
)\ast_{\mathcal{S}_{M}}\mathbf{a}(x)$$
It is trivial to see that the standard Nishimura algebroid$_{4}$ $\mathcal{S}_{M}$ over $M$ is a *Nishimura algebroid*$_{5}$ *over* $M$, since $\mathbf{a}$ is the identity transformation.
Let $G$ be a groupoid over $M$. It is easy to see that the Nishimura algebroid$_{4}$ $\mathcal{A}G$ over $M$ is a Nishimura algebroid$_{5}$ over $M$. It is also easy to see that a homomorphism $\varphi:G\rightarrow
G^{\prime}$ of groupoids over $M$ naturally gives rise to a homomorphism $\mathcal{A}\varphi:\mathcal{A}G\rightarrow\mathcal{A}G^{\prime}$of Nishimura algebroids$_{5}$ over $M$. Thus we obtain a functor $\mathcal{A}$ from the category of groupoids over $M$ to the category of Nishimura algebroids$_{5}$ over $M$.
The following proposition should be obvious.
Let $\varphi:\mathcal{A}\rightarrow\mathcal{A}^{\prime}$ be a homomorphism of Nishimura algebroids$_{5}$ over $M$. Then its kernel at each $m\in M$, denoted by $\mathrm{\ker}_{m}\varphi$, assigning $(\mathrm{\ker}_{m}\varphi
)^{\mathcal{D}}=\{x\in\mathcal{A}^{\mathcal{D}}\mid\varphi(x)=0_{m}^{\mathcal{D}}\}$ to each object $\mathcal{D}$ in $\mathbf{Simp}$ and assigning the restriction $(\mathrm{\ker}_{m}\varphi)^{f}:(\mathrm{\ker}_{m}\varphi)^{\mathcal{D}^{\prime}}\rightarrow(\mathrm{\ker}_{m}\varphi)^{\mathcal{D}}$ of $\mathcal{A}^{f}:\mathcal{A}^{\mathcal{D}^{\prime}}\rightarrow\mathcal{A}^{\mathcal{D}}$ to each morphism $f:\mathcal{D\rightarrow D}^{\prime}$ in $\mathbf{Simp}$ is naturally a Nishimura algebroid$_{5}$ over a single point.
*Nishimura Algebroids*$_{6}$
----------------------------
Let $\mathcal{A}$ be a Nishimura algebroid$_{5}$ over $M$. Since the anchor natural transformation $\mathbf{a}_{\mathcal{A}}:\mathcal{A}\rightarrow
\mathcal{S}_{M}$ is really a homomorphism of Nishimura algebroids$_{5}$ over $M$, its kernel $\mathrm{\ker}_{m}\mathbf{a}_{\mathcal{A}}$ at each $m\in M$ is a Nishimura algebroid$_{5}$ over a single point by dint of the last proposition of the previous subsection. By collecting $\mathrm{\ker}_{m}\mathbf{a}_{\mathcal{A}}$ over all $m\in M$, we obtain a bundle of Nishimura algebroids$_{5}$ over a single point, which is called the *inner subalgebroid* of $\mathcal{A}$ and which is denoted by $\mathbf{I}\mathcal{A} $. The reader should note that the inner subalgebroid $\mathbf{I}\mathcal{A}$ of $\mathcal{A}$ can naturally be reckoned as a Nishimura algebroid$_{5}$ over $M$ (as a subalgebroid of $\mathcal{A}$ in a natural sense). In the next definition we will consider the frame groupoid of Nishimura algebroids$_{5}$ over a single point for $\mathbf{I}\mathcal{A}$, which is denoted by $\Phi_{Nishi_{5}}(\mathbf{I}\mathcal{A})$.
A Nishimura algebroid$_{5}$ $\mathcal{A}$ over $M$ is called a *Nishimura algebroid*$_{6}$ *over* $M$ providing that it is endowed with a homomorphism $\mathrm{ad}_{\mathcal{A}}$ (usually written simply $\mathrm{ad}$) of Nishimura algebroids$_{5}$ over $M$ from $\mathcal{A}$ to $\mathcal{A}(\Phi_{Nishi_{5}}(\mathbf{I}\mathcal{A}))$ abiding by the following condition:
1. We have $$\mathrm{ad}(x)(d_{1})\circ\mathrm{ad}(y)(d_{2})=(\mathrm{ad}((\mathrm{ad}(x)(d_{1}))(y)))(d_{2})\circ\mathrm{ad}(x)(d_{1})$$ for any objects $\mathcal{D}_{1},\mathcal{D}_{2}$ in $\mathbf{Simp}$, any $d_{1}\in\mathcal{D}_{1}$, any $d_{2}\in\mathcal{D}_{2}$, any $x\in
\mathcal{A}^{\mathcal{D}_{1}}$ and any $y\in(\mathbf{I}\mathcal{A})^{\mathcal{D}_{2}}$ with $\pi(x)=\pi(y)$.
2. Given $x,y\in(\mathbf{I}\mathcal{A)}^{1}$ with $\pi(x)=\pi(y)$, we have $$(\mathrm{ad}(x))(d)(y)-y=d[x,y]$$ for any $d\in D$.
Since the inner subalgebroid $\mathbf{I}\mathcal{S}_{M}$ of the standard Nishimura algebroid$_{5}$ $\mathcal{S}_{M}$ is trivial, $\mathcal{S}_{M}$ is trivially a *Nishimura algebroid*$_{6}$ *over* $M$.
Let $G$ be a groupoid over $M$. By assigning a mapping $$y\in(\mathbf{I}G)_{\alpha x}\mapsto xyx^{-1}\in(\mathbf{I}G)_{\beta x}$$ to each $x\in G$, we get a homomorphism of groupoids over $M$ from $G$ to $\Phi_{grp}(\mathbf{I}G)$, which naturally gives rise to a homomorphism of groupoids over $M$ from $G$ to $\Phi_{Nishi_{5}}(\mathcal{A}(\mathbf{I}G))$. Since $\mathcal{A}(\mathbf{I}G)$ and $\mathbf{I}(\mathcal{A}G)$ can naturally be identified, we have a homomorphism of groupoids over $M$ from $G $ to $\Phi_{Nishi_{5}}(\mathbf{I}(\mathcal{A}G))$, to which we apply the functor $\mathcal{A}$ so as to get the desired $\mathrm{ad}_{\mathcal{A}G}$ as a homomorphism of Nishimura algebroids$_{5}$ over $M$ from $\mathcal{A}G$ to $\mathcal{A}(\Phi_{Nishi_{5}}(\mathbf{I(}\mathcal{A}G\mathcal{)}))$.
Totally Intransitive Nishimura Algebroids
=========================================
A Nishimura algebroid $\mathcal{A}$ over $M$ is said to be totally intransitive providing that its anchor natural transformation $\mathbf{a}_{\mathcal{A}}$ is trivial, i.e., $$\mathbf{a}_{\mathcal{A}}(x)=\mathbf{0}_{m}^{\mathcal{D}}$$ for any $m\in M$, any object $\mathcal{D}$ in $\mathbf{Simp}$ and any $x\in\mathcal{A}_{m}^{\mathcal{D}}$.
A totally intransitive Nishimura algebroid $\mathcal{A}$ over $M$ can naturally be regarded as a bundle of Nishimura algebroids over a single point over $M$.
In this section an arbitrarily chosen totally intransitive Nishimura algebroid $\mathcal{A}$ over $M$ shall be fixed.
Given $x\in\mathcal{A}^{\mathcal{D}_{1}}$ and $y\in\mathcal{A}^{\mathcal{D}_{2}}$ with $\pi(x)=\pi(y)$, we define $x\circledast y\in\mathcal{A}^{\mathcal{D}_{1}\times\mathcal{D}_{2}}$ to be $$(d\in\mathcal{D}_{2}\mapsto x)\ast y$$
For any $x\in\mathcal{A}^{\mathcal{D}_{1}}$, $y\in\mathcal{A}^{\mathcal{D}_{2}}$ and $z\in\mathcal{A}^{\mathcal{D}_{3}}$ with $\pi(x)=\pi(y)=\pi(z) $, we have $$x\circledast(y\circledast z)=(x\circledast y)\circledast z$$
This follows simply from the fourth condition in Definition \[d2.4\].
By this proposition we can omit parentheses in a combination by $\circledast$.
The following proposition is the Nishimura algebroid counterpart of Proposition 3 (§3.2) of Lavendhomme [@l1].
\[tr1\]Let $x\in\mathcal{A}^{1}$. Then we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D(2)\longmapsto d_{1}+d_{2}\in D)}(x)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D(2)\longmapsto(d_{1},d_{2})\in D^{2})}(x\circledast x)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D(2)\longmapsto(d_{2},d_{1})\in D^{2})}(x\circledast x)\end{aligned}$$
Let $z=\mathcal{A}^{((d_{1},d_{2})\in D(2)\longmapsto(d_{1},d_{2})\in D^{2})}(x\circledast x)$. Then we have $$\begin{aligned}
& \mathcal{A}^{(d\in D\longmapsto(d,0)\in D(2))}(z)\\
& =\mathcal{A}^{(d\in D\longmapsto(d,0)\in D^{2})}(x\circledast x)\\
& =x\end{aligned}$$ and $$\begin{aligned}
& \mathcal{A}^{(d\in D\longmapsto(0,d)\in D(2))}(z)\\
& =\mathcal{A}^{(d\in D\longmapsto(0,d)\in D^{2})}(x\circledast x)\\
& =(d\in D\longmapsto x)(0)\\
& =x\end{aligned}$$ Therefore the desired first equality follows at once from the quasi-colimit diagram in Proposition 6 (§2.2) of Lavendhomme [@l1]. The desired second equality can be dealt with similarly.
The following proposition is the Nishimura algebroid counterpart of Proposition 6 (§3.2) of Lavendhomme [@l1].
\[tr2\]Let $x,y\in\mathcal{A}^{1}$ with $\pi(x)=\pi(y)$. Then we have $$\begin{aligned}
& x+y\\
& =\mathcal{A}^{(d\in D\longmapsto(d,d)\in D^{2})}(y\circledast x)\\
& =\mathcal{A}^{(d\in D\longmapsto(d,d)\in D^{2})}(x\circledast y)\end{aligned}$$
Let $z=\mathcal{A}^{(d_{1},d_{2})\in D(2)\longmapsto(d_{1},d_{2})\in D^{2})}(y\circledast x)$. Then we have $$\begin{aligned}
& \mathcal{A}^{(d\in D\longmapsto(d,0)\in D(2))}(z)\\
& =\mathcal{A}^{(d\in D\longmapsto(d,0)\in D^{2})}(y\circledast x)\\
& =x\end{aligned}$$ and $$\begin{aligned}
& \mathcal{A}^{(d\in D\longmapsto(0,d)\in D(2))}(z)\\
& =\mathcal{A}^{(d\in D\longmapsto(0,d)\in D^{2})}(y\circledast x)\\
& =y\end{aligned}$$ Therefore it follows from the quasi-colimit diagram in Proposition 6 (§ 2.2) of Lavendhomme [@l1] that $$\begin{aligned}
& x+y\\
& =\mathcal{A}^{(d\in D\longmapsto(d,d)\in D(2))}(z)\\
& =\mathcal{A}^{(d\in D\longmapsto(d,d)\in D^{2})}(y\circledast x)\end{aligned}$$ which establishes the first desired equality. The second desired equality follows similarly.
\[tr3\]Given $x,y\in\mathcal{A}^{1}$ with $\pi(x)=\pi(y)$, there exists a unique $z\in\mathcal{A}^{1}$ with $\pi(x)=\pi(y)=\pi(z)$ such that $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{1}d_{2}\in D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast x\circledast y\circledast x)\end{aligned}$$
We will show that $$\begin{aligned}
& \mathcal{A}^{(d\in D\mapsto(d,0)\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast
x\circledast y\circledast x)\\
& =\mathbf{0}_{\pi(x)}^{D}$$ and $$\begin{aligned}
& \mathcal{A}^{(d\in D\mapsto(0,d)\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast
x\circledast y\circledast x)\\
& =\mathbf{0}_{\pi(y)}^{D}$$ Then the desired result will follow from the quasi-colimit diagram in Proposition 7 (§2.2) of Lavendhomme [@l1]. Now we deal with the first desired identity. Since the composition of $d\in D\mapsto(d,0)\in D^{2}$ and $(d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4}$ is equal to the composition of $d\in D\mapsto(d,d)\in D^{2}$ and $(d_{1},d_{2})\in D^{2}\longmapsto(d_{1},0,-d_{2},0)\in D^{4}$, we have $$\begin{aligned}
& \mathcal{A}^{(d\in D\mapsto(d,0)\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast
x\circledast y\circledast x)\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},0,-d_{2},0)\in D^{4})}(y\circledast
x\circledast y\circledast x)\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}(\mathcal{A}^{(d_{2}\in
D\longmapsto(-d_{2},0)\in D^{2})}(y\circledast x)\circledast\mathcal{A}^{(d_{1}\in D\longmapsto(d_{1},0)\in D^{2})}(y\circledast x))\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}((-x)\circledast x)\\
& =x-x\text{ \ \ \ \ \ [by Proposition \ref{tr2}]}\\
& =\mathbf{0}_{\pi(x)}^{D}$$ Now we turn to the second desired identity. Since the composition of $d\in
D\mapsto(0,d)\in D^{2}$ and $(d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4}$ is equal to the composition of $d\in
D\mapsto(d,d)\in D^{2}$ and $(d_{1},d_{2})\in D^{2}\longmapsto(0,d_{1},0,-d_{2})\in D^{4}$, we have $$\begin{aligned}
& \mathcal{A}^{(d\in D\mapsto(0,d)\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast
x\circledast y\circledast x)\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(0,d_{1},0,-d_{2})\in D^{4})}(y\circledast
x\circledast y\circledast x)\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}(\mathcal{A}^{(d_{2}\in
D\longmapsto(0,-d_{2})\in D^{2})}(y\circledast x)\circledast\mathcal{A}^{(d_{1}\in D\longmapsto(0,d_{1})\in D^{2})}(y\circledast x))\\
& =\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}((-y)\circledast y)\\
& =y-y\text{ \ \ \ \ \ [by Proposition \ref{tr2}]}\\
& =\mathbf{0}_{\pi(y)}^{D}$$ The proof is now complete.
We will denote the above $z$ by $[x,y]$.
\[tr5\]Given $x,y\in\mathcal{A}^{1}$ with $\pi(x)=\pi(y)$, we have $$\lbrack y,x]=-[x,y]$$
Let $m=\pi(x)=\pi(y)$. We have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{1}d_{2}\in D)}\circ\mathcal{A}^{(d\in D\mapsto(d,d)\in D^{2})}([x,y]\circledast\lbrack
y,x])\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1}d_{2},d_{1}d_{2})\in
D^{2})}([x,y]\circledast\lbrack y,x])\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1},d_{2})\in D^{4})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4})\in D^{4}\mapsto(d_{1}d_{2},d_{3}d_{4})\in D^{2})}([x,y]\circledast\lbrack y,x])\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1},d_{2})\in D^{4})}(\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{1}d_{2}\in D)}([x,y])\circledast\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto
d_{1}d_{2}\in D)}([y,x]))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1},d_{2})\in D^{4})}(\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{1}d_{2}\in D)}([x,y])\circledast\\
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{1}d_{2}\in
D)}([y,x]))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1},d_{2})\in D^{4})}(\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast x\circledast y\circledast
x)\circledast\\
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(x\circledast y\circledast x\circledast y))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1},d_{2})\in D^{4})}(\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(y\circledast x\circledast y\circledast
x)\circledast\\
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2},-d_{1})\in D^{4})}(x\circledast y\circledast x\circledast y))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1},d_{2})\in D^{4})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4})\in D^{4}\mapsto(d_{2},d_{1},-d_{2},-d_{1},d_{3},d_{4},-d_{3},-d_{4})\in D^{8})}\\
& (y\circledast x\circledast y\circledast x\circledast x\circledast
y\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2},-d_{1},d_{1},d_{2},-d_{1},-d_{2})\in D^{8})}(y\circledast x\circledast
y\circledast x\circledast x\circledast y\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2},d_{1},d_{2},-d_{1},-d_{2})\in D^{7})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4},d_{5},d_{6},d_{7})\in D^{7}\mapsto(d_{1},d_{2},d_{3},-d_{4},d_{4},d_{5},d_{6},d_{7})\in D^{8})}\\
& (y\circledast x\circledast y\circledast x\circledast x\circledast
y\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2},d_{1},d_{2},-d_{1},-d_{2})\in D^{7})}\\
& (y\circledast x\circledast y\circledast\mathcal{A}^{(d\in D\mapsto(-d,d)\in
D^{2})}(x\circledast x)\circledast y\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2},d_{2},-d_{1},-d_{2})\in D^{6})}(y\circledast x\circledast y\circledast
y\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},d_{2},-d_{1},-d_{2})\in D^{5})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4},d_{5})\in D^{5}\longmapsto(d_{1},d_{2},-d_{3},d_{3},d_{4},d_{5})\in D^{6})}\\
& (y\circledast x\circledast y\circledast y\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},d_{2},-d_{1},-d_{2})\in D^{5})}(y\circledast x\circledast\mathcal{A}^{(d\in
D\mapsto(-d,d)\in D^{2})}(y\circledast y)\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{1},-d_{2})\in D^{4})}(y\circledast x\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2})\in
D^{3})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\longmapsto(d_{1},d_{2},-d_{2},d_{3})\in D^{4})}(y\circledast x\circledast x\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},-d_{2})\in
D^{3})}(y\circledast\mathcal{A}^{(d\in D\mapsto(d,-d)\in D^{2})}(x\circledast
x)\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{2})\in D^{2})}(y\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{2}\in D)}\circ\mathcal{A}^{(d\in D\mapsto(d,-d)\in D^{2})}(y\circledast y)\\
& =\mathbf{0}_{m}^{D^{2}}$$
\[tr4\]Given $x,y\in\mathcal{A}^{1}$ with $\pi(x)=\pi(y)$, we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y\circledast x)\nonumber\\
& =\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast y) \label{tr4.3}$$ and $$[x,y]=y\circledast x\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in
D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast y)
\label{tr4.4}$$
Our proof is the proof of Proposition 8 (§3.4) of Lavendhomme [@l1] in disguise. In order to show the identity (\[tr4.3\]), it suffices, by dint of the quasi-colimit diagram in Proposition 6 (§2.2) of Lavendhomme [@l1], to show that $$\begin{aligned}
& \mathcal{A}^{(d\in D\longmapsto(d,0)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y\circledast
x)\nonumber\\
& =\mathcal{A}^{(d\in D\longmapsto(d,0)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}\circ\nonumber\\
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast y) \label{tr4.1}$$ and $$\begin{aligned}
& \mathcal{A}^{(d\in D\longmapsto(0,d)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y\circledast
x)\nonumber\\
& =\mathcal{A}^{(d\in D\longmapsto(0,d)\in D\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}\circ\nonumber\\
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast y) \label{tr4.2}$$ Since the composition of $d\in D\longmapsto(d,0)\in D\oplus D$ and $(d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2}$ is equal to $d\in
D\longmapsto(d,0)\in D^{2}$, and since the composition of $d\in D\longmapsto
(d,0)\in D\oplus D$, $(d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2}$ and $(d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2}$ is equal to $d\in D\longmapsto(0,d)\in D^{2}$, it is easy to see that both sides of the identity (\[tr4.1\]) are equal to $x$ by the second condition in Definition \[d2.4\]. The identity (\[tr4.2\]) can be established similarly. Let $$z=\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\longmapsto(d_{2},d_{3},d_{1})\in D^{3})}\mathcal{(}x\circledast\lbrack x,y]\circledast y)$$ Then we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},0)\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},0)\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\longmapsto(d_{2},d_{3},d_{1})\in D^{3})}\mathcal{(}x\circledast\lbrack
x,y]\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},0,d_{1})\in D^{3})}\mathcal{(}x\circledast\lbrack x,y]\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\circ\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},0,d_{2})\in
D^{3})}\mathcal{(}x\circledast\lbrack x,y]\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}(x\circledast\mathcal{A}^{(d\in D\longmapsto(d,0)\in D^{2})}([x,y]\circledast y))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}(x\circledast y)\end{aligned}$$ while we have $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\longmapsto(d_{2},d_{3},d_{1})\in D^{3})}\mathcal{(}x\circledast\lbrack
x,y]\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1}d_{2},d_{1})\in
D^{3})}\mathcal{(}x\circledast\lbrack x,y]\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1}d_{2},d_{1})\in
D^{3})}\mathcal{(}x\circledast\mathcal{A}^{(d\in D\mapsto-d\in D)}\circ\mathcal{A}^{(d\in D\mapsto-d\in D)}([x,y])\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1}d_{2},d_{1})\in
D^{3})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},-d_{2},d_{3})\in D^{3})}(x\circledast\mathcal{A}^{(d\in D\mapsto-d\in
D)}([x,y])\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{1}d_{2},d_{1})\in D^{3})}(x\circledast\lbrack y,x]\circledast y)\\
& \text{[By Proposition \ref{tr5}]}\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{2},d_{1},d_{1})\in D^{4})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4})\in D^{4}\mapsto(d_{1},d_{2}d_{3},d_{4})\in D^{3})}(x\circledast\lbrack y,x]\circledast
y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{2},d_{1},d_{1})\in D^{4})}(x\circledast\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto
d_{1}d_{2}\in D)}([y,x])\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{2},d_{1},d_{1})\in D^{4})}(x\circledast\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}(x\circledast y\circledast
x\circledast y)\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{2},d_{1},d_{1})\in D^{4})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4})\in D^{4}\mapsto(d_{1},d_{2},d_{3},-d_{2},-d_{3},d_{4})\in D^{6}})\\
& (x\circledast x\circledast y\circledast x\circledast y\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},-d_{2},d_{1},d_{2},-d_{1},d_{1})\in D^{6})}(x\circledast x\circledast y\circledast
x\circledast y\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},d_{2},d_{1})\in D^{4})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3},d_{4})\in
D^{4}\longmapsto(d_{1},-d_{1},d_{2},d_{3},-d_{4},d_{4})\in D^{6})}\\
& (x\circledast x\circledast y\circledast x\circledast y\circledast y)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},d_{2},d_{1})\in D^{4})}(\mathcal{A}^{(d\in D\mapsto(-d,d)\in D^{2})}\mathcal{(}x\circledast x)\circledast y\circledast x\circledast\\
& \mathcal{A}^{(d\in D\mapsto(d,-d)\in D^{2})}\mathcal{(}y\circledast y))\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1},d_{2},d_{1})\in D^{4})}(\mathbf{0}_{m}^{D}\circledast y\circledast x\circledast
\mathbf{0}_{m}^{D})\\
& =y\circledast x\end{aligned}$$ Therefore we have $$\begin{aligned}
& y\circledast x\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in
D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast y)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}(z)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\longmapsto(d_{2},d_{3},d_{1})\in
D^{3})}\mathcal{(}x\circledast\lbrack x,y]\circledast y)\\
& =[x,y]\end{aligned}$$ This completes the proof.
1. Given $x\in\mathcal{A}^{1}$ and $y,z\in\mathcal{A}^{2}$with $\pi
(x)=\pi(y)=\pi(z)$, if we have $$\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y)=\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(z)$$ then we have $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in D\times(D\oplus D)\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(y\circledast x)=\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D\times(D\oplus D)\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(z\circledast x)$$ and $$z\circledast x\overset{\cdot}{\underset{1}{-}}y\circledast x=(z\overset{\cdot
}{-}y)\circledast x$$
2. Given $x,y\in\mathcal{A}^{2}$ and $z\in\mathcal{A}^{1}$with $\pi
(x)=\pi(y)=\pi(z)$, if we have $$\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(x)=\mathcal{A}^{((d_{1},d_{2})\in D\oplus D\mapsto(d_{1},d_{2})\in D^{2})}(y)$$ then we have $$\mathcal{A}^{((d_{1},d_{2},d_{3})\in(D\oplus D)\times D\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(z\circledast x)=\mathcal{A}^{((d_{1},d_{2},d_{3})\in(D\oplus D)\times D\mapsto(d_{1},d_{2},d_{3})\in D^{3})}(z\circledast y)$$ and $$\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{2},d_{1})\in D^{2})}(z\circledast y\overset{\cdot}{\underset{3}{-}}z\circledast x)=z\circledast
(y\overset{\cdot}{-}x)$$
This follows simply from Proposition \[t2.6\].
Given $x,y,z\in\mathcal{A}^{1}$ with $\pi(x)=\pi(y)=\pi(z)$, let it be the case that $$\begin{aligned}
u_{123} & =z\circledast y\circledast x\\
u_{132} & =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in D^{3})}\mathcal{(}y\circledast z\circledast x)\\
u_{213} & =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{2},d_{1},d_{3})\in D^{3})}(z\circledast x\circledast y)\\
u_{231} & =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{2},d_{3},d_{1})\in D^{3})}(x\circledast z\circledast y)\\
u_{312} & =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{3},d_{1},d_{2})\in D^{3})}\mathcal{(}y\circledast x\circledast z)\\
u_{321} & =\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{3},d_{2},d_{1})\in D^{3})}\mathcal{(}x\circledast y\circledast z)\end{aligned}$$ Then the right-hands of the following three identities are meaningful, and all the three identities hold: $$\begin{aligned}
\lbrack x,[y,z]] & =(u_{123}\overset{\cdot}{\underset{1}{-}}u_{132})\overset{\cdot}{-}(u_{231}\overset{\cdot}{\underset{1}{-}}u_{321})\\
\lbrack y,[z,x]] & =(u_{231}\overset{\cdot}{\underset{2}{-}}u_{213})\overset{\cdot}{-}(u_{312}\overset{\cdot}{\underset{2}{-}}u_{132})\\
\lbrack z,[x,y]] & =(u_{312}\overset{\cdot}{\underset{3}{-}}u_{321})\overset{\cdot}{-}(u_{123}\overset{\cdot}{\underset{3}{-}}u_{213})\end{aligned}$$
Here we deal only with the first identity, leaving the other two identities to the reader. We have $$\begin{aligned}
& [x,[y,z]]\\
& =[y,z]\circledast x\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in
D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast\lbrack y,z])\\
& =\{z\circledast y\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in
D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}y\circledast
z)\}\circledast x\\
& \overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto
(d_{2},d_{1})\in D^{2})}\mathcal{(}x\circledast\{z\circledast y\overset{\cdot
}{-}\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}y\circledast z)\})\\
& =\{z\circledast y\circledast x\overset{\cdot}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}y\circledast z)\circledast x\}\\
& \overset{\cdot}{-}\mathcal{\{}x\circledast z\circledast y\overset{\cdot
}{\underset{3}{-}}x\circledast\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}y\circledast z)\}\\
& =\{z\circledast y\circledast x\overset{\cdot}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in
D^{3})}\mathcal{(}y\circledast z\circledast x)\}\\
& \overset{\cdot}{-}\mathcal{\{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto
(d_{2},d_{3},d_{1})\in D^{3})}(x\circledast z\circledast y)\\
& \overset{\cdot}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{2},d_{3},d_{1})\in D^{3})}(x\circledast\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}y\circledast z))\}\\
& =\{z\circledast y\circledast x\overset{\cdot}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in
D^{3})}\mathcal{(}y\circledast z\circledast x)\}\\
& \overset{\cdot}{-}\mathcal{\{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto
(d_{2},d_{3},d_{1})\in D^{3})}(x\circledast z\circledast y)\\
& \overset{\cdot}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in
D^{3}\mapsto(d_{2},d_{3},d_{1})\in D^{3})}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\longmapsto(d_{2},d_{1},d_{3})\in D^{3})}(x\circledast
y\circledast z)\}\\
& =\{z\circledast y\circledast x\overset{\cdot}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto(d_{1},d_{3},d_{2})\in
D^{3})}\mathcal{(}y\circledast z\circledast x)\}\\
& \overset{\cdot}{-}\mathcal{\{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto
(d_{2},d_{3},d_{1})\in D^{3})}(x\circledast z\circledast y)\overset{\cdot
}{\underset{1}{-}}\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{3}\mapsto
(d_{3},d_{2},d_{1})\in D^{3})}\mathcal{(}x\circledast y\circledast z)\}\\
& =(u_{123}\overset{\cdot}{\underset{1}{-}}u_{132})\overset{\cdot}{-}(u_{231}\overset{\cdot}{\underset{1}{-}}u_{321})\end{aligned}$$
Given $m\in M$, the Jacobi identity holds for $\mathcal{A}_{m}^{1}$ with respect to the Lie bracket $[\cdot,\cdot]$. I.e., we have $$\lbrack x,[y,z]]+[y,[z,x]]+[z,[x,y]]=\mathbf{0}$$ for any $x,y,z\in\mathcal{A}_{m}^{1}$.
From Nishimura Algebroids to Lie Algebroids
===========================================
Let $\mathcal{A}$ be a *Nishimura algebroid over* $M$. It is very easy to see that
\[lt2\]By assigning $\Gamma(\mathcal{A)}^{\mathcal{D}}\mathcal{=}\Gamma(\mathcal{A}^{\mathcal{D}}\mathcal{)}$ to each object $\mathcal{D}$ in $\mathbf{Simp}$ and assigning $\Gamma(\mathcal{A)}^{f}:\gamma\in$ $\Gamma(\mathcal{A}^{\mathcal{D}^{\prime}}\mathcal{)\mapsto A}^{f}\circ
\gamma\in\Gamma(\mathcal{A}^{\mathcal{D}}\mathcal{)}$ to each morphism $f:\mathcal{D\rightarrow D}^{\prime}$ in $\mathbf{Simp}$, we have a *Nishimura algebroid*$_{2}$ $\Gamma(\mathcal{A)}$ *over* a single point, where $\Gamma(\mathcal{A}^{\mathcal{D}}\mathcal{)}$ denotes the space of global sections of the bundle $\mathcal{A}^{\mathcal{D}}$ over $M$. Endowed with the trivial anchor natural transformation, it is a *Nishimura algebroid*$_{3}$ *over* a single point.
Given $X\in\Gamma(\mathcal{A}^{\mathcal{D}_{1}}\mathcal{)}$ and $Y\in
\Gamma(\mathcal{A}^{\mathcal{D}_{2}}\mathcal{)}$, we define $Y\circledcirc
X\in\Gamma(\mathcal{A}^{\mathcal{D}_{1}\times\mathcal{D}_{2}}\mathcal{)}$ to be $$(Y\circledcirc X)_{m}=(Y\circ\mathbf{a}(X_{m}))\ast X_{m}$$ for any $m\in M$.
Now we have
\[lt1\]Given $X\in\Gamma(\mathcal{A}^{\mathcal{D}_{1}}\mathcal{)}$, $Y\in\Gamma(\mathcal{A}^{\mathcal{D}_{2}}\mathcal{)}$ and $Z\in\Gamma
(\mathcal{A}^{\mathcal{D}_{3}}\mathcal{)}$, we have $$Z\circledcirc(Y\circledcirc X)=(Z\circledcirc Y)\circledcirc X$$
Let $m\in M$. We have $$\begin{aligned}
& (Z\circledcirc(Y\circledcirc X))_{m}\\
& =(Z\circ\mathbf{a}((Y\circledcirc X)_{m}))\ast(Y\circledcirc X)_{m}\\
& =(Z\circ(\mathbf{a}(Y)\circledcirc\mathbf{a}(X))_{m})\ast\{(Y\circ
\mathbf{a}(X_{m}))\ast X_{m}\}\\
& =[\{m^{\prime}\in M\mapsto(Z\circ\mathbf{a}(Y_{m^{\prime}}))\ast
Y_{m^{\prime}}\}\circ\mathbf{a}(X_{m})]\ast X_{m}\\
& \text{[By the fourth condition in Definition \ref{d2.4}]}\\
& =((Z\circledcirc Y)\circledcirc X)_{m}$$
By this proposition we can omit parentheses in a combination by $\circledcirc$.
\[lt3\]By adopting $\circledcirc$ as $\ast_{\Gamma(\mathcal{A)}}$, our *Nishimura algebroid*$_{3}$ $\Gamma(\mathcal{A)}$ *over* a single point is a *Nishimura algebroid*$_{4}$ *over* a single point.
The fourth condition in Definition \[d2.4\] follows from Proposition \[lt1\]. The other three conditions follow trivially.
Therefore all the discussions of the previous section hold. In particular, we have
\[lt4\]Given $X,Y\in\Gamma(\mathcal{A}^{1}\mathcal{)}$, we can define $[X,Y]\in\Gamma(\mathcal{A}^{1}\mathcal{)}$ to be the unique one satisfying $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto d_{1}d_{2}\in D)}\circ\lbrack X,Y]\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{1},d_{2},-d_{1},-d_{2})\in D^{4})}\circ(Y\circledcirc X\circledcirc Y\circledcirc X)\text{,}$$ with respect to which $\Gamma(\mathcal{A}^{1}\mathcal{)}$ is a Lie algebra.
\[lt5\]Given $X,Y\in\Gamma(\mathcal{A}^{1}\mathcal{)}$ and $f\in
\mathbb{R}^{M}$, we have $$Y\circledcirc fX=f\underset{1}{\cdot}(Y\circledcirc X)$$
Let $m\in M$. $$\begin{aligned}
& (Y\circledcirc fX)_{m}\\
& =(Y\circ\mathbf{a}(f(m)X_{m}))\ast(f(m)X_{m})\\
& =(Y\circ(f(m)\mathbf{a}(X_{m})))\ast(f(m)X_{m})\\
& =f(m)\underset{1}{\cdot}(Y\circledcirc X)_{m}$$
\[lt6\]Given $X,Y\in\Gamma(\mathcal{A}^{1}\mathcal{)}$ and $f\in
\mathbb{R}^{M}$, we have $$fY\circledcirc X\overset{\cdot}{-}f\underset{2}{\cdot}(Y\circledcirc
X)=\mathbf{a}(X)(f)Y$$
Let $m\in M$. We define $\mu\in\mathcal{A}_{m}^{D^{2}\oplus D}$ to be $$\mu=\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2}f(m)+d_{3}\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc X)_{m})$$ where $\mathbf{a}(X_{m})(f)$ is the Lie derivative of $f$ with respect $\mathbf{a}(X_{m})$. It is easy to see that $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}(\mu)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},d_{1}d_{2})\in
D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus
D\mapsto(d_{1},d_{2}f(m)+d_{3}\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc
X)_{m})\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2}f(m)+d_{1}d_{2}\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc X)_{m})\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2}f(m)+d_{1}d_{2}\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc X)_{m})\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2}f(\mathbf{a}(X_{m})(d_{1})))\in D^{2})}((Y\circledcirc X)_{m})\\
& =(fY\circledcirc X)_{m}\\
& \text{\lbrack By the third condition in Definition \ref{d2.4}]}$$ It is also easy to see that $$\begin{aligned}
& \mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}(\mu)\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2},0)\in D^{2}\oplus
D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2}f(m)+d_{3}\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc X)_{m})\\
& =\mathcal{A}^{((d_{1},d_{2})\in D^{2}\mapsto(d_{1},d_{2}f(m))\in D^{2})}((Y\circledcirc X)_{m})\\
& =(f\underset{2}{\cdot}(Y\circledcirc X))_{m}\\
& \text{\lbrack By the third condition in Definition \ref{d2.4}]}$$ Therefore we have $$\begin{aligned}
& (fY\circledcirc X)_{m}\overset{\cdot}{-}(f\underset{2}{\cdot}(Y\circledcirc
X))_{m}\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}(\mu)\\
& =\mathcal{A}^{(d\in D\mapsto(0,0,d)\in D^{2}\oplus D)}\circ\mathcal{A}^{((d_{1},d_{2},d_{3})\in D^{2}\oplus D\mapsto(d_{1},d_{2}f(m)+d_{3}\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc X)_{m})\\
& =\mathcal{A}^{(d\in D\mapsto(0,d\mathbf{a}(X_{m})(f))\in D^{2})}((Y\circledcirc X)_{m})\\
& =\mathbf{a}(X_{m})(f)Y_{m}$$ This completes the proof.
\[lt7\]Given $X,Y\in\Gamma(\mathcal{A}^{1}\mathcal{)}$ and $f\in
\mathbb{R}^{M}$, we have $$\lbrack X,fY]=f[X,Y]+\mathbf{a}(X)(f)Y$$
We have $$\begin{aligned}
& [X,fY]\\
& =fY\circledcirc X\overset{\cdot}{-}\mathcal{A}^{((d_{1},d_{2})\in
D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}\mathcal{(}X\circledcirc fY)\\
& =\{fY\circledcirc X\overset{\cdot}{-}f\underset{2}{\cdot}(Y\circledcirc
X)\}-\\
& \{\mathcal{A}^{((d_{1},d_{2})\in D^{2}\longmapsto(d_{2},d_{1})\in D^{2})}(f\underset{1}{\cdot}(X\circledcirc Y))\overset{\cdot}{-}f\underset{2}{\cdot}(Y\circledcirc X)\}\\
& \text{[By Proposition \ref{lt6}]}\\
& =\mathbf{a}(X)(f)Y+f[X,Y]\end{aligned}$$ This completes the proof.
Given a *Nishimura algebroid* $\mathcal{A}$ *over* $M$, $\mathcal{A}^{1}$ is a Lie algebroid over $M$.
This follows directly from Theorem \[lt4\] and Proposition \[lt7\].
[99]{}
Borceux, F.:Handbook of Categorical Algebra, 3 vols., Cambridge University Press, Cambridge, 1984.
Grabowski, J. and Urbański:Algebroids—general differential calculi on vector bundles, Journal of Geometry and Physics, **31** (1999), 111-141.
Grabowski, J.:Quasi-derivations and QD-algebroids, Reports on Mathematical Physics, **52** (2003), 445-451.
Kock, A.: Synthetic Differential Geometry, London Mathematical Society Lecture Note Series, **51**, Cambridge University Press, Cambridge, 1981.
Lavendhomme, R.: Basic Concepts of Synthetic Differential Geometry, Kluwer, Dordrecht, 1996.
Lawvere, W.:Functorial semantics of algebraic theories and some algebraic problems in the context of functorial semantics of algebraic theories, Repr. Theory Appl. Categ., 5 (2004), 1-121.
Mackenzie, K. C. H.:General Theory of Lie Groupoids and Lie Algebroids, London Mathematical Society Lecture Note Series, **213**, Cambridge University Press, Cambridge, 2005.
Makkai, M. and Paré, R.:Accessible Categories:The Foundations of Categorical Model Theory, Contemporary Mathematics, **104** (1989), American Mathematical Society.
Moerdijk, I. and Reyes, G. E.: Models for Smooth Infinitesimal Analysis, Springer-Verlag, New York, 1991.
Nishimura, H.:Theory of microcubes, International Journal of Theoretical Physics, **36** (1997), 1099-1131.
Nishimura, H.:General Jacobi identity revisited, International Journal of Theoretical Physics, **38** (1999), 2163-2174.
Nishimura, H.:Synthetic differential geometry of higher-order total differentials, Cahiers de Topologie et Géométrie Differéntielle Catégoriques, (2006), 207-232.
Nishimura, H.:The Lie algebra of the group of bisections, Far East Journal of Mathematical Sciences, **24** (2007), 329-342.
Schubert, H.:Categories, Springer-Verlag, Berlin and Heidelberg, 1972.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Pursuit-Evasion Games (in discrete time) are stochastic games with nonnegative daily payoffs, with the final payoff being the cumulative sum of payoffs during the game. We show that such games admit a value even in the presence of incomplete information and that this value is uniform, i.e. there are $\epsilon$-optimal strategies for both players that are $\epsilon$-optimal in any long enough prefix of the game. We give an example to demonstrate that nonnegativity is essential and expand the results to Leavable Games.'
author:
- 'Ori Gurel-Gurevich [^1]'
title: 'Pursuit-Evasion Games with Incomplete Information in Discrete Time'
---
**Key words:** pursuit-evasion games, incomplete information, zero-sum stochastic games, recursive games, nonnegative payoffs.
Introduction
============
Games of Pursuit and Evasion are two-player zero-sum games involving a Pursuer (P) and an Evader (E). P’s goal is to capture E, and the game consist of the space of possible locations and the allowed motions for P and E. These games are usually encountered within the domain of differential games, i.e., the location space and the allowed motions have the cardinality of the continuum and they tend to be of differentiable or at least continuous nature.
The subject of Differential Games in general, and Pursuit-Evasion Games in particular, was pioneered in the 50s by Isaacs (1965). These games evolved from the need to solve military problems such as airfights, as opposed to classical game theory which was oriented toward solving economical problems. The basic approach was akin to differential equations techniques and optimal control, rather than standard game theoretic tools. The underlying assumption was that of complete information, and optimal *pure* strategies were searched for. Conditions were given, under which a pure strategies saddle point exists (see, for example, Varaiya and Lin (1969)). Usually the solution was given together with a value function, which assigned each state of the game its value. Complete information was an essential requirement in this case. For a thorough introduction to Pursuit-Evasion and Differential Games see Basar and Olsder (1999).
A complete-information continuous-time game “intuitively” shares some relevant features with *perfect-information* discrete-time games. The latter are games with complete knowledge of past actions and without simultaneous actions. Indeed, if one player decides to randomly choose between two pure strategies which differ from time $t_0$ and on, his opponent will discover this “immediately” after $t_0$, thus enabling himself to respond optimally almost instantly. Assuming the payoff is continuous, the small amount of time needed to discover the strategy chosen by the opponent should affect the payoff negligibly. A well-known result of Martin (1975, 1985) implies that every perfect-information discrete-time game has $\epsilon$-optimal pure strategies (assuming a Borel payoff function) and so should, in a sense, continuous time games.
Another reason to restrict oneself to pure strategies is that unlike discrete-time games, there is no good formal framework for continuous-time games. By framework we mean a way to properly define the space of pure strategies and the measurable $\sigma$-algebra on them. There are some approaches but none is as general or complete as for discrete-time games. This kind of framework is essential when dealing with a general incomplete information setting.
This paper will therefore deal with discrete-time Pursuit-Evasion Games. We hope that our result will be applied in the future to discrete approximations of continuous-time games. Pursuit-Evasion Games in discrete time are formalized and discussed in Kumar and Shiau (1981).
Pursuit-Evasion Games are generally divided into two categories: *Games of Kind* and *Games of Degree*. Games of Kind deal with the question of *capturability*: whether a capture can be achieved by the Pursuer or not. In a complete-information setting this is a yes-or-no question, completely decided by the rules of the game and the starting positions. With incomplete information incorporated, we simply assign a payoff of 1 for the event of capture and payoff 0 otherwise. Games of Degree have the Pursuer try to minimize a certain payoff function such as the time needed for capture. The question of capturability is encountered here only indirectly: if the Evader have a chance of escaping capture indefinitely, the expected time of capture is infinity. The payoff, in general, can be any function, such as the minimal distance between the Evader and some target set.
What unites the two categories is that the payoff function in both is positive and cumulative. The maximizing player, be it the Pursuer or the Evader, gains his payoff and never loses anything. This is in contrast with other classes of infinitely repeated games, such as undiscounted stochastic games, where the payoff is the limit of the averages of daily payoffs.
Discrete-time stochastic games were introduced by Shapley (1953) who proved the existence of the discounted value in two-player zero-sum games with finite state and action sets. Recursive games were introduced by Everett (1957). These are stochastic games, in which the payoff is 0 except for absorbing states, when the game terminates. Thus, absorbing states are as happens in Pursuit-Evasion Games, where the payoff is obtained only when the game terminates. The game is said to have a uniform value if $\epsilon$-optimal strategies exist that are also $\epsilon$-optimal in any long enough prefix of the game. Everett proved the existence of the uniform value for two-player, zero-sum recursive games.
We shall now formally define Pursuit-Evasion Games to be two-player zero-sum games with cumulative and positive payoffs. To avoid confusion, the players will be called the Maximizer and the Minimizer, and their respective goals should be obvious.
Our main result is the existence of uniform value for Pursuit-Evasion Games with incomplete-information and finite action and signal sets, followed by a generalization for arbitrary signal sets. In section 4 we present a different class of games to which our proof also applies. In section 5 we show that the positiveness requirement is indispensable by giving an appropriate counterexample.
Definitions and the main Theorem
================================
A *cumulative game with complete information* is given by:
- Two finite sets $A^1$ and $A^2$ of actions.
Define $H_n=(A^1\times A^2)^n$ to be the set of all histories of length $n$, and $H=\cup_{n=0}^\infty H_n$ to be the set of all finite histories.
- A daily payoff function $f:H\rightarrow\mathbb{R}$.
Let $\widetilde{H}=(A^1\times A^2)^{\aleph_0}$ be the set of all infinite histories. The daily payoff function induces a payoff function $\rho:\widetilde{H}\rightarrow\mathbb{R}$ by $\rho(h)=\sum_{n=0}^\infty f(h_n)$, where $h_n$ is the length $n$ prefix of $h$. In the sequel we will only study the case in which $f$ is nonnegative, so that $\rho$ is well defined (though it may be infinite).
The game is played in stages as follows. The initial history is $h_0=\emptyset$. At each stage $n\geq 0$ both players choose simultaneously and independently actions $a\in A$ and $b\in B$, and each player is informed of the other’s choice. The new game history is $h_{n+1}=h_n\frown<a,b>$, i.e., the concatenation of $<a,b>$ to the current history. The infinite history of the game, $h$, is the concatenation of all pairs of actions chosen throughout the game. The payoff is $\rho(h)$, the goal of the Maximizer is to maximize the expectation of $\rho(h)$, and that of the Minimizer is to minimize it.
If all the values of $f$ are nonnegative, we call the game *nonnegative*. A *complete information Pursuit-Evasion Game* is a nonnegative cumulative game.
As cumulative games are a proper superset of recursive games (see Everett (1957)), Pursuit-Evasion Games are a proper superset of nonnegative recursive games.
As is standard in game theory, the term “complete information” is used to denote a game with complete knowledge of the history of the game, and not the lack of simultaneous actions (which is termed “perfect information”).
A *cumulative game with incomplete information* is given by:
- Two finite sets $A^1$ and $A^2$ of actions.
Define $H_n$ and $H$ as before.
- A daily payoff function $f:H\rightarrow\mathbb{R}$.
- Two measure spaces $S^1$ and $S^2$ of signals.
- $\forall h\in H$ two probability distributions $p_h^1 \in
\Delta(S^1)$ and $p_h^2 \in \Delta(S^2)$.
Define $\widetilde{H}$ and $\rho$ as before. In particular, the signals are not a parameter of the payoff function.
An incomplete-information cumulative game is played like a complete information cumulative game, except that the players are not informed of each other’s actions. Instead, a signal pair $<s^1,s^2>\in S^1\times S^2$ is randomly chosen with distribution $p_h^1\times p_h^2$, $h$ being the current history of the game, with player $i$ observing $s^i$. An *incomplete-information Pursuit-Evasion Game* is an incomplete-information nonnegative cumulative game.
Define $H^i_n$ to be $(A^i\times S^i)^n$. This is the set of private histories of length $n$ of player $i$. Similarly, define $H^i=\cup_{n=0}^\infty H^i_n$, the set of all private finite histories, and $\widetilde{H}^i=(A^i\times S^i)^{\aleph_0}$ the set of all private infinite histories.
In a complete-information cumulative game a behavioral *strategy* for player $i$ is a function $\sigma^i:H\rightarrow \Delta(A^i)$. In an incomplete-information cumulative game a (behavioral) *strategy* for player $i$ is a function $\sigma^i:H^i\rightarrow\Delta(A^i)$. Recall that by Kuhn’s Theorem (Kuhn (1953)) the set of all behavioral strategies coincides with the set of all mixed strategies, which are probability distributions over pure strategies.
Denote the space of all behavioral strategies for player $i$ by $\Omega^i$. A *profile* is a pair of strategies, one for each player. A profile $<\sigma^1,\sigma^2>$, together with $\{p_h^i\}$, induces, in the obvious manner, a probability measure $\mu_{\sigma^1,\sigma^2}$ over $\widetilde{H}$ equipped with the product $\sigma$-algebra.
The *value* of a strategy $\sigma^1$ for the Maximizer is $val(\sigma^1)=\inf_{\sigma^2 \in \Omega^2}
E_{\mu_{\sigma^1,\sigma^2}}(\rho(h))$.
The *value* of a strategy $\sigma^2$ for the Minimizer is $val(\sigma^2)=\sup_{\sigma^1 \in \Omega^1}
E_{\mu_{\sigma^1,\sigma^2}}(\rho(h))$.
When several games are discussed we will explicitly denote the value in game $G$ by $val_G$.
The *lower value* of the game is $\underline{val}(G)=\sup_{\sigma^1 \in \Omega^1} val(\sigma^1)$.
The *upper value* of the game is $\overline{val}(G)=\inf_{\sigma^2 \in \Omega^2} val(\sigma^2)$.
If $\underline{val}(G)=\overline{val}(G)$, the common value is the *value* of the game $val(G)=\underline{val}(G)=\overline{val}(G)$. Observe that $\underline{val}(G)$ and $\overline{val}(G)$ always exist, and that $\underline{val}(G)\leq\overline{val}(G)$ always holds.
A strategy $\sigma^i$ of player $i$ is *$\epsilon$-optimal* if $|val(\sigma^i)-val(G)|<\epsilon$. A strategy is *optimal* if it is 0-optimal.
A cumulative game is *bounded* if its payoff function $\rho$ is bounded, i.e. $\exists B\in\mathbb{R}\forall h \in
\widetilde{H} \ \ \ -B<\rho(h)<B$.
Let $G=<A^1,A^2,f>$ be a cumulative game. Define $f_n$ to be equal to $f$ for all histories of length up to $n$ and zero for all other histories. Define $G_n=<A^1,A^2,f_n>$. Thus, $G_n$ is the restriction of $G$ to the first $n$ stages. Let $\rho_n$ be the payoff function induced by $f_n$.
A game $G$ is said to have a *uniform value* if it has a value and for each $\epsilon>0$ there exist $N$ and two strategies $\sigma^1,\sigma^2$ for the two players that are $\epsilon$-optimal for every game $G_n$ with $n>N$.
The first main result is:
Every bounded Pursuit-Evasion Game with incomplete-information and finite signal sets has a uniform value. Furthermore, an optimal strategy exists for the Minimizer.
Let $G$ be a bounded Pursuit-Evasion Game with incomplete-information . Let $G_n$ be defined as above. Since $A^1,A^2,S^1,S^2$ are all finite, there are only a finite number of private histories of length up to $n$. $G_n$ is equivalent to a finite-stage finite-action game, and therefore it has a value $v_n$. From the definition of $G_n$ and since $f$ is nonnegative $$\forall h\in\widetilde{H} \ \
\rho_n(h)\leq\rho_{n+1}(h)\leq\rho(h)$$ which implies that for all $\sigma^1\in\Omega^1$ $$val_{G_n}(\sigma^1)\leq val_{G_{n+1}}(\sigma^1) \leq
val_G(\sigma^1)
\label{eq0}$$ so that $$\underline{val}(G_n)\leq \underline{val}(G_{n+1}) \leq
\underline{val}(G) .$$
Therefore, $v_n$ is a nondecreasing bounded sequence and $\underline{val}(G)$ is at least $v=\lim_{n\rightarrow\infty}v_n$.
On the other hand, define $K_n=\{\sigma^2\in\Omega^2 \mid
val_{G_n}(\sigma^2)\leq v\}$. Since $val(G_n)=v_n\leq v$, $K_n$ cannot be empty.
$K_n$ is a compact set, since the function $val_{G_n}(\sigma^2)$ is continuous over $\Omega^2$, which is compact, and $K_n$ is the preimage of the closed set $(-\infty,v]$.
For all $\sigma^2\in\Omega^2$ $val_{G_n}(\sigma^2)\leq
val_{G_{n+1}}(\sigma^2)$, so that $K_n\supseteq K_{n+1}$. Since the sets $K_n$ are compact, their intersection is nonempty.
Let $\sigma^2$ be a strategy for the Minimizer in $\cap_{n=0}^\infty K_n$. Let $\sigma^1$ be any strategy for the Maximizer. From $\rho(h)=\lim_{n\rightarrow\infty}\rho_n(h)$ and since $\rho$ is bounded, we get by the monotone convergence Theorem $$E_{\mu_{\sigma^1,\sigma^2}}(\rho(h))
=\lim_{n\rightarrow\infty}E_{\mu_{\sigma^1,\sigma^2}}(\rho_n(h)) .$$ Since $\sigma^2$ belongs to $K_n$, $E_{\mu_{\sigma^1,\sigma^2}}(\rho_n(h))\leq v$ and therefore $E_{\mu_{\sigma^1,\sigma^2}}(\rho(h))\leq v$. Since $\sigma^1$ is arbitrary $val(\sigma^2) \leq v$, so that $\overline{val}(G) \leq
v$. Consequentially, $v$ is the value of $G$.
Notice that any $\sigma^2 \in \cap_{n=0}^\infty K_n$ has $val_G(\sigma^2)=v$ and is therefore an optimal strategy for the Minimizer.
Given $\epsilon>0$ choose $N$ such that $v_{N}>v-\epsilon$. Let $\sigma^1$ be an optimal strategy for the Maximizer in $G_N$, and let $\sigma^2 \in \cap_{n=0}^\infty K_n$. By (\[eq0\]) $$\forall n>N \ \ v_n-\epsilon\leq v-\epsilon <
v_N=val_{G_N}(\sigma^1)\leq val_{G_n}(\sigma^1)$$ so that $\sigma^1$ is $\epsilon$-optimal in $G_n$. As $\sigma^2\in
K_n$ one has $val_{G_n}(\sigma^2)\leq v < v_n+\epsilon$ so that $\sigma^2$ is $\epsilon$-optimal in $G_n$.
These strategies are $\epsilon$-optimal in all games $G_n$ for $n>N$. Thus, the value is uniform.
Remark: Most of the assumption on the game $G$ are irrelevant for the proof of the theorem and were given only for the simplicity of description.
1. The action sets $A^i$ and the signal sets $S^i$ may depend respectively on the private histories $H^i_n$.
2. The signals $<s^1,s^2>$ may be correlated, i.e. chosen from a common distribution $p_h \in \Delta(S^1 \times S^2)$.
3. The game can be made stochastic simply by adding a third player, Nature, with a known behavioral strategy. The action set for Nature can be countable, since it could always be approximated by large enough finite sets. The action sets for the Maximizer can be infinite as long as the signals set $S^2$ is still finite (so the number of pure strategies for the Minimizer in $G_n$ is still finite).
4. Since the bound on payoffs was only used to bound the values of $G_n$, one can drop the boundedness assumption, as long as the sequence $\{v_n\}$ is bounded. If they are unbounded then $G$ has infinite uniform value in the sense that the Maximizer can achieve as high a payoff as he desires.
Arbitrary signal sets
=====================
Obviously, the result still hold if we replace the signal set $S$ by a sequence of signal sets $S_n$, all of which are finite, such that the signals for histories of length $n$ belong to $S_n$. The signal sets, like the action sets can change according to past actions, but since there are only finitely many possible histories of length $n$, this is purely semantical.
What about signals chosen from an infinite set? If the set $S$ is countable than we can approximate it with finite sets $S_n$, chosen such that for any history $h$ of length $n$ the chance we get a signal outside $S_n$ is negligible. We won’t go into details because the next argument applies for both the countable and the uncountable cases.
A cumulative game $G$ is *$\epsilon$-approximated* by a game $G'$ if $G'$ has the same strategy spaces as $G$ and for any pair of strategies $\sigma,\tau$ $$|\rho_G(\sigma,\tau)-\rho_{G'}(\sigma,\tau)|<\epsilon .$$
If $G$ is a bounded Pursuit-Evasion Game with incomplete information then $G$ can be $\epsilon$-approximated by a Pursuit Evasion Game with incomplete information with the same action sets and payoffs which can be simulated using a sequence of finite signal sets.
Let $G$ be such a game. Assume, w.l.o.g., that the payoff function $\rho$ is bounded by 1. Fix a positive $\epsilon$. Let $\epsilon_n=\epsilon/{2^n}$. Define $p_n^i=\sum_{h\in
H_n}p_h^i/|H_n|$, the mean distribution of the signals at stage $n$. Every distribution $p_h^i$ of time $n$ is absolutely continuous with respect to $p_n^i$. By Radon-Nykodim theorem, a density function $f_h^i$ exists such that $p_h^i(E)=\int_E f_h^i
dp_n^i$. Clearly, $f_h^i$ is essentially bounded by $|H_n|$.
Let $S'^i_n$ be $\{0,\epsilon_n,2\epsilon_n,3\epsilon_n,...,
\lfloor |H_n|/\epsilon_n \rfloor \epsilon_n\}^{|H_n|}$. For $h\in
H_n$ define $f'^i_h$ to be $f_h^i$ rounded down to the nearest multiple of $\epsilon_n$. Define $F'^i_n:S^i\rightarrow S'^i_n$ by $F'^i_n(s)=\{f'^i_h(s)\}_{h\in H_n}$. Let $G'$ be the same game as $G$ except that the players observe the signals $F'^i_n(s^i)\in
S'^i_n$ where $s^i$ is the original signal with density $f^i_h$.
Given a signal $s'^i$ in $S'^i_n$ one can project it back onto $S^i$ by choosing from a uniform distribution (with respect to the measure $p_n^i$) over the set $E(s'^i)={F'^i_n}^{-1}(s'^i)$. Let $G''$ be the game $G$ except that the signals are chosen with the distribution just described. Denote their density function by $f''^i_h$. This game can be simulated using only the signals in $G'$ and vice versa so they are equivalent.
$G$ and $G''$ have exactly the same strategy spaces. The only difference is a different distribution of the signals. But the way the signals in $G''$ were constructed it is obvious that the density function $f''^i_h$ do not differ from $f^i_h$ by more than $\epsilon_n$ for any history $h$ of length $n$. Given a profile $<\sigma^1,\sigma^2>$ denote the generated distributions on $\widetilde{H}$ in $G$ and $G''$ by $\mu$ and $\mu''$. The payoffs are $\rho_G(\sigma^1,\sigma^2)=\int \rho d\mu$ and $\rho_{G''}(\sigma^1,\sigma^2)=\int \rho d\mu''$ . But the distance, in total variation metric, between $\mu$ and $\mu''$ cannot be more than the sum of distances between the distributions of signals at each stage, which is no more than $\sum_{i=1}^\infty
\epsilon_i=\epsilon$. By definition of total variation metric, the difference between $\int \rho d\mu$ and $\int \rho d\mu''$ cannot be more than $\epsilon$.
If $G$ is as in lemma and have bounded nonnegative payoffs, it has a uniform value.
Let $G$ be such a game, and for any $\epsilon$ let $G_\epsilon$ be an $\epsilon$-approximation of $G$ produced by the lemma. $G_\epsilon$ is equivalent to a game with finite signal sets and therefore has a value according to Theorem 1, denoted $v_\epsilon$. It is immediate from the definition of $\epsilon$-approximation that $\underline{v}$, the lower value of $G$ cannot be less than $v_\epsilon - \epsilon$, and likewise $\overline{v}$ is no more than $v_\epsilon + \epsilon$. $\overline{v}-\underline{v}$ is therefore less than $2\epsilon$. But $\epsilon$ was chosen arbitrarily, so that $\overline{v}=\underline{v}$.
Given $\epsilon>0$ let $\sigma^1$ and $\sigma^2$ be $\epsilon/2$-optimal strategies in $G_{\epsilon/2}$ that are also $\epsilon/2$-optimal in any prefix of $G_{\epsilon/2}$ longer than $N$. Clearly, these strategies are $\epsilon$-optimal in any $G_n$ with $n>N$. Thus, the value is uniform.
Leavable games
==============
Leavable games are cumulative games in which one of the players, say the Maximizer, but not his opponent is allowed to leave the game at any stage. The obvious way to model this class of games would be to add a “stopping” stage between any two original stages, where the Maximizer will choose to either “stop” or “continue” the game. However, we would also like to force the Maximizer to “stop” at some stage. Unfortunately, it is impossible to do so and still remain within the realm of cumulative games, so we will have to deal with it a bit differently.
Leavable games were introduced by Maitra and Sudderth (1992) as an extension to similar concepts in the theory of gambling. They proved that a leavable game with complete information and finite action sets has a value. We will prove that the same is true for leavable games with incomplete information.
Let $G$ be a cumulative game with incomplete information. A *stop rule* for player $i$ is a function $s:\widetilde{H}^i\rightarrow \mathbb{N}$ such that if $s(h)=n$ and $h'$ coincides with $h$ in the first $n$ coordinates, then $s(h')=n$. A *leavable game with incomplete information* $L(G)$ is given by a cumulative game with incomplete information $G$ but is play differently, as follows. Instead of playing in stages, both players choose their behavioral strategies simultaneously with the Maximizer also choosing a stop rule $s$. The game is played according to these strategies and the payoff is $\rho(h^1)=\sum_{i=0}^{s(h^1)} f(h_n)$ where $h^1$ is the Maximizer’s private infinite history.
A bounded leavable game with incomplete information and finite signal sets has a value and that value is uniform. Furthermore, an optimal strategy exists for the Minimizer.
The proof is essentially identical to the proof of Theorem 1. $L_n$ is Defined to be the game where the Maximizer is forced to choose a stop rule $\leq n$. $L_n$ is thus equivalent to $G_n$ in the proof of Theorem 1.
The major point we should observe is that if $A^1$ and $S^1$ are finite, any stop rule $s:\widetilde{H}^1\rightarrow \mathbb{N}$ is uniformly bounded: $\exists B \forall h\in\widetilde{H}^1 \ \
s(h)<B$. This implies that any pure strategy for the Maximizer in $L$ actually belongs to some $L_n$. Therefore, a strategy $\sigma^2$ for the Minimizer with $val_{L_n}(\sigma^2)\leq v$ for all $n$, has $val_L(\sigma^2)\leq v$.
Counterexamples
===============
The question arises whether positiveness is an essential or just a technical requirement. Both our proof and the alternative proof outlined need the positiveness in an essential way, but still is it possible that every cumulative game have a value?
The answer is Negative. We shall provide a simple counterexample of a cumulative game (actually a stopping game, see Dynkin (1969)) with incomplete information without a value.
The game is as follows: at the outset of the game a bit (0 or 1) $b$ is chosen randomly with some probability $p>0$ to be 1 and probability $1-p$ to be 0. the Maximizer is informed of the value of $b$ but not the Minimizer. Then the following game is played. At each odd stage the Maximizer may opt to “stop” the game and the payoff is -1 if $b=0$ and 1 if $b=1$. At each even stage the Minimizer may opt to “stop” the game and the payoff is -1 if $b=0$ and some $A>\frac{1}{p}$ if $b=1$.
The payoff before and after someone decides to “stop” the game is zero.
This is a very simple stopping game with only one “unknown” parameter, yet, as we now argue, it has no value.
The upper value of this game is $p$
To see that $\overline{val}(G)\leq p$ let the Minimizer’s strategy be to continue at all stages. The Maximizer cannot gain more than $p 1 + (1-p) 0=p $ against this strategy, so the upper value cannot be higher than $p$.
On the other hand, let $\sigma$ be a strategy for the Minimizer. It consists of $\{\sigma_i\}_{i=1}^\infty$ the probabilities of stopping at stage $i$ and $\sigma_\infty=1-\sum_{i=1}^\infty
\sigma_i$ the probability of never choosing “stop”.
Fix $\epsilon>0$ and let $N$ be an odd integer such that $\sum_{i=N+1}^\infty \sigma_i < \epsilon$. Let $\tau$ be the following strategy for the Maximizer: if $b=0$ never stop, if $b=1$ stop at stage $N$. The payoff under $<\sigma,\tau>$ is: $$p\sum_{i=1}^N \sigma_i A + p (\sum_{i=N+1}^\infty \sigma_i +
\sigma_\infty) 1 + (1-p) \sum_{i=1}^\infty \sigma_i (-1) + (1-p)
\sigma_\infty 0$$ $$= p (\sum_{i=1}^\infty \sigma_i + \sigma_\infty) + \sum_{i=1}^N
\sigma_i (pA-1) + \sum_{i=N+1}^\infty \sigma_i (p-1) \geq
p-\epsilon$$ where the last inequality holds since $pA-1>0$ and $\sum_{i=N+1}^\infty \sigma_i < \epsilon$.
Therefore $\overline{val}(G)\geq p$.
The lower value of this game is $p-\frac{1-p}{A}$.
Let the Maximizer play the following strategy: If $b=1$ stop at time 1 with probability $1-\frac{1-p}{Ap}$ and continue otherwise. If the Minimizer never decides to stop the payoff will be $p(1-\frac{1-p}{Ap})1+(1-p)0=p-\frac{1-p}{A}$. If the Minimizer decides to stop at any stage, the payoff will be $p(1-\frac{1-p}{Ap})1 + p\frac{1-p}{Ap}A +
(1-p)(-1)=p-\frac{1-p}{A}$. Clearly any mix of these pure strategies will also result in payoff of exactly $p-\frac{1-p}{A}$.
To see that the Maximizer cannot guarantee more assume to the contrary that there exist a strategy $\sigma$ for the Maximizer with $val(\sigma)>p-\frac{1-p}{A}$. This strategy consists of the probabilities $\{\sigma^0_i\}_{i=1}^\infty$ of stopping at stage $i$ if $b=0$, and $\{\sigma^1_i\}_{i=1}^\infty$ if $b=1$.
By our assumption, the payoff against any strategy for the Minimizer should be more than $p-\frac{1-p}{A}$. Let the Minimizer always choose to continue. The expected payoff in that case is $$p(\sum_{i=1}^\infty \sigma^1_i)1 + (1-p)(\sum_{i=1}^\infty
\sigma^0_i)(-1)>p-\frac{1-p}{A} ,$$ which implies $$\sum_{i=1}^\infty \sigma^1_i > 1 - \frac{1-p}{Ap} .$$ Let $N$ be sufficiently large such that $\sum_{i=1}^N \sigma^1_i >
1 - \frac{1-p}{Ap}$. Consider the following strategy for the Minimizer: continue until stage $N$ and then stop. The payoff will be $$p(\sum_{i=1}^N \sigma^1_i)1 + p(1-\sum_{i=1}^N \sigma^1_i)A +
(1-p)(-1)$$ $$=p + p(1-\sum_{i=1}^N \sigma^1_i)(A-1) + (1-p)(-1)$$ $$<p+p\frac{1-p}{Ap}(A-1)+p-1=p-\frac{1-p}{A} ,$$ a contradiction.
[99]{}
[^1]: Weizmann Institute of Science, Rehovot, 76100, Israel. e-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper aims to select features that contribute most to the pose estimation in VO/VSLAM. Unlike existing feature selection works that are focused on efficiency only, our method significantly improves the accuracy of pose tracking, while introducing little overhead. By studying the impact of feature selection towards least squares pose optimization, we demonstrate the applicability of improving accuracy via good feature selection. To that end, we introduce the [*Max-logDet*]{} metric to guide the feature selection, which is connected to the conditioning of least squares pose optimization problem. We then describe an efficient algorithm for approximately solving the NP-hard [*Max-logDet*]{} problem. Integrating [*Max-logDet*]{} feature selection into a state-of-the-art visual SLAM system leads to accuracy improvements with low overhead, as demonstrated via evaluation on a public benchmark.'
author:
- 'Yipu Zhao$^{1}$ and Patricio A. Vela$^{1}$[^1] [^2]'
bibliography:
- 'draft.bib'
title: '**Good Feature Selection for Least Squares Pose Optimization in VO/VSLAM** '
---
Introduction
============
Feature Selection in Least Squares Pose Optimization
====================================================
Good Feature Selection Metrics
==============================
Efficient Max-logDet Subset Selection
=====================================
Experimental Results on Real-time VSLAM
=======================================
Conclusion
==========
[^1]: $^{1}$Yipu Zhao [[email protected]]{} and Patricio A. Vela [[email protected]]{} are with Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, Georgia, USA.
[^2]: This work was supported, in part, by the National Science Foundation under Grant No. 1400256 and 1544857.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'B.S. Vivek, R. Venkatesh Babu, [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'main.bib'
title: 'Regularizers for Single-step Adversarial Training'
---
[Shell : Bare Demo of IEEEtran.cls for Computer Society Journals]{}
Neural Networks (DNNs) achieve impressive performance on various tasks in Computer Vision. However, the susceptibility of these networks to adversarial samples [@intriguing-iclr-2014] (samples with crafted noise that can manipulate the model’s output) is an important issue. Further, Szegedy [*et al*. ]{}[@intriguing-iclr-2014] showed that these adversarial samples are transferable across models of the same or different architectures, and this property enables an attacker to launch attacks on the deployed models in a black-box setting ([@papernot2017practical; @delving-iclr-2017; @dong2018boosting]): where partial or no knowledge of the deployed model is available to the attacker. These properties of adversarial samples pose challenges for the deployment of DNNs in the real world. A plethora of works have proposed various methods to defend against adversarial attacks, such as input transformations ([@dziugaite2016study; @guo2018countering]), adversarial training ([@explainingharnessing-iclr-2015; @atscale-iclr-2017; @madry-iclr-2018; @ensembleAT-iclr-2018]), detection ([@feinman2017detecting; @xu2017feature]), etc. In this direction, adversarial training method shows promising results, where mini-batches are augmented with adversarial samples, and typically these samples are generated by the model being trained. Further, adversarial samples can be generated by non-iterative ([@explainingharnessing-iclr-2015; @atscale-iclr-2017]) or iterative methods ([@deepfool-cvpr-2016; @madry-iclr-2018]). In order to scale adversarial training to large datasets, non-iterative methods such as Fast Gradient Sign Method (FGSM) [@explainingharnessing-iclr-2015] are used. However, models trained using single-step adversarial training methods (adversarial samples are generated using non-iterative methods) are susceptible to iterative attacks in a white-box setting [@atscale-iclr-2017] (complete knowledge of the deployed model is available to the attacker), and to non-iterative and iterative attacks in a black-box setting ([@ensembleAT-iclr-2018; @dong2018boosting]).
Tramer [*et al*. ]{}[@ensembleAT-iclr-2018] demonstrated that models trained using a single-step adversarial training method converge to degenerative minima, and this causes gradient masking i.e., the linear approximation of loss becomes unreliable for generating adversarial samples using a non-iterative method. Non-iterative methods such as FGSM generate adversarial perturbations based on the first-order approximation of the loss function i.e., perturbation is in the direction of the sign of the gradient of the loss with respect to the input image. Further, it is implicitly assumed that the model’s loss increases for a large perturbation in this direction. This assumption is valid for normally trained models and is not valid for models trained using single-step adversarial training. Madry [*et al*. ]{}[@madry-iclr-2018] showed that it is possible to learn robust models, by including adversarial perturbations that maximize the loss while training, and further show that the maximization of loss can be achieved by generating adversarial samples using iterative methods. Unlike non-iterative methods, iterative methods perturb image slightly at each step and this prevents models from exhibiting gradient masking. Though iterative methods help to learn robust models, they are computationally expensive and cause training time to increase drastically. In this work, we propose three different types of regularizers that help to learn robust models using single-step adversarial training methods. The proposed regularizers harness the salient properties of models trained using iterative methods, such as loss surface smoothness and loss monotonicity, and incorporate these properties into models trained using single-step methods. Following are the major contributions of this work:
- We bring out the salient properties that differentiate a robust model from that of a pseudo robust model such as loss monotonicity.
- Harnessing on the above properties, we propose three different types of regularizers to learn robust models using single-step adversarial methods. The resultant models are robust against both non-iterative and iterative attacks, and achieve on par results when compared to models trained using computationally expensive multi-step adversarial training methods.
The paper is organised as follows: section \[sec:notations\] introduces the notation followed in this paper, section \[sec:related\_works\] discusses the related works, section \[sec:proposed\_approach\] presents the proposed approach, section \[sec:experiments\] hosts the experiments, and section \[sec:discussion\_conclusion\] concludes the paper.
Notations {#sec:notations}
=========
In this section we define the notations followed throughout this paper:
- $x:$ clean image from the dataset.
- $y_{true}:$ ground truth label corresponding to the image $x$.
- $f(\cdot):$ neural network that maps input image $x$ to the class probabilities.
- $g(\cdot):$ pre-softmax output of the neural network.
- $\theta:$ parameters of the neural network.
- $J:$ loss function e.g., cross-entropy loss.
- $\nabla_{x}J:$ gradient of the loss with respect to the input image $x$
- $m:$ mini-batch size.
- $\epsilon:$ perturbation size of the crafted noise.
- $x_{fgsm}:$ potential adversarial sample corresponding to the image $x$, generated using FGSM [@explainingharnessing-iclr-2015].
- $x_{ifgsm}:$ potential adversarial sample corresponding to the image $x$, generated using IFGSM [@physicalworld-arxiv-2016].
- $x_{rfgsm}:$ potential adversarial sample corresponding to the image $x$, generated using RFGSM [@ensembleAT-iclr-2018].
Related works {#sec:related_works}
=============
Following the findings of Szegedy [*et al*. ]{}[@intriguing-iclr-2014], various image specific (e.g. [@explainingharnessing-iclr-2015; @physicalworld-arxiv-2016; @robustness-arxiv-2016; @dong2018boosting]) and image agnostic (e.g. [@universal-cvpr-2017; @gduap-mopuri-2018]) attacks have been proposed. Various defense methods ([@defensivedistillation-arxiv-2015; @explainingharnessing-iclr-2015; @buckman2018thermometer; @ma2018characterizing; @guo2018countering; @s.2018stochastic; @xie2018mitigating; @song2018pixeldefend; @samangouei2018defensegan; @kannan2018adversarial; @S_2019_CVPR_Workshops]) have been proposed to defend against adversarial attacks. In this direction, adversarial training approach [@explainingharnessing-iclr-2015] shows promising results. Kurakin [*et al*. ]{}[@atscale-iclr-2017] observed that models trained using single-step adversarial training methods were susceptible to multi-step adversarial attacks in a white-box setting. Further, Tramer [*et al*. ]{}[@ensembleAT-iclr-2018] demonstrated that the pseudo robustness of these models is due to the gradient masking. Gradient masking causes the first-order approximation of the loss function to become unreliable for generating adversarial samples using non-iterative methods, and this results in the exclusion of useful adversarial samples during training.\
Madry [*et al*. ]{}[@madry-iclr-2018] demonstrated that it is possible to learn models that are robust to single-step and multi-step attacks, if perturbations crafted while training maximize the loss. An iterative method named Projected Gradient Descent (PGD) is used to generate such adversarial samples. Further, Zhang [*et al*. ]{}[@Zhang2019theoretically] proposed a regularizer for multi-step adversarial training to encourage the output of the classifier to be smooth. Other line of works such as [@wong2017provable; @raghunathan2018certified] provide defense certification for norm-bound attacks. However such methods are hard to scale for large datasets and attack perturbation sizes. Whereas in this work, we propose three different types of regularizes to learn robust models using single-step adversarial training methods. The proposed regularizers help to mitigate the effect of gradient masking during single-step adversarial training.
Adversarial Sample Generation Methods {#sub_sec:adversarial-sample-generation-methods}
-------------------------------------
In this subsection we explain methods for generating adversarial samples.\
**Fast Gradient Sign Method (FGSM) [@explainingharnessing-iclr-2015]**: Generates adversarial samples based on the first order approximation of the loss function, via performing simple gradient ascent.\
$$x^* = x + \epsilon. sign\big(\nabla_{x}J(f(x;\theta), y_{true})\big)$$ **Random + Fast Gradient Sign Method (RFGSM) [@ensembleAT-iclr-2018]**: This method adds small random noise before generating an adversarial sample using the FGSM method. $$\begin{aligned}
x' &=& x + \alpha . sign\big(\mathcal{N}(0^d,I^d)) \\
x^* &=& x' + (\epsilon-\alpha). sign\big(\nabla_{x'}J(f(x';\theta), y_{true})\big)\end{aligned}$$ Where, $\mathcal{N}$ represents normal distribution\
\
**Iterative Fast Gradient Sign Method (IFGSM) [@physicalworld-arxiv-2016]**: In this method, FGSM is applied in an iterative fashion with a small step size ($\alpha$). In our experiments we use $\alpha=\epsilon/steps$. $$\begin{aligned}
x^{0} & = & x\\
x^{N+1} & = & x^{N} + \alpha . sign\big(\nabla_{x^{N}}J(f(x^{N};\theta), y_{true})\big)\end{aligned}$$ **Projected Gradient Descent (PGD) [@madry-iclr-2018]**: Here the perturbation is initialized with a random point within the allowed $l_\infty$ norm ball and then IFGSM is applied with re-projection. $$\begin{aligned}
x^{0} & = & x + \mathcal{U}\big(-\epsilon_{step},\epsilon_{step},shape(x)\big) \\
x^{N+1} & = & x^{N} + \epsilon_{step} . sign\big(\nabla_{x^{N}}J(f(x^{N};\theta), y_{true})\big) \\
x^{N+1} & = & clip\big(x^{N+1},min=x-\epsilon,max=x+\epsilon\big) \end{aligned}$$ Where, $\mathcal{U}$ represents uniform distribution.\
\
**Projected Gradient Descent with CW loss (PGD-CW)**: Variant of PGD attack which uses C&W [@robustness-arxiv-2016] loss instead of cross-entropy loss.\
\
**Momentum Iterative Fast Gradient Sign Method (MI-FGSM) [@dong2018boosting]**: Introduces momentum into the IFGSM formulation. Here, $\mu$ represents the momentum term and $\alpha$ is set to $\epsilon/steps$. $$\begin{aligned}
x^{0} & = & x,~~~p^{0}=0\\
p^{N+1} &=& \mu.p^{N} + \frac{\nabla_{x^{N}}J(f(x^{N};\theta), y_{true})}{||\nabla_{x^{N}}J(f(x^{N};\theta), y_{true})||_1}\\
x^{N+1} & = & x^{N} + \alpha . sign\big(p^{N+1}\big)\end{aligned}$$\
**DeepFool**: An iterative method proposed by [@deepfool-cvpr-2016]. The method generates an adversarial perturbation based on the linear approximation of the model, that would cause the sample to cross the decision boundary.\
\
**Carlini and Wagner (C&W)**: An iterative method proposed by [@robustness-arxiv-2016], aims at generating perturbation with a minimum $l_2$ norm that is sufficient to fool the model i.e., the optimization objective is to find an adversarial perturbation with a minimum $l_2$ norm.
Adversarial training
--------------------
**FGSM Adversarial Training (FGSM-AT) [@atscale-iclr-2017]:** During FGSM adversarial training, a portion (typically 50%) of clean samples in the mini-batch are replaced with their corresponding FGSM adversarial samples. This method is fast and simple, but the resultant models are not robust to multi-step attacks.\
**Ensemble Adversarial Training (EAT) [@ensembleAT-iclr-2018]:** During training, FGSM adversarial samples are generated by the model being trained or by one of the models from a fixed set of normally trained models, chosen at random. Models trained using this method show improvement against adversarial attacks in a black-box setting. Further, models are still susceptible to multi-step attacks in a white-box setting.\
**PGD Adversarial Training (PGD-AT) [@madry-iclr-2018]:** During training, all the clean samples in the mini-batch are replaced with their corresponding PGD adversarial samples.\
**TRADES [@Zhang2019theoretically]:** During training, an augmented mini-batches containing clean samples and their corresponding adversarial samples are created. These adversarial samples are generated using PGD method with a surrogate loss instead of cross-entropy loss. Models trained using TRADES and PGD-AT are robust against both single-step and multi-step attacks. However, training time is significantly large when compared to single-step adversarial training methods i.e., FGSM-AT and EAT.
Loss surface {#subsection:loss_surface}
------------
In this work, we obtained the plot of loss surface [@ensembleAT-iclr-2018] around the data points to illustrate the effect of gradient masking. Loss surface is obtained by varying the input to the model using Eq.(\[equ:loss\_surface\_eq\]). $$\begin{aligned}
\label{equ:loss_surface_eq}
x^* &=& x + \epsilon_1 . \delta_1 + \epsilon_2 . \delta_2 \end{aligned}$$ Where, $\delta_1$ is the sign of the gradient of loss with respect to the input sample, and $\delta_2$ is the sign of the random noise sampled from Normal distribution ($\mathcal{N}$). $\epsilon_1$ and $\epsilon_2$ represent the perturbation size. Loss surface is a 3D plot, where x and y axes represent the perturbation size, and z-axis represents the loss.
Proposed Approach {#sec:proposed_approach}
=================
In this section, we explain the criteria for learning robust models using the adversarial training method [@madry-iclr-2018]. We show that this criterion is not satisfied during the single-step adversarial training. Further, we explain the salient properties that differentiate a robust model from that of a pseudo robust model. Harnessing on these properties, we propose three different types of regularizes that help to learn robust models using single-step adversarial training methods.
Criteria for learning robust models {#subsec:criteria_adv}
-----------------------------------
Madry [*et al*. ]{}[@madry-iclr-2018] demonstrated that it is possible to learn robust models using the adversarial training method, if perturbations crafted while training maximize the loss. This objective can be formulated as a mini-max optimization problem (Eq. \[equ:pgd\_formaulation\]). $$\begin{aligned}
\label{equ:pgd_formaulation}\displaystyle{\min_{\theta} \Bigg[E_{(x,y)\in D}\bigg[\displaystyle{\max_{\delta\in S}}~J\big(f(x+\delta;\theta),y_{true}\big)\bigg]\Bigg]}\end{aligned}$$ Where, $D$ is the training dataset, and $S$ is the feasible set $S=\{\delta: ||\delta||_\infty\le\epsilon\}$. At each iteration, we need to find an adversarial perturbation ($||\delta||_\infty\le\epsilon$) that maximizes the model’s loss, and further we need to update the model’s parameters ($\theta$) so as to minimize the loss on adversarial samples. Madry [*et al*. ]{}[@madry-iclr-2018] solves the inner maximization problem by generating adversarial samples using the Projected Gradient Descent method (iterative method). Single-step adversarial training is a special case of mini-max optimization problem (Eq. \[equ:pgd\_formaulation\]), where the inner maximization is assumed to be achieved by adversarial samples generated by single-step methods. Iterative methods such as PGD, ensure that the generated perturbations will increase the model’s loss, since at each step of the generation process, perturbation with small $\epsilon$ is added to the image. The increase in the loss is not guaranteed when perturbation with high $\epsilon$ is added to the image in a single step. In the next subsection, we show that during the initial stages of single-step adversarial training, the extent of loss maximization achieved by the adversarial samples generated using single-step and multi-step methods are similar. Further, we show that in the later stages of training, single-step adversaries are not able to maximize the loss due to gradient masking effect.
Gradient masking effect {#subsection:effect_of_gradient_masking}
-----------------------
In this subsection, we empirically show that the extent of maximization of loss achieved by FGSM adversaries during the initial stages of single-step adversarial training, is similar to that achieved by PGD (iterative method) adversaries. Further, we show that as training progress, the ability of FGSM samples to maximize the loss diminishes. We train WideResNet-28-10 on CIFAR-10 dataset using FGSM adversarial training method. We obtain the plot of cross-entropy loss versus perturbation size ($\epsilon$) of FGSM and PGD attacks, during the initial (at iteration 40 ($\times$100)) and final stages (at iteration 600 ($\times$100)) of training. Bottom-left and bottom-right plots of Fig. \[figure:fgsm\_l2\_distance\] shows the obtained plots. It can be observed that during the initial stage of training, the difference between the average loss on FGSM and PGD adversaries is small (see bottom-left plot of Fig. \[figure:fgsm\_l2\_distance\] for $\epsilon$=8). This implies that the extent of loss maximization achieved by FGSM samples is similar to that achieved by PGD samples. Whereas during the later stage of training, the difference between the average loss on FGSM and PGD adversaries is large (see bottom-right plot of Fig. \[figure:fgsm\_l2\_distance\] for $\epsilon$=8) i.e., the generated FGSM samples are not able to maximize the training loss. This large difference is due to the gradient masking. During single-step adversarial training, when the model starts to mask the gradient, its decision surface exhibits a sharp curvature near the data points [@ensembleAT-iclr-2018]. This sharp curvature obfuscates the adversarial direction. Single-step adversarial sample generation methods such as FGSM generate adversarial samples based on the linear approximation of the loss function, and gradient masking causes the linear approximation to become unreliable for generating adversarial samples. To illustrate the gradient masking effect, we obtain the loss surface plots. Fig. \[figure:fgsm\_decision\_surface\] shows the loss surface plots obtained during the initial and final stages of FGSM adversarial training. From the left plot of Fig. \[figure:fgsm\_decision\_surface\], it can be observed that there is no sharp curvature in the loss surface plot of the model obtained during the initial stage of training. Whereas, a sharp curvature can be observed in the loss surface plot of the model obtained during the later stage of training, and this curvature artifact obfuscates the adversarial direction.
Unlike FGSM adversarial training, during PGD adversarial training, the difference in the average loss on the FGSM and PGD samples is small during the initial and final stages of training (see bottom-left and bottom-right plot of Fig. \[figure:pgd\_l2\_distance\]). Further, from Fig. \[figure:pgd\_decision\_surface\] it can be observed that there is no sharp curvature in the loss surface plots obtained during the initial and final stages of training. Note that, iterative methods such as PGD generate adversarial samples by adding a small perturbation to the image at every step, and this ensures that the added perturbation will increase the loss.
Salient properties of robust models {#subsec:salient_properties}
-----------------------------------
In this subsection, we bring out the salient properties that differentiate a robust model from that of a pseudo robust model. We train WideResNet-28-10 on CIFAR-10 dataset using FGSM-AT and PGD-AT methods. During training, we obtain the average Euclidean distance between (i) [pre-softmax output]{}of FGSM and IFGSM adversaries, and (ii) [pre-softmax output]{}of FGSM and RFGSM adversaries. After training, we obtain the plot of the average cross-entropy loss versus perturbation size of FGSM attack. Following are the salient properties observed in robust models:\
\
**(i) Loss increases monotonically with the increase in perturbation size:** Column-2 of Fig. \[figure:main\_explaination\_monotonicity\] shows the plot of the average cross-entropy loss versus perturbation size of FGSM attack obtained for the model trained using the PGD-AT method. It can be observed that the average loss increases monotonically with the increase in perturbation size. From column-1 of Fig. \[figure:main\_explaination\_monotonicity\] it can be observed that for the model trained using the FGSM-AT method, the average loss does not increase monotonically with the increase in perturbation size.\
**(ii) Similar [pre-softmax output]{}for adversarial samples generated using different methods:** The top plot of Fig. \[figure:pgd\_l2\_distance\] shows the average Euclidean distance between [pre-softmax output]{}of adversarial samples generated using different methods, obtained during PGD-AT. It can be observed that for the entire training duration, the average Euclidean distance is relatively small. Whereas during FGSM-AT, these distances are initially small and become relatively large after a few iterations (Fig. \[figure:fgsm\_l2\_distance\], top). Note that, during the initial stage of FGSM-AT, the average Euclidean distance between [pre-softmax output]{}of (i) FGSM and IFGSM samples, and (ii) FGSM and RFGSM samples, are small. Further, these Euclidean distances start to increase when the model starts to mask the gradients (Fig. \[figure:fgsm\_l2\_distance\], top).
Proposed single-step adversarial training with regularization term
------------------------------------------------------------------
In the previous subsection, we have shown the salient properties that differentiate a robust model from a pseudo robust model. Harnessing on these observations, we propose three different types of regularizers which help to learn robust models using single-step adversarial training methods. The proposed regularizers penalize the model for masking gradients, and this enables the inclusion of useful single-step adversarial samples during the entire training process.
### Single-step Adversarial Training with Regularizer-1 (SAT-R1) {#subsubsection:sat_r1}
In subsection \[subsec:salient\_properties\], we observed that during single-step adversarial training, the average Euclidean distance between the [pre-softmax output]{}of FGSM and IFGSM adversaries increase drastically. Whereas during PGD-AT, this distance is relatively small and does not increase as training proceeds. Based on these observations, we include a penalty term in the training loss (Eq. \[equ:proposed\_penalty\_r1\]) to minimize the distance between [pre-softmax output]{}of FGSM and IFGSM adversaries of clean samples during single-step adversarial training. $$\begin{split}
\mathcal{L}oss=& \frac{1}{m}\sum_{i=1}^{m}J(f(x^i_{fgsm};\theta),y^i_{true}) \\&+ \lambda\frac{1}{k}\sum_{j=1}^{k}\big\|{g(x^j_{fgsm}) - g(x^j_{ifgsm})}\big\|_{2}^{2} \end{split}
\label{equ:proposed_penalty_r1}$$ In Eq. (\[equ:proposed\_penalty\_r1\]), the first term corresponds to the classification task, and the second term represents the proposed regularization. Further, $\lambda$ represents the regularization weighting factor and $k$ represents the number of adversaries generated using IFGSM. During training, when the model starts to mask the gradients, the proposed penalty term causes training loss to increase (since the distance between [pre-softmax output]{}of FGSM and IFGSM adversarial pair increases). This behavior of the proposed penalty term helps in mitigating the effect of gradient masking, and thus enables the generation of stronger adversaries while training. In section \[sec:experiments\] we show that $k$=1 (i.e., penalty is imposed on one FGSM and IFGSM adversarial pair of a clean sample in the mini-batch) is sufficient to learn robust models. This means that only *one* adversarial sample in the mini-batch is generated using an iterative method and the remaining adversarial samples are generated using non-iterative method. Further, we show that adversarial training with mini-batches containing $one$ IFGSM and $m$ FGSM samples without the proposed regularizer, does not improve the model’s robustness significantly. The result of this ablation experiment is shown in section \[sec:experiments\].
### Single-step Adversarial Training with Regularizer-2 (SAT-R2) {#subsubsection:sat_r2}
In section \[subsec:salient\_properties\], we showed that when the model starts to mask gradients, then the Euclidean distance between [pre-softmax output]{}of FGSM and RFGSM adversaries of a clean sample becomes large. Based on this observation, we introduce a regularization term in the training loss (Eq.\[equ:proposed\_penalty\_r2\]) that penalizes the effect of gradient masking during single-step adversarial training. $$\begin{split}
\mathcal{L}oss=& \frac{1}{m}\sum_{i=1}^{m}J(f(x^i_{fgsm};\theta),y^i_{true}) \\&+ \lambda\frac{1}{m}\sum_{j=1}^{m}\big\|{g(x^j_{fgsm}) - g(x^j_{rfgsm})}\big\|_{2}^{2}
\end{split}
\label{equ:proposed_penalty_r2}$$ In Eq.(\[equ:proposed\_penalty\_r2\]), the first term corresponds to the classification loss, and the second term represents the proposed regularization. During training, if the model starts to mask gradients, then the Euclidean distance between [pre-softmax output]{}of FGSM and RFGSM adversaries of clean samples increases, and this in turn causes the training loss (Eq. \[equ:proposed\_penalty\_r2\]) to increase. This behavior of the proposed regularizer prevents the model from masking gradients. Note that, adversarial training with RFGSM or with both RFGSM and FGSM samples does not improve the model’s robustness significantly. The results of these experiments are shown in section \[sec:experiments\].
### Single-step Adversarial Training with Regularizer-3 (SAT-R3) {#subsubsection:sat_r3}
In subsection \[subsec:salient\_properties\], we demonstrated that for a robust model, loss on the FGSM adversarial samples increases monotonically with the increase in perturbation size, and this behavior is not observed in the model trained using the single-step adversarial training method. Based on this observation, we propose a regularisation term which enforces the model’s loss to increase monotonically with the increase in perturbation size. Eq. (\[equ:proposed\_penalty\_r3\]) represents the training loss. $$\mathcal{L}oss = loss_{\epsilon_{High}} + \lambda.max\big(0, loss_{\epsilon_{Low}} - \tau.{loss_{\epsilon_{High}}}\big)
\label{equ:proposed_penalty_r3}$$ Where, $loss_{\epsilon_{Low}}$ and $loss_{\epsilon_{High}}$ represent the average loss on FGSM adversarial samples with perturbation size of $\epsilon_{Low}$ and $\epsilon_{High}$ respectively. The first term corresponds to the classification task, and the second term represents the monotonic loss constraint. During training, we ensure $loss_{\epsilon_{Low}} < loss_{\epsilon_{High}}$ by enforcing ($loss_{\epsilon_{Low}}/loss_{\epsilon_{High}}) < \tau$, where $\tau<1$. For stability purpose, we consider ($loss_{\epsilon_{Low}} - \tau.loss_{\epsilon_{High}}) < 0$. During training, if the model starts to mask gradients, the second term in Eq. (\[equ:proposed\_penalty\_r3\]) becomes greater than zero, and this causes training loss to increase. This behavior of the proposed loss constraint, explicitly prevents the model from generating weaker adversaries during adversarial training. Also, note that the monotonicity of loss is enforced for the allowed perturbation range \[0, $\epsilon_{High}$\] i.e., the maximum value of $\epsilon_{High}$ is restricted based on perceptual constraints.
Experiments {#sec:experiments}
===========
In this section, we show the performance of models trained using the proposed training methods against adversarial attacks in white-box and black-box settings. We perform sanity tests described in [@obfuscated-gradients; @carlini2019evaluating] to ensure that models trained using the proposed regularizers are robust, and do not exhibit obfuscated gradients. Since, models exhibiting gradient masking or obfuscated gradients are not robust against adversarial attacks [@obfuscated-gradients]. Code for the proposed approach is available at <https://github.com/val-iisc/SAT-Rx>.\
\
**Dataset**: We show results on MNIST [@lecun1998mnist], CIFAR-10 [@cifar_10_dataset] and ImageNet-subset (100 classes) [@imagenet-ijcv-2015] datasets. For ImageNet-subset, we randomly choose 100 classes. We use LeNet+ (refer to table \[table:eat\_mnsit\_architecure\]), WideResNet-28-10 (WRN-28-10) [@BMVC2016_87] and ResNet-18 [@resnet-2015] for MNIST, CIFAR-10 and ImageNet-subset datasets respectively. Images are pre-processed to be in \[0,1\] range. For data-augmentation, horizontal flip and random crop are performed for CIFAR-10 and ImageNet-subset datasets.\
**Training methods:** We compare the proposed training methods with Normal training (NT), FGSM Adversarial Training (FGSM-AT) [@atscale-iclr-2017], Ensemble Adversarial Training (EAT) [@ensembleAT-iclr-2018], PGD Adversarial Training (PGD-AT) [@madry-iclr-2018], and TRADES [@Zhang2019theoretically]. Refer table \[table:ensemble\_setup\] for details on the experimental setup used for EAT.\
**Attacks**: We show the performance of models trained using different training methods against $l_\infty$ and $l_2$ attacks. For $l_\infty$ norm-bounded attacks, we use FGSM [@explainingharnessing-iclr-2015], IFGSM [@physicalworld-arxiv-2016], MI-FGSM [@dong2018boosting], PGD [@madry-iclr-2018] and PGD-CW attacks. For $l_2$ attacks, we use DeepFool [@deepfool-cvpr-2016] and C&W [@robustness-arxiv-2016] attacks. We follow Madry [*et al*. ]{}[@madry-iclr-2018] for attack parameters. For $l_\infty$ attacks, we limit $\epsilon$ to 0.3, 8/255 and 8/255 for MNIST, CIFAR-10, and ImageNet-subset datasets respectively. Note that, DeepFool and C&W attacks measure the robustness of the model based on the $l_2$ norm of the generated adversarial perturbation.\
**Hyper-parameters**: For SAT-R1, we set ($\lambda$, $k$) to (0.2, 1), (0.2, 1) and (0.05, 1) for MNIST, CIFAR-10 and ImageNet-subset datasets respectively. For SAT-R2, we set $\lambda$ to 5, 25 and 3 for MNIST, CIFAR-10 and ImageNet-subset datasets respectively. For SAT-R3, we set ($\lambda$, $\tau$) to (1, 0.4), (1, 0.6) and (1, 0.6) for MNIST, CIFAR-10 and ImageNet-subset datasets respectively.
Performance against $l_\infty$ attacks {#subsec:l_infty_attacks}
--------------------------------------
We train models on MNIST, CIFAR-10, and ImageNet-subset datasets using the proposed single-step adversarial training methods SAT-R1, SAT-R2, and SAT-R3. Further, we also train models using NT, FGSM-AT, EAT, PGD-AT, and TRADES methods. Models are trained for 20, 100 and 100 epochs on MNIST, CIFAR-10, and ImageNet-subset datasets respectively. Table \[table:mnist\_linfty\_performance\], \[table:cifar10\_linfty\_performance\] and \[table:imgenet\_linfty\_performance\] shows the performance of these models against single-step and multi-step attacks in white-box and black-box settings. For black-box attacks, a normally trained model is used for generating adversarial samples, and these generated adversarial samples are tested on the target model. Typically, the model used for generating adversarial samples is referred to as a “source model" or “substitute model".\
**White-box setting:** From tables \[table:mnist\_linfty\_performance\], \[table:cifar10\_linfty\_performance\] and \[table:imgenet\_linfty\_performance\], it can be observed that models trained using single-step adversarial training methods (i.e., FGSM-AT, EAT) are susceptible to multi-step attacks. Whereas models trained using PGD-AT, TRADES, SAT-R1, SAT-R2, and SAT-R3 are robust against both single-step and multi-step attacks. Note that, PGD-AT and TRADES use iterative methods for the generation of adversarial samples, due to which training time is significantly high. Unlike PGD-AT and TRADES, the proposed methods SAT-R1, SAT-R2, and SAT-R3 use non-iterative method for crafting adversarial samples. Further, we obtain the plot of test-set recognition accuracy of models trained using the proposed regularizers for PGD attack with increasing steps. Fig. \[figure:acc\_vs\_pgdsteps\] shows the obtained plot, it can observed that the model’s accuracy saturates with increase in steps of PGD attack. We obtain this plot to verify that the model’s performance does not degrade significantly with the increase in the number of iteration/steps of adversarial attack [@engstrom2018evaluating].\
**Black-box setting:** The last four columns of table \[table:mnist\_linfty\_performance\], \[table:cifar10\_linfty\_performance\] and \[table:imgenet\_linfty\_performance\] shows the performance of models in a black-box setting. It can be observed that the performance of models trained using PGD-AT, TRADES, SAT-R1, SAT-R2, and SAT-R3, in a black-box setting is better than that in a white-box setting. Note that, the model trained on CIFAR-10 dataset using FGSM-AT is more susceptible to adversarial attack in a black-box setting than in a white-box setting.
Performance against $l_2$ attacks {#subsec:l_2_attacks}
---------------------------------
DeepFool and C&W attacks belongs to a class of attacks that generate adversarial perturbations without norm constraints. These attacks aim to generate adversarial perturbations with a minimum $l_2$ norm, that is just sufficient to fool the classifier. The average $l_2$ norm of the generated perturbations indicates the robustness of the model. For an undefended classifier, perturbations with a small $l_2$ norm is sufficient to fool the classifier. Whereas, for a robust classifier, perturbations with relatively large $l_2$ norm, are required to fool the classifier. Table \[table:deepfool\_cw\] shows the performance of models against DeepFool and C&W attacks. Fooling Rate (FR) represents the percentage of test-set samples that are misclassified. It can be observed that for models trained using PGD-AT, SAT-R1, SAT-R2, and SAT-R3, the average $l_2$ norm of the generated perturbations is relatively high.
Ablation study {#subsec:ablation_study}
--------------
We perform ablation study to show the significance of the proposed regularizers. For the ablation study, we train LeNet+ on MNIST dataset.
### Ablation study on SAT-R1 {#subsubsec:ablation_sat_r1}
**Ablation-R1-1:** SAT-R1 with $\lambda$=0. We train LeNet+ on MNIST dataset using SAT-R1 with $\lambda$=0, and cross-entropy loss imposed on both IFGSM and FGSM samples. We perform this experiment to show that the gain in the robustness of the model is due to the proposed regularizer, and not due to the inclusion of one IFGSM sample. Table \[table:ablations\] shows the performance of the model trained using this method, and it can be observed that the model is not robust to multi-step attacks.\
**Ablation-R1-2:** SAT-R1 with mini-batch containing $k$ IFGSM and $m$ (mini-batch size) FGSM samples. We train LeNet+ using SAT-R1 with different values of $k$, to show that it is sufficient to impose the proposed regularizer on a single pair of FGSM and IFGSM sample in a mini-batch. Column-1 of Fig. \[figure:ablation\_r1\] shows the plot of accuracy of the model on the PGD validation set, trained using SAT-R1 with different values of $k$. It can be observed that $k$=1 is sufficient to learn robust models using SAT-R1.\
**Ablation-R1-3:** Adversarial training with mini-batches containing $k$ IFGSM and $m$ FGSM samples without the proposed regularizer. In this experiment, cross-entropy loss is imposed on FGSM and IFGSM adversarial samples. We train LeNet+ on MNIST dataset and during training, we generate $k$ (expressed in terms of original mini-batch size $m$) IFGSM and $m$ FGSM samples. We perform this experiment to show that without the proposed regularizer, at least 40% of the samples should be generated using IFGSM method, so as to learn robust models. Column-2 of Fig. \[figure:ablation\_r1\] shows the plot of accuracy of the model on the PGD validation set, for different values of $k$. It can be observed that for $k$= 40%, there is a significant improvement in the model’s robustness. This implies that at least 40% of the samples should be generated using the IFGSM method so as to learn robust models without the proposed regularizer.
### Ablation study on SAT-R2 {#subsubsec:ablation_sat_r2}
**Ablation-R2-1:** SAT-R2 with $\lambda$=0. We train LeNet+ on MNIST dataset using SAT-R2 with $\lambda$=0, and cross-entropy loss is imposed on both FGSM and RFGSM samples. We perform this experiment to show that the gain in the robustness of models trained using SAT-R2 is not due to the inclusion of RFGSM samples.\
**Ablation-R2-2:** Adversarial training with RFGSM samples. We train LeNet+ on MNIST dataset using the adversarial training method, and during training, adversarial samples are generated using RFGSM. We perform this experiment to show that adversarial training with RFGSM samples does not improve the model’s robustness significantly.
### Ablation study on SAT-R3 {#subsubsec:ablation_sat_r3}
**Ablation-R3-1:** SAT-R3 with $\lambda$=0. We train LeNet+ on MNIST dataset using SAR-R3 with $\lambda$=0. We perform this experiment to show that adversarial training with only FGSM samples does not improve the model’s robustness.\
**Ablation-R3-2:** SAT-R3 with $\lambda$=0 and cross-entropy is imposed on FGSM samples with the perturbation size of $\epsilon_{Low}$ and $\epsilon_{High}$. We train LeNet+ on MNIST dataset using SAT-R3 with $\lambda$=0, and cross-entropy imposed on FGSM samples with the perturbation size of $\epsilon_{Low}$ and $\epsilon_{High}$. We perform this experiment to show that adversarial training with mini-batches containing FGSM samples with the perturbation size of $\epsilon_{Low}$ and $\epsilon_{High}$, does not improve the model’s robustness.
Table \[table:ablations\] shows the performance of models trained using the above training methods, and it can be observed that there is no significant improvement in the model’s robustness against multi-step attacks.
Sanity tests {#subsec:obfuscated_gradeint_detection}
------------
We perform sanity test described in Carlini [*et al*. ]{}[@carlini2019evaluating] to verify the robustness of models trained using the proposed method, and to rule out obfuscated gradients. Athalye [*et al*. ]{}[@obfuscated-gradients] showed that certain defense methods unintentionally or intentionally cause models to exhibit obfuscated gradients. Further, a method to break such defense methods was proposed. This implies that models exhibiting obfuscated or masked gradients are not robust. We perform the following sanity tests to verify the robustness of models trained using the proposed regularizers:\
**(i) Verify multi-step attacks perform better than single-step attacks**: For robust models, iterative attacks should be stronger than non-iterative attacks in a white-box setting. Table \[table:mnist\_linfty\_performance\], \[table:cifar10\_linfty\_performance\] and \[table:imgenet\_linfty\_performance\], shows the performance of models against single-step and multi-step attacks, and it can be observed that iterative attacks (IFGSM and PGD) are stronger than non-iterative attack (FGSM) for models trained using SAT-R1, SAT-R2, and SAT-R3 methods.\
**(ii) Verify white-box attacks perform better than black-box attacks**: For any model, white-box attacks should be stronger than black-box attacks. In black-box setting, partial or no knowledge of the deployed model is available to the attacker, and hence black-box attacks should be weaker than white-box attacks. From table \[table:mnist\_linfty\_performance\], \[table:cifar10\_linfty\_performance\] and \[table:imgenet\_linfty\_performance\], it can be observed that white-box attacks are stronger than black-box attacks on models trained using SAT-R1, SAT-R2, and SAT-R3 methods.\
**(iii) Verify for large perturbation size, model’s accuracy reach levels of random guessing**: Typically, the model’s performance degrades drastically for attacks with a large perturbation size ($\epsilon$). We obtain the plot of recognition accuracy (%) of the model on PGD test set for different values of attack perturbation size. Fig. \[figure:proposed\_acc\_vs\_eps\] shows the plot obtained for models trained using the proposed regularizers, it can be observed that the recognition accuracy (%) of the model is zero for PGD attack with large perturbation size.\
**(iv) Verify increase in the perturbation size strictly increases attack success rate**: From Fig. \[figure:proposed\_acc\_vs\_eps\], it can be observed that the PGD attack success rate (success rate = 100 - accuracy) increases with the increase in the distortion bound i.e., perturbation size of PGD attack.
Loss trend {#subsec:loss_vs_eps_plots}
----------
In this subsection, we obtain the plot of average loss on the test set versus perturbation size of FGSM and PGD attacks. Fig. \[figure:proposed\_loss\_vs\_fgsm\_pgd\_eps\] shows the plots obtained for models trained using the proposed regularizers. It can be observed that the average loss increases monotonically with the increase in perturbation size of FGSM and PGD attacks. Further, it can be observed that the difference between the loss on FGSM and PGD samples is relatively small, even for higher perturbation size ($\epsilon$). Whereas, this difference would be large for models exhibiting masked gradients.
Loss surface {#subsec:decision_surface_plots}
------------
In this subsection, we obtain the loss surface plots for models trained using the proposed regularizers. Fig. \[figure:proposed\_loss\_vs\_fgsm\_pgd\_eps\] shows the obtained plots. It can be observed that there is no sharp curvature near the decision points. Whereas, sharp curvature near data points can be observed for models exhibiting masked gradients (see column-2 plot of Fig. \[figure:fgsm\_decision\_surface\]).
Complexity {#subsec:complexity}
----------
In this subsection, we discuss the complexity of different adversarial training methods. The adversarial sample generation is the major bottleneck for adversarial training methods. Typically, adversarial sample generation involves computation of gradient of the loss with respect to the input image. This requires one or more forward and backward propagation through the network. We define a metric $FBP$, which corresponds to one Forward and Backward propagation through the network. Further, we use $FBP$ to express the complexity of the adversarial sample generation process. Table \[table:training\_complexity\] summaries the complexity involved in generating adversarial samples during different adversarial training methods. For FGSM-AT, which uses single-step method for generating adversaries, requires only one forward and backward propagation through the network to generate adversaries. For EAT, a fixed set of pre-trained source models are required along with the model being trained for generating FGSM adversarial samples. Therefore, additional training of source models is involved. PGD-AT and TRADES methods, which use iterative methods for generating adversaries, require multiple forward and backward propagation through the network, and this causes a significant increase in the training time. The proposed training methods use single-step methods for generating adversaries, and this requires only one forward and backward propagation through the network, except for SAT-R1 where one sample in the mini-batch is generated using an iterative method. Among the proposed methods, complexity of SAT-R1$>$SAT-R2$>$SAT-R3. SAT-R1 requires one IFGSM and ‘m’ FGSM samples. The generation of one adversarial sample using iterative method is the bottleneck for SAT-R1. SAT-R2 requires generation of ‘m’ FGSM and ‘m’ R-FGSM samples. Whereas, SAT-R3 requires ‘m’ FGSM samples with perturbation size of $\epsilon_{Low}$ and ‘m’ FGSM samples with perturbation size of $\epsilon_{High}$. We use computational trick to reduce the time required for generating these adversarial samples i.e., sign of the gradient of loss required for generating adversarial samples with perturbation size of $\epsilon_{High}$ and $\epsilon_{Low}$ is same. Therefore computational complexity of SAT-R3 is lesser than SAT-R2.
Comparison of SAT-R1, SAT-R2 and SAT-R3 {#subsec:comparison}
---------------------------------------
In terms of computational complexity, SAT-R1$>$SAT-R2$>$SAT-R3. The performance of models trained using SAT-R1 and SAT-R2 are similar. Whereas the performance of models trained using SAT-R3 is relatively less superior than compared to models trained using SAT-R1 and SAT-R2. Among the proposed regularizers, SAT-R1 and SAT-R2 are suitable when performance is of importance, and SAT-R3 is suitable for cases where training time is the bottleneck e.g., training on large datasets.
Further, we demonstrate that it is possible to boost the performance of models trained using SAT-R3, by either pre-training or fine-tuning these models using SAT-R1 or SAT-R2 for few epochs. Table \[table:sequential\_training\] shows the performance of LeNet+ trained on MNIST dataset using this approach. It can be observed that there is an improvement in the model’s robustness against multi-step attacks (PGD-40 and PGD-100) when compared to the model trained using SAT-R3 only.
Discussion and conclusion {#sec:discussion_conclusion}
=========================
Adversarial training, a straightforward solution to defend model against adversarial attacks, shows promising results. However, models trained using the existing single-step adversarial training converge to degenerative minima where the model appears to be (pseudo) robust. Though multi-step adversarial training methods such as PGD adversarial training and TRADES methods help to learn adversarially robust models, they are computationally expensive.
In this work, we have proposed three different types of regularizers for single-step adversarial training. The proposed regularizers harness the salient properties of robust models to mitigate gradient masking effect, and help to learn robust models using single-step methods in a computationally efficient manner. Unlike models trained using the existing single-step adversarial training methods, models trained using the proposed methods are robust against both single-step and multi-step attacks.
[B.S. Vivek]{} is a research student at Video Analytics Lab, CDS, Indian Institute of Science, Bangalore. He received his B.E. from M.S. Ramaiah Institute of Technology, Bangalore. His research interest includes computer vision and machine learning.
[\
\
R. Venkatesh Babu]{} received his Ph.D from Dept. of Electrical Engineering, IISc, Bangalore. Thereafter, he held postdoctoral positions at NTNU, Norway and IRISA/INRIA, France. Subsequently, he worked as a research fellow at NTU, Singapore. He is currently an Associate Professor at Dept. of CDS and convener of VAL, IISc. His interests span vision, image/video processing, ML, and multimedia.
[^1]: The authors are with the Video Analytics Lab, Department of Computational and Data Sciences, Indian Institute of Science, Bangalore, India.\
E-mail: [email protected] and [email protected]\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Functional limit theorems are presented for the rescaled occupation time fluctuation process of a critical finite variance branching particle system in $\mathbb{R}^{d}$ with symmetric $\alpha$-stable motion starting off from either a standard Poisson random field or from the equilibrium distribution for intermediate dimensions $\alpha<d<2\alpha$. The limit processes are determined by sub-fractional and fractional Brownian motions, respectively.'
author:
- |
Piotr Miłoś\
Institute of Mathematics\
Polish Academy of Sciences\
Warsaw
title: Occupation time fluctuations of Poisson and equilibrium finite variance branching systems
---
AMS subject classification: primary 60F17, 60G20, secondary 60G15\
Key words: Functional central limit theorem; Occupation time fluctuations; Branching particles systems; Fractional Brownian motion; Sub-fractional Brownian motion; equilibrium distribution.
Introduction
============
Consider a system of particles in $\mathbb{R}^{d}$ starting off at time $t=0$ from a certain distribution (a standard Poisson and equilibrium fields are investigated in this paper). They evolve independently, moving according to a symmetric $\alpha$-stable Lévy process and undergoing finite variance branching at rate $V$ ($V>0$). We obtain functional limit theorems for the rescaled occupation time fluctuations of this system when $\alpha<d<2\alpha$. This is an extension of [@BGT2 Theorem 2] where the starting distribution is a Poisson field and the branching law is critical and binary.
Branching law\[sub:Branching-law\]
----------------------------------
\
In the [@BGT1; @BGT2; @BGT3] the law of branching is critical and binary. In this paper an extended model is investigated. The particles branch according to the law given by a moment generating function $F$. $F$ fulfills two requirements:
1. $F'(1)=1$, which means that the law is critical (the expected number of particles spawning from one particle is $1$),
2. $F''(1)<+\infty$, which states that the second moment exists.
(Note here that the branching law in [@BGT2] is given by $F(s)=\frac{1}{2}\left(1+s^{2}\right)$ and obviously fulfills the two requirements.) Although constraints imposed on $F$ are not very restrictive and quite natural (so that the class of the branching laws satisfying them is broad) still there remain other interesting cases to be investigated. One of them is the class of branching laws in the domain of attraction of the $\left(1+\beta\right)-$stable law (i.e., the moment generating function is $F\left(s\right)=s+\frac{1}{2}\left(1+s\right)^{1+\beta}$), the case studied in [@BGT4; @BGT5]. A remarkable feature of the latter case is that the limit processes are stable ones and not Gaussian as it occurs in the finite variance case.
Equilibrium distribution
------------------------
\
Another concept naturally related to particle systems is an equilibrium distribution. It has been shown that in certain circumstances the system converges to the equilibrium distribution [@GW2]. It is both an interesting and important question whether the theorems shown by Bojdecki et al still hold in the case when the equilibrium state is taken as the initial condition. A conjecture in [@BGT1] states that the temporal structure of the limit is given by fractional Brownian motion. It is of interest to notice that the limit is different from the one in the case of the system starting off from the Poisson field (where temporal structure is sub-fractional Brownian motion). We study behavior of the system for a branching law given by $F$. But there is still broad area for further studies. No attempt has been made to develop more general theory concerning systems with a general starting distribution (or a large class of distributions).
General concepts and notation\[sub:General-concepts-and\]
---------------------------------------------------------
\
Let us denote $N_{t}^{Poiss}$ and $N_{t}^{eq}$, the empirical processes for the system starting off from the Poisson field with Lebesgue intensity measure and the equilibrium respectively. For a measurable set $A\subset\mathbb{R}^{d}$, $N_{t}^{Poiss}\left(A\right)$, $N_{t}^{eq}\left(A\right)$, respectively are the numbers of particles of the system in set $A$ at time $t$. Note that they are measure-valued processes but we will consider them as processes with values in $\mathcal{S}'$ (the space of tempered distributions) because this space has good analytical properties.
The equilibrium distribution is defined by $$\lim_{t\rightarrow+\infty}N_{t}^{Poiss}=N_{eq},$$ where the limit is understood in weak sense. The Laplace functional of the equilibrium distribution is given by $$\mathbb{E}\exp\left\{ -\left\langle N_{eq},\varphi \right\rangle \right\} =\exp\left\{ \left\langle \lambda,e^{-\varphi}-1\right\rangle +V\int_{0}^{\infty}\left\langle \lambda,H\left(j\left(\cdot,s\right)\right)\right\rangle ds\right\} ,\label{eq: laplace_equilibrum}$$ where $$j\left(x,l\right):=\mathbb{E}\exp\left(-\left\langle N_{l}^{x},\varphi \right\rangle \right)
\label{def: h}$$ $H(s)=F\left(s\right)-s$, $\varphi:\mathbb{R}^d\rightarrow\mathbb{R}_{+}$, $\varphi \in \mathcal{L}^1(\mathbb{R}^d)\cap C(\mathcal{R}^d)$ and $j$ satisfies the integral equation$$j\left(x,l\right)=\mathcal{T}_{l}e^{-\varphi}\left(x\right)+V\int_{0}^{l}\mathcal{T}_{l-s}H\left(j\left(\cdot,s\right)\right)\left(x\right)ds,$$ This equations can be obtained in the same way as [@GW2 (2.4)]. Note that in [@GW2] function $\varphi$ is continuous with compact support. We approximate $\varphi\in \mathcal{L}^1$ using functions $\varphi_n$ with compact support $\varphi_n\nearrow \varphi$. Using Lebesgue’s monotone convergence theorem it is easy to obtain the above equations for $\varphi$ ($H$ is decreasing because of the criticality of the branching law).
For an empirical process $N_{t}$ the rescaled occupation time fluctuation process is defined by$$X_{T}\left(t\right)=\frac{1}{F_{T}}\int_{0}^{Tt}\left(N_{s}-\mathbb{E}N_{s}\right)ds,\: t\geq0,\label{eq occupation}$$ where $T>0$ and $F_{T}$ is a suitable norming. We are interested in the weak functional limit of $X_T$ when time is accelerated (i.e., $T$ tends to $\infty$).
The $\alpha$-stable process starting from $x$ will be denoted by $\eta_{t}^{x}$ its semigroup by $\mathcal{T}_{t}$ and its infinitesimal operator by $\Delta_{\alpha}$. The Fourier transform of $\mathcal{T}_{t}$ is $$\widehat{\mathcal{T}_{t}}\varphi\left(z\right)=e^{-t\left|z\right|^{\alpha}}\widehat{\varphi}\left(z\right).\label{eq: Fourier T_t}$$ For brevity let us denote $$K=\frac{V\Gamma\left(2-h\right)}{2^{d-1}\pi^{d/2}\alpha\Gamma\left(d/2\right)h\left(h-1\right)},\label{def: K}$$ where $$h=3-d/\alpha$$ (in this paper we always assume that $\alpha<d<2\alpha$ so $h>1$) and $$M=F''\left(1\right).\label{eq: M}$$ We will now introduce two centered Gaussian processes. One of them is sub-fractional Brownian motion with parameter $h$ with the covariance function $C_{h}$$$C_{h}\left(s,t\right)=s^{h}+t^{h}-\frac{1}{2}\left[\left(s+t\right)^{h}+\left|s-t\right|^{h}\right]\label{def: C - subfractional}$$ and the second one is fractional Brownian motion with parameter $h$ and the covariance function $c_{h}$ $$c_{h}\left(s,t\right)=\frac{1}{2}\left(s^{h}+t^{h}-\left|s-t\right|^{h}\right).\label{def: c - fractional}$$
Space-time method
-----------------
\
The space-time method is a very convenient technique for investigating the weak convergence in the $C\left(\left[0,\tau\right],\mathcal{S}'\left(\mathbb{R}^{d}\right)\right)$ space. It was developed by Bojdecki et al and can be found in [@BGR]. If $X=\left(X\left(t\right)\right)_{t\in\left[0,\tau\right]}$ is a continuous $\mathcal{S}'\left(\mathbb{R}^{d}\right)$-valued process we define a random element $\tilde{X}$ of $\mathcal{S}'\left(\mathbb{R}^{d+1}\right)$ by $$\left\langle \tilde{X},\Phi\right\rangle =\int_{0}^{\tau}\left\langle X\left(t\right),\Phi\left(\cdot,t\right)\right\rangle dt,\label{eq:space-time}$$ where $\Phi\in\mathcal{S}\left(\mathbb{R}^{d+1}\right)$. In order to prove that $X_{T}$ converges weakly to $X$ in $C\left(\left[0,\tau\right],\mathcal{S}'\left(\mathbb{R}^{d}\right)\right)$ it suffices to show that $$\left\langle \tilde{X}_{T},\Phi\right\rangle \Rightarrow\left\langle \tilde{X},\Phi\right\rangle ,\,\forall_{\Phi\in\mathcal{S}\left(\mathbb{R}^{d+1}\right)}$$ and that the family $X_{T}$ is tight.
Convergence theorems
====================
We will present two theorems. In the first of them (which is a direct extension of [@BGT2 Theorem 2.2]) we study the occupation time fluctuation process for the branching system starting off from the Poisson field with Lebesgue intensity measure (denoted by $\lambda$) with the branching law given by a moment generating function as described in Section \[sub:Branching-law\]. The result is very similar to the one obtained in [@BGT2 Theorem 2.2] - namely, the limit process is the same up to constants.
\[thm:For-the-branching\]Assume that $\alpha<d<2\alpha$ and let $X_{T}$ be the occupation time fluctuation process defined by (\[eq occupation\]) for the branching system $N^{Poiss}$, and $F_{T}=T^{\left(3-\frac{d}{\alpha}\right)/2}$. Then $X_{T}\Rightarrow X$ in $C\left(\left[0,\tau\right],\mathcal{S}'\left(\mathbb{R}^{d}\right)\right)$ as $T\rightarrow+\infty$ for any $\tau>0$, where $\left(X\left(t\right)\right)_{t\geq0}$ is a centered $\mathcal{S}'$-valued, Gaussian process with covariance function: $$Cov\left(\left\langle X\left(s\right),\varphi\right\rangle ,\left\langle X\left(t\right),\psi\right\rangle \right)=KM\left\langle \lambda,\varphi\right\rangle \left\langle \lambda,\psi\right\rangle C_{h}\left(s,t\right),\label{def:kowariancja_tw1}$$ where$\ \varphi,\psi\in\mathcal{S}\left(\mathbb{R}^{d}\right)$.
The second theorem concerns the case where the system starts from the equilibrium distribution. As it was mentioned hereinabove the theorem is interesting because the limit has a different time structure from the one in [@BGT2 Theorem 2.2] and Theorem \[thm:For-the-branching\].
\[thm:equlibiurim\]Assume that $\alpha<d<2\alpha$ and let $X_{T}$ be the occupation time fluctuation process defined by (\[eq occupation\]) for the branching system $N^{eq}$, and $F_{T}=T^{\left(3-\frac{d}{\alpha}\right)/2}$. Then $X_{T}\Rightarrow X$ in $C\left(\left[0,\tau\right],\mathcal{S}'\left(\mathbb{R}^{d}\right)\right)$ as $T\rightarrow+\infty$ for any $\tau>0$, where $\left(X\left(t\right)\right)_{t\geq0}$ is a centered Gaussian process with the covariance function $$Cov\left(\left\langle X\left(s\right),\varphi\right\rangle ,\left\langle X\left(t\right),\psi\right\rangle \right)=KM\left\langle \lambda,\varphi\right\rangle \left\langle \lambda,\psi\right\rangle c_{h}\left(s,t\right),\label{def:kowariancja_tw2}$$ where $\varphi,\psi\in\mathcal{S}\left(\mathbb{R}^{d}\right)$.
The limit processes above can be represented as follows:\
For Theorem \[thm:For-the-branching\]$$X=\left(MK\right)^{1/2}\lambda\beta^{h}$$ and for Theorem \[thm:equlibiurim\]$$X=\left(MK\right)^{1/2}\lambda\xi^{h},$$ where $\beta^{h}$ and $\xi^{h}$ are respectively sub-fractional and fractional Gaussian processes defined in Section \[sub:General-concepts-and\]. In both cases the limit process $X$ has a trivial spatial structure (Lebesgue measure), whereas the time structure is complicated, with long range dependence.
The occupation time fluctuation processes of particle systems form an area that receives a lot of research attention. We would like to mention some other related work. Firstly the case of non-branching systems has been studied in [@BGT2 Theorem 2.1]. The result is analogous, both to Theorem \[thm:For-the-branching\] and \[thm:equlibiurim\] because the Poisson field is the equilibrium distribution for the system. The limit process is essentially the same as in Theorem \[thm:equlibiurim\]. For the critical $d=2\alpha$ and large dimensions $d>2\alpha$, there is no long range dependence and the results can be found in [@BGT3]. In [@BZ] the fluctuations of the occupation time of the origin are studied for a critical binary branching random walks on the $d$-dimensional lattice, $d\geq3$, including also the equilibrium case. The convergence results are analogous to those in [@BGT2; @BGT3] and in this paper but the proofs are substantially different. A similar model with $\alpha=2$ was investigated in [@DW] (ie. with particles moving according to Brownian motion).
Proofs
======
The main idea used in both of the proofs is to study the Laplace functional of a process given by the space-time method. The Fourier transform is used for this purpose. This is similar to the method in [@BGT2]. In the case of Theorem \[thm:For-the-branching\] the proof follows the same principle as [@BGT2 Theorem 2.2]. The moment generating function can be represented using Taylor’s expansion and two following statements need to be proved. Firstly, one has to check that the method used in [@BGT2] can still be applied. Secondly, it needs to be shown that terms of order higher then $2$ play no role in the limit. The proof of Theorem \[thm:equlibiurim\] requires more work. The Laplace formula contains a function that is a solution of a differential equation. This makes the computations more cumbersome. Some expressions in this proof had to be examined more carefully than in Theorem \[thm:For-the-branching\]. It should be noted that Theorem \[thm:equlibiurim\] covers all branching laws described in Section \[sub:Branching-law\].\
Now we introduce some notation and facts used further on.
For a generating function $F$ we define $$G\left(s\right)=F\left(1-s\right)-1+s.\label{def: G}$$ The following fact describes basic properties of $G$ which are straightforward consequences of the properties of $F$.
\[fact: properties of G\]
1. $G\left(0\right)=F\left(1\right)-1=0$,
2. $G'\left(0\right)=-F'\left(1\right)+1=0$ since $F'\left(1\right)=1$,
3. $G''\left(0\right)=F''\left(1\right)<+\infty$,
4. $G\left(v\right)=\frac{M}{2}v^{2}+g\left(v\right)v^{2}$ where $M$ is defined by (\[eq: M\]) and $\lim_{v\rightarrow0}g\left(v\right)=0$.
The next simple fact will be useful in proving some inequalities
\[G>=3D0\]$G\left(v\right)\geq0$ for $v\in\left[0,1\right]$.
$F''\left(1-v\right)\geq0$ which is an obvious consequence of the fact that all of the coefficients in the expansion of $F''$ are non-negative and $1-v\in\left[0,1\right]$. $G''\left(v\right)=F''\left(1-v\right)\geq0$. We also know that $G'\left(0\right)=0$ so $G'\left(v\right)\geq0$ for $v\in\left[0,1\right]$. The proof is complete since $G\left(0\right)=0$ and $G$ is non-decreasing.
The existence of the second moment of the moment generating function $F$ implies also that $G$ is comparable with function $v^{2}$.
\[G(v)/v2\]We have $$\sup_{v\in\left[0,1\right]}\frac{G\left(v\right)}{v^{2}}<+\infty$$
Since both $G\left(v\right)$ and $v^{2}$ are continuous we only have to check that the limit of the quotient at $v=0$ is finite. This becomes obvious when we recall Taylor’s expansion of $G\left(v\right)$ from Fact \[fact: properties of G\], property 4.
Let us now introduce some notation used throughout the rest of the paper. $\Phi$ will denote a positive function from $\mathcal{S}\left(\mathbb{R}^{d+1}\right)$. [@BGT2 Lemma in Section 3.2] explains why without loss of generality it can be assumed $\Phi\geq0$. We denote$$\Psi\left(x,s\right)=\int_{s}^{1}\Phi\left(x,t\right)dt,$$ $$\Psi_{T}\left(x,s\right)=\frac{1}{F_{T}}\Psi\left(x,\frac{s}{T}\right).$$ To make computations less cumbersome we will sometimes assume that $\Phi$ is of the form $\Phi\left(x,t\right)=\varphi\left(x\right)\psi\left(t\right)$ for $\varphi\in\mathcal{S}\left(\mathbb{R}^{d}\right),\,\psi\in\mathcal{S}\left(\mathbb{R}\right)$ and hence $$\Psi_{T}\left(x,t\right)=\varphi_{T}\left(x\right)\chi_{T}\left(t\right),
\label{eq:uproszczenie}$$ where $\varphi_{T}\left(x\right)=\frac{1}{F_{T}}\varphi\left(x\right)$, $\chi\left(t\right)=\int_{t}^{1}\psi\left(s\right)ds$, $\chi_{T}=\chi\left(\frac{t}{T}\right)$. Notice that $\varphi\geq0,\chi\geq0$ as $\Phi\geq0$.\
Let us introduce now an important function which will appear as a part of the Laplace functional of the occupation time fluctuation processes $$v_{\Psi}\left(x,r,t\right)=1-\mathbb{E}\exp\left\{ -\int_{0}^{t}\left\langle N_{s}^{x},\Psi\left(\cdot,r+s\right)\right\rangle ds\right\} ,$$ where $N_{s}^{x}$ denotes the empirical measure of the particle system with the initial condition $N_{0}^{x}=\delta_{x}$. Let us note here that due to the fact that $\Psi\geq0$ we have $v_{\Psi}\in\left[0,1\right]$. We also write$$n_{\Psi}\left(x,r,t\right)=\int_{0}^{t}\mathcal{T}_{t-s}\Psi\left(\cdot,r+t-s\right)\left(x\right)ds.\label{def: n}$$ For simplicity of notation, we write $$v_{T}\left(x,r,t\right)=v_{\Psi_{T}}\left(x,r,t\right),\label{eq:v_T notacja}$$ $$n_{T}\left(x,r,t\right)=n_{\Psi_{T}}\left(x,r,t\right),\label{eq:n_T notacja}$$ $$v_{T}\left(x\right)=v_{T}\left(x,0,T\right),\label{eq: v_T_x notacja}$$ $$n_{T}\left(x\right)=n_{T}\left(x,0,T\right)\label{eq: n_T_x notacja}$$ when no confusion can arise.
Now we obtain an integral equation for $v$ which will play a crucial role in the next proofs. Note that similar computations can be found also in [@GW1].
$v_{\Psi}$ satisfies the equation$$v_{\Psi}\left(x,r,t\right)=\int_{0}^{t}\mathcal{T}_{t-s}\left[\Psi\left(\cdot,r+t-s\right)\left(1-v_{\Psi}\left(\cdot,r+t-s,s\right)\right)-VG\left(v_{\Psi}\left(x,r+t-s,s\right)\right)\right]\left(x\right)ds.\label{calkowe}$$
Firstly let us investigate
$$w\left(x,r,t\right)\equiv w_{\Psi}\left(x,r,t\right)=\mathbb{E}\exp\left(-\int_{0}^{t}\left\langle N_{s}^{x},\Psi(\cdot,r+s)\right\rangle ds\right)=1-v_{\Psi}\left(x,r,t\right),$$ We assume $\Psi\geq0$ hence we have $w\left(x,r,t\right)\in\left[0,1\right]$. By conditioning on the time of the first branching we obtain the following equation $$\begin{aligned}
w\left(x,r,t\right)= & e^{-Vt}\mathbb{E}\left(-\int_{0}^{t}\Psi(\eta_{s}^{x},r+s)ds\right)\\
& +V\int_{0}^{t}e^{-Vs}\mathbb{E}\exp\left(-\int_{0}^{s}\Psi(\eta_{u}^{x},r+u)du\right)F\left(w\left(\eta_{s}^{x},r+s,t-s\right)\right),\end{aligned}$$ where $t\geq0,\, r\geq0$.\
Using Feynman-Kac formula one can obtain the following equation for $w$ (for details see [@BGT2 (3.13)-(3.17)]) $$\left\{ \begin{array}{l}
\frac{\partial}{\partial t}w\left(x,r,t\right)=\left(\Delta_{\alpha}+\frac{\partial}{\partial r}-\Psi\left(x,r\right)\right)w\left(x,r,t\right)+V\left[F\left(w\left(x,r,t\right)\right)-w\left(x,r,t\right)\right],\\
w\left(x,r,0\right)=1.\end{array}\right.$$ $v\left(x,r,t\right)=v_{\Psi}\left(x,r,t\right)=1-w_{\Psi}\left(x,r,t\right)$ so $v$ satisfies the equation$$\left\{ \begin{array}{l}
\frac{\partial}{\partial t}v\left(x,r,t\right)=\left(\Delta_{\alpha}+\frac{\partial}{\partial r}\right)v\left(x,r,t\right)+\Psi\left(x,r\right)\left(1-v\left(x,r,t\right)\right)-VG\left(v\left(x,r,t\right)\right),\\
v\left(x,r,0\right)=0.\end{array}\right.$$ Its integral version is \[calkowe\] (note that in [@BGT2] $G\left(t\right)=\frac{1}{2}t^{2}$). $$v\left(x,r,t\right)=\int_{0}^{t}\mathcal{T}_{t-s}\left[\Psi\left(\cdot,r+t-s\right)\left(1-v\left(\cdot,r+t-s,s\right)\right)-VG\left(v\left(x,r+t-s,t\right)\right)\right]\left(x\right)ds.$$
$$v_{\Psi}\left(x,r,t\right)\leq n_{\Psi}\left(x,r,t\right)\label{ineq: v<n}$$
This is a direct consequence of the equation (\[calkowe\]), the fact that $1\geq v\geq0$ and Fact \[G>=3D0\].
For the system $N_{t}^{Poiss}$ the covariance function is given by$$\begin{aligned}
Cov\left(\left\langle N_{u}^{Poiss},\varphi\right\rangle ,\left\langle N_{v}^{Poiss},\psi\right\rangle \right) & =\left\langle \lambda,\varphi\mathcal{T}_{v-u}\psi\right\rangle \label{eq:kowariancja Poisson}\\
& F''\left(1\right)\cdot V\int_{0}^{u}\left\langle \lambda,\varphi\mathcal{T}_{u+v-2r}\psi\right\rangle dr,\,\,\, u\leq v,\nonumber \end{aligned}$$ where$\ \varphi,\psi\in\mathcal{S}\left(\mathbb{R}^{d}\right)$.
The proof of the fact follows from a simple computation which can be carried on using [@GR formula (3.14)], therefore we omit it.
Proof of theorem \[thm:For-the-branching\]
------------------------------------------
### Tightness
The first step required to establish the weak convergence is to prove tightness of $X_{T}$. By the Mitoma theorem [@M Mitoma 1983] it is sufficient to show tightness of the real processes $\left\langle X_{T},\phi\right\rangle $ for all $\phi\in\mathcal{S}\left(\mathbb{R}^{d}\right)$. This can be done using a criterion [@B Theorem 12.3]. Detailed examination of the proof in [@BGT2] reveals that only the covariance function of the $N_{t}^{Poiss}$ is needed [@BGT2 Section 3.1]. One can see that the covariance function (\[eq:kowariancja Poisson\]) is essentially the same as for the binary branching. Hence the proof from [@BGT2] still holds for the new family of processes.
### The Laplace functional
\
The second step uses the space-time method. According to (\[eq:space-time\]) we define $\tilde{X}_{T}$ (from now on $\tau=1$). To establish the convergence we use Laplace functional. By the Poisson initial condition we have (this equation is the same as [@BGT2 (3.10)] $$\mathbb{E}\exp\left\{ -\left\langle \tilde{X}_{T},\Phi\right\rangle \right\} =\exp\left\{ \int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,s\right)dsdx\right\} \exp\left\{ \int_{\mathbb{R}^{d}}-v_{T}\left(x,0,T\right)dx\right\} ,\label{eq:Laplace_transform_with_w}$$ Now we make similar computations to [@BGT2 (3.21)-(3.23)]. By combining (\[eq:Laplace\_transform\_with\_w\]) and (\[calkowe\]) we obtain:$$\begin{aligned}
\mathbb{E}\exp\left\{ -\left\langle \tilde{X}_{T},\Phi\right\rangle \right\} & =\exp\left\{ \int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,s\right)dsdx\right\} \\
& \,\,\,\,\,\cdot\exp\left\{ -\int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,T-s\right)\left(1-v_{T}\left(x,T-s,s\right)\right)-VG\left(v_{T}\left(x,T-s,s\right)\right)dsdx\right\} \\
& =\exp\left\{ \int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,T-s\right)v_{T}\left(x,T-s,s\right)+VG\left(v_{T}\left(x,T-s,s\right)\right)dsdx\right\} \end{aligned}$$ The last expression can be rewritten as:$$\mathbb{E}\exp\left\{ -\left\langle \tilde{X}_{T},\Phi\right\rangle \right\} =\exp\left\{ V\left(I_{1}\left(T\right)+I_{2}\left(T\right)\right)+I_{3}\left(T\right)\right\} ,\label{eq: Laplace_functional_decomposition}$$ where $$\begin{aligned}
I_{1}\left(T\right)= & \int_{0}^{T}\int_{\mathbb{R}^{d}}\frac{M}{2}\left(\int_{0}^{s}\mathcal{T}_{u}\Psi_{T}\left(\cdot,T+u-s\right)\left(x\right)du\right)^{2}dxds,\nonumber \\
I_{2}\left(T\right)= & \int_{0}^{T}\int_{\mathbb{R}^{d}}\left[G\left(v_{T}\left(x,T-s,s\right)\right)-\frac{M}{2}\left(\int_{0}^{s}\mathcal{T}_{u}\Psi_{T}\left(\cdot,T+u-s\right)\left(x\right)du\right)^{2}\right]dxds,\label{def: I2}\\
I_{3}\left(T\right)= & \int_{0}^{T}\int_{\mathbb{R}^{d}}\Psi_{T}\left(x,T-s\right)v_{T}\left(x,T-s,s\right)dxds.\nonumber \end{aligned}$$ To complete the proof we have to compute limits as $T\rightarrow+\infty$. We claim$$I_{1}\left(T\right)\rightarrow\frac{MK}{2V}\int_{0}^{1}\int_{0}^{1}\int_{\mathbb{R}^{d}}\int_{\mathbb{R}^{d}}\Phi\left(x,t\right)\Phi\left(y,s\right)dxdyC_{h}\left(s,t\right)dsdt,\label{zb:1}$$ $$I_{2}\left(T\right)\rightarrow0,$$ $$I_{3}\left(T\right)\rightarrow0,$$ Combining (\[eq: Laplace\_functional\_decomposition\]) with the above limits we obtain$$\lim_{T\rightarrow+\infty}\mathbb{E}\exp\left\{ -\left\langle \tilde{X}_{T},\Phi\right\rangle \right\} =\exp\left\{ \frac{MK}{2}\int_{0}^{1}\int_{0}^{1}\int_{\mathbb{R}^{d}}\int_{\mathbb{R}^{d}}\Phi\left(x,t\right)\Phi\left(y,s\right)dxdyC_{h}\left(s,t\right)dsdt\right\} \label{eq:limit_<X_T,PHI>}$$ hence the limit process $X_{T}$ is a Gaussian process with covariance (\[def:kowariancja\_tw1\]).
### Convergence proofs
\
$I_{1}\left(T\right)$ does not depend on $F$ so it can be evaluated in the same way as in [@BGT2 (3.32)-(3.34)].\
Let us now deal with $I_{3}\left(T\right)$. By using (\[ineq: v<n\]) we obtain$$I_{3}\left(T\right)\leq\int_{0}^{T}\int_{\mathbb{R}^{d}}\Psi_{T}\left(x,T-s\right)\int_{0}^{s}\mathcal{T}_{u}\Psi_{T}\left(\cdot,T-u\right)dudxds\leq$$ $$\frac{C}{F_{T}^{2}}\int_{0}^{T}\int_{\mathbb{R}^{d}}\varphi\left(x\right)\int_{0}^{s}\mathcal{T}_{u}\varphi\left(x\right)dudxds$$ Now the rest of the proof goes along the same lines as in [@BGT2].\
We will turn to $I_{2}\left(T\right)$ which is a little more intricate. Combining (\[def: I2\]) and property $4$ from Fact \[fact: properties of G\]$$\begin{aligned}
& I_{2}\left(T\right)=\int_{0}^{T}\int_{\mathbb{R}^{d}}\left[\frac{M}{2}\left[v_{T}\left(\ldots\right)^{2}-\left(\int_{0}^{s}\mathcal{T}_{u}\Psi_{T}\left(\cdot,T+u-s\right)\left(x\right)du\right)^{2}\right]+g\left(v_{T}\left(\ldots\right)\right)v_{T}\left(\ldots\right)^{2}\right]dxds=\\
& =\frac{M}{2}I'_{2}\left(T\right)+I''_{2}\left(T\right),\end{aligned}$$ where $$I'_{2}\left(T\right)=\int_{0}^{T}\int_{\mathbb{R}^{d}}v_{T}\left(x,T-s,s\right)^{2}-\left(\int_{0}^{s}\mathcal{T}_{u}\Psi_{T}\left(\cdot,T+u-s\right)\left(x\right)du\right)^{2}dxds,$$ $$I''_{2}\left(T\right)=\int_{0}^{T}\int_{\mathbb{R}^{d}}g\left(v_{T}\left(x,T-s,s\right)\right)v_{T}\left(x,T-s,s\right)^{2}dxds\label{eq: I2pp}$$ By inequality (\[ineq: v<n\]) we have $$0\leq-I'_{2}\left(T\right)=\int_{0}^{T}\int_{\mathbb{R}^{d}}\left[\left(n_{T}\left(x,T-s,s\right)\right)^{2}-\left(v_{T}\left(x,T-s,s\right)\right)^{2}\right].$$ Combining (\[calkowe\]) and (\[def: n\]) yields$$\begin{aligned}
0\leq & n_{T}\left(x,T-s,s\right)-v_{T}\left(x,T-s,s\right)=\\
& \int_{0}^{s}\mathcal{T}_{s-u}\left[\Psi_{T}\left(\cdot,T-u\right)v_{T}\left(\cdot,T-u,u\right)+VG\left(v_{T}\left(\cdot,T-u,u\right)\right)\right]\left(x\right)du=\left(*\right).\end{aligned}$$
We have $\mathcal{T}_{s}\Psi\geq0$ for $\Psi\geq0$ which is a direct consequence of the fact that $\mathcal{T}$ is the semigroup of a Markov process. By Fact \[G(v)/v2\] we have $c(F)$ such that $F(v)\leq \frac{c(F)}{2}v^2$. Hence $$\begin{aligned}
\left(*\right) & \leq\int_{0}^{s}\mathcal{T}_{s-u}\left[\Psi_{T}\left(\cdot,T-u\right)v_{T}\left(\cdot,T-u,u\right)+c\left(F\right)\frac{V}{2}v_{T}\left(\cdot,T-u,u\right)^{2}\right]\left(x\right)du\\
& \leq\max\left(1,c\left(F\right)\right)\int_{0}^{s}\mathcal{T}_{s-u}\left[\Psi_{T}\left(\cdot,T-u\right)v_{T}\left(\cdot,T-u,u\right)+\frac{V}{2}v_{T}\left(\cdot,T-u,u\right)^{2}\right]\left(x\right)du\\
& \leq\max\left(1,c\left(F\right)\right)\int_{0}^{s}\mathcal{T}_{s-u}\left[\Psi_{T}\left(\cdot,T-u\right)n_{T}\left(\cdot,T-u,u\right)+\frac{V}{2}n_{T}\left(\cdot,T-u,u\right)^{2}\right]\left(x\right)du.\end{aligned}$$ Except of the constant $c\left(F\right)$ the last expression does not depend on $F$.\
Next we consider$$n_{T}\left(x,T-s,s\right)+v_{T}\left(x,T-s,s\right)\leq2n_{T}\left(x,T-s,s\right)\leq2\int_{0}^{s}\mathcal{T}_{s-u}\Psi\left(\cdot,T-u\right)\left(x\right)du.$$ The rest of the proof goes along the lines of the proof in [@BGT2 inequalities (3.39)-(3.42)] and hence we acquire $I_{2}'\left(T\right)\rightarrow0$.\
Before proving the convergence of $I_{2}''\left(T\right)$ we state two facts:
$n_{T}\left(x,T-s,s\right)\rightarrow0$ in uniformly $x\in\mathbb{R}^{d}$, $s\in\left[0,T\right]$ as $T\rightarrow+\infty$. \[fact: uniformaly\]
\
$$n_{T}\left(x,T-s,s\right)=\int_{0}^{s}\mathcal{T}_{s-u}\Psi_{T}\left(\cdot,T-u\right)du=$$ $$\frac{1}{F_{T}}\int_{0}^{s}\mathcal{T}_{s-u}\varphi\left(x\right)\chi\left(\frac{T-u}{T}\right)du\leq$$ $$\frac{C}{F_{T}}\int_{0}^{+\infty}\mathcal{T}_{u}\varphi\left(x\right)du=\frac{C_{1}}{F_{T}}\int_{\mathbb{R}^{d}}\frac{\varphi\left(y\right)}{\left|x-y\right|^{d-\alpha}}dy\leq\frac{C_{2}}{F_{T}}\rightarrow0.$$ The last line contains the definition of the potential operator of the semigroup $\mathcal{T}_{t}$ which is bounded in respect to $x$ (this can be found in [@I Lemma 5.3]).
The following convergence holds:
$$\int_{0}^{T}\int_{\mathbb{R}^{d}}v_{T}\left(x,T-s,s\right)^{2}\rightarrow c'\left(\Psi\right)\textrm{ as }T\rightarrow+\infty.$$
One easily checks that$$2\frac{I_{1}\left(T\right)}{M}+I_{2}'\left(T\right)=\int_{0}^{T}\int_{\mathbb{R}^{d}}v_{T}\left(x,T-s,s\right)^{2}.$$ Hence the result follows from (\[zb:1\]) and $I'_{2}\left(T\right)\rightarrow0$ as $T\rightarrow0$.
It is now easy to prove the convergence of $I''_{2}$. From Fact \[fact: properties of G\] property $4$ we know that for given $\epsilon>0$ we can choose such $\delta$ that $\forall_{x\in\left(-\delta,\delta\right)}\left|g\left(x\right)\right|\leq\epsilon$. Fact \[fact: uniformaly\] provides us with $T_{0}$ such that $\forall_{T\geq T_{0}}\, n_{T}\left(x,T-s,s\right)<\delta$. Combining this with (\[ineq: v<n\]) we obtain $\forall_{T\geq T_{0}}\, g\left(v_{T}\left(x,T-s,s\right)\right)\leq\epsilon$. Hence for $T>T_{0}$ holds:$$\begin{aligned}
\left|I_{2}''\left(T\right)\right| & \leq\epsilon\int_{0}^{T}\int_{\mathbb{R}^{d}}v_{T}^{2}\left(x,T-s,s\right)dxds\rightarrow\epsilon c'\left(\Psi\right).\end{aligned}$$ Since $\epsilon$ was chosen arbitrary we have convergence: $I_{2}''\left(T\right)\rightarrow0$ hence also $I_{2}\left(T\right)\rightarrow0$ as $T\rightarrow+\infty$.
Thus we obtained the limits for $I_{1},I_{2}$ and $I_{3}$ and the proof of Theorem \[thm:For-the-branching\] is completed.
Proof of Theorem \[thm:equlibiurim\]
------------------------------------
### Tightness
\
We begin by claiming that the family $\left\{ X_{T}\right\} _{T>0}$ is tight. Close examination of [@BGT2 Section 3.1] reveals that only the covariance function of the underlying system is significant for the proof. By [@BGT1 (3.16)] we know that the covariance function of the branching system is of the same form as the covariance function of the non-branching system with the Poisson initial condition. From this we conclude that $X_{T}$ is tight.
### Laplace functional for $\tilde{X}_{T}$
\
We consider $\tilde{X}_{T}$ defined by (\[eq:space-time\]). Using (\[eq occupation\]) and interchanging the order of integration we obtain$$\left\langle \tilde{X}_{T},\Phi\right\rangle =\frac{T}{F_{T}}\left[\int_{0}^{1}\left\langle N_{Ts},\Psi\left(\cdot,s\right)\right\rangle ds-\left\langle \lambda,\int_{0}^{1}\Psi\left(\cdot,s\right)ds\right\rangle \right].$$ To prove the convergence of $\tilde{X}_{T}$ to $\tilde{X}$ we will use its Laplace functional$$\begin{aligned}
& \mathbb{E}\exp\left\{ -\left\langle \tilde{X}_{T},\Phi\right\rangle \right\} = & \exp\left\{ \int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,t\right)dtdx\right\} \label{eq: Laplace tildeX}\\
& & \mathbb{E}\exp\left\{ -\int_{0}^{T}\left\langle N_{s},\Psi_{T}\left(\cdot,s\right)\right\rangle ds\right\} ,\nonumber \end{aligned}$$ It is easy to check that$$\mathbb{E}\left(\exp\left\{ -\int_{0}^{T}\left\langle N_{s},\Psi_{T}\left(\cdot,s\right)\right\rangle ds\right\} |N_{0}=\mu\right)=\exp\left\{ \left\langle \mu,\ln w_{T}\right\rangle \right\} ,\label{eq: conditionallaplace}$$ where $$w_{T}\left(x\right)=\mathbb{E}\exp\left\{ -\int_{0}^{T}\left\langle N_{s}^{x},\Psi_{T}\left(\cdot,t\right)\right\rangle dt\right\}$$\
Now we check that $0\leq-\ln(w_T)$ is integrable. For $T$ big enough by Fact \[fact: uniformaly\] and inequity (\[ineq: v<n\]) we have $0\leq v_T \leq c <1$. Hence there exists a constant $C$ such that we have $-\ln(w_T)=-\ln(1-v_T)\leq C v_T \leq C n_T$. A trivial verifications shows that $n_T \in \mathcal{L}^1(\mathbb{R}^d)$ so by (\[eq: laplace\_equilibrum\]) and (\[eq: conditionallaplace\]) we obtain$$\mathbb{E}\exp\left\{ -\int_{0}^{T}\left\langle N_{s},\Psi_{T}\left(\cdot,s\right)\right\rangle ds\right\} =\mathbb{E}\left(\mathbb{E}\left(\exp\left\{ -\int_{0}^{T}\left\langle N_{s},\Psi_{T}\left(\cdot,s\right)\right\rangle ds\right\} |N_{0}\right)\right)=$$ $$\exp\left\{ \left\langle \lambda,w_{T}-1\right\rangle +V\int_{0}^{+\infty}\left\langle \lambda,H\left(W_{T}\left(\cdot,s\right)\right)\right\rangle ds\right\} ,$$ where $W_{T}$ satisfies the equation $$W_{T}\left(x,l\right)=\mathcal{T}_{l}w_{T}\left(x\right)+V\int_{0}^{l}\mathcal{T}_{l-s}H\left(W_{T}\left(\cdot,s\right)\right)\left(x\right)ds$$\
It will be a bit easier to deal with $V_{T}\left(x,l\right)=1-W_{T}\left(x,l\right)$. The equations have the form (let us recall that $G$ is defined by (\[def: G\])) $$\mathbb{E}\exp\left\{ -\int_{0}^{T}\left\langle N_s,\Psi_{T}\left(\cdot,s\right)\right\rangle ds\right\} =\exp\left\{ \left\langle \lambda,-v_{T}\right\rangle +V\int_{0}^{+\infty}\left\langle \lambda,G\left(V_{T}\left(\cdot,s\right)\right)\right\rangle ds\right\} ,
\label{eq: inner}$$ and $$V_{T}\left(x,l\right)=\mathcal{T}_{l}v_{T}\left(x\right)-V\int_{0}^{l}\mathcal{T}_{l-s}G\left(V_{T}\left(\cdot,s\right)\right)\left(x\right)ds,\label{main_equation}$$ $W_T$ is defined by (\[def: h\]) with $\varphi(x)=-\ln w_T(x)$ ($w_T\in[0,1]$ hence $\varphi$ is positive). One can easily see that the definition implies that $W_T\in[0,1]$. Consequently $V_T\in[0,1]$ which together with Fact \[G>=3D0\] yields $G(V_T)\geq 0$. Hence we obtain an inequality $$V_{T}\left(x,l\right)\leq\mathcal{T}_{l}v_{T}\left(x\right),\,\forall_{x\in\mathbb{R}^{d},l\geq0}.
\label{neq: WT<TvT}$$ Combining (\[eq: Laplace tildeX\]) and (\[eq: inner\]) we obtain$$\begin{aligned}
\mathbb{E}\exp\left\{ -\left\langle \tilde{X}_{T},\Phi\right\rangle \right\} & = & \exp\left\{ \int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,t\right)dtdx\right\} \exp\left\{ -\int_{\mathbb{R}^{d}}v_{T}\left(x\right)dx\right\} \\
& & \exp\left\{ V\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}G\left(V_{T}\left(x,t\right)\right)dxdt\right\} =A\left(T\right)\cdot B\left(T\right),\end{aligned}$$ where $$A\left(T\right)=\exp\left\{ \int_{\mathbb{R}^{d}}\int_{0}^{T}\Psi_{T}\left(x,t\right)dtdx\right\} \exp\left\{ -\int_{\mathbb{R}^{d}}v_{T}\left(x\right)dx\right\} ,$$ $$B\left(T\right)=\exp\left\{ V\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}G\left(V_{T}\left(x,t\right)\right)dxdt\right\} .$$ Let us note that $A$ is the same as (\[eq:Laplace\_transform\_with\_w\]) in the first proof hence we know that its limit is given by (\[eq:limit\_<X\_T,PHI>\]).
### Limit of B
\
To complete the proof the limit $\lim_{T\rightarrow+\infty}B\left(T\right)$ has to be calculated. It suffices to consider $$\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}G\left(V_{T}\left(x,t\right)\right)dxdt.\label{exp(...)}$$ Using Fact \[fact: properties of G\], property 4, we split it in the following way $$\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}G\left(V_{T}\left(\cdot,t\right)\right)dxdt=\frac{M}{2}\left(B_{1}\left(T\right)+B_{2}\left(T\right)+B_{3}\left(T\right)\right)+B_{4}\left(T\right),$$ where $$B_{1}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}V_{T}\left(x,t\right)^{2}-\left(\mathcal{T}_{t}v_{T}\left(x\right)\right)^{2}dxdt,$$ $$B_{2}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(\mathcal{T}_{t}v_{T}\left(x\right)\right)^{2}-\left(\mathcal{T}_{t}n_{T}\left(x\right)\right)^{2}dxdt,$$ $$B_{3}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(\mathcal{T}_{t}n_{T}\left(x\right)\right)^{2}dxdt,$$ $$B_{4}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}g\left(V_{T}\left(x,t\right)\right)V_{T}\left(x,t\right)^{2}.$$ We will prove the following limits (let us recall that we assume (\[eq:uproszczenie\]) for simplicity) $$B_{1}\left(T\right)\rightarrow0,$$ $$B_{2}\left(T\right)\rightarrow0,$$ $$B_{3}\left(T\right)\rightarrow\frac{K}{2V}\left\langle \lambda,\varphi\right\rangle ^{2}\int_{0}^{1}\int_{0}^{1}\left\{ -u_{1}^{h}-u_{2}^{h}+\left(u_{1}+u_{2}\right)^{h}\right\} \psi\left(u_{1}\right)\psi\left(u_{2}\right)du_{1}du_{2},$$ $$B_{4}\left(T\right)\rightarrow0,$$ as $T\rightarrow+\infty$.
#### Limit of $B_{1}$ {#limit-of-b_1 .unnumbered}
\
By (\[neq: WT<TvT\]) we obtain $$0\leq-B_{1}\left(T\right)=\int_{0}^{+\infty}\left(\int_{\mathbb{R}^{d}}\left(\mathcal{T}_{t}v_{T}\left(x\right)\right)^{2}-V_{T}\left(x,t\right)^{2}dx\right)dt=$$ $$\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(\mathcal{T}_{t}v_{T}\left(x\right)-V_{T}\left(x,t\right)\right)\left(\mathcal{T}_{t}v_{T}\left(x\right)V+_{T}\left(x,t\right)\right)dxdt\leq$$ Combining this with inequality (\[neq: WT<TvT\]) and equation (\[main\_equation\]) we have$$\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(V\int_{0}^{t}\mathcal{T}_{t-t'}G\left(V_{T}\left(\cdot,t'\right)\right)\left(x\right)dt'\right)\left(2\mathcal{T}_{t}v_{T}\left(x\right)\right)dxdt=$$ Taking into account the form of $G$ (Fact \[fact: properties of G\], property 4)$$B_{11}\left(T\right)+B_{12}\left(T\right),$$ where $$B_{11}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(V\frac{M}{2}\int_{0}^{t}\mathcal{T}_{t-t'}V_{T}\left(\cdot,t'\right)^{2}\left(x\right)dt'\right)\left(2\mathcal{T}_{t}v_{T}\left(x\right)\right)dxdt,$$ $$B_{12}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(V\int_{0}^{t}\mathcal{T}_{t-t'}g\left(V_{T}\left(\cdot,t'\right)\right)V_{T}\left(\cdot,t'\right)^{2}\left(x\right)dt'\right)\left(2\mathcal{T}_{t}v_{T}\left(x\right)\right)dxdt$$ Once again we use inequality (\[neq: WT<TvT\])$$B_{11}\left(T\right)\leq VM\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(\int_{0}^{t}\mathcal{T}_{t-t'}\left(\mathcal{T}_{t'}v_{T}\left(\cdot\right)\right)^{2}\left(x\right)dt'\right)\left(\mathcal{T}_{t}v_{T}\left(x\right)\right)dxdt=$$ $$VM\int_{0}^{+\infty}\int_{0}^{t}\int_{\mathbb{R}^{d}}\mathcal{T}_{t-t'}\left(\mathcal{T}_{t'}v_{T}\left(\cdot\right)\right)^{2}\left(x\right)\mathcal{T}_{t}v_{T}\left(x\right)dxdt'dt\leq$$ Applying (\[ineq: v<n\]) twice$$VM\int_{0}^{+\infty}\int_{0}^{t}\int_{\mathbb{R}^{d}}\mathcal{T}_{t-t'}\left(\mathcal{T}_{t'}n_{T}\left(\cdot\right)\right)^{2}\left(x\right)\mathcal{T}_{t}n_{T}\left(x\right)dxdt'dt=$$
$$MV\int_{0}^{+\infty}\int_{0}^{t}\int_{\mathbb{R}^{d}}\mathcal{T}_{t'}n_{T}\left(x\right)\mathcal{T}_{t'}n_{T}\left(x\right)\mathcal{T}_{2t-t'}n_{T}\left(x\right)dxdt'dt=$$ We use the Plancherel formula and (\[eq: Fourier T\_t\])$$\frac{MV}{\left(2\pi\right)^{2d}}\int_{0}^{+\infty}\int_{0}^{t}\int_{\mathbb{R}^{2d}}\widehat{\mathcal{T}_{t'}n_{T}}\left(z_{1}\right)\widehat{\mathcal{T}_{t'}n_{T}}(z_{2})\overline{\widehat{\mathcal{T}_{2t-t'}n_{T}}}\left(z_{1}+z_{2}\right)dz_{1}dz_{2}dt'dt=$$ $$\frac{MV}{\left(2\pi\right)^{2d}}\int_{0}^{+\infty}\int_{0}^{t}\int_{\mathbb{R}^{2d}}e^{-t'\left|z_{1}\right|^{\alpha}}\widehat{n}_{T}\left(z_{1}\right)e^{-t'\left|z_{2}\right|^{\alpha}}\widehat{n}_{T}(z_{2})e^{-(2t-t')\left|z_{1}+z_{2}\right|^{\alpha}}\overline{\widehat{n}_{T}}\left(z_{1}+z_{2}\right)dz_{1}dz_{2}dt'dt=$$ $$\frac{MV}{\left(2\pi\right)^{2d}}\int_{\mathbb{R}^{2d}}\widehat{n}_{T}\left(z_{1}\right)\widehat{n}_{T}(z_{2})\overline{\widehat{n}_{T}}\left(z_{1}+z_{2}\right)\int_{0}^{+\infty}\int_{0}^{t}e^{-t'\left|z_{1}\right|^{\alpha}}e^{-t'\left|z_{2}\right|^{\alpha}}e^{-(2t-t')\left|z_{1}+z_{2}\right|^{\alpha}}dt'dtdz_{1}dz_{2}=$$ $$\frac{MV}{\left(2\pi\right)^{2d}}\int_{\mathbb{R}^{2d}}\frac{1}{2\left|z_{1}+z_{2}\right|^{\alpha}\left(\left|z_{1}\right|^{\alpha}+\left|z_{2}\right|^{\alpha}+\left|z_{1}+z_{2}\right|^{\alpha}\right)}\widehat{n}_{T}\left(z_{1}\right)\widehat{n}_{T}(z_{2})\overline{\widehat{n}_{T}}\left(z_{1}+z_{2}\right)dz_{1}dz_{2}=\left(*\right)$$ Before proceeding further we will estimate $\widehat{n}_{T}$ $$\left|\widehat{n}_{T}\left(z,r,t\right)\right|=\left|\widehat{\int_{0}^{t}\mathcal{T}_{t-s}\Psi_{T}\left(\cdot,r+t-s\right)ds(z)}\right|=$$ $$\left|\frac{1}{F_{T}}\int_{0}^{t}e^{-\left(t-s\right)\left|z\right|^{\alpha}}\widehat{\varphi}\left(z\right)\chi_{T}\left(r+t-s\right)ds\right|\leq$$ $$\frac{\sup\chi}{F_{T}}\left|\widehat{\varphi}\left(z\right)\right|\int_{0}^{t}e^{-\left(t-s\right)\left|z\right|^{\alpha}}ds\leq$$ Hence$$\left|\widehat{n}_{T}\left(z,r,t\right)\right|\leq\frac{C}{F_{T}}\frac{\left|\widehat{\varphi}\left(z\right)\right|}{\left|z\right|^{\alpha}}\left[1-e^{-t\left|z\right|^{\alpha}}\right]\label{neq: n(z,r,t)< strong}$$ and this immediately implies (see (\[eq: n\_T\_x notacja\])) $$\left|\widehat{n}_{T}\left(z\right)\right|\leq\frac{C}{F_{T}}\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-T\left|z\right|^{\alpha}}\right].\label{ineq: hat(n_T)}$$ Here, and in what follows, $C$ denotes a generic constant.\
Coming back to $\left(*\right)$ and using the last inequality we obtain$$\begin{aligned}
\left|\left(*\right)\right|\leq\frac{C}{F_{T}^{3}}\int_{\mathbb{R}^{2d}}\frac{1}{2\left|z_{1}+z_{2}\right|^{\alpha}\left(\left|z_{1}\right|^{\alpha}+\left|z_{2}\right|^{\alpha}+\left|z_{1}+z_{2}\right|^{\alpha}\right)}\frac{1}{\left|z_{1}\right|^{\alpha}}\left[1-e^{-T\left|z_{1}\right|^{\alpha}}\right]\\
\frac{1}{\left|z_{2}\right|^{\alpha}}\left[1-e^{-T\left|z_{2}\right|^{\alpha}}\right]\frac{1}{\left|z_{1}+z_{2}\right|^{\alpha}}\left[1-e^{-T\left|z_{1}+z_{2}\right|^{\alpha}}\right]dz_{1}dz_{2} & =\end{aligned}$$ Substituting $T^{1/\alpha}z_{1}=y_{1}$ and $T^{1/\alpha}z_{2}=y_{2}$ yields$$\begin{aligned}
\frac{CT^{5}}{F_{T}^{3}T^{2\frac{d}{\alpha}}}\int_{\mathbb{R}^{2d}}\frac{1}{\left|y_{1}+y_{2}\right|^{\alpha}\left(\left|y_{1}\right|^{\alpha}+\left|y_{2}\right|^{\alpha}+\left|y_{1}+y_{2}\right|^{\alpha}\right)}\frac{1}{\left|y_{1}\right|^{\alpha}}\left[1-e^{-\left|y_{1}\right|^{\alpha}}\right]\\
\frac{1}{\left|y_{2}\right|^{\alpha}}\left[1-e^{-\left|y_{2}\right|^{\alpha}}\right]\frac{1}{\left|y_{1}+y_{2}\right|^{\alpha}}\left[1-e^{-\left|y_{1}+y_{2}\right|^{\alpha}}\right]dy_{1}dy_{2} & \leq\end{aligned}$$ $$B'_{11}\left(T\right)\cdot B''_{11},$$ where$$B'_{11}\left(T\right)=\frac{C'T^{5}}{F_{T}^{3}T^{2\frac{d}{\alpha}}}$$ $$\begin{aligned}
B''_{11}=\int_{\mathbb{R}^{2d}}\frac{1}{\left|y_{1}+y_{2}\right|^{\alpha}\left(\left|y_{1}\right|^{\alpha}+\left|y_{2}\right|^{\alpha}+\left|y_{1}+y_{2}\right|^{\alpha}\right)}\frac{1}{\left|y_{1}\right|^{\alpha}}\left[1-e^{-\left|y_{1}\right|^{\alpha}}\right]\\
\frac{1}{\left|y_{2}\right|^{\alpha}}\left[1-e^{-\left|y_{2}\right|^{\alpha}}\right]\frac{1}{\left|y_{1}+y_{2}\right|^{\alpha}}\left[1-e^{-\left|y_{1}+y_{2}\right|^{\alpha}}\right]dy_{1}dy_{2}\end{aligned}$$ The integral $B''_{11}$ is finite which will be proved in Fact \[fact: calka1\]. The expression $B'_{11}\left(T\right)$ can be evaluated$$B'_{11}\left(T\right)=T^{\frac{10-3\left(3-\frac{d}{\alpha}\right)-4\frac{d}{\alpha}}{2}}=T^{\frac{1-\frac{d}{\alpha}}{2}}$$ and as $1-\frac{d}{\alpha}<0$ $B'_{11}\left(T\right)\rightarrow0$ hence the convergence: $B_{11}\left(T\right)\rightarrow0$ is obtained too.\
From Fact \[fact: uniformaly\] and inequalities (\[ineq: v<n\]) and (\[neq: WT<TvT\]) we know $V_{T}(x,l)\rightarrow0$ uniformly as $T\rightarrow0$ and so $g\left(V_{T}\left(x,l\right)\right)\leq\epsilon$ for $T$ sufficiently large hence $$B_{12}\left(T\right)\leq\epsilon\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(V\int_{0}^{t}\mathcal{T}_{t-t'}V_{T}\left(\cdot,t'\right)^{2}\left(x\right)dt'\right)\left(2\mathcal{T}_{t}v_{T}\left(x\right)\right)dxdt\leq\frac{2\epsilon}{M}B_{11}\left(T\right)$$ thus $B_{12}\left(T\right)\rightarrow0$ and $B_{1}\left(T\right)\rightarrow0$ too.
#### Limit of $B_{2}$ {#limit-of-b_2 .unnumbered}
\
Let us first estimate expression $n_{T}-v_{T}$ using (\[calkowe\]) and (\[def: n\]) $$\begin{aligned}
& n_{T}\left(x\right)-v_{T}\left(x\right)= & \int_{0}^{T}\mathcal{T}_{T-u}\Psi_{T}\left(\cdot,T-u\right)\left(x\right)du\\
& & -\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)\left(1-v_{T}\left(\cdot,T-u,u\right)\right)-VG\left(v_{T}\left(\cdot,T-u,u\right)\right)\right]\left(x\right)du\end{aligned}$$ $$n_{T}\left(x\right)-v_{T}\left(x\right)=\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)v_{T}\left(\cdot,T-u,u\right)+VG\left(v_{T}\left(\cdot,T-u,u\right)\right)\right]\left(x\right)du\leq$$ Applying Fact \[G(v)/v2\]$$\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)v_{T}\left(\cdot,T-u,u\right)+Vc\left(v_{T}\left(\cdot,T-u,u\right)\right)^{2}\right]\left(x\right)du,$$ where $c$ is a constant. By inequality (\[ineq: v<n\])$$n_{T}\left(x\right)-v_{T}\left(x\right)\leq\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)n_{T}\left(\cdot,T-u,u\right)+Vc\left(n_{T}\left(\cdot,T-u,u\right)\right)^{2}\right]\left(x\right)du\label{ineq: n-v}$$ We have$$0\leq-B_{2}\left(T\right)=\int_{0}^{+\infty}\left(\int_{\mathbb{R}^{d}}\left(\mathcal{T}_{t}n_{T}\left(x\right)\right)^{2}-\left(\mathcal{T}_{t}v_{T}\left(x\right)\right)^{2}dx\right)dt=$$ $$\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\left(\mathcal{T}_{t}\left(n_{T}\left(\cdot\right)-v_{T}\left(\cdot\right)\right)\left(x\right)\right)\left(\mathcal{T}_{t}\left(v_{T}\left(\cdot\right)+n_{T}\left(\cdot\right)\right)\left(x\right)\right)dxdt\leq$$ Applying (\[ineq: v<n\]) and (\[ineq: n-v\])$$\begin{aligned}
2\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}\mathcal{T}_{t}\left\{ \int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)n_{T}\left(\cdot,T-u,u\right)+Vc\left(n_{T}\left(\cdot,T-u,u\right)\right)^{2}\right]du\right\} \left(x\right)\\
\mathcal{T}_{t}n_{T}\left(x\right)dxdt=\end{aligned}$$ Now we apply the Plancherel formula$$\begin{aligned}
\frac{2}{\left(2\pi\right)^{d}}\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}e^{-2t\left|z\right|^{\alpha}}\widehat{\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)n_{T}\left(\cdot,T-u,u\right)+Vc\left(n_{T}\left(\cdot,T-u,u\right)\right)^{2}\left(\cdot\right)\right]\left(z\right)du}\\
\widehat{n}_{T}\left(z\right)dzdt & =\end{aligned}$$ Interchanging the order of integration and integrating with respect to $t$ we get$$\begin{aligned}
\frac{1}{\left(2\pi\right)^{d}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\widehat{\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)n_{T}\left(\cdot,T-u,u\right)+Vc\left(n_{T}\left(\cdot,T-u,u\right)\right)^{2}\left(\cdot\right)\right]\left(z\right)du}\\
\widehat{n}_{T}\left(z\right)dz=c'\left(B_{21}\left(T\right)+B_{22}\left(T\right)\right),\end{aligned}$$ where $$B_{21}\left(T\right)=\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\left\{ \widehat{\int_{0}^{T}\mathcal{T}_{T-u}\left[\Psi_{T}\left(\cdot,T-u\right)n_{T}\left(\cdot,T-u,u\right)\right]\left(z\right)du}\right\} \widehat{n}_{T}\left(z\right)dz,$$ $$B_{22}\left(T\right)=\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\left\{ \widehat{\int_{0}^{T}\mathcal{T}_{T-u}\left[Vc\left(n_{T}\left(\cdot,T-u,u\right)\right)^{2}\left(\cdot\right)\right]\left(z\right)du}\right\} \widehat{n}_{T}\left(z\right)dz.$$ We shall compute $\lim_{T\rightarrow+\infty}B_{21}\left(T\right)$ first. We have $$B_{21}(T)=\int_{\mathbb{R}^{d}}\frac{1}{|z|^{\alpha}} \left\{ \int_{0}^{T} e^{-(T-u)|z|^\alpha}
\widehat{\Psi_{T}\left(\cdot,T-u\right)}*\widehat{n_{T}\left(\cdot,T-u,u\right)\left(z\right)}
\right\} \widehat{n}_T(z)dz.$$ The inner convolution can be estimated using inequality (\[neq: n(z,r,t)< strong\]) and simplification (\[eq:uproszczenie\]) $$\left|\widehat{\Psi_{T}\left(\cdot,T-u\right)}*\widehat{n_{T}\left(\cdot,T-u,u\right)}\left(z\right)\right|=
\left|\chi_{T}\left(T-u\right)\widehat{\varphi_{T}}\left(\cdot\right)*\widehat{n_{T}\left(\cdot,T-u,u\right)\left(z\right)}\right|=$$ $$\left|\chi_{T}\left(T-u\right)\int_{\mathbb{R}^{d}}\widehat{\varphi}_{T}\left(z-x\right)\widehat{n}_{T}\left(x,T-u,u\right)dx\right|\leq$$ $$\frac{c\left(\chi\right)}{F_{T}^{2}}\chi_{T}\left(T-u\right)\int_{\mathbb{R}^{d}}\left|\widehat{\varphi}\left(z-x\right)\widehat{\varphi}\left(x\right)\right|\frac{1}{\left|x\right|^{\alpha}}dx\leq\frac{C}{F_{T}^{2}}$$ In the last inequality we use the fact that $\hat{\varphi}$ is bounded and $\frac{\widehat{\varphi}\left(x\right)}{\left|x\right|^{\alpha}}$ is integrable. Hence we have inequality$$\left|\widehat{\Psi_{T}\left(\cdot,T-u\right)}*\widehat{n_{T}\left(\cdot,T-u,u\right)}\left(z\right)\right|\leq\frac{C}{F_{T}^{2}}\label{neq:convultion1}$$ Thus $B_{21}$ satisfies$$\left|B_{21}\left(T\right)\right|\leq\frac{C}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\int_{0}^{T}e^{-\left(T-u\right)\left|z\right|^{\alpha}}du\cdot\widehat{n}_{T}\left(z\right)dz\leq$$ Using inequality (\[ineq: hat(n\_T)\]) and integrating with respect to $u$$$C'\frac{1}{F_{T}^{3}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-T\left|z\right|^{\alpha}}\right]\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-T\left|z\right|^{\alpha}}\right]dz=$$ Substituting $zT^{1/\alpha}=y$$$C'\frac{T^{3}}{F_{T}^{3}T^{\frac{d}{\alpha}}}\int_{\mathbb{R}^{d}}\frac{1}{\left|y\right|^{\alpha}}\frac{1}{\left|y\right|^{\alpha}}\left[1-e^{-\left|y\right|^{\alpha}}\right]\frac{1}{\left|y\right|^{\alpha}}\left[1-e^{-\left|y\right|^{\alpha}}\right]dy\leq B_{21}'\left(T\right)\cdot B''_{21},$$ where$$B_{21}'\left(T\right)=C''\frac{T^{3}}{F_{T}^{3}T^{\frac{d}{\alpha}}}$$ $$B_{21}''=\int_{\mathbb{R}^{d}}\frac{1}{\left|y\right|^{\alpha}}\frac{1}{\left|y\right|^{\alpha}}\left[1-e^{-\left|y\right|^{\alpha}}\right]\frac{1}{\left|y\right|^{\alpha}}\left[1-e^{-\left|y\right|^{\alpha}}\right]dy$$ Is is clear that integral $B_{21}''$ in the last expression is finite since in a neighborhood of $0$ the integrated expression is proportional to $\frac{1}{\left|y\right|^{\alpha}}$ and it is $O\left(\frac{1}{\left|y\right|^{3\alpha}}\right)$ as $\left|y\right|\rightarrow+\infty$ (recall that $\alpha<d<2\alpha$). Now only $B_{21}'$ needs to be evaluated$$B_{21}'\left(T\right)=C''T^{\frac{6-3\left(3-\frac{d}{\alpha}\right)-2\frac{d}{\alpha}}{2}}=C''T^{\frac{-3+\frac{d}{\alpha}}{2}}.$$ Hence it is obvious that $B_{21}'\left(T\right)\rightarrow0$ as $T\rightarrow0$ and so $\lim_{T\rightarrow0}B_{21}\left(T\right)=0$.\
Before proceeding to $B_{22}$ we will make the following estimation using inequality (\[neq: n(z,r,t)< strong\]) $$\left|\widehat{\left(n_{T}\left(\cdot,T-u,u\right)\right)^{2}}\right|\left(z\right)=\left|\int_{\mathbb{R}^{d}}\widehat{n}_{T}\left(x,T-u,u\right)\widehat{n}_{T}\left(z-x,T-u,u\right)dx\right|\leq$$ $$\frac{C}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|x\right|^{\alpha}}\left[1-e^{-u\left|x\right|^{\alpha}}\right]\frac{1}{\left|z-x\right|^{\alpha}}\left[1-e^{-u\left|z-x\right|^{\alpha}}\right]dx\leq$$ Substitution $xu^{1/\alpha}=y$ yields$$u^{2-\frac{d}{\alpha}}\frac{C}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|y\right|^{\alpha}}\left[1-e^{-\left|y\right|^{\alpha}}\right]\frac{1}{\left|zu^{1/\alpha}-y\right|^{\alpha}}\left[1-e^{-\left|zu^{1/\alpha}-y\right|^{\alpha}}\right]dy\leq\frac{C'}{F_{T}^{2}}u^{2-\frac{d}{\alpha}}$$ since the integral can be regarded as a convolution of $\mathcal{L}^{2}$ functions so it is bounded. This clearly implies$$\left|B_{22}\left(T\right)\right|\leq\frac{C'}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\int_{0}^{T}e^{-\left(T-u\right)\left|z\right|^{\alpha}}u^{2-\frac{d}{\alpha}}du\cdot\left|\widehat{n}_{T}\left(z\right)\right|dz\leq$$ $$C'\frac{T^{2-\frac{d}{\alpha}}}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\int_{0}^{T}e^{-\left(T-u\right)\left|z\right|^{\alpha}}du\cdot\left|\widehat{n}_{T}\left(z\right)\right|dz\leq$$ Using inequality (\[ineq: hat(n\_T)\]) we obtain$$C''\frac{T^{2-\frac{d}{\alpha}}}{F_{T}^{3}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\frac{1}{\left|z\right|^{\alpha}}\left(1-e^{-T\left|z\right|^{\alpha}}\right)\frac{1}{\left|z\right|^{\alpha}}\left(1-e^{-T\left|z\right|^{\alpha}}\right)dz=$$ Substituting $zT^{1/\alpha}=y$ we can rewrite the last expression as$$C''\frac{T^{5-\frac{d}{\alpha}}}{F_{T}^{3}T^{\frac{d}{\alpha}}}\int_{\mathbb{R}^{d}}\frac{1}{\left|y\right|^{\alpha}}\frac{1}{\left|y\right|^{\alpha}}\left(1-e^{-\left|y\right|^{\alpha}}\right)\frac{1}{\left|y\right|^{\alpha}}\left(1-e^{-\left|y\right|^{\alpha}}\right)dy.$$ The integral is finite (the same proof as for $B_{21}''$) and $$\frac{T^{5-\frac{d}{\alpha}}}{F_{T}^{3}T^{\frac{d}{\alpha}}}=T^{\frac{10-2\frac{d}{\alpha}-3\left(3-\frac{d}{\alpha}\right)-2\frac{d}{\alpha}}{2}}=T^{\frac{1-\frac{d}{\alpha}}{2}}$$ which yields $B_{22}\left(T\right)\rightarrow0$ as $T\rightarrow+\infty$
#### Limit of $B_{3}$ {#limit-of-b_3 .unnumbered}
\
Applying the Plancherel formula to $B_{3}\left(T\right)$ we get$$B_{3}\left(T\right)=\frac{1}{\left(2\pi\right)^{d}}\int_{0}^{\infty}\int_{\mathbb{R}^{d}}e^{-2t\left|z\right|^{\alpha}}\left(\widehat{n}_{T}\left(z\right)\right)^{2}dzdt=$$ $$\frac{1}{\left(2\pi\right)^{d}}\int_{\mathbb{R}^{d}}\left(\widehat{n}_{T}\left(z\right)\right)^{2}\int_{0}^{\infty}e^{-2t\left|z\right|^{\alpha}}dtdz=$$ $$\frac{1}{2\left(2\pi\right)^{d}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\left(\widehat{n}_{T}\left(z\right)\right)^{2}dz=$$ $$\frac{1}{2\left(2\pi\right)^{d}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\left(\int_{0}^{T}e^{-\left(T-u\right)\left|z\right|^{\alpha}}\widehat{\varphi}_{T}\left(z\right)\chi_{T}\left(T-u\right)du\right)^{2}dz=$$ $$\frac{1}{2\left(2\pi\right)^{d}}\frac{1}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\left(\int_{0}^{T}e^{-u\left|z\right|^{\alpha}}\widehat{\varphi}\left(z\right)\chi_{T}\left(u\right)du\right)^{2}dz=$$ Substituting $u'=u/T$$$\frac{1}{2\left(2\pi\right)^{d}}\frac{T^{2}}{F_{T}^{2}}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\left(\int_{0}^{1}e^{-Tu'\left|z\right|^{\alpha}}\widehat{\varphi}\left(z\right)\chi\left(u'\right)du'\right)^{2}dz=$$ $$\frac{1}{2\left(2\pi\right)^{d}}\frac{T^{2}}{F_{T}^{2}}\int_{0}^{1}\int_{0}^{1}\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}e^{-T\left(u_{1}+u_{2}\right)\left|z\right|^{\alpha}}\left(\widehat{\varphi}\left(z\right)\right)^{2}\chi\left(u_{1}\right)\chi\left(u_{2}\right)du_{1}du_{2}dz=$$ Let $z=\left[T\left(u_{1}+u_{2}\right)\right]^{-\frac{1}{\alpha}}y$$$\begin{aligned}
\frac{1}{2\left(2\pi\right)^{d}}\frac{T^{3-\frac{d}{\alpha}}}{F_{T}^{2}}\int_{0}^{1}\int_{0}^{1}\int_{\mathbb{R}^{d}}\left(u_{1}+u_{2}\right)\frac{1}{\left|y\right|^{\alpha}}e^{-\left|y\right|^{\alpha}}\left(\widehat{\varphi}\left(\left[T\left(u_{1}+u_{2}\right)\right]^{-\frac{1}{\alpha}}y\right)\right)^{2}\\
\left(u_{1}+u_{2}\right)^{-\frac{d}{\alpha}}\chi\left(u_{1}\right)\chi\left(u_{2}\right)du_{1}du_{2}dy\end{aligned}$$ Therefore by Lebesgue’s dominated convergence theorem we obtain the limit of $B_{3}\left(T\right)$$$\lim_{T\rightarrow+\infty}B_{3}\left(T\right)=\frac{1}{2\left(2\pi\right)^{d}}\int_{0}^{1}\int_{0}^{1}\int_{\mathbb{R}^{d}}\left(u_{1}+u_{2}\right)^{1-\frac{d}{\alpha}}\frac{1}{\left|y\right|^{\alpha}}e^{-\left|y\right|^{\alpha}}\left(\widehat{\varphi}\left(0\right)\right)^{2}\chi\left(u_{1}\right)\chi\left(u_{2}\right)du_{1}du_{2}dy=$$ $$\frac{\Gamma\left(\frac{d}{\alpha}-1\right)}{2^{d}\alpha\Gamma\left(\frac{d}{2}\right)\pi^{\frac{d}{2}}}\left\langle \lambda,\varphi\right\rangle ^{2}\int_{0}^{1}\int_{0}^{1}\left(u_{1}+u_{2}\right)^{1-\frac{d}{\alpha}}\chi\left(u_{1}\right)\chi\left(u_{2}\right)du_{1}du_{2}=$$ Integrating by parts$$\frac{K}{2V}\left\langle \lambda,\varphi\right\rangle ^{2}\int_{0}^{1}\int_{0}^{1}\left\{ -u_{1}^{h}-u_{2}^{h}+\left(u_{1}+u_{2}\right)^{h}\right\} \psi\left(u_{1}\right)\psi\left(u_{2}\right)du_{1}du_{2}$$
#### Limit of $B_{4}$ {#limit-of-b_4 .unnumbered}
\
Firstly, let us notice that $$B_{1}\left(T\right)+B_{2}\left(T\right)+B_{3}\left(T\right)=\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}V_{T}\left(x,t\right)^{2},$$ and hence $$\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}V_{T}\left(x,t\right)^{2}\rightarrow_{T\rightarrow+\infty}C.$$ Secondly by Fact \[fact: uniformaly\] and inequalities (\[neq: WT<TvT\]) and (\[ineq: v<n\]) we know $V_{T}\left(x\right)\rightarrow0$ uniformly as $T\rightarrow0$. Hence $g\left(W_{T}\left(x\right)\right)\leq\epsilon$ for $T$ sufficiently large so $$\left|B_{4}\left(T\right)\right|\leq\epsilon\int_{0}^{+\infty}\int_{\mathbb{R}^{d}}V_{T}\left(x,t\right)^{2},$$ which clearly implies that $B_{4}\left(T\right)\rightarrow0$ as $T\rightarrow+\infty$.
#### Putting the results together {#putting-the-results-together .unnumbered}
\
Combining the previous results we conclude$$\lim_{T\rightarrow+\infty}B\left(T\right)=\exp\left\{ \frac{MK}{4}\left\langle \lambda,\varphi\right\rangle ^{2}\int_{0}^{1}\int_{0}^{1}\left\{ -u_{1}^{h}-u_{2}^{h}+\left(u_{1}+u_{2}\right)^{h}\right\} \psi\left(u_{1}\right)\psi\left(u_{2}\right)du_{1}du_{2}\right\}$$ And finally by (\[eq:limit\_<X\_T,PHI>\])$$\lim_{T\rightarrow+\infty}A\left(T\right)B\left(T\right)=\exp\left\{ \frac{MK}{2}\left\langle \lambda,\varphi\right\rangle ^{2}\int_{0}^{1}\int_{0}^{1}c_{h}\left(u_{1},u_{2}\right)\psi\left(u_{1}\right)\psi\left(u_{2}\right)du_{1}du_{2}\right\} ,$$ where $c_{h}$ is the covariance function of fractional Brownian motion defined by (\[def: c - fractional\]). This Laplace functional defines a process $\tilde{X}_{T}$ corresponding to the Gaussian process $X_{T}$ with the covariance (\[def:kowariancja\_tw2\]) hence Theorem \[thm:equlibiurim\] is proved.
Appendix
========
The appendix contains a technical fact used in the main proof.
\[fact: calka1\] $$\begin{aligned}
& & \int_{\mathbb{R}^{2d}}\frac{1}{\left|y_{1}+y_{2}\right|^{\alpha}\left(\left|y_{1}\right|^{\alpha}+\left|y_{2}\right|^{\alpha}+\left|y_{1}+y_{2}\right|^{\alpha}\right)}\frac{1}{\left|y_{1}\right|^{\alpha}}\left[1-e^{-\left|y_{1}\right|^{\alpha}}\right]\\
& & \frac{1}{\left|y_{2}\right|^{\alpha}}\left[1-e^{-\left|y_{2}\right|^{\alpha}}\right]\frac{1}{\left|y_{1}+y_{2}\right|^{\alpha}}\left[1-e^{-\left|y_{1}+y_{2}\right|^{\alpha}}\right]dy_{1}dy_{2}<+\infty\end{aligned}$$
Substituting $x=y_{1}+y_{2}$ and $z=y_{2}$ we get$$\int_{\mathbb{R}^{2d}}\frac{1}{\left|x\right|^{\alpha}\left(\left|x\right|^{\alpha}+\left|z\right|^{\alpha}+\left|x-z\right|^{\alpha}\right)}\frac{1}{\left|x-z\right|^{\alpha}}\left[1-e^{-\left|x-z\right|^{\alpha}}\right]\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-\left|z\right|^{\alpha}}\right]\frac{1}{\left|x\right|^{\alpha}}\left[1-e^{-\left|x\right|^{\alpha}}\right]dxdz=$$ $$\int_{\mathbb{R}^{2d}}\frac{1}{\left|x\right|^{\alpha}}\frac{1}{\left|x\right|^{\alpha}}\left[1-e^{-\left|x\right|^{\alpha}}\right]\int_{\mathbb{R}^{d}}\frac{1}{\left|x\right|^{\alpha}+\left|z\right|^{\alpha}+\left|x-z\right|^{\alpha}}\frac{1}{\left|x-z\right|^{\alpha}}\left[1-e^{-\left|x-z\right|^{\alpha}}\right]\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-\left|z\right|^{\alpha}}\right]dzdx=\left(*\right)$$ Let us investigate now $$\int_{\mathbb{R}^{d}}\frac{1}{\left|x\right|^{\alpha}+\left|z\right|^{\alpha}+\left|x-z\right|^{\alpha}}\frac{1}{\left|x-z\right|^{\alpha}}\left[1-e^{-\left|x-z\right|^{\alpha}}\right]\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-\left|z\right|^{\alpha}}\right]dz \leq%\label{def:int_in_fact1}%$$ $$\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-\left|z\right|^{\alpha}}\right]\frac{1}{\left|x-z\right|^{\alpha}}\left[1-e^{-\left|x-z\right|^{\alpha}}\right]dz \leq c\int_{\mathbb{R}^{d}}\frac{1}{\left|z\right|^{\alpha}}\frac{1}{\left|z\right|^{\alpha}}\left[1-e^{-\left|z\right|^{\alpha}}\right]dz$$
The last integral is finite since in the neighborhood of $0$ the integrated function is $O\left(\frac{1}{\left|z\right|^{\alpha}}\right)$ and for big $\left|z\right|$ is $O\left(\frac{1}{\left|z\right|^{2\alpha}}\right)$. Going back to $\left(*\right)$ we obtain$$\left(*\right)\leq c_{2}\int_{\mathbb{R}^{d}}\frac{1}{\left|x\right|^{\alpha}}\frac{1}{\left|x\right|^{\alpha}}\left[1-e^{-\left|x\right|^{\alpha}}\right]<c_{3},$$ by the same reason as above.
The author would like to thank his supervisor - prof. Tomasz Bojdecki - for much appreciated help given in general introduction to the branching systems theory and in writing this paper. The author wishes to thank also prof. Luis Gorostiza for several helpful comments.
P. Billingsley, Convergence of Probability Measures., John Wiley&Sons, New York, 1968. T. Bojdecki, L.G. Gorostiza and S. Ramaswami, Convergence of $\mathcal{S}'$-valued processes and space time random fields, J. Funct. Anal. 66 (1986), pp. 21-41. T. Bojdecki, L.G. Gorostiza and A. Talarczyk, Sub-fractional Brownian motion and its relation to occupation times, Statist. Probab. Lett. 69 (2004), pp. 405-419. T. Bojdecki, L.G. Gorostiza and A. Talarczyk, Limit theorems for occupation time fluctuations of branching systems I: Long-range dependence, Stoch. Proc. Appl. 116 (2006), pp. 1-18. T. Bojdecki, L.G. Gorostiza and A. Talarczyk, Limit theorems for occupation time fluctuations of branching systems II: Critical and large dimensions Functional, Stoch. Proc. Appl. 116 (2006), pp. 19-35. T. Bojdecki, L.G. Gorostiza and A. Talarczyk, A long range dependence stable process and an infinite variance branching system, www.arxiv.org, math.PR/0511739 (2005). T. Bojdecki, L.G. Gorostiza and A. Talarczyk, Occupation time fluctuations of an infinite variance branching systems in large dimensions, www.arxiv.org, math.PR/0511745 (2005). M. Birkner and I. Zähle, Functional central limit theorems for the occupation time of the origin for branching random walks in $d\geq3$, Weierstraß Insitut für Angewandte Analysis und Stochastik, Berlin, preprint No. 1011 (2005). J.D. Deuschel and K. Wang, Large deviations for the occupation time of a Poisson system of independent Brownian particles, Stoch. Proc. Appl. 52 (1994), pp. 183-209. L.G. Gorositza and E.R. Rodrigues, A stochastic model for transport of particulate matter in air: an asymptotic analysis. Acta Appl. Math. 59 (1999), pp. 21-43. L.G. Gorostiza and A. Wakolbinger, Long time behavior of critical branching particle systems and its applications, CRM Proc. and Lect. Notes Vol. 5 (1994), pp. 119-137. L.G. Gorostiza and A. Wakolbinger, Persistence criteria for a class of critical branching particle systems in continuous time. Ann. Probab. 19 (1991), pp. 266-288. I. Iscoe, A weighted occupation time for a class of measure-valued branching processes, Probab. Th. Rel. Fields 71 (1986), pp. 85-116. I. Mitoma, Tightness of probabilities on $C\left(\left[0,1\right],\mathcal{S}'\right)$ and $D\left(\left[0,1\right],\mathcal{S}'\right)$, Ann. Probab. 11 (1983), pp. 989-999.
|
{
"pile_set_name": "ArXiv"
}
|
**The typical quantum Markovian processes**
**do not support the standard ensemble interpretation**
Jasmina Jekni' c-Dugi' c$^a$, Momir Arsenijevi' c$^b$, Miroljub Dugi' c$^{\ast b}$
$^a$University of Niš, Faculty of Science and Mathematics, Višegradska 33, 18000 Niš, Serbia
$^b$University of Kragujevac, Faculty of Science, Radoja Domanovi' ca 12, 34000 Kragujevac, Serbia
[**Abstract**]{} Assuming the unitary-only dynamics for the ”open system plus environment” composite quantum system, for certain typical Markovian models we find that the standard ignorance-based ensemble interpretation of the open system’s dynamics is not appropriate. Our analysis is purely technical and within the standard quantum-mechanical formalism without employing or calling for any additional assumptions or the formalism interpretation. Conceptual implications are presented in comparison with some related results, which include the so-called master equation unraveling in quantum optics, the continuous spontaneous localization theory as well as the concept of mixed states applied to single quantum systems.
Keywords: open quantum systems; quantum ensembles; quantum foundations
[**1. Introduction**]{}
Physical interpretation of mixed quantum states (statistical operators, i.e. density matrices) is a deep problem intimately linked with the foundations and interpretations of quantum theory as well as with the diverse applications in quantum optics, quantum metrology, complex quantum systems, quantum thermodynamics and the emerging quantum technology [@breuer; @brojer1; @Nilsen] (and the references therein). The problem remains even if the subtle conceptual mathematical problems on interpretation of probability [@Hrenikov; @ZurekNJP; @Landsman] are ignored.
Traditionally, mixed quantum states for a closed system are called ”proper mixtures” if the related mixed ensemble is simply a union (a ”mixture”) of certain pure subensembles thus giving rise to the subjective ignorance of the pure state of individual elements of the mixed ensemble [@breuer; @vonneuman; @Despa]. On the other hand, quantum states of systems quantum correlated with other quantum systems (or fields) are referred to as ”improper mixtures” due to impossibility to define quantum state of such systems [@Despa]. In this sense, the improper mixtures step aside of the standard, ignorance-based ensemble conceptualization thus presenting an aspect of the quantum measurement problem and the problem of the ”transition from quantum to classical” [@Giulini; @Slosi; @JACAWS] as well as of the quantum formalism interpretation [@ZurekNJP; @vonneuman; @Despa; @MW; @Bohm; @Ruth; @Fuchs]. As an alternative to the standard, ignorance-based interpretation it is sometimes proposed to assume that every single element of a mixed ensemble in the mixed state $\hat\rho$ is also in the mixed state $\hat\rho$ [@Nilsen; @ZurekNJP].
This kinematical picture regards the mixed quantum states in arbitrary instant of time. Its dynamical side however is even more complicated, going to both extremes, of statistical irrelevance of the system dynamics as appearing in the context of the so-called ”statistical typicality” [@Durr] on the one hand, and the necessity of explicit dynamical description of the mixed state of the system as typical for the open systems theory [@breuer] and the objective collapse theories [@Bassi] (and the references therein) on the other hand.
In this paper we perform a dynamical analysis of some typical Markovian processes [@breuer; @rivas] that regard the open systems quantum-correlated with their environments, in the context of the standard, ignorance-based ensemble interpretation of mixed states. Our analysis is purely technical and without calling for or assuming any interpretation of the standard formalism of the open quantum systems theory.
Unitary dynamics for the composite system ”system+environment ($S+E$)” is assumed, that implies Markovian dynamics for the $S$ system, whose density matrix $\hat\rho$ is exact, that is, not assuming any more-fundamental level of the system’s dynamics. In this context, the standard kinematical definition of the ”statistical ensemble” as a mixture of the pure subensembles is here extended for the dynamical purposes by the following assumption (actually requirement):
(EI) In order to have a [*dynamically*]{} complete ensemble interpretation of a mixed quantum state $\hat\rho$, it is required that every pure state $\vert\psi_k\rangle$, that admits a decomposition $\hat\rho = \sum_k p_k \vert\psi_k\rangle\langle\psi_k\vert$ ($\sum_kp_k=1$), is subject of the same dynamical map as the mixed state itself while remaining pure in every instant of time.
The requirement (EI) assumes existence of at least one decomposition of a mixed state $\hat\rho$ into pure states, which are both subject of the dynamical map valid for the mixed state $\hat\rho$ as well as remaining pure in every instant of time, formally: $\hat\rho(t)=\sum_k p_k(t)\vert\psi_k(t)\rangle\langle\psi_k(t)\vert, \sum_kp_k(t)=1,\forall{t}$. That is, the assumption (EI) imposes the requirement of the possibility dynamically to ”track” the pure subensembles in the Hilbert state space of the system for at least one decomposition of $\hat\rho$ into pure states.
The analysis presented in the following sections of this paper does not require the solution of the master equation. Actually, we slightly extend and use a condition for the pure-state-dynamics for the Markovian processes [@Sandu]. While this condition is known for a relatively long time and has been applied in the phenomenological modelling of the damped harmonic oscillator [@isar] (and the references therein), its use in the present context is, to the best of our knowledge, here presented for the first time.
Our findings are somewhat unexpected: for the considered models, there appear inevitable, [*a priori*]{} arguments against the standard (ignorance-based) ensemble picture of the system dynamics as established by (EI). For the initially pure states, the considered models impose necessity to consider every single element of the ensemble to be in a mixed state, which is the state of the mixed ensemble. While the standard ensemble picture is thus lost, a connection to the problems of conceptualization of probability and randomness [@Hrenikov; @ZurekNJP; @Landsman] is made requiring a separate careful analysis.
In Section 2, we introduce the main concepts and the basis of our considerations. Section 3 presents the main results of this paper from the technical side. Section 4 is discussion and we conclude in Section 5.
[**2. The pure-state conditions for the Markovian processes**]{}
[*Ensemble interpretation*]{} (EI) fully adopts the kinematical aspect of the standard ensemble interpretation of mixed states. That is, an instantaneous mixed state (density matrix) $\hat\rho$ admits a (nonunique) decomposition of the mixed state, $\hat\rho =\sum_i p_i \vert\psi_i\rangle\langle \psi_i\vert$, into the imagined pure subensembles presented by the normalized pure states $\vert \psi_i\rangle$, with the statistical weights $p_i$, while $\sum_i p_i=1$. The ensemble consists of the individual elements in pure states, each pure state $\vert\psi_i\rangle$ being chosen with the related probability $p_i$. If the mixed state dynamics is sufficiently [*smooth*]{} and presented by a [*linear*]{} dynamical map, $\mathcal{G}_{(t, t_{\circ})}: \hat\rho(t_{\circ}) \to \hat\rho(t)$, then (EI) assumes that every pure subensemble (and therefore every single element of the ensemble) undergoes the [*same*]{} dynamics, in the most general form of $\mathcal{G}_{(t,t_{\circ})}: p_i(t_{\circ})) \vert \psi_i(t_{\circ})\rangle\langle\psi_i(t_{\circ})\vert \to p_i(t) \vert \psi_i(t)\rangle\langle\psi_i(t)\vert$, while $\sum_i p_i(t)=1,\forall{t\ge t_{\circ}}$.
The unitary dynamics admits (EI). If the mixed state is (nonuniquely) initially decomposed as $\hat\rho=\sum_k p_k \vert\psi_k\rangle\langle\psi_k\vert$, then linearity of the unitary evolution ”sustains” the ensemble decomposition, $\hat\rho(t)=\sum_k p_k \hat U(t)\vert\psi_k\rangle\langle\psi_k\vert \hat U^{\dag}(t)\equiv \sum_k p_k \vert\psi_k(t)\rangle\langle\psi_k(t)\vert$. That is, every pure subensemble undergoes the same unitary dynamics, which is supposed for the mixed state $\hat\rho$, and remains to be pure in every instant of time. For the linear non-unitary dynamical maps, one might expect an analogous possibility. However, as we show below, this is not the case already for some quantum Markovian processes. That is, for the considered Markovian processes, we prove that there are not pure states that would stay pure in the course of the dynamics, which is supposed for the mixed state $\hat\rho$. As emphasized above, in this sense we point out inadequacy of (EI) for the considered Markovian processes.
By Markovian process we assume a differentiable and CP-divisible dynamical map [@rivas] that admits the Lindblad form of the master equation of the general form (in the Schr" odinger picture):
$${d\hat\rho(t)\over dt} \equiv \mathcal{L}[\hat\rho(t)] = -{i\over \hbar} [\hat H,\hat\rho(t)] + \sum_i \gamma_i \left(\hat L_i \hat\rho(t) \hat L_i^{\dag} - {1\over 2}\{\hat L_i^{\dag} \hat L_i,\hat\rho(t)\}\right).$$
In general, the damping factors $\gamma_i\ge 0$ and all the Lindblad operators ($\hat L_i$) may carry time dependence, while the time dependence of the Hamiltonian $\hat H$ may be allowed in the case of the week external field(s) [@rivas].
For a sufficiently smooth dynamics, arbitrary decomposition $\hat\rho(t) = \sum_k p_k(t)$ $\vert\psi_k(t)\rangle$ $\langle\psi_k(t)\vert$, $\sum_i p_k(t)=1, \forall{t}$, in general, implies:
$${d\hat\rho(t)\over dt} =
\sum_k \left(
{{dp_k(t)\over dt} \vert\psi_k(t)\rangle\langle\psi_k(t)\vert
+
p_k(t) {d\over dt} (\vert\psi_k(t)\rangle\langle\psi_k(t)\vert)}
\right).$$
The knowledge of the (not necessarily orthogonal but normalized) $\vert\psi_k\rangle$s determines also the ”populations” $p_k$ thus presenting eq.(2) as a mathematically well defined problem.
Placing $p_k=1$ and $p_{k'}=0,\forall{k'\neq k}$ in eq.(1) gives rise to the pure-state dynamics, $d\hat\rho/dt = d(\vert\psi_k\rangle\langle\psi_k\vert)/dt$, which, [*if assumed to preserve purity*]{} of the states in the course of the dynamics, admits for eq.(1) the following form for a pure state in every instant of time [@Sandu]:
$${d\over dt} (\vert\psi\rangle\langle\psi\vert) = -\imath (\mathcal{H}\vert\psi\rangle\langle\psi\vert - \vert\psi\rangle\langle\psi\vert \mathcal{H}^{\dag}),$$
where, for simplicity, we dropped the index $k$, and:
$$\mathcal{H} = \hat H + \imath \sum_i \gamma_i \left(
\langle \hat L_i^{\dag}\rangle \hat L_i - {1\over 2} \langle \hat L_i^{\dag}\hat L_i\rangle - {1\over 2} \hat L_i^{\dag} \hat L_i
\right),$$
while $\langle\ast\rangle \equiv \langle\psi\vert\ast\vert\psi\rangle = tr(\vert\psi\rangle\langle\psi\vert\ast)$ and we assume $\hbar=1$. A derivation of equations (3) and (4) for the semigroup maps can be found in [@Sandu]; the main steps in the derivation, also applying for the general Markovian case, are presented in Appendix A. Solutions to eq.(3)–if such exist–determine also the ”populations” $p_k$ in eq.(2) as presented in Appendix A. Although probably obvious, it is worth emphasizing: eq.(3) regards the [*same*]{} dynamical map as eq.(1) [*while*]{} assuming preservation of purity of the initially pure state of the system.
Now, the knowledge of at least two pairs $(p_k(t),\vert\psi_k(t)\rangle)$, $k=1,2,...$ ”unravels” the imagined (although non-unique) ensemble composition of the mixed state $\hat\rho$, which is subject of the dynamics eq.(1). That is the idea we follow in this paper: to investigate dynamics of certain Markovian processes so as to see if any ensemble ”structure”, i.e. the set $\{p_k(t), \vert\psi_k(t)\rangle,k=1,2,...\}$, of the state $\hat\rho$ could in principle appear.
In Discussion section we emphasize that our analysis is more stringent than the related analysis based on the requirement of excluding the superluminal communication in quantum information science [@Gisin1; @Gisin2; @Bassi2] as well as it does not have much in common with the so-called piece-wise-deterministic processes (PDP) unraveling [@breuer; @Carmichel].
[**2.1 The pure-state condition**]{}
For the pure states applies $\hat\rho^2 = \hat\rho$, which is equivalent with $tr\hat\rho^2=1$. Hence the equality $d tr(\hat\rho^2)/dt=0$, i.e. $tr(\hat\rho d\hat\rho/dt)=0$. We are interested in the Markovian processes, which are the completely positive and trace preserving processes, i.e. leave invariant (represent contractions on) the Banach space of the statistical operators [@rivas]. Therefore, with the use of eq.(1), we obtain the following condition for the pure-state (i.e. for the case $\hat\rho^2(t)=\hat\rho(t),\forall{t}$) dynamics for the Markovian processes:
$$tr\left(\hat\rho {d\hat\rho\over dt}\right)=tr(\hat\rho\mathcal{L}[\hat\rho])= \sum_i \gamma_i \left( tr(\hat\rho \hat L_i\hat\rho \hat L_i^{\dag}) - tr(\hat\rho \hat L_i^{\dag}\hat L_i) \right) = 0,$$
where we used the commutation under the tracing out operation. Equivalently, applying the same procedure to eq.(3) follows $\langle\mathcal{H}\rangle = \langle\mathcal{H}^{\dag}\rangle$, which leads to eq.(5).
Equation (5) is well known for the dynamical semigroups, e.g., [@Sandu; @isar] (and the references therein) and here is extended to every Markovian process when the time dependence may appear for both $\gamma$s as well as for the Lindblad operators in eq.(1). Equation (5) is often written in the form of:
$$\sum_i \gamma_i \langle\psi\vert \hat L_i\vert\psi\rangle\langle\psi\vert \hat L_i^{\dag} \vert \psi\rangle = \sum_i \gamma_i \langle\psi\vert \hat L_i^{\dag}\hat L_i\vert\psi\rangle,$$
and applies for any instant in time, $t\ge 0$, as well as for every Markovian process [@Sandu; @isar]. That is, equation (6) is [*necessary*]{} in order to have the pure-state [*continuous*]{}-in-time dynamics, $\hat\rho^2(t)=\hat\rho(t)$, for eq.(1) as well as for eq.(3) [*in every instant of time*]{}.
Then unraveling a mixed state, which satisfies eq.(1), in accordance with (EI) assumes continuous-in-time dynamics for a set of pure states $\vert\psi_k(t)\rangle$, which satisfy eq.(3), and remain pure in the course of the system dynamics.
[**2.2 The ensemble-interpretation consistency**]{}
Another derivation of the expression eq.(6) follows from the requirement of consistency of the ensemble interpretation (EI). That is, assuming validity of eq.(3) for every pure state $\vert\psi_k\rangle$ in a decomposition $\hat\rho = \sum_k p_k(t) \vert\psi_k(t)\rangle\langle\psi_k(t)\vert$ gives rise to eq.(6) as follows.
Substitution of eqs.(3) and (4) into eq.(2), after some algebra, gives (we assume $\hbar=1$): $$\begin{aligned}
{}&\dot{\hat\rho} = -\imath [\hat H, \hat \rho] - {1\over 2} \sum_{j} \gamma_j\{\hat L_j^{\dag}\hat L_j,\hat\rho \} + \sum_k \dot{p}_k \vert\psi_k\rangle\langle\psi_k\vert+\\
&+\sum_{k,j} p_k\gamma_j \left[\left(\langle\hat L_j^{\dag}\rangle_k\hat L_j-\frac{\langle\hat L_j^{\dag}
\hat L_j\rangle_k}{2}\right)\vert \psi_k\rangle\langle \psi_k\vert+\vert\psi_k\rangle\langle\psi_k\vert\left(\langle\hat L_j^{\dag}\rangle_k\hat L_j - \frac{\langle\hat L_j^{\dag} \hat L_j\rangle_k}{2}\right)
\right],
\end{aligned}$$
where $\langle\ast\rangle_k\equiv\langle\psi_k\vert\ast\vert\psi_k\rangle$ and the curly brackets denote the anticommutator.
Now, due to eq.(1), validity of eq.(7) implies the equality:
$$\begin{aligned}
&{}\sum_j \gamma_j \hat L_j\hat\rho\hat L_j^{\dag} = \sum_k \dot{p}_k \vert\psi_k\rangle\langle\psi_k\vert +\sum_{k,j} p_k
\gamma_j \left[
\left(\langle\hat L_j^{\dag}\rangle_k\hat L_j - {1\over 2}\langle\hat L_j^{\dag} \hat L_j\rangle_k\right)\vert \psi_k\rangle\langle \psi_k\vert\right.+\\ &\left.+\vert\psi_k\rangle\langle\psi_k\vert \left(\langle\hat L_j^{\dag}\rangle_k\hat L_j - {1\over 2}\langle\hat L_j^{\dag} \hat L_j\rangle_k\right)
\right],
\end{aligned}$$
which is the set of equations for the statistical weights $p_k$.
By tracing out eq.(8) and rearranging the terms, with the use of the normalization condition $\sum_k p_k(t)=1,\forall{t}$, it follows the condition for the ensemble-interpretation consistency:
$$\sum_{j,k} \gamma_j p_k\left(
\langle\hat L_j^{\dag}\hat L_j\rangle_k - \langle\hat L_j^{\dag}\rangle_k \langle\hat L_j\rangle_k\right) = 0.$$
It is probably obvious that the condition eq.(9) reduces to equation (6): by returning the index $k$ into eq.(6) and multiplying eq.(6) by $p_k$, summation over the index $k$ gives eq.(9). That is, satisfiability of eq.(6) (i.e. of the set of equations of the form of eq.(6) for every $k$) directly leads to validity of eq.(9).
[**2.3 Some general remarks**]{}
Introduce an instantaneous normalized state $\vert\psi_k\rangle$ and define the following two vectors: $\vert u\rangle=\vert\psi_k\rangle$ and $\vert v\rangle=\hat L_i\vert\psi_k\rangle$. Then the Cauchy-Schwarz (CS) inequality, $\vert\langle u\vert v\rangle\vert^2\le \langle u\vert u\rangle \langle v\vert v\rangle$, implies:
$$\langle\psi_k\vert \hat L_j\vert \psi_k\rangle \langle\psi_k\vert \hat L_j^{\dag}\vert\psi_k\rangle \le \langle\psi_k\vert \hat L_j^{\dag}\hat L_j\vert\psi_k\rangle.$$
Bearing in mind that, for Markovian processes, $\gamma_j(t)\ge0, \forall{j,t}$, eq.(6) can be satisfied if and only if (iff) equality in eq.(10) is satisfied for every index $k$ and every index $j$ in every instant of time $t$.
Due to the CS inequality, the equality sign in eq.(10) may appear iff the two vectors are collinear, i.e. iff $\hat L_j\vert\psi_k\rangle = l_{jk}\vert\psi_k\rangle$ with the eigenvalues $l_{jk}$–for every $k$ and every $j$. That is, eq.(6) can be satisfied iff all the pure states $\vert\psi_k\rangle$ are common eigenstates of all the Lindblad operators $\hat L_j$ in every instant of time $t$. Needless to say, this is a stringent [*necessary*]{} condition for the Markovian processes to allow for the ensemble interpretation (EI).
We tried to devise sufficient conditions for the (EI) for Markovian processes without success. Accordingly, we believe that such conditions cannot be unambiguously formulated. For this reason, in the next section, we focus our considerations to some typical models of the Markovian processes thus exhibiting the condition eq.(6) at work.
[**3. Analysis of some Markovian processes**]{}
Consider some typical examples of the homogeneous Markovian processes in the context of the condition eq.(6).
[**3.1 The two-level system**]{}
Choose the system’s basis states denoted $\vert \pm\rangle$ and construct the observable $\hat\sigma_z \vert\pm\rangle=\pm\vert\pm\rangle$. Introduce also the nonhermitian operators $\hat\sigma_-=\vert -\rangle\langle +\vert =\hat\sigma_+^{\dag}$.
The homogeneous Markovian master equation (in the interaction picture) for such system is given in the general form [@breuer; @rivas]:
$${d\hat\rho\over dt} = \gamma_1 \left(\hat\sigma_+ \hat\rho \hat\sigma_- -{1\over 2} \{\hat\sigma_-\hat\sigma_+,\hat\rho\}\right) + \gamma_2 \left(\hat\sigma_- \hat\rho \hat\sigma_+ -{1\over 2} \{\hat\sigma_+\hat\sigma_-,\hat\rho\}\right)$$
where the curly brackets denote the anticommutator. Physically, for $\gamma_1=0$, we have the case of a single-qubit amplitude damping process [@Nilsen], or the model of the two-level atomic system in contact with the environment on the absolute zero temperature [@breuer; @rivas]. For the atomic system,
$$\gamma_1=\gamma_{\circ}N, \gamma_2 = \gamma_{\circ}(N+1),$$
where
$$N=(e^{\beta\hbar\omega_{\circ}}-1)^{-1}\ge 0,$$
is the mean number of quanta for the environmental mode of frequency $\omega_{\circ}$ on the inverse temperature $\beta$, while the real $\gamma_{\circ}>0$.
Let us separately consider the two cases: when $\gamma_1=0$, and when $\gamma_i\neq 0, i=1,2$.
[*The case $\gamma_1=0$*]{}. For this case of the absolute zero, $T=0$ (and $N=0$), the master equation (11) reduces to the second term, $\gamma_2=\gamma_0$, with the only one Lindblad operator, $\hat L=\hat\sigma_-$. Then eq.(6) easily gives the pure-state dynamics condition to read:
$$p_+(1-p_+)=p_+,$$
where $p_+\equiv \vert\langle +\vert \psi\rangle\vert^2$. That is, eq.(14) reveals the only one possible solution: $p_+=0$ and therefore the only one pure state allowed for every instant of time–the ground state $\vert -\rangle$; in accordance with the CS inequality (cf. Section 2.3), this state is also an eigenstate of the Lindblad operator $\hat\sigma_-$.
[*The case*]{} $\gamma_i\neq 0,i=1,2$. Now consider the full master equation (8) with both nonzero damping factors $\gamma_i$. In this case there are the two Lindblad operators, $\hat\sigma_-$ and $\hat\sigma_+$. Due to Section 2.3 and nonexistence of even a single common eigenstate for $\hat\sigma_+$ and $\hat\sigma_-$, we know in advance that the pure-state condition eq.(6) cannot be satisfied. Nevertheless, for the sake of completeness as well as of demonstrating the condition (6) for this model, we proceed as follows.
A straightforward application of eq.(6) with the definitions eq.(12) and (13) gives the pure-state dynamics condition:
$$(2N+1)p_+^2 - 2Np_+ + N = 0,$$
which leads to the solutions for the $p_+$:
$$p_{+1,2} = {N\pm \sqrt{-N^2-N}\over2N+1}.$$
Again, from eq.(16) follows the [*only one*]{} solution: $N=0$, which implies $p_+=0,\forall{t}$. Needless to say, $N=0$ is equivalent with $T=0$–which is analysed above–thus emphasizing the absence of even a single pure state satisfying eq.(6) for the finite temperature.
[**3.2 The damped linear harmonic oscillator**]{}
The one-dimensional (i.e. one-mode) harmonic oscillator damped by the bosonic heat bath is described by the widely used master equation (in the interaction picture) [@breuer; @rivas]:
$${d\hat\rho\over dt} = \gamma_{\circ}(N+1) \left(
\hat a\hat\rho \hat a^{\dag} - {1\over 2} \{\hat a^{\dag}\hat a,\hat\rho\}
\right) + \gamma_{\circ}N \left(
\hat a^{\dag}\hat\rho \hat a- {1\over 2}\{\hat a \hat a^{\dag},\hat\rho\}
\right),$$
where appear the standard bosonic operators, $[\hat a, \hat a^{\dag}]=\hat I$, with the $N$ defined by eq.(13), while the real $\gamma_{\circ}>0$.
Again, for the bath on the absolute zero, the second term on the rhs of eq.(17) disappears.
[*The case $T=0$*]{}. Then $N=0$ and hence the master equation reduces to the first term with only one Lindblad operator, $\hat L=\hat a$. For this case the condition eq.(6) reads:
$$\langle\psi\vert \hat a\vert\psi\rangle \langle\psi\vert \hat a^{\dag}\vert\psi\rangle = \langle\psi\vert \hat a^{\dag} \hat a\vert\psi\rangle.$$
The solution to eq.(18) is obvious, also following from the CS inequality: every eigenstate of the Lindblad operator $\hat a$ is the solution to eq.(15). That is, every ”coherent state” $\vert \alpha\rangle$ satisfying the eigen-problem $\hat a\vert\alpha\rangle = \alpha\vert\alpha\rangle$ also satisfies the condition eq.(18) for the pure-state dynamics. In the context of (EI), this means that the only possibility is to have a continuous transition between the ”coherent states”, of the form $\vert\alpha(t)\rangle$, such that $\hat a\vert\alpha(t)\rangle = \alpha(t)\vert\alpha(t)\rangle, \forall{t}$.
Placing a coherent state $\vert\alpha(t)\rangle$ into eq.(3) leads to non-satisfiability of eq.(3). Possibly the simplest way to show this is to employ the Sudarshan-Glauber $P$-representation, $\hat\rho(t) = \int d^2\alpha P(\alpha,\alpha^{\ast},t) \vert\alpha\rangle\langle\alpha\vert$, with the quasi-probability distribution $P(\alpha,\alpha^{\ast},t)$ satisfying $\int d^2\alpha P(\alpha,\alpha^{\ast},t)=1$. Eq.(3) assumes the constraint of purity of the state, $P(\alpha,\alpha^{\ast},t)=\delta^{(2)}(\alpha-\alpha(t)), \forall{t}$; $\delta^{(2)}(\ast)$ is a complex-valued Dirac delta-function. Now performing the standard procedure for the $P$-representation leads essentially to a repetition of the well-known procedure that gives rise to the conclusion that the distribution $P(\alpha,\alpha^{\ast},t)$ cannot be a Dirac delta-function for any instant of time $t$, except for the initial instant of time $t=0$; to this end see e.g. eq.(3.337) in [@breuer]. The only exception to this finding is the initially zero state, $\vert\alpha_{\circ}\rangle\equiv\vert \alpha(t=0)=0\rangle$, for the case $\hat H=\omega\hat a^{\dag}\hat a$,. However, this state returns the rhs of eq.(3) to equal zero, thus being a stationary state for the process ($\hbar=1$):
$$\mathcal{H}\vert\alpha_{\circ}\rangle=\left(\omega \hat a^{\dag}\hat a + \imath\gamma \langle\alpha_{\circ}\vert\hat a^{\dag}\vert \alpha_{\circ}\rangle \hat a
-{\imath\gamma\over 2} \langle\alpha_{\circ}\vert\hat a^{\dag}\hat a\vert \alpha_{\circ}\rangle -
{\imath\gamma\over 2} \hat a^{\dag}\hat a
\right)
\vert\alpha_{\circ}\rangle =0.$$
Therefore, there is not even a single dynamically-evolving pure state satisfying eq.(3) and eq.(6) for the process on $T=0$.
[*The case $T>0$*]{}. Then both terms in the master equation are nonzero ($N>0$) giving rise to the two Lindblad operators, $\hat a$ and $\hat a^{\dag}$. It is easy to show that the pure-dynamics condition (6) now gives:
$$\langle \hat a^{\dag} \hat a\rangle + {N\over(2N+1)} = \langle \hat a\rangle \langle \hat a^{\dag}\rangle.$$
According to the Cauchy-Schwarz inequality, $\langle \hat a^{\dag} \hat a\rangle \nless \langle \hat a\rangle \langle \hat a^{\dag}\rangle$, so we conclude that it is necessary to have the condition $N=0$ satisfied–in contradiction with the assumption of nonzero temperature of the thermal bath. That is, for the finite temperature, there is not a single pure state that could fulfill the condition eq.(6) for the pure-state dynamics of the damped harmonic oscillator.
[**3.3 Comments**]{}
Explicit use of the conditions eq.(6) and eq.(3) is performed in Appendix B for some other models of the quantum Markovian processes known in the literature. For all the considered models we obtain the same conclusion drawn from Sections 3.1 and 3.2: there is not even a single nontrivial (e.g. nonstationary) state that could satisfy eq.(6) and eq.(3). Therefore we find the strong and [*a priori*]{} arguments against the (EI) for the considered dynamical models without resorting to any interpretation of the formalism.
Equation (6) is necessary for equation (3) to be valid. Therefore nonsatisfiability of eq.(6), or of eq.(3) if eq.(6) is satisfied (cf. the case of $T=0$ in Section 3.2 and also eq.(B.13) in Appendix B), exhibits that an initially pure state (which is not a stationary state for the process) dynamically becomes mixed. This finding, of course, also follows from the general solution of the related master equation but the use of eq.(6) is both universal and easy to perform compared to the task of solving the (Markovian) master equations, especially for the continuous variable systems with the infinite number of the degrees of freedom.
It is worth stressing albeit probably obvious: our findings may be expected to fail for the non-Markovian processes, for which some damping factors $\gamma_i$ in eq.(1) may take negative values, at least for some instants of time.
[**4. Discussion**]{}
Alternatives to the dynamical ensemble interpretation (EI), Introduction section, that would incorporate the kinematical, ignorance-based, ensemble interpretation, might be possible. Nevertheless, alternative(s) would discard the ensemble ”unraveling”, which is formally presented by $\hat\rho(t) = \sum_i p_i(t)\vert\psi(t)\rangle\langle\psi(t)\vert$. In such a case we are not aware of a sense in which it would be justified to deal with the pure subensembles of a mixed ensemble in the mixed state $\hat\rho$. For this reason we stick to (EI) in our considerations.
Our results are concordant with the general task of defining the ”pointer basis states” (PBS) in quantum decoherence theory [@Giulini; @Slosi; @zanardi; @boixo]. Exact PBSs do not change in time and represent exact stationary states for the open system’s dynamics. Approximate PBSs known for some Markovian models [@Giulini; @Slosi; @JACAWS] are readily described by approximate equalities in eq.(6) and eq. (3). While, in the context of our [*exact*]{} considerations, those states vary with time and approximately preserve the purity, in the context of our considerations their time dependence becomes irrelevant: within the margin of the approximation (of the purity), those states are stationary states analogous with the exact stationary states and therefore of no interest for the task of the mixed state (dynamical) unraveling.
It may be argued that equation (3) as a nonlinear deterministic Schr" odinger equation is physically irrelevant, since such kind of dynamics can lead to superluminal communication [@Gisin1; @Gisin2; @Bassi2]. The superluminal communication may be performed due to the observably distinguishable nonlinear dynamics of a pair of different pure-state decompositions of the mixed state. However, this argument [@Gisin1; @Gisin2; @Bassi2] is essentially kinematic in the sense it does not answer the following question: may it be [*dynamically*]{} allowed to have [*only one*]{} physically relevant decomposition of $\hat\rho$ into pure states? To this end, Section 3 gives a clear and definite answer for the considered Markovian models: there is not even a single dynamically relevant decomposition of $\hat\rho$ into pure states. Therefore the superluminal communication–which requires at least two physically relevant dynamical decompositions–is [*a priori*]{} excluded.
While the task of ”unraveling” the mixed states subject of Markovian master equations may sound like the program of the master equations ”unraveling” in the context of the so-called piece-wise-deterministic processes (PDP) models [@breuer] (and the references therein), the similarity is only superficial and partially in the narrative [@breuer; @brojer1; @Carmichel]. Actually, the PDP unraveling regards the completely different physical situation, which endows the continuous-in-time dynamics eq.(1) by the occasional instantaneous state collapses. This is physically a tripartite and hybrid model: the open system interacts with its environment, which, in turn, is monitored by a [*classical*]{} apparatus, whose actions on the environment (the ”probe” of the measurement scheme) induce the collapse to the open system’s dynamics [@breuer]. In contrast to this picture, we are concerned with the unitary-only dynamics. The formal side of this discrepancy is also interesting.
Our considerations regard [*every*]{} Markovian (not necessarily ”homogeneous”) process so that the pure-state condition eq.(6) is satisfied along with the dynamics eq.(3), which is equivalent with the Schr" odinger-like dynamics ($\hbar=1$):
$${d\vert\psi(t)\rangle\over dt}= -\imath \mathcal{H}\vert\psi(t)\rangle,$$
and $\mathcal{H}$ is defined by eq.(4). On the other hand, the deterministic part of the PDP unraveling [@breuer] is also defined by eq.(21) [*but*]{} with the different ”Hamiltonian”, which reads:
$$\mathcal{H}'= \hat H-{\imath\over 2}\sum_i \gamma_i \hat L_i^{\dag}\hat L_i + {\imath\over 2}\sum_i \gamma_i\langle \hat L_i^{\dag} \hat L_i\rangle
.$$
Compared to eq.(4), eq.(22) is missing the term $\imath\sum_i\langle \hat L_i^{\dag}\rangle \hat L_i$ while having the opposite sign for the last term.
Therefore we may say that the possibility to ”unravel” the master equations for the PDP processes [@breuer] does not have much in common with our considerations based on eq.(3). That is, our conclusions are not limited by the PDP method or by the usefulness of its formalism [@breuer; @brojer1; @Carmichel].
On the ensemble-interpretation side, the physical picture of the PDP models is well defined. Every pure state remains to be pure in the course of the system dynamics thus defining the ”true” (”real”) pure ensembles, whose statistically-weighted, i.e. statistically averaged, union represents the mixed quantum state $\hat\rho$, which is subject of a linear master equation. Then, in contrast to (EI), the pure states do [*not*]{} undergo the same dynamics as the density matrix for the process. This is a common feature for the continuous spontaneous localization (CSL) theories [@Bassi] (and the references therein) that also assume stochastic processes capable of producing the state collapse. Hence the same ensemble interpretation for the PDP and the CSL theories: ignorance of the pure state dynamics regards the ”hidden” level of the system dynamics very much like the ignorance of the individual molecules velocities in a classical gas; in both cases statistical averaging provides the effective dynamics on the level of the statistical operator, i.e. of the collective behavior of the gas, respectively.
Our considerations assume unitary dynamics for the composite system ”system+environment” thus making the dynamics eq.(1) exact, i.e. without any additional [@breuer; @brojer1; @Carmichel] or underlying [@Gisin1; @Gisin2; @Bassi2] stochastic processes. Accordingly, the desired ensemble picture of the process is exact, in the sense that it does not follow as an average of some more fundamental ensemble. This is precisely what is established by the requirement (EI), Introduction section. Nonexistence of the solutions of eq.(3) and eq.(6) implies, that for the considered Markovian models, the initial pure states dynamically become mixed and remain to be mixed during the time evolution–except possibly in the asymptotic limit $t\to\infty$ (cf., e.g., the case of $T=0$ in Section 3.1).
Thence our findings suggest a nonstandard ensemble interpretation of the considered Markovian processes. Actually, by following the standard wisdom, for the initially pure state $\vert\psi\rangle$ every element of the ensemble is assumed to be in the same pure state $\vert\psi\rangle$. Now, the fact that no pure state can remain pure in the course of the open system’s dynamics, and, in the context of the unitary-only dynamics for the $S+E$ system, every single system in the ensemble of the open system $S$ undergoes the [*same*]{} dynamics imposed by eq.(1), it seems unavoidable to assume that every single element of the initially pure ensemble is in a mixed state $\hat\rho(t)$, which is also the state of the mixed ensemble in a later instant of time $t$. That is, instead of a mixture $\{\vert k\rangle, p_k, k=1,2,...\}$ of pure (instantaneous) states, the ensemble described by the mixed state $\hat\rho$ is ”homogeneous” in the sense that every single element of the ensemble is in the same mixed state $\hat\rho$. In other words, the total ”ignorance” carried by the mixed state is equally ”imprinted” in every [*single*]{} system of the mixed ensemble.
At this point our analysis tackles the interpretational issues of both mathematical [@Hrenikov; @ZurekNJP; @Landsman] as well as of quantum-physical foundations [@ZurekNJP; @vonneuman; @Despa; @MW; @Bohm; @Ruth; @Fuchs] that are out of the scope of the present paper.
[**5. Conclusion**]{}
Traditionally, statistical meaning of the mixed quantum states (statistical operators, i.e. the density matrices) is mainly a matter of interpretation. In this paper we provide a conclusive technical finding that, for some typical and often used models of the quantum Markovian processes, the statistical operator does not allow the standard, ignorance-based ensemble interpretation. The interpretive implications of this finding remain open yet.
[**Acknowledgements**]{}. This work is financially supported by Ministry of Science Serbia, contract no 171028, and in part for MD by the ICTP–-SEENET-MTP project NT-03 Cosmology–Classical and Quantum Challenges. We are indebted to Aurelian Isar for the discussion on the subject.
[**Appendix A**]{}
From a decomposition $\hat\rho = \sum_k p_k \vert\psi_k\rangle\langle\psi_k\vert$ follows:
$$\rho_{ii} = \sum_k p_k \vert\langle i\vert\psi_k \rangle\vert^2, \tag{A.1}$$
where $\rho_{ii} = \langle i\vert\hat\rho\vert i\rangle$ for an orthonormalized basis $\vert i\rangle$ of the system’s Hilbert state space; $i=1,2,3,...,n$. Since existence of the $\rho_{ii}$s is guaranteed by eq.(1), and the states $\vert \psi_k\rangle, k=1,2,3,...,m\le n$, are solutions to eq.(3) (assuming that such exist), equation (A.1) is well-defined algebraic equation for the unknown ”populations” $p_k$ for every instant in time. This establishes consistency of the mathematical task posed by eq.(2), without a need to refer to the derivatives of $p_k$ appearing in eq.(2).
For completeness, we outline the main steps leading to eqs.(3) and (4) as it can be found in [@Sandu].
Assume the pure state $\hat\rho(t)=\vert\varphi(t)\rangle\langle\varphi(t)\vert, \forall{t}$. Introduce arbitrary pure state (a time-independent vector) $\vert\theta\rangle$. Then $\hat\rho\vert\theta\rangle=\langle\varphi\vert\theta\rangle \vert\varphi\rangle$ so it is straightforward to obtain:
$${d\over dt}\hat\rho\vert\theta\rangle = \mathcal{L}[\hat\rho] \hat\rho \vert\theta\rangle + \hat\rho\mathcal{L}[\hat\rho] \vert\theta\rangle. \tag{A.2}$$
Now the use of $\hat\rho \ast \hat\rho = (tr \ast\hat\rho) \hat\rho$ and the, possibly time dependent, Liouvillian $\mathcal{L}$ from eq.(1) easily lead to the Schr" odinger-like equation:
$${d\vert\varphi\rangle\over dt} = -{\imath\over\hbar}\mathcal{H} \vert\varphi\rangle, \tag{A.3}$$
which is equivalent with eq.(3) and eq.(4).
[**Appendix B**]{}
Following the considerations in Section 3, we analyse some standard Markovian master equations. Our analysis formally includes the master equations that regard the isolated quantum systems, cf. the models [*H*]{} and [*I*]{} below. The findings justify the findings of Sections 3.1 and 3.2.
[*A. A three-level atom*]{}
A three-level atom with the (non-degenerate) energies $E_1 < E_2 < E_3$ is endowed by the dipole transitions which exclude the transitions between the two lower levels. Defining the operators $\hat \sigma_{ij}\equiv\vert i\rangle\langle j\vert, i\neq j=1,2,3$, the master equation (in the interaction picture) follows from the quantum-optical master equation [@breuer; @Gardiner]:
$$\nonumber
\dot{\hat\rho} = \gamma_1 (N_1+1) \left(\hat\sigma_{13}\hat\rho\hat\sigma_{31}-{1\over 2}\{\hat\sigma_{31}\hat\sigma_{13},\hat\rho\}\right) +
\gamma_1 N_1 \left(\hat\sigma_{31}\hat\rho\hat\sigma_{13}-{1\over 2}\{\hat\sigma_{13}\hat\sigma_{31},\hat\rho\}\right) +$$
$$\gamma_2 (N_2+1) \left(\hat\sigma_{23}\hat\rho\hat\sigma_{32}-{1\over 2}\{\hat\sigma_{32}\hat\sigma_{23},\hat\rho\}\right) +
\gamma_2 N_2 \left(\hat\sigma_{32}\hat\rho\hat\sigma_{23}-{1\over 2}\{\hat\sigma_{23}\hat\sigma_{32},\hat\rho\}\right). \tag{B.1}$$
In eq.(B.1): $N_i\equiv N(\omega_i)=(e^{\omega_i/k_BT}-1)^{-1}$, with the transition frequencies $\omega_i, i=1,2$; we assume ($\hbar=1$).
Therefore the four Lindblad operators (for the finite temperature): $\hat\sigma_{13}, \hat\sigma_{31}$, $\hat\sigma_{23}, \hat\sigma_{32}$; generalization to the degenerate case is straightforward. Then the pure-state condition eq.(6) of the main text gives the equality:
$$\gamma_1N_1(p_1+p_3-2p_1p_3) + \gamma_1(p_3-p_1p_3) + \gamma_2N_2(p_2+p_3-2p_2p_3) + \gamma_2(p_3-p_2p_3)=0, \tag{B.2}$$
where $p_i\equiv\vert\langle i\vert\psi\rangle\vert^2$.
[*The case $T=0$.*]{} For $T=0$, $N_1 = 0 = N_2$ and eq.(B.2) requires $p_3=0$, with arbitrary $p_i,i=1,2$. That is, every pure state $\vert\psi\rangle = c_1\vert 1\rangle + c_2\vert 2 \rangle$ is allowed. However, since $\hat\sigma_{i3}\vert\psi\rangle=0, \forall{i=1,2,3}$, placing $N_1=0=N_2$ into eq.(B.1) reveals that every pure state $\vert\psi\rangle$ (i.e. for arbitrary $c_i, i=1,2$) is a stationary state. That is, such states do not evolve in time and hence there is not even a single pure state that could be used to unravel the mixed state dynamics eq.(B.1).
[*The case $T>0$.*]{} Then $N_1\neq 0 \neq N_2$ and the pure-state condition reads:
$$\gamma_1(N_1+1)(p_3-p_1p_3) + \gamma_1N_1(p_1-p_1p_3)+\gamma_2(N_2+1)(p_3-p_2p_3)+\gamma_2N_2(p_2-p_2p_3)=0, \tag{B.3}$$
with the constraint $p_1+p_2+p_3=1$. Without loss of generality, choose (like for the models of the atomic dark states and induced transparency [@Gardiner; @Scully]) $\gamma_1=1=100\gamma_2$ and $N_1=0.4=1000N_2$. By inspection it can be seen that eq.(B.3) does not return any relevant solutions. For example, place $p_3=1-p_1-p_2$ and solve eq.(B.3) for $p_1$. For $p_3=0$, eq.(B.3) implies a negative value for either $p_1$ or $p_2$. For nonzero $p_3$, Fig.1 exhibits that the sum $p_1+p_2>1$; the minimum value for $p_2$ that returns a real value for $p_1$ is larger than $0.83$.
{width="50.00000%"}
Analogous elimination of the other terms (either $p_1$ or $p_2$) leads to the similar conclusions. Therefore it is found, that even a single pure state cannot be found to fulfill the pure-state condition.
[*B. A multimode system*]{}
Consider a linear system of $n$ harmonic oscillators or more generally modes in contact with a thermal bath on the temperature $T$. In any case this system can be transformed into a set of mutually uncoupled normal coordinates, i.e. of uncoupled modes, here presented by the commuting Bose annihilation operators $\hat a_i, i=1,2,...,n$; $[\hat a_i,\hat a_j^{\dag}]=\delta_{ij}$. Then a generalization of eq.(14) of the main text is straightforward:
$$\dot{\hat\rho} = \sum_{i=1}^n \gamma_i \left(
(N(\omega_i)+1)\left(\hat a_i\hat\rho \hat a_i^{\dag}-{1\over 2}\{\hat a_i^{\dag}\hat a_i,\hat\rho\}\right) + N(\omega_i) \left(\hat a_i^{\dag}\hat\rho \hat a_i-{1\over 2}\{\hat a_i\hat a_i^{\dag},\hat\rho\}\right)
\right). \tag{B.4}$$
The Lindblad operators are all the Bose operators, $\hat a_i$ and $\hat a_i^{\dag}$. Then eq.(6) takes the form:
$$\sum_i \gamma_i \left( (N(\omega_i)+1)
\left(\langle \hat a_i^{\dag} \hat a_i \rangle - \langle \hat a_i\rangle\langle \hat a_i^{\dag}\rangle\right) +
N(\omega_i) \left(\langle \hat a_i \hat a_i^{\dag}\rangle - \langle \hat a_i^{\dag}\rangle \langle \hat a_i\rangle \right)
\right)=0. \tag{B.5}$$
Due to the Cauchy-Schwarz inequality (cf. the main text), as well as to $\gamma_i, N(\omega_i) \ge 0, \forall{i}$, all the terms in eq.(B.5) are non-negative. Therefore the only possibility to fulfill the equality in (B.5) is already recognized in Section 3.2 of the main text:
$$N(\omega_i) = 0, \quad \langle \hat a_i^{\dag}\hat a_i \rangle - \langle \hat a_i\rangle\langle \hat a_i^{\dag}\rangle=0, \tag{B.6}$$
for every mode $i$. Mutual independence of the modes implies the conclusion drawn in Section 3.2 for every mode $i$ as well as for the solution, $\otimes_i\vert\alpha_i\rangle$, of eq.(B.5) (where $\alpha_i$ states for the $i$th ”coherent state” for the $i$th mode).
[*C. The phase damped harmonic oscillator*]{}
This model [@Gardiner] is comprised of a single harmonic oscillator in contact with a thermal bath such that the bosonic number operator, $\hat N=\hat a^{\dag}\hat a$, is coupled with the bath’s variable(s). The effective master equation (in the Schr" odinger picture) is ($\hbar=1$):
$$\dot{ \hat\rho }= -\imath \omega_{\circ} [\hat N,\hat\rho] +\gamma \left(\hat N\hat\rho \hat N - {1\over 2}\{{\hat N}^2,\hat\rho\}\right) \tag{B.7}$$
The only Lindblad operator, i.e. the number operator, $\hat N$ commutes with the system Hamiltonian and hence the pure-state condition:
$$(\Delta \hat N)^2=\langle\hat N^2\rangle - \langle\hat N\rangle^2=0, \tag{B.8}$$
determines its eigenstates $\vert n\rangle$, $\hat N\vert n\rangle = n\vert n\rangle$, as the solutions. However, those states are the exact ”pointer basis” states [@Giulini] that do not evolve in time. Thus there is not a single pure state whose dynamics could describe the decoherence dynamics eq.(B.7). This conclusion applies to all the similar Markovian decoherence-models for a qubit or a continuous-variable (CV) systems [@breuer; @Giulini; @Slosi].
[*D. The one-qubit generalized depolarizing channel*]{}
The so-called generalized one-qubit depolarizing channel is modelled by the following master equation (in the interaction picture) [@BJP17]:
$$\dot {\hat\rho} = \sum_{i=x}^z \gamma_i (\hat\sigma_i\hat\rho \hat\sigma_i - \hat\rho), \tag{B.9}$$
where appear the standard Pauli operators $\hat\sigma_i, i = x,y,z$.
Therefore there are three Lindblad operators, $\hat L_i=\hat\sigma_i$, which give rise to the pure-state condition:
$$\sum_i \gamma_i (1-\langle \hat\sigma_i\rangle^2) = 0. \tag{B.10}$$
It is obvious that eq.(B.10) implies $\langle\hat\sigma_i\rangle=1,\forall{i}$, which cannot be fulfilled for any pure state. This result may be expected due to the fact that this is a unital channel, i.e., dynamics preserving the fully mixed state, $\hat I/2$, where $\hat I$ is the identity operator . This is a situation also for all the unital maps for which the Lindblad operators do not have even a single common eigenstate [@breuer].
[*E. Decay of a two-level atom into a squeezed field vacuum*]{}
This is another standard model in quantum optics for a two-level atom that is described by the following master equation [@breuer; @Gardiner]:
$$\dot {\hat\rho} = \gamma_{\circ} \left(\hat
C\hat\rho\hat C^{\dag} - {1\over 2} \{\hat C^{\dag}\hat C,\hat\rho\}
\right), \tag{B.11}$$
where $\hat C = \cosh(r) \hat\sigma_- + e^{\imath \theta}\sinh(r) \hat\sigma_+$, and the environmental squeeze parameters $r$ and $\theta$, while $\hat\sigma_-=\vert g\rangle\langle e\vert$ for the excited ($e$) and the ground ($g$) atomic states.
Since there is only one Lindblad operator, $\hat L=\hat C$, the pure-state condition reads: $$\cosh^2(r) p_e + \sinh^2(r)(1-p_e) - p_e(1-p_e)(\cosh(2r)+\sinh(2r)\cos(\omega)) = 0; \tag{B.12}$$
in eq.(B.12): $p_e=\vert \langle e\vert\psi\rangle\vert^2$, $\omega = \theta+2\delta$, with the arbitrary phase $\delta$.
Solutions to eq.(B.12) are readily obtained: $(1 + e^{\imath \omega} \coth(r))^{-1}$ and $1-(1+e^{\imath\omega}\tanh(r))^{-1}$. The complex term can be eliminated for either $\omega=0$ or $\omega=\pi$. For the latter the negative values for $p_e$ are obtained. Therefore the only possibility is $\omega=0$ (i.e. $\delta=-\theta/2$) when the two solutions become equal. Hence the unique solution for $p_e = (1 + \coth(r))^{-1}$, which is independent of $\theta$. Thus, for the environmental state with the fixed $r$ and $\theta$ parameters, the [*unique*]{} state of the two-level system is found to read:
$$\vert\psi\rangle = (1 + \coth(r))^{-1/2} \vert e\rangle + e^{-\imath\theta/2} (1-(1 + \coth(r))^{-1})^{1/2} \vert g\rangle, \tag{B.13}$$
which is the eigenstate of the operator $\hat C$ with the eigenvalue $e^{\imath\theta/2}\sqrt{\sinh(2r)/2}$, as it can be easily checked.
Certainly, the single pure state eq.(B.13) cannot unravel the mixed-state dynamics. Furthermore, as we already know from Section 3.2, the condition (6) is only necessary but, in general, not sufficient for dynamical preservation of purity of a state as presented by eq.(3). Going to the matrix representation of the dissipator, $\mathcal{D}$, in eq.(B.11) and therefore to the map $e^{\mathcal{D}t}$, it can be shown that dynamics of the state eq.(B.13) takes the form of $e^{\mathcal{D}t}[\vert\psi\rangle\langle\psi\vert] = (\hat I + \sum_i n_i(t) \hat\sigma_i)/2$, where (for simplicity we choose $\gamma_{\circ}=1$):
$$\begin{aligned}
&\nonumber&
n_1(t) =
\sqrt{(1 - p) p}
\cos[\theta/2] (e^{ t (-\cosh[2 r] - \sinh[2 r])/4} + e^{ t (-\cosh[2 r] + \sinh[2 r])/4} -\\&&\nonumber
e^{ t (-\cosh[2 r] - \sinh[2 r])/4} \cos[\theta] +
e^{ t (-\cosh[2 r] + \sinh[2 r])/4} \cos[\theta])-\\&&\nonumber
(e^{ t (-\cosh[2 r] - \sinh[2 r])/4} +
e^{t (-\cosh[2 r] + \sinh[2 r])/4}) \sqrt{(1 - p) p} \sin[\theta] \sin[\theta/2],
\\&&\nonumber
n_2(t) =
-\sqrt{(1 - p) p} (e^{ t (-\cosh[2 r] - \sinh[2 r])/4} + e^{t (-\cosh[2 r]
\sinh[2 r])} +\\&&\nonumber
e^{ t (-\cosh[2 r] - sinh[2 r])/4} \cos[\theta] -
e^{ t (-\cosh[2 r] + \sinh[2 r])/4} \cos[\theta]) \sin[\theta/
2]\\&&\nonumber - (e^{ t (-\cosh[2 r] - \sinh[2 r])/4} - e^{t (-\cosh[2 r] + \sinh[2 r])/4}) \sqrt{(1 - p) p}
\cos[\theta/2] \sin[\theta],
\\&&\nonumber
n_3(t) =
e^{- t \cosh[2r]/2} (2p-1)-
(1-e^{- t \cosh[2r]/2} )
\operatorname{sech}[2r],\end{aligned}$$
{width="40.00000%"} {width="40.00000%"}
and $\hat\sigma_i, i=1,2,3$, represent the standard Pauli operators. The condition of purity of state, $P=n_1^2+n_2^2+n_3^2=1$, is fulfilled for the initial instant of time $t=0$, and, from Fig. 2, for $t>0$ [*only*]{} for $r=0$, which is the case $\gamma_1=0$ investigated in Section 3.1. Therefore there is not even a single pure state satisfying eq.(3), i.e., remaining pure in the course of the dynamics eq.(B.11).
[*F. A nonadiabatic Markovian model*]{}
For an externally driven damped harmonic oscillator, the following Markovian master equation applies (in the interaction picture) [@Dann]:
$$\dot{\hat\rho} = \vert\xi(t)\vert^2 \gamma(t) \left(
\hat F_+\hat\rho \hat F_- - {1\over 2}\{\hat F_- \hat F_+,\hat\rho\} + e^{-\hbar\alpha(t)/k_BT}
(\hat F_-\hat\rho \hat F_+ - {1\over 2}\{\hat F_+\hat F_-,\hat\rho\})
\right), \tag{B.14}$$
where $\hat F_+= A \hat x + B \hat p = \hat F_-^{\dag}$ and $A=(1+\imath\mu/\kappa)/2, B=\imath/m\omega(0)\kappa$, with all the positive parameters, $m,\omega(0),\kappa >0$ as well as $\gamma(t), \alpha(t)\ge 0, \forall{t}$.
From (B.14), the two Lindblad operators are found: $\hat L_1 = \hat F_+$ and $\hat L_2=e^{-\hbar\alpha(t)/2k_BT}\hat F_-$. Then applying eq.(6) from the main text gives:
$$\left(
\langle \hat F_-\hat F_+\rangle - \langle \hat F_-\rangle \langle \hat F_+\rangle
\right)(1+ e^{-\hbar\alpha(t)/k_BT}) + e^{-\hbar\alpha(t)/k_BT} \langle [F_+,F_-]\rangle=0. \tag{B.15}$$
With the use of the commutator, $[\hat F_+,\hat F_-]=\hbar/m\omega(0)\kappa$, eq.(B.15) becomes a sum of the non-negative terms:
$$\left(
\langle \hat F_-\hat F_+\rangle - \langle \hat F_-\rangle \langle \hat F_+\rangle
\right)(1+ e^{-\hbar\alpha(t)/k_BT}) + {\hbar e^{-\hbar\alpha(t)/k_BT}\over m\omega(0)\kappa}=0. \tag{B.16}$$
Since the second term on the rhs of (B.16) cannot equal zero, we conclude that there does not exist even a single pure state for the master equation (B.14) unraveling.
[*G. The Walls-Collet-Milburn model*]{}
Consider the dynamical model of quantum measurement performed on a two-mode system (a system $S$ with the mode $\hat a$ and the ”meter” system $M$ with the mode $\hat b$) [@Walls], such that only one of the coupled modes is in contact with the thermal bath of modes $\hat c_i$ on some temperature $T$. Assuming that interaction of the object of measurement with meter system is given by a four-wave-mixing interaction:
$$\hat H_{SM} = - {\imath\over 2} \hat a^{\dag} \hat a \otimes (\epsilon^{\ast}\hat b - \epsilon\hat b^{\dag}) \tag{B.17}$$
and that the interaction of the meter with the interaction is bilinear:
$$\hat H_{ME}= \hat b\otimes \hat C^{\dag} + \hat b^{\dag} \otimes \hat C, \tag{B.18}$$
where $\hat C^{\dag} = \sum_j\kappa_j \hat c_j$, it can be shown that, for longer times, the master equation for the object of measurement (in the interaction picture) is of the Markovian form:
$$\dot{\hat{\rho}}_S = {\vert\epsilon\vert^2\over\gamma}\left(\hat N\hat\rho\hat N-{1\over 2} \{\hat N,\hat\rho_S\}\right) \tag{B.19}$$
where $\hat N \equiv \hat a^{\dag}\hat a$ and the real parameter $\gamma>0$.
In eq.(B.19) appears only one Lindblad operator, $\hat L\equiv\hat N$, which placed in eq.(6) gives rise to the following condition for the pure-state dynamics of the object of measurement:
$$(\Delta\hat N)^2 \equiv \langle\hat N^2\rangle - \langle\hat N\rangle^2 = 0. \tag{B.20}$$
Of course, every eigenstate, but no linear combination of the eigenstates, of the number operator $\hat N$ satisfies the condition eq.(B.20). However, as it can be easily seen, those states also represent the stationary states for the process, thus not being subject of a dynamical change in time. In other words, a pure (sub)ensemble in the state $\vert n\rangle$ does not evolve in time. Therefore, by definition, there is not possibility of the ensemble unraveling for the quantum measurement modeled by eq.(B.19).
[*H. Quantum mechanics with spontaneous localization model*]{}
Consider the Ghirardi-Rimini-Weber master equation [@Bassi]:
$$\dot{\hat{\rho}} = \lambda \left(
\sqrt{{\alpha\over\pi}} \int ds e^{-\alpha(\hat x-s)/2}\hat\rho e^{-\alpha(\hat x-s)/2} - \hat\rho
\right) \tag{B.21}$$
It is worth emphasizing that equation (B.21) is assumed to describe dynamics of a [*closed*]{}, not of an open, system.
There is a Lindblad operator, $\hat L_s\equiv e^{-\alpha(\hat x-s)/2}$, for every value of the continuous real parameter $s$. Then the condition eq.(6) obviously implies:
$$\Delta (e^{-\alpha(\hat x-s)^2/2}) = 0, \forall{s} \tag{B.22}$$
for the standard deviation of the Lindblad operators.
Needless to say, there is not even a single pure state (in the Hilbert space) that could fulfill eq.(B.22) and hence the ensemble unraveling is not possible for this model.
[*I. Continuous spontaneous localization model*]{}
The master equation of interest that is assumed to describe dynamics of a [*closed*]{} system reads [@Bassi]:
$$\dot{\hat{\rho}} = \lambda \sum_{i,k} \left(
\hat N_i^{(k)} \hat\rho \hat N_i^{(k)} - {1\over 2} \{\hat N_i^{(k)2},\hat\rho\}
\right), \tag{B.23}$$
with the Hermitian Lindblad operators $\hat N_i^{(k)}$, which represent appropriate bosonic number operators for the model. Then eq.(6) leads to:
$$\sum_{i,k} (\Delta\hat N_i^{(k)})^2 = 0, \tag{B.24}$$
where appear the standard deviations for $\hat N_i^{(k)}$s. Certainly, eq.(B.24) can be only identically fulfilled, i.e. only if $\Delta \hat N_i^{(k)}=0, \forall{i,k}$. Therefore, the only pure states allowing in principle the ensemble unraveling are the common eigenstates for all $\hat N_i^{(k)}$s. However, it can be easily shown than the eigenstates for those operators, denoted $\vert n_1^{(k)} n_2^{(k)} n_3^{(k)} ... \rangle$, when placed into eq.(B.23) return the rhs of eq.(B.23) to be equal to zero. That is, such states are the stationary states, again emphasizing impossibility of the ensemble unraveling.
[9]{}
Breuer H-P and Petruccione F 2002 [*The Theory of Open Quantum Systems*]{} (Oxford, Clarendon Press)
Carnio E G, Breuer H-P and Buchleitner A 2019 Wave-Particle Duality in Complex Quantum Systems [*Phys. Chem. Lett.*]{} [**10**]{} 2121
Nielsen M A and Chuang I L 2000 [*Quantum Computation and Quantum Information*]{} (Cambridge, Cambridge University Press)
Khrennikov A 2016 [*Probability and Randomness. Quantum versus Classical*]{} (London, Imperial College Press)
Deffner S and Zurek W H 2016 Foundations of statistical mechanics from symmetries of entanglement [*New J. Phys.*]{} [**18**]{} 063013
Landsman K 2019 [*Randomness? What randomness?*]{}. http://philsci-archive.pitt.edu/16273/.
von Neumann 1955 [*Mathematical Foundations of Quantum Mechanics*]{} (Princeton, Princeton University Press)
d’Espagnat B 1999 [*Conceptual Foundations of Quantum Mechanics*]{} (New York, Perseus; 2nd edition)
Joos E, Zeh H D, Kiefer C, Giulini D J W, Kupsch J and Stamatescu I-O 2003 [*Decoherence and the Appearance of a Classical World in Quantum Theory*]{} (Berlin, Springer)
Schlosshauer M 2009 [*Decoherence and the Quantum-to-Classical Transition*]{} (Heidelberg/Berlin, Springer)
2017 Eds., R. Kastner, J. Jekni' c-Dugi' c, G. Jaroszkiewicz, (Singapore, World Scientific)
2010 Eds., S. Saunders, J. Barrett, A. Kent & D. Wallace (Oxford, Oxford University Press)
D" urr D, Goldstein S and Zanghi N 2012 [*Quantum Physics Without Quantum Philosophy*]{} (Heidelberg/Berlin, Springer)
Kastner R E 2012 [*The Transactional Interpretation of Quantum Mechanics: The Reality of Possibility*]{} (Cambridge, Cambridge University Press)
Fuchs C A and Schack R 2013 Quantum-Bayesian coherence [*Rev. Mod. Phys.*]{} [**85**]{} 1693
D" urr D and Struyve W 2019 Typicality in the foundations of statistical physics and Born’s rule [*in ”Do wave functions jump? Perspectives on the work of GC Ghirardi”*]{}, eds.: V. Allori, A. Bassi, D. D" urr and N. Zanghì (Springer International Publishing)
Bassi A and Ghirardi G C 2003 Dynamical reduction models, [*Phys. Rep.*]{} [**379**]{} 257
Rivas A and Huelga S F 2011 [*Open Quantum Systems. An Introduction*]{} (Heidelberg/Berlin, Springer-Briefs in Physics)
Sandulescu A and Scutary H 1987 Open Quantum Systems and the Damping of Collective Modes in Deep Inelastic Collisions [*Ann. Phys.*]{} [**173**]{} 277
Isar A, Sandulescu A and Scheid W 1999 Purity and decoherence in the theory of a damped harmonic oscillator [*Phys. Rev. E*]{} [**60**]{} 6371
Gisin N 1989 Stocahstic quantum dynamics and relativity [*Helv. Phys. Acta*]{} [**62**]{} 363.
Gisin N and Rigo M 1995 Relevant and irrelevant nonlinear Schr" odinger equations, [*J. Phys. A: Math. Gen.*]{} [**28**]{} 7375.
Bahrami M et al 2014 The Schr" odinger-Newton equation and its foundations [*New J. Phys.*]{} [**16**]{} 115007.
Carmichael B 1993 [*An Open Systems Approach to Quantum Optics*]{} (Heidelberg/Berlin, Springer Verlag)
Zanardi P and Rasetti M 1997 Noiseless Quantum Codes [*Phys. Rev. Lett.*]{} [**79**]{} 3306
Boixo S, Viola L and Ortiz G 2007 Generalized Coherent States as Preferred States of Open Quantum Systems [*Eurphys. Lett.*]{} [**79**]{} 40003
Gardiner C W and Zoller P 2000 Quantum Noise (Heidelberg/Berlin, Springer)
Scully M O and Zubairy S 2001 Quantum Optics (Cambridge, Cambridge University Press)
Arsenijevi' c M, Jekni' c-Dugi' c J and Dugi' c M 2017 Generalized Kraus Operators for the One-Qubit Depolarizing Quantum Channel [*Braz. J. Phys.*]{} [**47**]{} 339
Dann R, Levy A and Kosloff R 2018 Time-dependent Markovian quantum master equation [*Phys. Rev. A*]{} [**98**]{} 052129
Walls D F, Collet M J and Milburn G J 1985 Analysis of a quantum measurement [*Phys. Rev. D*]{} [**32**]{} 3208
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We used infrared spectroscopic ellipsometry to investigate the electronic properties of LaTiO$_{3}$/SrTiO$_{3}$ superlattices (SLs). Our results indicated that, independent of the SL periodicity and individual layer-thickness, the SLs exhibited a Drude metallic response with sheet carrier density per interface $\approx3\times10^{14}$ cm$^{-2}$. This is probably due to the leakage of $d$-electrons at interfaces from the Mott insulator LaTiO$_{3}$ to the band insulator SrTiO$_{3}$. We observed a carrier relaxation time $\approx35\ $fs and mobility $\approx35\
$cm$^{2}$V$^{-1}$s$^{-1}$ at 10 K, and an unusual temperature dependence of carrier density that was attributed to the dielectric screening of quantum paraelectric SrTiO$_{3}$.
author:
- 'S. S. A. Seo'
- 'W. S. Choi'
- 'H. N. Lee'
- 'L. Yu'
- 'K. W. Kim'
- 'C. Bernhard'
- 'T. W. Noh'
title: 'Optical Study of the Free Carrier Response of LaTiO$_{3}$/SrTiO$_{3}$ Superlattices'
---
Recent advances in the growth of atomic-scale multilayers of perovskites have opened up new avenues for tailoring their electromagnetic properties. For example, Ohtomo *et al.* grew superlattices (SLs) consisting of an LaTiO$_3$ (LTO) Mott insulator and an SrTiO$_3$ (STO) band insulator with atomically abrupt interfaces (IFs) [@Ohtomo]. They observed an interesting charge modulation involving electron transfer across the IF from the LTO to the STO layers. This had a decay length of 1.0$\pm$0.2 nm, which is about one order of magnitude larger than expected for conventional Thomas-Fermi screening. Subsequent transport measurements have been interpreted in terms of an unusual metallic state with quasi two-dimensional properties. This intriguing experimental observation has stimulated a number of theoretical investigations [@Okamoto_nature; @Okamoto1; @theories; @Kancharla], which confirm that so-called “electronic reconstruction" [@Okamoto_nature] can indeed give rise to a unique interfacial metallic state.
 (a) RHEED intensity oscillation during the initial growth of the (LTO)2(STO)10 SL. (b) atomic force microscope topographic image (5$\times$5 $\mu$m$^2$) of an (LTO)1(STO)10 SL 100 nm thick with single unit cell terraces conserved. (c) X-ray *$\theta$*-2*$\theta$* scan of (LTO)1(STO)10. Peaks marked with (\*) in (c) are reflected from the STO substrate.](Fig1.eps){width="3.3in"}
There is now a great deal of demand for experimental confirmation and direct investigation of such interfacial metallic states. Takizawa *et al.* recently measured photoemission spectra of LTO/STO SLs with a topmost STO layer of variable thickness [@Takizawa]. They observed a metallic Fermi edge, indicating the formation of a metallic IF. However, to date there have been no reports of quantitative experimental information on the electrodynamic properties of the itinerant electrons. Here, we present spectroscopic ellipsometry measurements of the infrared (IR) dielectric properties for a set of (LTO)$\alpha$/(STO)10 SLs with $\alpha$=1, 2, 4, and 5 unit cells of LTO layer and 10 unit cells of STO. Our IR optical data yield reliable information about the intrinsic electrodynamic properties of these SLs. First, they clearly demonstrate that these SLs contain a sizeable concentration of itinerant charge carriers. Furthermore, they establish that the sheet carrier density is proportional to the number of IFs, and its absolute value agrees closely with the theoretical predictions. Our data also highlight an unusually strong temperature ($T$) dependence of the carrier density, which was unexpected for bulk metals, but can be explained in terms of the $T$-dependent dielectric screening of the STO layer controlling charge transfer across the LTO/STO IF.
We grew high-quality LTO/STO SLs $\sim$100 nm thick with atomically flat surfaces and abrupt IFs on single crystalline STO (001) substrates. To do this, we used pulsed laser deposition (PLD) at $T$=720 $^\circ$C in P$_{O2}$=10$^{-5}$ Torr with *in situ* monitoring of the specular spot intensity of reflection high-energy electron diffraction (RHEED) (see Ref. [@HNLee] for more details on PLD growth). The RHEED intensity oscillations in Fig. 1(a) confirmed the controlled growth of alternating LTO and STO layers. Doing this at a higher oxygen pressure would result in the growth of unwanted La$_2$Ti$_2$O$_7$ phases, as reported previously [@LTO227]. After growth, all samples were exposed immediately to a higher pressure (P$_{O2}$=10$^{-2}$ Torr) for *in situ* postannealing at growth *T* for 5 min, and then cooled to room $T$. No changes in the RHEED specular spot pattern were observed after postannealing. The atomic force microscope topographic image in Fig. 1(b) shows that the SLs retained their single unit cell terrace steps even after deposition of $\sim$250 unit cells of LTO and STO. Figure 1(c) shows x-ray *$\theta$*-2*$\theta$* diffraction, exhibiting well-defined satellite-peaks, confirming the high IF quality and SL periodicity of LTO and STO layers. The full-width half-maximum $\leq0.04^{\circ}$ in *$\omega$*-scan, which is almost identical to the substrate scan, confirmed the high crystallinity of the SLs.
 (a) Optical conductivity spectra of (LTO)$\alpha$/(STO)10 at 10 K (with vertical offsets), containing the coherent (Drude) and incoherent (Lorentz) contributions. (b) SL real dielectric constant spectra as a function of photon wavenumber at 10 K. The arrow indicates the screened plasma frequency $\omega_{p}^{*}[=\omega_{pD}/\sqrt{\varepsilon_{\infty}}]$. (c) $T$-dependent $\sigma_1(\omega)$ of (LTO)2/(STO)10.](Fig2.eps){width="3.3in"}
We measured the $T$-dependent IR optical properties of the SLs using ellipsometry in the range of 80-5000 cm$^{-1}$ (10-600 meV). The far-infrared (FIR) measurements for 80-700 cm$^{-1}$ were performed with a home-built ellipsometer attached to a Bruker 66V Fourier transform IR (FT-IR) spectrometer at the IR beamline of the ANKA synchrotron at FZ Karlsruhe, Germany [@Bernhard]. For the mid-infrared (MIR) range of 400-5000 cm$^{-1}$, we used a home-built ellipsometer in combination with the glow-bar source of a Bruker 113V FT-IR spectrometer. The FIR (MIR) measurement was performed with an angle of incidence of the linearly polarized light of 82.5$^{\circ}$ (80$^{\circ}$), $i.e.$ close to the pseudo-Brewster angle of the SLs. Details about the ellipsometry technique are given in Ref. [@Bernhard; @Azzam]. Here, we only point out that the ellipsometry is a self-normalizing technique that directly measures the complex dielectric function $\tilde{\varepsilon}$($\omega$)\[=*$\varepsilon$*$_1$(*$\omega$*)+*$i\varepsilon$*$_2$(*$\omega$*)\], without the need for Kramers-Kronig analysis. We also measured the $T$-dependent spectroscopic response of a bare STO substrate that was treated thermally under the same growth and annealing conditions of $T$ and P$_{O2}$ as the SLs. Then we used a uniaxially anisotropic single-layer model [@Azzam] to obtain the in-plane component [@c-axis] of *$\tilde{\varepsilon}$*(*$\omega$*), and the related optical conductivity *$\tilde{\sigma}$*(*$\omega$*)\[=*$\tilde{\varepsilon}$*(*$\omega$*)$\omega$/4$\pi$*i*=*${\sigma}$*$_1$(*$\omega$*)-*i$\varepsilon$*$_{1}$(*$\omega$*)$\omega$/4$\pi$\] of the SLs. As the thickness of the SLs is well below our IR wavelength, the entire SL film can be treated as a single layer according to the effective medium theory. The resulting effective dielectric functions thus correspond to the volume-averaged dielectric functions of the components in the SL. Our optical technique can provide intrinsic values of the bulk properties of these SLs, which can be compared with the preexisting results of transport measurements [@Ohtomo].
---------- ------------------------------- -------------------------- ------------- -------------------------- ----------------------
$\alpha$ $\omega^{2}_{pD}$ (cm$^{-1}$) $\Gamma_{D}$ (cm$^{-1}$) $S$ $\omega_{0}$ (cm$^{-1}$) $\Gamma$ (cm$^{-1}$)
1 $6.0\times10^7$ $128(\pm3)$ $75(\pm11)$ $920(\pm95)$ $1980(\pm580)$
2 $6.6\times10^7$ $144(\pm8)$ $51(\pm22)$ $980(\pm240)$ $1530(\pm1200)$
4 $5.1\times10^7$ $105(\pm6)$ $87(\pm20)$ $780(\pm200)$ $1630(\pm1000)$
5 $6.1\times10^7$ $122(\pm2)$ $126(\pm8)$ $750(\pm65)$ $1740(\pm350)$
---------- ------------------------------- -------------------------- ------------- -------------------------- ----------------------
: \[tab:table1\]Drude-Lorentz fitting parameters for optical spectra of (LTO)$\alpha$/(STO)10 SLs measured at 10 K.
Figures 2(a) and 2(b) show the low-$T$ spectra of $\sigma_{1}(\omega)$ and $\varepsilon_{1}(\omega)$ for the series of (LTO)$\alpha$/(STO)10 SLs with $\alpha$=1, 2, 4, and 5. Significantly, all the spectra exhibited a prominent Drude-like peak at low frequency in *${\sigma}$*$_1$(*$\omega$*) and a strongly inductive response in *${\varepsilon}$*$_1$(*$\omega$*), which provided unambiguous evidence that these SLs contained sizeable concentrations of itinerant charge carriers. The strongly conducting response also seemed to damp out IR active phonons in the spectra. For quantitative analysis we applied the Drude-Lorentz fitting function: $$\tilde{\varepsilon}(\omega)=
\varepsilon_{\infty}-\frac{\omega^{2}_{pD}}{\omega^{2}+i\omega\Gamma_{D}}
+\frac{S\omega^{2}_{0}}{\omega^{2}_{0}-\omega^{2}-i\omega\Gamma}.
\label{eq:eq1}$$ This is shown in Fig. 2(a) by the solid and dashed lines. The parameters of the Drude term are the scattering rate $\Gamma_{D}$, and the plasma frequency $\omega^{2}_{pD}=4\pi{n}{e}^{2}/m^{*}$, where *n*, *e*, and $m^{*}$ are the density, charge, and effective mass of the itinerant charge carriers, respectively. The parameters of the Lorentz term that accounts for the broad MIR band are the oscillator strength $S$, the width $\Gamma$, and the resonant frequency $\omega_{0}$. A broad MIR band is commonly observed in conducting oxides and has been interpreted in terms of either the inelastic interaction of the itinerant carriers, a second type of carrier in a bound state, or low-lying interband transitions. As LTO is well-known to have weak broad MIR bands, the Lorentz peak with a small value of $S$ can be reasonably interpreted to be the lowest intersite $d$-$d$ ($i.e.$, U-3J) transition of LTO [@JSLee]. An overview of the fitting parameters for the low-$T$ spectra is given in Table 1.
Most importantly, we noted that all the SLs have surprisingly high $\omega_{pD}$, and thus high densities of free carriers. This observation motivated us to follow up on suggestion of Okamoto and Millis [@Okamoto_nature] of a quasi-two-dimensional metallic state that develops in the LTO/STO IFs. We derived the sheet carrier density per IF, $n_{sheet}=n\frac{d}{N_{I}}$, with the total SL thickness $d$ and the number of LTO/STO IFs $N_{I}$ of the SL [@LTO1]. Figure 3(a) shows $n_{sheet}$ values within error-bars dependent on the detailed assignment of $m^{*}$ based on the analysis of our ellipsometric data in the framework of the extended Drude-model. This directly yields the renormalized mass $m^{*}/m_{e}=(\omega_{p}/\omega)^{2}(\varepsilon_{0}-\varepsilon_{1})/((\varepsilon_{0}-\varepsilon_{1})^{2}-\varepsilon_{2}^{2})$. Note that all of our SLs exhibit a very similar value of $n_{sheet}\approx3\times10^{14}$ cm$^{-2}$ with $m^{*}=1.8m_e$ [@Okuda], which agrees well with theoretical predictions [@Okamoto1; @Kancharla; @integral].
 (a) $n_{sheet}$ per IF. The solid line indicates the value of an ideal 0.5 $e^-$ per unit cell area, and the dashed and dotted lines represent theoretically predicted values reported by Kancharla [@Kancharla] and Okamoto [@Okamoto1], respectively. The vertical arrow shows the range of $n_{sheet}$ in LaAlO$_3$/SrTiO$_{3-\delta}$ IF from Ref. [@Kalakukhov]. (b) Comparison of SL $\omega_{pD}$ with solid-solution [@Fujishima] and the ideal doping of \[La\]=*n* as a function of La and Sr composition.](Fig3.eps){width="3.3in"}
Here, we should mention that the Drude term is only a phenomenological description, and its microscopic origin is not clear. One possible explanation is in terms of a chemical intermixing of La and Sr ions across the IFs because La$_x$Sr$_{1-x}$TiO$_3$ solid solutions are metallic [@Tokura]. As indicated by the solid circles in Fig. 3(b), the experimental data of bulk La$_x$Sr$_{1-x}$TiO$_3$ solid solutions reported by Fujishima *et al.* [@Fujishima] show that the $\omega_{pD}$ value increases from x = 0.1 to 0.7, while it decreases above 0.7 due to the strong electron correlation effect as a function of increasing $x$. However, our SL data follow a different path although the value of $\omega_{pD}$ is expected to increase when going from $\alpha$ = 1 to $\alpha$ = 5, as shown by the solid squares in Fig. 3(b). Moreover, the well-defined superstructure peaks in our x-ray diffraction data and the abrupt IFs seen on cross-sectional transmission electron microscopy (not shown) confirm that the extent of La and Sr intermixing is negligible.
As other candidates of extrinsic origin, it is also necessary to consider the possibility of oxygen excess LaTiO$_{3+\delta}$ [@Schmehl] or La vacancy La$_{1-x}$TiO$_3$ [@Crandles2] of LTO, and oxygen deficiency of STO [@Crandles1]. The former possibility can be excluded as our data for $\omega_{pD}$ do not exhibit a corresponding change as the relative volume fraction of the LTO layers is altered by a factor of 5 between $\alpha$=1 and 5. On the other hand, serious consideration should be given to the latter possibility of oxygen deficiency and thus metallic SrTiO$_{3-\delta}$ layers. It is noteworthy that numerous recent controversial studies and some debate in oxide electronics circles have centered on the discontinuous IF polarity of LaAlO$_3$/STO [@Ohtomo_LAO_STO] especially regarding the origin of conduction due to the oxygen vacancies [@Eckstein]. As indicated by the arrow in Fig. 3(a), the range of $n_{sheet}$ values of LaAlO$_3$/SrTiO$_{3-\delta}$ grown at low oxygen pressure [@Kalakukhov] is so wide that it also overlaps our experimental value of $n_{sheet}\approx3\times10^{14}$ cm$^{-2}$ as well as with the theoretical values [@Okamoto1; @Kancharla] of electronic reconstruction at the LTO/STO IF. However, according to the recent report by Herranz *et al.*, LaAlO$_3$/STO films were all insulating or highly resistive when cooled in a high oxygen pressure environment from the deposition $T$ to room $T$ [@Herranz]. We performed a similar *in situ* oxygen annealing process to compensate for possible oxygen vacancies, but our LTO/STO SLs still exhibited metallic behavior.
![\[fig4\] (color online) $T$ dependence of (a) the carrier density, and (b) the relaxation time and mobility of free carriers for (LTO)$\alpha$/(STO)10 SLs.](Fig4.eps){width="3.3in"}
Another interesting aspect concerns the strong $T$ dependence of $n(T)$. Figure 2(c) shows the $T$ dependence of $\sigma_{1}(\omega)$ for (LTO)2/(STO)10, which are characteristic of all SLs. Figure 4(a) shows that $n(T)$ increased significantly by almost 30$\%$ between 300 K and 10 K. The corresponding $n(T)$ in a conventional normal metal is typically of the order of 1$\%$, even for oxides with strongly correlated electrons. For the cuprate high-$T_c$ superconductors, for example, it is smaller than 10$\%$ [@Toschi]. Even the opposite $T$ dependence is observed for doped semiconductors where $n(T)$ decreases as the carriers become trapped at low $T$. Thus, the increase of $n(T)$ at low $T$ also cannot be explained by the electron-doping effect of oxygen vacancies of STO. However, we note that the quantum paraelectric behavior of incipient ferroelectric STO exhibits a significant increase in dielectric permittivity ($\epsilon_{0}^{STO}$) as $T$ decreases. Therefore, we can explain the increase of $n(T)$ in the context of electronic reconstruction at an IF. The charge transfer across the IF of LTO/STO (and thus $n_{sheet}$) should be affected by the dielectric screening of the STO layer. The larger value of $\epsilon_{0}^{STO}$ thus gives rise to the larger screening length of the electronic charges, and this allows more charges to be transferred across the LTO/STO IF. (See also the theoretical prediction of the coherent carrier density as a function of $\epsilon_{0}$ by Kancharla [@Kancharla].)
Figure 4(b) displays the relaxation time of free carriers $\tau(T)[=\Gamma_{D}^{-1}(T)]$. As all of our LTO/STO SLs showed a significantly reduced $\Gamma_{D}$, which is at least one order of magnitude smaller than a doped Mott system [@RCaTiO3], fairly high mobility $\mu$(10 K) $\approx$ 35 cm$^2V^{-1}s^{-1}$ was obtained according to the relation $\mu=e\tau/m^{*}$ with $m^{*}=
1.8m_{e}$. It would be interesting to consider whether signatures of novel phenomena, such as the quantum Hall effect, may be observable in the LTO/STO SLs. The measured value of $\tau\approx3.5\times10^{-14}$ s at 10 K yields $\omega_{c}\tau=0.01-0.02$, where the cyclotron frequency $\omega_c=eB/m^{*}$ with the magnetic field $B\approx20$ T and $m^{*}=1.0-2.5 m_{e}$. From the obtained carrier density and mobility of our LTO/STO SLs, we determined that the mean free path of the conducting electrons is about 10 nm ($\approx$ 25 unit cells long) at 10 K. It is remarkable that there was a recent observation of the quantum Hall effect in ZnO-based heterostructures below 1 K [@QHE; @ZnO], in which the coherence length of the electron was considerably longer than those in other oxides.
In summary, our use of IR spectroscopic ellipsometry showed that all the LTO/STO SLs exhibited a Drude-like metallic response regardless of the SL periodicity and the individual layer-thicknesses, even after *in situ* post annealing in oxygen. Our data yielded a nearly constant $n_{sheet}$ per IF of $\sim3\times10^{14}$ cm$^{-2}$ with small $\Gamma_{D}$ of $\sim120$ cm$^{-1}$, and a sizeable mean free path of $\sim$25 unit cells at 10 K. We note that the unusually strong $T$ dependence of $n$ can be interpreted qualitatively as quantum paraelectric behavior of STO with theoretical prediction of the leakage of $d$-electrons from LTO to STO across the IF. A picture of electron-doping by the oxygen vacancy of STO, which has been the subject of recent debate on similar systems of LaAlO$_3$/STO, is insufficient to explain our observations in LTO/STO SLs. Hence, we note the difference between an LTO Mott insulator and an LaAlO$_3$ band insulator, and suggest that the strongly correlated $d$-electrons of LTO should play an important role in electronic reconstruction and the resultant metallic state at a polar (LTO) and nonpolar (STO) oxide IF.
The authors thank S. Okamoto, A.J. Millis, H.Y. Hwang, G.W.J. Hassink, J.S. Ahn, D.-W. Kim, Y.S. Lee, J. Yu, K. Char, K.H. Kim J.H. Park, J.Y. Kim, J.S. Kim, A.V. Boris, and B. Keimer for valuable discussions, as well as Y.L. Mathis for support at IR-BL of ANKA. This work was supported by the Creative Research Initiatives (Functionally Integrated Oxide Heterostructure) of KOSEF, the Swiss National Science Foundation (SNF project 200021-111690/1), and the Division of Materials Sciences and Engineering, U.S. Department of Energy (H.N.L.). Experiments in PLS were supported by MOST and POSTECH.
[27]{} A. Ohtomo *et al.*, Nature **419**, 378 (2002).
S. Okamoto and A. J. Millis, Nature **428**, 630 (2004).
S. Okamoto and A. J. Millis, Phys. Rev. B **70**, 241104(R) (2004).
S. Okamoto and A. J. Millis, Phys. Rev. B **70**, 075101 (2004), S. Okamoto and A. J. Millis, *ibid.* **72**, 235108 (2005), S. Okamoto and A. J. Millis, Physica B-Cond. Matt. **359**, 1378 (2005), S. Okamoto, A. J. Millis, and N. A. Spaldin, Phys. Rev. Lett. **97**, 056802 (2006), Z. S. Popovic and S. Satpathy, *ibid.* **94**, 176805 (2005).
S. S. Kancharla and E. Dagotto, Phys. Rev. B **74**, 195427 (2006).
M. Takizawa *et al.*, Phys. Rev. Lett. **97**, 057601 (2006).
H. N. Lee *et al.*, Nature **433**, 395 (2005).
A. Ohtomo *et al.*, Appl. Phys. Lett. **80**, 3922 (2002).
C. Bernhard, J. Humlicek, and B. Keimer, Thin Solid Films **455-456**, 143 (2004).
R. M. A. Azzam and N. M. Bashara, *Ellipsometry and Polarized Light* (North-Holland, Amsterdam, 1987).
Okamoto *et al.* [@Okamoto1] suggested sharp optical transitions along the $c$-axis due to bound states of quantum well-like structures in IR energy region (with $t\sim0.3$ eV). However, the large refractive indices ($>$10) due to conducting IFs make the refraction angle $\leq5^{\circ}$. Therefore, the electric field of infrared waves is almost parallel to the in-plane direction, and it allows discussion of the in-plane response only.
J. S. Lee, M. W. Kim, and T. W. Noh, New J. of Phys. **7**, 147 (2005).
For (LTO)1/(STO)10 and (LTO)1/(STO)6, it is rather ambiguous to define the number of IF. Here, we count two IFs (both sides) per LTO layer for consistency with other SL samples. Hence, the $n_{sheet}$ of (LTO)1/(STO)10 and (LTO)1/(STO)6 may be reduced due to overestimated $N_{I}$.
Note $m^{*}=1.62m_e$ from the thermoelectric experiment for La$_{x}$Sr$_{1-x}$TiO$_{3}$ by T. Okuda *et al.*, Phys. Rev. B **63**, 113104 (2001).
The values from the theoretical calculations of $\sim6.5\times10^{14}$ cm$^{-2}$ and $\sim3.4\times10^{14}$ cm$^{-2}$ are obtained by integrating the coherent $d$-electron density curves displayed in Figs. 2 and 2(a) of Refs. [@Okamoto1] and [@Kancharla], respectively.
Y. Tokura *et al.*, Phys. Rev. Lett. **70**, 2126 (1993).
Y. Fujishima *et al.*, Phys. Rev. B **46**, 11167 (1992).
A. Schmehl *et al.*, App. Phys. Lett. **82**, 3077 (2003).
D. A. Crandles *et al.*, Phys. Rev. B **49**, 16207 (1994).
D. A. Crandles *et al.*, Phys. Rev. B **59**, 12842 (1999).
A. Ohtomo and H. Y. Hwang, Nature **427**, 423 (2004).
J. N. Eckstein, Nat. Mater. **6**, 473 (2007).
A. Kalabukhov *et al.*, Phys. Rev. B **75**, 121404(R) (2007).
G. Herranz *et al.*, Phys. Rev. Lett. **98**, 216803 (2007).
A. Toschi *et al.*, Phys. Rev. Lett. **95**, 097002 (2005).
T. Katsufuji, Y. Okimoto, and Y. Tokura, Phys. Rev. Lett. **75**, 3497 (1995).
A. Tsukazaki *et al.*, Science **315**, 1388 (2007).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We hypothesize that dynamical systems concepts used to study the transition to turbulence in shear flows are applicable to other transition phenomena in fluid mechanics. In this paper, we consider a finite air bubble that propagates within a Hele-Shaw channel containing a depth-perturbation. Recent experiments revealed that the bubble shape becomes more complex, quantified by an increasing number of transient bubble tips, with increasing flow rate. Eventually the bubble changes topology, breaking into two or more distinct entities with non-trivial dynamics. We demonstrate that qualitatively similar behaviour to the experiments is exhibited by a previously established, depth-averaged mathematical model; a consequence of the model’s intricate solution structure. For the bubble volumes studied, a stable asymmetric bubble exists for all flow rates of interest, whilst a second stable solution branch develops above a critical flow rate and transitions between symmetric and asymmetric shapes. The region of bistability is bounded by two Hopf bifurcations on the second branch. By developing a method for a numerical weakly nonlinear stability analysis we show that unstable periodic orbits emanate from the Hopf bifurcation at the lower flow rate and, moreover, that these orbits are edge states that influence the transient behaviour of the system.'
author:
- 'J. S. Keeler'
- 'A. B. Thompson'
- 'G. Lemoult'
- 'A. Juel'
- 'A. L. Hazel'
bibliography:
- 'hele\_shaw\_paper.bib'
title: 'The influence of invariant solutions on the transient behaviour of an air bubble in a Hele-Shaw channel.'
---
Introduction {#sec:intro}
============
A Hele-Shaw channel consists of two parallel glass plates, separated by a distance much smaller than the width of the channel. If a trapped viscous fluid is extracted at a constant flux from one end of the channel and an air bubble is placed at the other end, then the bubble will propagate and change shape as it does so. For sufficiently large bubbles the only stable solution is for the bubble to propagate symmetrically along the centreline of the channel; a solution analogous to the symmetric semi-infinite air finger that develops when one end of the channel is left open to the atmosphere [@saffman1958penetration]. For higher flow rates in large aspect ratio channels, propagating air fingers develop complex patterns via multiple tip splitting events as well as side-branching [@Tabeling1987; @MooreJuel2002]. The onset of this complex interfacial dynamics appears to be a subcritical transition; a feature that it shares with the transition to turbulence in shear flows. Specifically, the steady symmetrically propagating solution is linearly stable for all values of the flow rate at which it has been computed, meaning that finite perturbations are required to initiate the complex dynamics. Moreover, the value of the critical dimensionless flow rate for the onset of patterns cannot be precisely determined and is very sensitive to the level of perturbation in the system: the transition occurs at lower dimensionless flow rates as the roughness of the channel walls is increased[@Tabeling1987].
In this paper, we concentrate on bubble propagation in a geometrically-perturbed Hele-Shaw cell: a rectangular prism is added to the base of the channel, as sketched in figure \[fig:dimdomain\]. By studying bubbles, rather than air fingers, and working in a co-moving frame, we can follow long-time evolution of the system in short computational domains. This relatively simple system still exhibits a wide variety of nonlinear dynamical phenomena [@de2009tube; @pailha2012oscillatory; @hazel2013multiple; @jisiou2014geometry; @thompson2014multiple; @franco2016sensitivity; @franco2017propagation; @franco2017bubble] and the recent experimental results of Franco-G[ó]{}mez *et al*[@franco2017propagation] have shown that the system contains regions of bistability in which finite perturbations can provoke a range of possibilities for the time-evolution of an initially centred bubble. For small volume fluxes the bubble will eventually settle towards a stable asymmetric state on one side (or the other) of the depth perturbation, but if the volume flux exceeds a critical threshold the bubble shape becomes increasingly deformed before eventually breaking up into two or more distinct parts, as sketched in figure \[fig:break\_up1\](b), despite theoretical predictions that a stable steady state exists for all flow rates; a feature preserved from the unperturbed Hele-Shaw channel. Thus, this model system allows us to explore whether the dynamical systems concepts applied in the study of transition to turbulence in shear flows also apply to the transient behaviour of other canonical problems in fluid mechanics.
We study the behaviour theoretically by finding invariant solutions, namely steady states and periodic orbits, of the system. We pursue the idea, hypothesised by Franco-G[ó]{}mez *et al* [@franco2017propagation], that when the flux is large enough the complex time-dependent behaviour of a single bubble can be interpreted as a transient exploration of the stable manifolds of weakly unstable *edge states*. We consider only the dynamics before the changes in topology when the bubble breaks up into two or more separate bubbles: systems with their own dynamics that we do not pursue here.
In the context of fluid mechanics an edge state is an invariant solution of the governing equations that has only a small number of unstable eigenvalues and whose stable manifold forms the ‘boundary’ between two qualitatively different dynamical outcomes. Edge tracking techniques have been utilised in a number of different scenarios including shear flow and pipe flow [@kerswell2005recent; @schneider2007turbulence; @schneider2008laminar; @eckhardt2008dynamical] and more recently droplet breakup [@gallino2018edge]. In these studies the edge state was found by direct numerical simulation of the governing equations and interval bisection of the initial conditions. This methodology is easy to implement numerically but can be computationally expensive and does not reveal whether the edge state corresponds to a steady state, periodic orbit or other invariant solution.
One advantage of the Hele-Shaw system is that the only nonlinearities in the system arise due to boundary conditions on the bubble because the reduced Reynolds number is extremely small. Thus the shape of the interface gives an immediate visual representation of nonlinear behaviour. Another advantage is that by assuming that the width of the channel is large compared to its height, the behaviour of the system can be described by a depth-averaged set of equations that is more amenable to analysis than the full Navier–Stokes equations[@saffman1958penetration; @taylor1959note]. These reduced equations (stated in § \[sec:form\]), often called Darcy or Hele-Shaw equations, have been used in most theoretical studies of this system, and lead to predictions for the bubble shape as viewed from above (sketched in figure \[fig:break\_up1\](b)). In the unperturbed system at zero surface tension, exact steady solutions can be found by conformal mapping techniques, and for each bubble volume, there is a two-parameter family of solutions described by the centroid offset and bubble speed [@tanveer1987surfacetension; @tanveer1987stability; @tanveer1987newsolutions]. The introduction of surface tension selects both a main, symmetric branch of linearly stable steady solutions which persists for all fluxes, along with a countably infinite sequence of unstable ‘exotic’ bubble shapes; these exist in both channels bounded by parallel side walls [@tanveer1987stability] and in unbounded channels [@green2017effect]. Figure \[fig:break\_up1\](a) shows the bubble shape and speed for the first three solution branches in channels with a rectangular cross-section. Introducing a depth-perturbation to the system (see sketch in figure \[fig:dimdomain\]) allows the solution branches to interact, which results in the diverse range of observed stable steady states and time-dependent behaviour [@de2009tube; @pailha2012oscillatory; @hazel2013multiple; @jisiou2014geometry; @thompson2014multiple; @franco2016sensitivity; @franco2017propagation; @franco2017bubble].
We will show that the bifurcation structure of the depth-averaged equations does indeed feature bistability between steady states, and that previously unknown unstable periodic orbits influence whether the bubble breaks up or returns to a stable configuration. We have sufficient knowledge of the solution structure that, rather than edge tracking, we can develop a general numerical procedure for weakly nonlinear analysis that allows us to determine an approximation to the periodic orbits and steady states. We emphasise that the method described in this paper, although applied to our specific set of equations, can readily be adapted for other problems and we provide a numerical recipe for this procedure. The results of the weakly nonlinear analysis, along with time-dependent calculations, show that the previously unknown unstable periodic orbits are indeed edge states of the system that influence the eventual fate of the bubble.
![Sketch of a bubble propagating in a Hele-Shaw channel with a depth-perturbation on the bottom, showing the 3D location of the interface, depth-perturbation and direction of propagation and the dimensional quantities, $W^*$, $H^*$, $V^*$ and $Q^*$.[]{data-label="fig:dimdomain"}](figure_dimdomain.eps)
![(a) The steady solution space for a bubble of fixed volume in an unoccluded channel with side walls. The solution measure $U_b$ is the speed of bubble relative to fluid ahead, while $Q$ is the dimensionless imposed flow rate. The insets indicate the bubble shapes denoted by crosses on the solution branches and are on a 1:1 scale with $x,y\in[-1,1]$. These solutions were calculated using the set of equations described in § \[sec:form\]. (b) A sketch of an initial bubble evolving in time to form two tips and eventually break up, yielding two separate bubbles.[]{data-label="fig:break_up1"}](figure_break3.eps)
The paper is set out as follows. Initially, in § \[sec:form\], we introduce the details of the system and describe the depth-averaged governing equations, non-dimensional parameters and numerical methods. In § \[sec:one\], we start our investigation with a range of initial-value calculations and compare these to the previous experiments [@franco2017propagation]. In § \[sec:two\], we present an extended bifurcation diagram for the invariant steady states and perform a linear stability analysis. In § \[sec:four\] we derive our general method for determining the weakly nonlinear approximation to the invariant periodic orbits of the system by employing the method of multiple scales near the bifurcation points. Finally, in § \[sec:six\], we summarise our results and briefly discuss the importance of the periodic orbits to the transient behaviour of the system and their interpretation as edge states.
Governing Equations {#sec:form}
===================
![The non-dimensional computational domain. (a) Top-view sketch of bubble propagation, which is the focus of the depth-averaged model. Calculations take place in a frame of reference moving with the bubble centroid, and the computational domain is truncated in $x\in[-L,L]$ where $L$ at a dimensionless distance, typically chosen to be $L=4$ in the simulations. (b) For the depth-averaged calculations, a smoothed version of the depth-perturbation is used, plotted here in the $(y,z)$ plane when $h=0.024,s=40,w=0.25$. \[fig:nondimdomain\] ](figure_nondimdomain.eps)
The physical situation is shown in figure \[fig:dimdomain\]. The channel has outer dimensional width $W^*$ and height $H^*$. The channel is filled with a viscous fluid of density $\rho$ and dynamic viscosity $\mu$, containing a single air bubble with known volume $V^*=(W^*)^2H^*V/4$, where $V$ is the non-dimensional volume. Fluid flow is driven by withdrawing fluid at a constant flux $Q^*$ far ahead of the bubble. The system is non-dimensionalised based on a length scale $W^*/2$ in the $x$ and $y$ directions, and on the velocity scale $U_0^* = Q^*/(W^*H^*)$. We define the non-dimensional flow rate $ Q = \mu U_0^*/\gamma, $ where $\gamma$ is the coefficient of surface tension at the air/fluid interface. Additionally we define the aspect ratio of the channel $\alpha=W^*/H^*$. Details of the derivation of the depth-averaged equations from the Navier-Stokes equations can be found in [@thompson2014multiple; @franco2017propagation]. The critical assumptions are that the channel aspect ratio $\alpha$ is large; the reduced Reynolds number $U_0^* W^*/(\rho \mu\alpha^2)$ and Bond number $\mbox{Bo} = \rho gL^{*2}/\gamma$ are small; the bubble occupies the full height of the channel; and the component of curvature as $z$ varies corresponds to a semi-circle filling the channel height [with the fluid perfectly wetting]{} the upper and lower walls. We neglect the thin film corrections proposed by Homsy [@homsy1987viscous] and Reinelt [@reinelt1987interface], because they do not change the qualitative comparison with the experiment, and it is not obvious how they should be modified due to the presence of the depth-perturbation and multiple solutions [@thompson2014multiple; @franco2016sensitivity; @franco2017propagation].
The nondimensional domain is shown in figure \[fig:nondimdomain\]. The depth-averaged velocity at any point within the fluid is given by $\mathbf{\hat{u}} = -b^2(y) \nabla
p$, where $p=p(x,y)$ is the fluid pressure and $b(y)$ is the height of the channel (see figure \[fig:nondimdomain\](b)). We work in a frame moving at dimensionless speed $U_b(t)$ chosen at each time step so that the $x$ component of the bubble centroid remains fixed in this moving frame. The equations to be solved are
$$\begin{aligned}
\nabla \cdot(b^3(y)\nabla p) = 0\qquad \mbox{in\ }\:&
\Omega,\label{exact_eqn1}\\ p_b - p = \frac{1}{3\alpha
Q}\left(\frac{1}{b(y)}+\frac{\kappa}{\alpha}\right)\qquad
\mbox{on\ }\:&
\Gamma,\label{exact_eqn2}\\ \textbf{n}\cdot\textbf{R}_t +
\textbf{n}\cdot \mathbf{e}_x U_b(t) + b^2(y)\textbf{n}\cdot\nabla
p =0\qquad \mbox{on\ }\:&
\Gamma,\label{exact_eqn3}\\ \frac{\partial p}{\partial y} =0
\qquad \mbox{on\ } \:& y =\pm 1, \label{exact_eqn4}\\
p=0 \qquad \mbox{on\ } \:& x =
-L, \label{inlet_equation}\\ \frac{\partial p}{\partial x} = -1
\qquad \mbox{on\ } \:& x = L, \label{exact_eqn5}\\ \int b(y)
\textbf{R}\cdot\textbf{n}\,\mbox{d}\Gamma =
V.&\label{exact_eqn6}
\end{aligned}$$
\[exact\_eqn\]
Here $\Omega$ denotes the two-dimensional fluid domain (the shaded region in figure \[fig:nondimdomain\](a)) and $\Gamma$ the bubble boundary, $p_b(t)$ is the unknown pressure inside the bubble, $\textbf{R} =
(x_b,y_b)$ is the position of the bubble interface, $\kappa$ is the curvature of this 2D interface, and $\mathbf{n}$ is a unit normal vector directed out of the bubble. Equation is the equation for the pressure in the fluid domain whilst equations and are the dynamic and kinematic conditions on the bubble surface. Our depth-averaged equations are second order in space, and hence we can apply only no-penetration conditions, ; likewise the condition derives from the normal component of the dynamic boundary condition at the bubble interface, and we neglect the tangential stress balance. [[Equation imposes the pressure to be zero far behind the bubble at $x=-L$ and equation achieves the constant volume flux at $x=L$ by imposing a constant pressure gradient.]{}]{} Additionally, the unknown value of $p_b$ is determined by ensuring the volume constraint, , is satisfied. The computational domain is truncated at a fixed distance $L$ ahead of and behind the bubble. Following previous papers, we model the depth-perturbation by a smoothed profile: b(y) = 1 - h. \[occlusion\] A typical example of this smoothed profile is shown in figure \[fig:nondimdomain\](b). The dimensionless depth-perturbation height $h$ can be viewed as a controlled, axially uniform perturbation of the rectangular channel. The topography profile $b(y)$ enters the equations through the bulk equation (which is no longer amenable to conformal mapping), the kinematic boundary conditions and the variable transverse curvature of the interface. The unknowns of the problem are $[p,\textbf{R},p_b,U_b]$ and the control parameters are $[Q,V,h,w,s,\alpha]$. An additional solution measure is the centroid of the bubble, denoted $\overline{y}$, which is useful in determining the symmetry of the system. are solved by a finite-element discretisation, utilising the open-source software `oomph-lib` [@heil2006oomph]. Details of the implementation can be found in [@thompson2014multiple] for the case of a propagating air-finger and [@franco2017propagation] for a finite air bubble.
It is convenient to keep $h,w$ and $\alpha$ constant and analyse the system mainly through variations in $Q$ and $V$, which are easily manipulated in a physical, experimental set-up. Unless otherwise specified the results presented in this paper are compared to the values used by Franco-G[ó]{}mez *et al* [@franco2017propagation] and [[hence a bubble of fixed volume $V$ is chosen so that the projected area is $A = \pi r^2$ with $r=0.46$,]{}]{} together with a perturbation height $h=0.024$, width $w=0.25$, sharpness $s=10$ and aspect ratio $\alpha = 40$.
Initial-value calculations {#sec:one}
==========================
![ []{data-label="fig:fourexp"}](experimental_pictures_v2.png)
![[]{data-label="fig:four"}](figure2_v17.eps)
[[Franco-G[ó]{}mez *et al* [@franco2017propagation] performed a range of experiments which showed that an initially circular bubble would settle to a stable state below a critical flow rate whilst above this threshold the bubble becomes increasingly deformed, resulting in a variety of transient outcomes, see figure \[fig:fourexp\], including topological break up, despite evidence to suggest that an asymmetric stable state existed for all flow rates. ]{}]{}
In order to investigate the potential range of evolution in our model we perform a number of numerical simulations designed to mimic these experiments. The nonlinear dynamics can be examined by recording the bubble shape as it propagates along the channel. We do this by solving the equations in [ with the initial condition of a circular bubble with radius $r$ centered at $(x_c,y_c) = (0.0,0.01)$ and released from rest]{}. Our simulations reveal a number of qualitatively distinct modes of propagation. The eventual propagation mode selected has a dependence on the flow rate $Q$ [and the radius $r$ which sets the bubble volume $V$]{}. A summary of the qualitative behaviour is shown in figure \[fig:four\], with regions of parameter space classified according to the characteristic shape first adopted by the bubble. As $Q$ and $V$ are varied, we find that the first emergent shape may have between one and four tips. The start of each different coloured region as $Q$ increases indicates the place in parameter space where the bubble first develops two, three or four tips for a given bubble volume. [Single tipped bubbles]{} evolve smoothly to an asymmetric steadily-propagating configuration (see inset labelled [<span style="font-variant:small-caps;">1</span>]{} in figure \[fig:four\]), while multiple-tipped bubbles may self-intersect in the simulations and thus break up into one or more fragments, in which case the simulations are terminated (see insets labelled [<span style="font-variant:small-caps;">2</span>]{},[<span style="font-variant:small-caps;">3</span>]{},[<span style="font-variant:small-caps;">4</span>]{} in figure \[fig:four\]) . In some regimes, the bubble may exhibit oscillatory behaviour before decaying to either a steady state (see inset [<span style="font-variant:small-caps;">5</span>]{} in figure \[fig:four\]) or stable periodic oscillations (see inset [<span style="font-variant:small-caps;">6</span>]{}). The nature of the oscillatory modes that tend to a stable state are dependent on the flow rate in that for lower flow rates ($Q\approx 0.05$) the bubble appears to oscillate around a symmetric steady state whilst for larger flow rates ($Q\approx 0.1$) the oscillations manifest as ‘waves’ above one edge of the depth-perturbation with the opposite side of the asymmetric bubble above the other edge of the depth-perturbation as seen in the [<span style="font-variant:small-caps;">3</span>]{} inset in figure \[fig:four\].
We can compare these numerical results to the experimental data of Franco-G[ó]{}mez *et al* [@franco2017propagation] shown in figure \[fig:fourexp\]. Although the boundaries between regions are quantitatively different in figures \[fig:fourexp\] and \[fig:four\], the phenomenological behaviour is in agreement, and the transition between different modes of propagation occur over similar ranges of $Q$ and $V$. For nearly all values of $Q$ explored in figure \[fig:four\], varying the bubble volume at fixed $Q$ allows us to access a range of qualitatively different modes of propagation.
As demonstrated by these numerical time simulations, there is a large range of possible outcomes from [an initially circular bubble]{} including oscillatory transients and multiple stable steady states. To understand this further we now perform a detailed analysis of the steady state bifurcation structure.
Bifurcation Structure for Steady States {#sec:two}
=======================================
In this section we analyse the steady solution space. We use the flow rate, $Q$, as a continuation parameter for a fixed bubble volume and record the bubble speed, $U_b$, and the centroid, $\overline{y}$ as convenient solution measures. We also analyse the linear stability of the numerical solutions as the solution branches are traced out. The solution space is shown in figure \[fig:b\_diag\_nonlinear\] in the $(Q,U_b)$ and $(Q,\overline{y})$ projections respectively, with stable (unstable) solution branches denoted by solid (dashed) lines. We note that the branches in this steady solution space are connected through several bifurcations, in contrast to the $h=0$ case (see figure \[fig:break\_up1\](a)), where the solution branches are disjoint.
![[]{data-label="fig:b_diag_nonlinear"}](figure_steady_v6.pdf){width="\textwidth"}
For small $Q$ there are three distinct branches as shown in figure \[fig:b\_diag\_nonlinear\]. The ‘upper’ branch (with largest $U_b$), AS1, is stable, characterised by an asymmetric bubble shape and persists for all values of $Q$ that were sampled. The other two branches, denoted as S1a and S1b, correspond to symmetric double-tipped solutions. As the value of $h$ is decreased to zero the S1a branch approaches the $m=0$ solution sketched in figure \[fig:break\_up1\](b) whilst the lower part becomes the double-tipped solution $m=1$ solution. The S1 branch only exists for a small range of the parameter $Q$, in contrast to the case with no depth-perturbation where the solution persists for all $Q$. Both parts of the S1 branch are unstable; the segment with larger $U_b$ has one unstable eigenmode, while the slower segment is doubly unstable. Note that although it appears from figure \[fig:b\_diag\_nonlinear\](a) that the S1 and AS1 branches intersect at small $Q$, these branches are actually disjoint as can be seen by the projection of the solution onto the $(Q,\overline{y})$ plane in figure \[fig:b\_diag\_nonlinear\](c).
For $Q\gtrapprox 0.01$, we are able to compute a three-tipped symmetric solution, S2a, which is unstable. As $Q$ increases this branch undergoes two fold bifurcations. [[The first fold, F1, (shown in figure \[fig:b\_diag\_nonlinear\](b)) separates the S2a and S2b branches and the second fold, F2, separates the S2b and S2c branches]{}]{}. By decreasing $h$ to zero, the S2a branch converges to a triple-tipped solution branch, the S2b branch becomes the double-tipped solution branch and the S2c branch becomes the stable solution branch (this transition sequence is likely to involve interaction between the AS1, S1 and S2 branches). All of the S2 branches are unstable but for this bubble volume and channel geometry, the solution finally stabilises via a Hopf bifurcation H1 just beyond the second fold point. The stable branch of steady solutions that results from this Hopf bifurcation is labelled S3.
As $Q$ increases further on the S3 branch, a supercritical pitchfork bifurcation (PF) occurs where the S3 solution breaks symmetry and two stable asymmetric branches, denoted AS2, emanate from the bifurcation, see figure \[fig:b\_diag\_nonlinear\]. Beyond the pitchfork, the unstable symmetric branch experiences multiple bifurcations as $Q$ is varied and becomes increasingly unstable. This collection of unstable branches is denoted by S4. [The most extreme shapes on the S4 branch, with $U_b$ close to 1, correspond to]{} the eight-tipped solution as seen in figure \[fig:b\_diag\_nonlinear\]. Note the loops on the S2 branch around $Q \approx 0.11$ where an asymmetric branch connects to the main S4 branch via two pitchfork bifurcations. This is reminiscent of the so called ‘snakes and ladders’ bifurcations seen in other nonlinear systems, see, for example, [@burke2007snakes; @schneider2010snakes]. This branch has at least two unstable eigenmodes and as we do not believe it affects the transient behaviour of the system we do not pursue the details of this branch here. For the largest $Q$ in our calculations ($Q = 0.15$), two distinct asymmetric states persist in the $(Q,U_b)$ projection, (four in the $(Q,\overline{y})$ projection). The AS1 branch is always stable but the stable AS2 branch experiences a Hopf bifurcation, denoted by H2, and becomes unstable, the resulting branch labelled AS3.
In contrast with the $h=0$ solution space, where only one solution is ever stable, in our case there is a finite-width region of bistability (with respect to steady states) between the H1 and H2 bifurcations. The AS1 solution is always stable but the nature of the second stable solution changes from symmetric (the S2 branch) to asymmetric (the AS2 branch), via the pitchfork bifurcation. [[The symmetric stable solution (S2) has been observed before in this system, see Franco-G[ó]{}mez *et al* [@franco2017propagation], but its transition to an asymmetric stable solution as $Q$ increases, and indeed the presence of the two Hopf bifurcations, is a new observation.]{}]{} The results presented here are for a fixed volume, $V$, and fixed depth-perturbation height, $h$ but we can use bifurcation tracking calculations to investigate the robustness of the location and order of the bifurcations F2, H1, PF and H2 that bound the bistable region. The locations of these bifurcations as functions of $V$ and $h$ are shown in figure \[fig:loci\_volume\]. We find that the ordering of the F2, H1, PF and H2 bifurcations are very robust to changes in volume and, with the exception of H2, their position is fixed, see figure \[fig:loci\_volume\](a). As $V$ decreases, H2 migrates to larger values of $Q$ and hence the width of the bistable region increases. All the bifurcations are more sensitive to $h$ than $V$, and when $h<h_c \approx 0.012$, the change in stability occurs at the fold, F2, instead of the Hopf, H1, see figure \[fig:loci\_volume\](b). A fold-Hopf bifurcation occurs when $h=h_c$. The dynamics of the system near this type of co-dimension two bifurcation are very complex and can lead to the appearance of invariant tori and [[heteroclinic orbits]{}]{}, see, for example, [@kuznetsov2013elements].
The time-dependent behaviour explored in § \[sec:one\] can now be interpreted in terms of the bifurcation structure. For all flow rates, there is an asymmetric stable mode of propagation, which corresponds to the AS1 branch. The region of bistability appears to coincide with the solid symbols in figure \[fig:four\] which denote that the bubble is evolving towards a steady solution. The multi-tipped modes of propagation may relate to the presence of the unstable S1, S2 and S4 branches, while the oscillating modes of propagation may be expected to arise due to the Hopf bifurcations. However, the linear stability analysis does not reveal the [centre]{}, nor the stability and size of the periodic orbits emanating from the Hopf bifurcations and thus the criticality of H1 and H2 remains unknown.
There are [ at least]{} three possible approaches to determine the periodic orbits. If the periodic orbit is stable, we expect that for suitable initial conditions, the periodic orbit is attracting and hence initial-value simulations will eventually converge towards the periodic orbit. These simulations may be expensive if the convergence rate is slow, and in any case will not capture unstable periodic orbits. The second approach is to calculate the periodic orbits directly by solving an extended system of equations as proposed by, for example, [@net2015continuation]. This direct solution will capture unstable periodic orbits, but leads to a significantly larger system of equations and requires a discretization in both space and time. The third approach is to perform a numerical weakly nonlinear stability analysis near the Hopf points, leading to an analytic normal form of the perturbation equations near the Hopf points with numerically computed coefficients, providing an approximate expression for the periodic orbits in terms of eigenfunctions. This latter approach is the least computationally expensive and has the advantage of providing semi-analytic approximations for the amplitude, period and location of the periodic orbits and we pursue this analysis in the next section.
![[]{data-label="fig:loci_volume"}](figure_loci.eps)
Weakly Nonlinear Stability {#sec:four}
==========================
The aim of this section is to perform a local analysis near the Hopf bifurcation points to obtain approximations for the location, stability and size of the periodic orbits. In linear stability theory the growth of the perturbation occurs on a single time-scale and is determined by the eigenvalues alone and not the amplitude of the perturbation. In weakly nonlinear stability theory near a marginally stable solution (i.e. bifurcation point) the growth/decay of a perturbation is assumed to happen over two time scales, the eigenvalue providing the growth/decay on the fast scale $t$, whilst the amplitude varies on a slower time-scale. The [[main]{}]{} objective of a weakly nonlinear analysis is to obtain an evolution equation for this unknown amplitude function. [[This method has been applied to a number of specific examples in physics, including shear flows, shallow water waves, thermoacoustics and magnetohydrodynamics, see, for example [@fujimura1989equivalence; @bera_poiseuille; @orchiniweaklynonlinear; @sanchez2006amplitude; @ghidaouiwaterwavenonlinear; @laroze2010amplitude; @laroze2009amplitude].]{}]{} [[In these examples the nature of the equations analysed means that the algebra often becomes very complicated.]{}]{} Our approach (inspired by [@sanchez2006amplitude]) is to perform the analysis for a general set of equations that can be applied to a wide range of systems so that an analytic approximation for the periodic orbits and steady states can be obtained. The analysis is based on a continuous set of equations that are independent of the nature of a discretisation procedure. We then describe a [[versatile]{}]{} numerical procedure that can be implemented to obtain semi-analytic expressions for the periodic orbits and steady states. We examine a set of equations of the form (,[u]{},) = 0, \[eq0\] where $\mathcal{R}$ is a nonlinear [[function]{}]{} that depends on the state variables $u\in \mathcal{U}$, where $\mathcal{U}$ is an appropriately defined Hilbert space, time derivatives $\dot{u}$ and a parameter $\beta\in\mathbb{R}$. The state variable, $u$, will in general depend continuously on spatial coordinates, $x$ and temporal coordinates $t$. To proceed further it is assumed that the set of equations can be separated so that (,[u]{},) \[[u]{}\] + \[[u]{}\] + \[[u]{}\]= 0, \[general\] where $\mathcal{M}$ is a linear mass operator and $\mathcal{F}$, $\mathcal{G}$ are nonlinear operators on the state variables [[independent of the parameter $\beta$]{}]{}. The form of the equations in , whilst not completely general, are representative of a large number of physical systems where time derivatives appear in linear combinations, including the equations that form the subject of this paper, .
We are interested in the nonlinear evolution of a perturbation near a Hopf point. A standard linearisation procedure about a steady solution, $u_s$, will yield a generalised eigenproblem that can be solved to find the eigenmodes, denoted $g$, and corresponding eigenvalues, denoted $s$. A Hopf point occurs at a particular value of the parameter, $\beta_c$ say, such that a single pair of eigenvalues are located on the imaginary axis, i.e. of the form $s=\pm\mathrm{i}\omega_c$. Near this point of marginal stability the oscillations will occur on a fast time scale and will be modulated by a time-dependent amplitude function that operates on a slower time scale. Following [@sanchez2006amplitude; @orchiniweaklynonlinear; @drazin2004hydrodynamic] we define a fast time scale, $t_0=t$ and a slow time scale, $t_1 =
\varepsilon^2 t$, where $\varepsilon\ll 1$ is a small unfolding parameter, and employ the method of multiple scales. The solution, $u$, is now a function of two different time scales as well as the spatial variables denoted by $x$. The solution, time derivative and parameter $\beta$ are therefore expanded in ascending powers of $\varepsilon$:
$$\begin{aligned}
{u} =&\: {u}_s(x) + \varepsilon{u}_1(t_0,t_1,x) +
\varepsilon^2{u}_2(t_0,t_1,x) +
\varepsilon^3{u}_3(t_0,t_1,x)+\mathcal{O}(\varepsilon^4),\label{asym1}\\ {\frac{\partial }{\partial t}}
=&\: {\frac{\partial }{\partial t_0}} + \varepsilon^2{\frac{\partial }{\partial t_1}} +
\mathcal{O}(\varepsilon^4), \label{asym2}\\ \beta =&\:\beta_c(1 +
\delta\varepsilon^2).\label{asym3}
\end{aligned}$$
\[asym\]
The parameter $\delta = \pm 1$ signifies which ‘side’ of the Hopf bifurcation is being analysed. Substituting the expressions in into yields a sequence of linear problems that can be solved at each order of $\varepsilon$.
Examining terms of $\mathcal{O}(\varepsilon^0)$ in the expansion gives the following equation to solve for the steady base solution, \_c\[[u]{}\_s\] + \[[u]{}\_s\]= 0. \[pence\] [[With $u_s$ known]{}]{}, we continue further and at $\mathcal{O}(\varepsilon)$ we have (,[u]{}\_1) \[[u]{}\_s\] + \[[u]{}\_s\][u]{}\_1 = 0, \[jack\] where the operator $\mathcal{J}[u_s]$ is the Jacobian operator defined by its action on $v_1\in \mathcal{U}$ by the Fréchet derivative, [[$$\mathcal{J}[{u_s}]v_1 \equiv
D_{v_1}\mathcal{K}[u_s] \equiv \lim_{k\to
0}\frac{\mathcal{K}[u_s+kv_1] - \mathcal{K}[u_s]}{k},
\label{frechjaco}$$]{}]{} [[where $\mathcal{K}[u]\equiv\beta_c\mathcal{F}[u] +
\mathcal{G}[u]$. ]{}]{}[[We separate $u_1$ into spatial and temporal parts so that ]{}]{}$u_1(t_0,t_1,x) = A(t_1)g(x)\mbox{e}^{st_0}$ and equation becomes a generalised eigenproblem. Its solution determines the linear stability of the steady solutions solved in . [[There are an infinite number of solutions to but as we are only interested in periodic solutions arising from a Hopf bifurcation, we choose $u_1$ to be]{}]{} \_1 = A(t\_1)[g]{}(x)\^[\_c t\_0]{}+, \[homog\] where $A(t_1)$ is the undetermined amplitude of the perturbation depending on the slow time scale $t_1$. The function ${g}(x)$ is a complex eigenfunction of the generalised eigenvalue problem, $\mathrm{i}\omega_c\mathcal{M}{g} = \mathcal{J}{g}$. Proceeding to $\mathcal{O}(\varepsilon^2)$ the problem to be solved is (,[u]{}\_2) = [-\[[u]{}\_1\] - \_c\[[u]{}\_s\]-\[[u]{}\_s\]([u]{}\_1,[u]{}\_1)]{}, \[secondorder\] [[where $\mathcal{H}[u_s]$ represents the bilinear Hessian operator defined its action on $v_1,v_2\in \mathcal{U}$ and using by $$\begin{aligned}
\mathcal{H}[{u_s}](v_1,v_2) \equiv
D_{v_2}(\mathcal{J}[u_s]v_1)\equiv\lim_{k\to
0}\frac{\mathcal{J}[u_s + kv_2]v_1 - \mathcal{J}[u_s]v_1}{k}.
\label{frechhess} \end{aligned}$$]{}]{}The terms on the right hand side [ of ]{} are either formed of products of $u_1$ or have no time dependence. It can be shown that the products of $u_1$, defined in , introduce time dependent terms proportional to $\mbox{exp}(2\mathrm{i}\omega_ct_0)$ [ in ]{} and these do not resonate with the homogeneous solutions of the operator $\mathcal{L}$. By examining the form of ${u}_1$ in a particular solution to is sought in the form $${u}_2 = A^2{\varphi}_0\,\mbox{e}^{2\mathrm{i}\omega_c t_0} +
\mbox{c. c.} + |A|^2{\varphi}_1+{\varphi}_2,
\label{form}$$ where the functions $\varphi_i$ are to be determined. Substituting into and then equating coefficients of $A^2$, $|A|^2$ and constant terms leads to three linear equations that can be solved to find the undetermined functions ${\varphi}_i$ in . These equations can be stated as
$$\begin{aligned}
& \left(2\mathrm{i}\omega_c\mathcal{M}[{u}_s] +
\mathcal{J}[{u}_s]\right){\varphi}_0
=\:-\mathrm{i}\omega_c\mathcal{M}[{g}]{g}-\frac{1}{2}\mathcal{H}[{u}_s]({g},{g}),\\ &\mathcal{J}[{u}_s]{\varphi}_1
=\:
-\frac{1}{2}\mathcal{H}[{u}_s]({g},{g}^*)-\frac{1}{2}\mathcal{H}[{u}_s]({g}^*,{g})
- \mathrm{i}\omega_c\mathcal{M}[{g}^*]{g} +
\mathrm{i}\omega_c\mathcal{M}[{g}]{g}^*,\\ &\mathcal{J}[{u}_s]{\varphi}_2
= \: - \delta \beta_c\mathcal{F}[{u}_s],
\end{aligned}$$
\[east\]
where the star denotes complex conjugation. The amplitude function $A(t_1)$ remains undetermined at this order. Therefore the analysis continues to the next order, $\mathcal{O}(\varepsilon^3)$. The corresponding equation to be solved is $$\begin{gathered}
\mathcal{L}\left({\frac{\partial {u}_3}{\partial t_0}},u_3\right) =
{-\mathcal{M}[{u}_2]{\frac{\partial {u}_1}{\partial t_0}} -
\mathcal{M}[{u}_1]{\frac{\partial {u}_2}{\partial t_0}} -
\mathcal{M}[{u}_s]{\frac{\partial {u}_1}{\partial t_1}}} \\ {-\delta
\beta_c\mathcal{J}_{\mathcal{F}}[{u}_s]{u}_1}
{-\frac{1}{2}\mathcal{H}[{u}_s]({u}_1,{u}_2) -
\frac{1}{2}\mathcal{H}[{u}_s]({u}_2,{u}_1)} {-
\frac{1}{6}\mathcal{T}[{u}_s]({u}_1,{u}_1,{u}_1)},
\label{errol}$$ where $\mathcal{J}_{\mathcal{F}}[u_s]u_1\equiv
D_{u_1}\mathcal{F}[u_s]$ and the operator $\mathcal{T}[u]$ is the trilinear third-order differential operator defined by its action on $v_1,v_2,v_3\in \mathcal{U}$ and using by $$\begin{aligned}
\mathcal{T}[{u}](v_1,v_2,v_3) \equiv
D_{v_3}(\mathcal{H}[u_s](v_1,v_2)) \equiv \lim_{k\to
0}\frac{\mathcal{H}[u_s+kv_3](v_1,v_2) -
\mathcal{H}[u_s](v_1,v_2)}{k}.
\label{frechtress}$$ At this order the products formed on the right-hand side of will result in time-dependent terms proportional to $\mbox{exp}(\mathrm{i}\omega_ct_0)$ and these *are* resonant with homogeneous solutions of the operator $\mathcal{L}$. Therefore a solvability condition is invoked. In linear operator theory, by the Fredholm alternative (see, for example, [@boyce1969elementary]), a linear system of the form $\mathcal{L}u = f$ will either have (a) a unique solution $u$, or (b) a non-trivial solution to $\mathcal{L}^{\dagger}v = 0$. For condition (a) to hold $\langle
v^{\dagger}, f\rangle = 0$. The dagger superscripts indicate the adjoint problem. The adjoint problem and inner product on the Hilbert space, $U$, are defined by $$\langle \mathcal{L}u,v\rangle = \langle
u,\mathcal{L}^{\dagger}v\rangle,\qquad \langle u,v \rangle =
\int_{0}^{T}\int_{x\in\Omega} u v^* \,\mbox{d} x\,\mbox{d}t_0$$ for any functions $u$ and $v$ belonging to the solution function space, where $\Omega$ is the domain of definition of the spatial variable $x$, and $T$ is the period of oscillation in the fast time scale $t_0$. Applying the Fredholm alternative to yields a solvability condition which provides a constraint on the amplitude function $A(t_1)$. This can be written as + A + A|A|\^2=0, \[ls2\] which is the weakly nonlinear Landau equation. The coefficients, $\hat{\nu}$, $\hat{\lambda}$ and $\hat{\mu}$ are defined by = \[[u]{}\_s\][g]{},[g]{}\^,= \_[k=0]{}\^3\_k,[g]{}\^, = \_[k=0]{}\^7\_k,[g]{}\^, \[storm\] where $g^{\dagger}$ is the eigenfunction corresponding to the adjoint problem, $\mathcal{L}^{\dagger}g^{\dagger} = 0$. The values of $\Lambda_k$ and $\Gamma_k$ are defined as $$\begin{split}
\Gamma_0 = \:& \delta \beta_c\mathcal{J}_{\mathcal{F}}[{u}_s]{g},
\quad \Gamma_1 = \:
\mathrm{i}\omega_c\mathcal{M}[{\varphi}_2]{g},\quad \Gamma_2 = \:
\frac{1}{2}\mathcal{H}[{u}_s]({g},{\varphi}_2), \quad \Gamma_3 =\:
\frac{1}{2}\mathcal{H}[{u}_s]({\varphi}_2,{g}),\\ \Lambda_0 =\:&
-\mathrm{i}\omega_c\mathcal{M}[{\varphi}_0]{g}^*,\quad\Lambda_1 = \:
\mathrm{i}\omega_c\mathcal{M}[{\varphi}_1]{g}, \quad\Lambda_2 = \:
2\mathrm{i}\omega_c\mathcal{M}[{g}^*]{\varphi}_0, \quad\Lambda_3 =
\: \frac{1}{2}\mathcal{H}[{u}_s]({g},{\varphi}_1),\\\Lambda_4 = \:&
\frac{1}{2}\mathcal{H}[{u}_s]({g}^*,{\varphi}_0), \quad\Lambda_5 =
\: \frac{1}{2}\mathcal{H}[{u}_s]({\varphi}_1,{g}),\quad \Lambda_6 =
\frac{1}{2}\mathcal{H}[{u}_s]({\varphi}_0,{g}^*),\\ \Lambda_7 = \:&
\frac{1}{6}\mathcal{T}[{u}_s]({g}^*,{g},{g})+\frac{1}{6}\mathcal{T}[{u}_s]({g},{g}^*,{g})+\frac{1}{6}\mathcal{T}[{u}_s]({g},{g},{g}^*).
\end{split}
\label{grizzly}$$ The values of $\hat{\nu},\hat{\lambda},\hat{\mu}$ can be obtained once the functions $\varphi_i$ and the eigenmodes $g$ and $g^\dagger$ have been calculated. For convenience the eigenmodes $g$ and $g^\dagger$ are normalised so that $|\langle \mathcal{M}g,g^\dagger\rangle| =
1$. Therefore, $|\nu| = 1$. Furthermore it is more convenient to write the Landau equation as = A + A|A|\^2, \[ely\] where $\lambda = -\hat{\lambda}/\hat{\nu}$ and $\mu =
-\hat{\mu}/\hat{\nu}$.
The analysis up to this point has been based on a spatially and temporally continuous set of equations. We now describe a procedure, independent of the choice of discretisation, appropriate for a spatially discretised version of the equations in . In the description of the algorithm that follows, $\textbf{J}$ and $\textbf{M}$ are the matrix representations of $\mathcal{J}$ and $\mathcal{M}$ respectively and bold symbols are the discretised vectors of their continuous counterpart. It is assumed that efficient numerical linear solvers (for example, SuperLu [@li2005overview]), generalised eigensolvers (for example, Trilinos [@herouxtrilnos]). and continuation algorithms are available. We note that for high-dimensional systems, these calculations are computationally demanding, but they need to be performed only once for each bifurcation point. The procedure is as follows:
1. Find a Hopf bifurcation as $\beta$ is varied by a continuation method as described by, for example, [@kuznetsov2013elements].
2. Calculate the solution of , $\textbf{u}_s$, using Newton’s method at $\beta = \beta_c$.
3. Calculate the eigenfunction $\textbf{g}$ and eigenvalue $s_c=\mathrm{i}\omega_c$ using [[an]{}]{} eigensolver to solve $\textbf{J}[\textbf{u}_s]\textbf{g}=s\textbf{M}[\textbf{u}_s]\textbf{g}$.
4. Calculate the Hessian products on the right hand side of . If an exact expression for the Jacobian matrix is available, the Hessian products can be numerically computed using the central-difference formula \[[u]{}\_s\](v\_1,v\_2) , \[hess\_num\] where $k_1$ is a small finite difference parameter.
5. Calculate the functions $\boldsymbol{\varphi}_0,\boldsymbol{\varphi}_1$ and $\boldsymbol{\varphi}_2$ in using a linear solver.
6. Calculate the adjoint eigenfunction, $g^{\dagger}$ by solving $\textbf{J}^T[\textbf{u}_s]\textbf{g}^{\dagger}=-s\textbf{M}^T[\textbf{u}_s]\textbf{g}^{\dagger}$.
7. Calculate the Hessian and third-order products in using by approximating the third-order operator (again assuming an exact Jacobian) as $$\begin{gathered}
\nonumber
\mathcal{T}[u_0](v_1,v_2,v_3)
\approx\\ \frac{\textbf{J}[\textbf{u}_s+k_2\textbf{v}_2+k_2\textbf{v}_3]\textbf{v}_1
- \textbf{J}[\textbf{u}_s + k_2\textbf{v}_2]\textbf{v}_1 -
\textbf{J}[\textbf{u}_s-k_2\textbf{v}_2+\textbf{v}_3]\textbf{v}_1
+
\textbf{J}[\textbf{u}_s-k_2\textbf{v}_2]\textbf{v}_1}{2k_2^2}.\end{gathered}$$
8. Finally combine the quantities in to calculate the values of $\hat{\nu},\hat{\lambda}$ and $\hat{\mu}$ given in , defining the discrete inner product as $\langle \textbf{f},\textbf{g}\rangle =
\textbf{f}\textbf{g}^{T}$.
[[We now return our focus to the specific equations in ]{}]{}. These are discretised by the finite-element method, utilising the open-source software `oomph-lib` [@heil2006oomph] in which the functions required for the above procedure have all been implemented. It is convenient to choose the independent parameter $\beta = Q^{-1}$ and, to ensure the finite difference formulas stated above converge, $k_1=10^{-4}$ and $k_2=10^{-3}$. When $V = \pi r^2$, with $r=0.46$, $\alpha=40$, $w=0.25$ and $h=0.024$, this procedure was implemented and the critical values of $Q_c$, $\omega_c$ and the Landau coefficients were found to 3 significant figures to be
$$\begin{aligned}
\mbox{H}1: Q_c = Q_{1} = 0.0301, \quad \omega_c=3.58,
\quad&\:\lambda = \delta(5.63 -10.7\mathrm{i}),\quad \mu =
1280-1200\mathrm{i}, \label{hopf1}\\ \mbox{H}2: Q_c = Q_{2} =
0.0896, \quad \omega_c=10.8, \quad&\:\lambda = \delta(-5.43 -
15.8\mathrm{i}),\quad \mu = -6150 -
1320\mathrm{i}. \label{hopf2}
\end{aligned}$$
\[hopf\]
[[We have checked that the quadrant in the complex plane of which $\lambda$ and $\mu$ lie (setting $\delta$ constant) does not change as parameters are varied thus not affecting the criticality of the bifurcation.]{}]{}
At this point, we wish to analyse solutions to . We note that a decomposition into modulus argument form, i.e. $A =
\hat{r}(t)\exp(\mathrm{i}\theta)$ converts the equation into a pair of ODEs: $${\frac{\mbox{d} \hat{r}}{\mbox{d} t_1}} = a\hat{r} + b\hat{r}^3,\qquad
{\frac{\mbox{d} \theta}{\mbox{d} t_1}} = c + d\hat{r}^2,
\label{jaco1}$$ where $a = \delta \Re(\lambda)$, $b = \Re(\mu)$, $c =
\delta\Im(\lambda)$ and $d = \Im(\mu)$. The invariant solutions occur when $\mbox{d}{\hat{r}}/\mbox{d}{t_1} = 0$. The solutions corresponding to $\hat{r}=0$ are the steady states of the fully nonlinear steady equations, while the solution corresponding to $\hat{r}=\sqrt{-a/b}$ is an invariant periodic orbit of the fully nonlinear system. By examining the signs of $a$ and $b$ at each Hopf point and performing a stability analysis on the equations in it is straightforward to show that H1 is subcritical; unstable periodic orbits exist when $Q>Q_1$ for $Q$ near to $Q_1$. Similarly, H2 is supercritical so stable periodic orbits exist when $Q>Q_2$.
[[Near the H1 and H2 points the steady state solution of , corresponding to $\hat{r}=0$ in , can be written in terms of the state variables as]{}]{} u = [u]{}\_s + \^2\_2 + (\^4). \[steadyapprox\]
![[]{data-label="fig:b_diag_wnl"}](figure_wnl.eps)
This weakly nonlinear approximation can be compared to the fully nonlinear steady state bifurcation structure of § \[sec:two\]. The dotted lines in the inset panels of figure \[fig:b\_diag\_wnl\] are the approximations given by . The approximation near the H2 bifurcation point is excellent [[in this projection]{}]{}; the fully nonlinear and weakly nonlinear curves are visually indistinguishable. For the H1 bifurcation point the weakly nonlinear steady state approximation deviates significantly away from the solution branch as we move away from H1 in the $(Q,U_b)$ projection. This difference [ is likely due to]{} the existence of the fold, F2, in the immediate vicinity of H1. As $h$ is varied, to approach the fold-Hopf point, a weakly nonlinear analysis should include the eigenmode associated with the zero eigenvalue and hence the form of the perturbation equation in would differ.
The periodic orbits, corresponding to $\hat{r}=\sqrt{-a/b}$, can be stated as = [u]{}\_s + [g]{}\^[(\_ct\_0+)]{}+ \^2(\_2-\_1-\_0\^[2(\_ct\_0+)]{}) + (\^3), \[limitapprox\] where $\theta=\theta_0$ at $t=0$ and $\Delta = \theta_0 +
\varepsilon^2(c-ad/b)t_0$ represents the $\mathcal{O}(\varepsilon^2)$ correction to the phase of the oscillations. The location of the centre of oscillations [[(ignoring time-dependent terms in )]{}]{} of these periodic orbits are shown as dotted lines in the main panel of figure \[fig:b\_diag\_wnl\]. The circular markers in this diagram indicate the amplitude of the periodic orbits. The initial phase $\theta_0$ will not alter the amplitude or period of the periodic orbit.
An example of the weakly nonlinear bubble shapes of the periodic orbits near the H1 point is shown in figure \[fig:periodic\_orbits\_h1\_2\]. The oscillations are characterised by an oscillation about the symmetric steady solution and the oscillations are visible on both sides of the bubble. It can be seen that after half a period the bubble has ‘flipped’ from its initial shape. The weakly nonlinear equations in predict that if the initial amplitude, $r_0$, of a perturbation is smaller than [a critical amplitude]{} $\hat{r}$, then the system will evolve towards $r=0$, i.e. the S3 steady state. In contrast if $r_0>\hat{r}$ then the trajectory will diverge to $r\to\infty$. The H1 periodic orbit can therefore be considered an edge state of the reduced system as it forms the boundary between stable and unstable behaviour.
The orbits near H1 are unstable so as noted in § \[sec:two\] we are unable to precisely capture the fully nonlinear periodic orbit by time-integration alone. However by choosing the weakly nonlinear solution in as an initial [[condition]{}]{} to the fully nonlinear time-dependent problem a qualitative comparison can be made. A characteristic feature of the unstable periodic orbits is that the perturbation oscillates all the way around the edge of the bubble (see figure \[fig:periodic\_orbits\_h1\_2\]). [[These oscillations will cause oscillations in the centroid of the bubble. Therefore a useful way to compare the fully nonlinear simulations and weakly nonlinear approximations is to measure the time signal of $\overline{y}$ so that the amplitude and period can be compared.]{}]{} A result of one such simulation when $\varepsilon=0.01$ is shown in figure \[fig:periodic\_orbits\_h1\_1\] in the fully nonlinear simulations (a), and weakly nonlinear simulations, (b). The period and amplitude of oscillations of both regimes are nearly identical in both regimes. At this value of $\epsilon$, we are very close to the H1 point and hence the oscillations in the fully nonlinear regime will take a long time to decay. For larger values of $\varepsilon$ we find that depending on the initial phase of the initial condition, the bubble will either break up, evolve to the AS1 solution or continue to oscillate towards the S3 solution, see figure \[fig:periodic\_orbits\_h1\_3\]. This is further evidence that the H1 orbit is an edge state in the fully nonlinear regime. In contrast to the weakly nonlinear regime there are three possible outcomes with the fully nonlinear H1 periodic edge state forming the dividing line between stable symmetric evolution and either breaking up or being attracted towards the stable asymmetric states.
An example of the weakly nonlinear bubble shapes of the periodic orbits near the H2 point is shown in figure \[fig:periodic\_orbits\_h2\_2\]; now the underlying steady state is asymmetric. The stable periodic orbit near the H2 point is characterised by oscillations appearing on the single side of the bubble that is over the edge of the depth-perturbation. These periodic orbits are linearly stable and hence time-integration of the fully nonlinear system should give a good indication of the period and amplitude of the orbit as $t$ progresses. As the centroid of the bubble shape does not vary significantly, more appropriate time-signals are $p_b$ and $U_b$, shown in figure \[fig:periodic\_orbits\_h2\_1\]. The period and amplitude of the weakly nonlinear periodic orbits demonstrate excellent agreement with the fully nonlinear time-dependent simulations.
![[]{data-label="fig:periodic_orbits_h1_2"}](figure_H1_eps_0_2.eps)
![[]{data-label="fig:periodic_orbits_h1_1"}](figure_H1_eps_0_01.eps)
![[]{data-label="fig:periodic_orbits_h1_3"}](figure_H1_eps_0_2437_v4.eps)
![[]{data-label="fig:periodic_orbits_h2_2"}](figure_H2_eps_0_2.eps)
![[]{data-label="fig:periodic_orbits_h2_1"}](figure_H2.eps)
Conclusion {#sec:six}
==========
We have investigated the transient dynamics and invariant solutions of a finite air bubble propagating in a perturbed Hele-Shaw channel, a model system in which to explore complex transition scenarios in fluid mechanics. In § \[sec:one\], time-dependent simulations confirmed that a number of different modes of propagation are possible, including multi-tipped solutions and oscillatory solutions as the flow-rate increases. These simulations and an analysis of the steady solution space revealed a finite width region of bistability (with respect to steady states). The demarcation of this bistable region by two Hopf bifurcations explains the oscillatory solutions observed from the time simulations. The first Hopf bifurcation, H1, is subcritical, resulting in the emergence of unstable periodic orbits and the second, H2, is supercritical, resulting in stable periodic orbits. [[An interesting feature is that the nature of the bistability changes as $Q$ increases. Initially ($Q$ immediately larger than $Q_1$) a slower symmetric steady state, S3, co-exists with an ever-present asymmetric steady state, AS1. A transition occurs at the pitchfork bifurcation as the stable AS2 asymmetric branch emerges. Finally due to the supercritical Hopf bifurcation, H2, a stable periodic orbit appears as an alternative stable invariant solution to the asymmetric AS1 branch.]{}]{}
The criticality of the Hopf bifurcations was determined and periodic orbits were approximated using a weakly nonlinear stability analysis in § \[sec:four\]. [[The method described here applies to a range of governing equations and is independent of the choice of discretisation. It has been implemented in a general form in the open-source library `oomph-lib`]{}]{}. The approximate periodic orbits were in good agreement with the fully nonlinear transient simulations. The oscillations of the unstable orbit near H1 manifest themselves as oscillations all the way around the bubble that cause an oscillatory change in the centroid, $\overline{y}$, of the bubble. For flow rates just above the H2 bifurcation point, stable periodic orbits exist. The oscillations now appear as ‘waves’ on the side of the bubble above the depth-perturbation only, with the lower edge of the bubble ‘resting’ on the lower limits of the depth-perturbation. In this case the offset of the bubble centroid remains non-zero and constant and the underlying steady state is asymmetric. These stable, asymmetric oscillations have similar characteristics to the oscillations found for air-fingers by [@de2009tube; @pailha2012oscillatory; @hazel2013multiple; @thompson2014multiple] and can be viewed as the finite-air bubble analogue. The oscillations are driven by interactions between the edge of the bubble and the edge of the constriction, which is typically only possible for asymmetric bubble shapes. In contrast the unstable periodic orbits near H1 have not been seen before and appear to be a new phenomena in this system. In this case the oscillations may be driven by underdamping in the mechanism that restores a symmetric bubble to the centre of the channel after perturbation; the generic mechanism requires an interaction between bubble deformation and surface tension and is described in [@franco2017propagation].
The periodic orbits near the H1 bifurcation point, although unstable, are important in understanding the full time-dependent behaviour of the system. Figure \[fig:periodic\_orbits\_h1\_3\](a) is intriguing evidence that these unstable periodic orbits are edge states of the system (see, for example [@kerswell2005recent; @duguet2008transition]). Slight changes in the initial condition can result in either the bubble returning to the steady S2 state or breaking up/evolving to the AS1 solution. The evidence presented here shows that the unstable periodic orbit is on the boundary between these two forms of transient behaviour. We have quantified the edge state by approximating the unstable periodic orbits by a weakly nonlinear approximation but others have calculated these by using edge-tracking techniques, including shear flow [@kerswell2005recent; @duguet2008transition], and droplet pinch-off, [@gallino2018edge] which rely on two distinct types of transient behaviour being present. An added complication in this case is that are three possible eventual outcomes; the bubble returning to the S2 steady state, breaking up, or evolving to the AS1 symmetric state. We would expect to see evidence of the periodic orbits and the transitions in the bistable region in any experimental results. The periodic orbit near H1, despite being unstable, is therefore highly influential in the underlying dynamics of a bubble and understanding the basin boundary and a full comparison of the sensitivity of bubble break up to initial conditions, especially close to bifurcation points, is the subject of a combined experimental and theoretical study currently underway.
Acknowledgments
===============
We acknowledge funding from the Engineering and Physical Sciences Research Council through grant number EP/P026044/1.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This review is based on an extensive work done in collaboration with N. Gorkavyi, J. Mather, and T. Taidakova, which aimed at the physical modelling of the interplanetary dust (IPD) cloud in the Solar system, [*i.e.*]{}, establishing a link between the observable characteristics of the zodiacal cloud and the dynamical and physical properties of the parent minor bodies. Our computational approach permits with modest computer resources to integrate the trajectories of hundreds of particles and to effectively store up to $10^{10-11}$ positions, which provides a high fidelity 3D distribution of the dust. Our numerical codes account for the major dynamical effects that govern the motion of IPD particles: the Poynting-Robertson (P-R) drag and solar wind drag; the solar radiation pressure; particle evaporation; gravitational scattering by the planets; and the influence of mean-motion resonances. The incorporation of secular resonances and collisions of dust particles (both mutual and with interstellar dust) is underway. We have demonstrated the efficacy of our codes by performing the following analyses: (i) simulation of the distribution of Centaurs (comets scattered in their journey from the Kuiper belt inward the Solar system) and revealing the effects of the outer planets in producing ‘cometary belts’; (ii) detailed inspection of a rich resonant structure found in these belts, which predicts the existence of gaps similar to the Kirkwood gaps in the main asteroid belt; (iii) a preliminary 3-D physical model of the IPD cloud, which includes three dust components – asteroidal, cometary, and kuiperoidal – and is consistent with the available data of Pioneers and Voyagers dust detectors; (iv) modeling of the IPD cloud, which provides a zodiacal light distribution in accord, to the order of 1%, with a subset of the COBE/DIRBE observations; and (v) showing that the resonant structure in dusty circumstellar disks of Vega and Epsilon Eridani is a signature of embedded extrasolar planets. Further improvements of our modelling and their importance for astronomy and cosmology are outlined.'
author:
- '**Leonid M. Ozernoy**'
title: Physical Modelling of the Zodiacal Dust Cloud
---
\#1[[*\#1*]{}]{} \#1[[*\#1*]{}]{} =
\#1 1.25in .125in .25in
1. Introduction {#introduction .unnumbered}
===============
The necessity to account accurately for zodiacal emission as part of the measurements of the cosmic infrared background has revived and strengthened a long-standing interest in the theoretical aspects of the dynamics, structure, and evolution of the interpanetary dust (IPD) cloud ([*e.g.*]{}, Hauser et al. 1998). Until recently, the main stumbling block to implementing the comprehensive study of IPD has been the absence of a physical model for the interplanetary dust cloud. Such a model would establish a link between the observable characteristics of the zodiacal cloud and the dynamical and physical properties of the parent minor bodies of the Solar system. Recently, we constructed a preliminary physical model of the IPD cloud based on combining new methods elaborated by the present authors (Gorkavyi, Ozernoy & Mather \[GOM hereinafter\] 1996; GOM 1997; Gorkavyi et al. \[GOMT hereinafter\] 1997a-c; GOMT 1998a-c, 2000a-c) [*and*]{} some features of previously used numerical and analytical methods (Haug 1958; Kessler 1981; Leinert et al. 1983; Dermott et al. 1996; Valsecchi & Manara 1997; Liou & Zook 1999). Specifically, our approach, which is based on the use of the kinetic equation for the density of dust in the space of orbital coordinates as well as in the ordinary space, combines analytical (kinetic or hydrodynamical) methods in conjunction with celestial mechanics orbit calculations and numerical computations. More recently, we have substantially strengthened the computational component of our approach by developing a new, more powerful technique described in the next Section, which enables us to get rid of a number of approximations used before and therefore considerably improve our physical modelling.
2. A New Computational Approach Employed {#a-new-computational-approach-employed .unnumbered}
========================================
The development of the physical model of the IPD cloud is complicated by:
- uncertainties in the distribution of comets or other bodies as the major sources of dust in planetary systems;
- difficulties in specifying accurately the level and times (positions) of dust contribution by the objects; and
- the wide variety of relevant physical processes, such as (i) radiation pressure and dissipative effects (Poynting-Robertson drag and stellar wind drag), (ii) resonant interaction with planets, (iii) gravitational scattering by planets, (iv) evaporation and sputtering of dust particles, (v) mutual collisions in the cometary and dust populations, and (vi) orbital variations in the dust production rate by minor bodies.
These complexities exclude attaining an analytical solution, so reliance must be on the numerical modelling. In our approach, two features are worth mentioning: (1) the techniques employed permit construction of high-quality 3D models of IPD clouds with the number of particles (strictly speaking, particle positions) as high as $10^{(10-11)}$; and (2) incorporation of an original stable numerical integrator suitable for both dissipationless and, with some modification, dissipational dynamics of minor bodies or particles.
2.1. Our computational approach {#our-computational-approach .unnumbered}
-------------------------------
Up until now, numerical models suffered from the limited number of particles that could be used in the computations. For instance, so far the best results in modeling the dynamics of dust from the Kuiper belt were obtained by Liou & Zook (1999) with $50-100$ particles of three different sizes using $\sim 10^4$ positions for each particle giving $\sim 10^6$ particle positions. Those authors used their results for 2-D projection of the zodiacal cloud with resolution 1 AU and typical statistics 200-400 particles/AU$^2$ (or 20-40 particles/AU$^3$ for a 3-D model). Such modeling cannot easily match the large-scale structure of the IPD cloud, the respected maps of the zodiacal emission, etc.
In our approach described in Ozernoy, Gorkavyi, & Taidakova \[OGT hereinafter\] (2000a,b) and Ozernoy, Gorkavyi, Mather & Taidakova \[OGMT hereinafter\] (2000), the particle-number limitation is substantially relaxed, which is decisive to provide reliable numerical simulations. In brief, our approach is as follows (for simplicity of understanding, we consider a stationary dust particle distribution in the frame co-rotating with the planet). The locus of the given particle’s positions (taken, say, as $6\cdot 10^3$ positions every revolution about the star) are recorded and considered as the positions of [*many other particles*]{} produced by the same source of dust but [*at a different time*]{}. After this particle ‘dies’ (as a result of infall or ejection from the system by a planet-perturber), its recorded positions sampled over its lifetime form a stationary distribution as if it were produced by [**many**]{} particles. Typically, each run includes $10^4-10^5$ revolutions, i.e. $\sim 10^8$ positions of a dust particle, which is equivalent, for a stationary distribution, to $10^8$ particles. If we allow for 100 sources of dust (in fact, we can include, if necessary, a larger number of sources), after 100 runs we deal with $\sim 10^{10}$ particle positions as if they were real particles. In the present project, we will not only keep information about the dynamical path of each particle (as we did in OGT 2000a,b), but in addition, we will immediately sort the information about the computed coordinates of each particle into $10^6-10^7$ spatial cells (each cell containing $10^3-10^4$ particles), thereby forming a 3D grid that models the dust cloud around the Sun (or a star) (OGMT 2000). An appreciable increase in statistics, compared to Liou & Zook (1999), brings a factor of $10^4$ improvement in the detail of a model and enables us to model the IPD cloud at a qualitatively new, 3-D level. Moreover, our approach makes it possible to study, besides stationary processes, certain non-stationary processes as well, e.g. evolution toward steady-state distributions, dust production from non-steady sources, decrease in particle size (due to evaporation and sputtering) and number (due to collisions), etc.
2.2. Numerical Integrators for Dissipationless and Dissipative Systems {#numerical-integrators-for-dissipationless-and-dissipative-systems .unnumbered}
----------------------------------------------------------------------
Our computational method as well as the use of an implicit second-order integrator (Taidakova 1997) appropriately adapted to achieve our goals (Taidakova & Gorkavyi 1999) are described in more detail in OGT (2000b); as shown there, the integrator for a dissipationless system provides the necessary accuracy of computations on the time scale of $0.5\cdot 10^9$ years. A big advantage of this integrator is its stability: an error in the energy (the Tisserand parameter) does not grow as the number of time steps increases if the value of the step remains the same. The latter situation is exemplified by a [*resonant particle*]{} – it does not approach too close to the planet so that the same time step can be taken. Meanwhile [*non-resonant particles*]{}, in due course of their gravitational scatterings, approach one or another planet from time to time, and therefore one has to change the time step near the planet. Obviously, whenever the time step diminishes near the planet, an error in the Tisserand parameter slowly grows together with the increased number of the smaller time steps. Nevertheless, in our simulations a fractional error in the Tisserand parameter typically does not exceed 0.001 during $3\cdot 10^6$ Neptune’s revolution (OGT 2000b), which amounts to 0.5 Gyrs. To increase the accuracy of the computations, we use a second iteration (OGT 2000a). While the 1st iteration yields the gravitational field between points $A$ and $B$ using an approximative formula based on the particle parameters at point $A$ (because those at point $B$ are still unknown), the 2nd iteration enables us to compute the gravitational field between $A$ and $B$ using a middle position between them because the position of $B$ is already given by the 1st iteration.
As for dissipative systems (e.g., with a P-R drag), a modified implicit second-order integrator has been elaborated (Taidakova and Gorkavyi 1999). It provides a necessary accuracy of integration, which remains stable if the time step of computations is changed in jumps, and not continuously, as the particle approaches the Sun. By applying this approach, one can compute the dynamical evolution of dust particles accounting for virtually [*all*]{} physical processes listed above.
3. Components of the Interplanetary Dust Cloud {#components-of-the-interplanetary-dust-cloud .unnumbered}
==============================================
3.1. Asteroidal Component of Dust {#asteroidal-component-of-dust .unnumbered}
---------------------------------
Under the bombardment of other asteroids and large grains, each asteroid serves as a source of dust. The asteroidal dust gradually approaches the Sun due to the P-R drag or escapes from the Solar system due to gravitational scattering by Jupiter and radiation pressure (for small particles) and solar wind (for small to moderate sized particles, depending on their charge, etc.).
The asteroidal component of the IPD has the following features: (i) its distribution is flat; (ii) the dust density run is expected to be $R^{-1}$ at $R\sim 1$ AU and have a cut-off at $R>2$ AU (GOMT 1997a), $R$ being heliocentric distance; (iii) the asteroidal dust is mainly responsible for the content of the Earth resonant ring (Jackson & Zook 1989, Dermott et al. 1994). Using the ‘dust bands’ data, Dermott et al. (1996) estimate the fraction of asteroidal particles in the IPD cloud to be about 1/3.
3.2. Cometary Component of Dust {#cometary-component-of-dust .unnumbered}
-------------------------------
The cometary component of dust originates from sublimation of comets and has the following features: (i) its distribution is relatively thick; (ii) the dust density run is expected to be $R^{-2.4}$ at $R\sim 1$ AU (GOMT 1997a); (iii) most of the cometary dust escapes from our planetary system due to perturbations by Jupiter and the solar radiation pressure. Except a few attempts (Liou et al. 1995, GOMT 1997a), so far there are no reliable estimations of the cometary dust fraction in the IPD cloud.
3.3. Dust From the Kuiper Belt Objects {#dust-from-the-kuiper-belt-objects .unnumbered}
--------------------------------------
The sources of the IPD cloud cannot be entirely reduced simply to comets (a part of which is also responsible for the observed dust tails) and to asteroids (a part of which assembled in asteroid families is also responsible for the observed ‘dust bands’ in the IPD emission) – a number of facts force us to suspect that additional sources of interplanetary dust must exist:
1\. Chemical analyses indicate that a part of IPD spent a much longer time in space than the typical asteroidal and cometary particles (Flynn 1994, 1996).
2\. Pioneer 10 and 11 data indicate that the dust particles of mass $10^{-(8-9)}$ g have approximately constant flux seen up to 18 AU (Humes 1980, Divine 1993). Similarly, Voyager 1 and 2 data indicate that the dust particles of mass $10^{-(11-12)}$ g are seen from 6 to 40 AU with approximately constant flux $(0.5-1)\cdot 10^{-3}$ particles/m$^2$/s (Gurnett et al. 1997). Both these results cannot be explained by the cometary and asteroidal sources giving an entirely different distribution of dust (GOMT 1997a).
3\. The total number of KBOs inferred by from available observations is $8\cdot 10^8$ (Jewitt 1999), which exceeds the number of the known Jupiter family comets by a factor of $10^6$. This indicates that the overall dust production rate from KBOs may be not negligible compared to that of comets and hence a third important component of the IPD cloud might be the ‘kuiperoidal’ dust, as it has recently been suspected (Backman et al. 1995).
In our opinion, the Kuiper belt influences the formation of the IPD cloud in two ways: (i) as a source of small-size particles slowly drifting toward the Sun under a combined action of the P-R drag and perturbations from the planets; and (ii) as a source of millions of comets between Jupiter and Neptune (Levison & Duncan 1997; OGT 2000a,b), which, in turn, serve as additional sources of dust. The dust can be produced due to evaporation of the volatile material from the KBO surface as a result of a variety of processes, such as the Solar wind and the heating by the Sun, micrometeor bombardment, mutual collisions of kuiperoids (e.g., Stern 2000), etc. Although these processes are very complicated, further work could enable us to put important constraints on the contribution of kuiperoidal dust in the overall dust balance.
So far there are no reliable estimations of the kuiperoidal dust fraction in the inner Solar system. Our working hypothesis quantitatively analyzed below is that [*KBOs and Centaurs*]{} (invisible comets mainly beyond Jupiter) [*could produce an important contribution to the dust content of the IPD cloud*]{}.
4. Simulating the Distributions of Dust Sources and Interplanetary Dust {#simulating-the-distributions-of-dust-sources-and-interplanetary-dust .unnumbered}
=======================================================================
4.1. The Simulated Distribution of Dust Sources in the Outer Solar System {#the-simulated-distribution-of-dust-sources-in-the-outer-solar-system .unnumbered}
-------------------------------------------------------------------------
The outer Solar system beyond the four giant planets includes the Kuiper belt and the Oort cloud, which contain raw material left since the formation of the system. The KBOs are thought to be responsible for progressive replenishment of the observable cometary populations, and gravitational scattering of these objects by the four giant planets can provide their transport from the trans-
\[!ht\]
Neptunian region all the way inward, down to Jupiter (Levison & Duncan 1997; Malhotra, Duncan & Levison 1999; OGT 2000a,b). An approach started in OGT 2000a,b makes emphasis on the structure of cometary populations between Neptune and Jupiter, both in phase space, [*i.e.*]{}, in the space of orbital coordinates {$a,e,i$}, and in real space. Using numerical simulations, we examined the structure of a cometary population near a massive planet, such as a giant planet of the Solar system, starting with one-planet approximation (the Sun plus one planet). By studying the distributions of comets in semimajor axis, eccentricity, pericenter, and apocenter distances, we revealed several interesting features in these distributions. The most remarkable ones include: (i) each giant planet dynamically controls a cometary population that we call the [*‘cometary belt’*]{}); and, (ii) avoidance of resonant orbits by comets. We then enhanced the calculations by determining how a cometary belt is modified when the influence of all eight planets is taken into consideration. To this end, we simulated a stationary distribution of comets, which results from the gravitational scattering of the Kuiper belt objects by mainly four giant planets and takes into account the effects of mean motion resonances. The objects start from the Kuiper belt and are typically traced until the bulk of comets ($\sim 90$%) are ejected from the Solar system (this happens on a time scale of $\lax 0.5$ Gyrs). Accounting for the influence of four giant planets makes the simulated cometary belts overlap (Fig. 1a), but nevertheless keeps almost all their basic features found in the one-planet approximation. In particular, the simulated belts maintain the gaps in the $(a,e)$- and $(a,i)$-space similar to the Kirkwood gaps in the main asteroid belt.
The simulated spatial accumulations of comets near the orbits of all four giant planets – the cometary belts – have a dynamical nature, because the comets belonging to the given planet’s belt are either in a resonance with the host or are gravitationally scattered predominantly by this planet. We conclude that the [*large-scale structure of the Solar system includes the four cometary belts*]{} expected to contain at present 20-30 million scattered comets. Only a tiny fraction of them is currently visible as Jupiter-, Saturn-, etc. family comets.
4.2. Simulations of Dust Distribution from Kuiper Belt Objects {#simulations-of-dust-distribution-from-kuiper-belt-objects .unnumbered}
--------------------------------------------------------------
Knowledge of the simulated distribution of sources of dust, along with the known sources, has enabled us to compute the structure of the asteroidal, cometary, and kuiperoidal components of the IPD cloud. Here, we describe the structure of the latter.
In accordance with the main dynamical factors, we would expect to get three major components of the kuiperoidal dust: i) ‘freely’ drifting particles, (ii) gravitationally scattered particles, and (iii) particles captured into resonances.
In the phase space, we find the dust distribution highly non-uniform, with most of the dust concentrating into the four belts associated with the orbits of the four giant planets, with the Neptune dust belt being the most dense and extended (Fig. 1b). As distinct from the simulated cometary belts described in Sec. 4.1, for which the dominating gravitational scattering results in avoidance of resonant orbits by comets, the dust belts, due to an additional factor – the P-R drag – reveal a rich and complex resonant structure of captured particles.
Using our approach, we have reconstructed the spatial structure of the IPD cloud in the Solar system between 0.5 and 100 AU. Our simulations offer a 3-D physical model of the kuiperoidal dust cloud based on $(2-8)\times 10^6$ cells
\[!htb\]
containing $(0.5-0.6)\times 10^{11}$ positions of dust particles. Here we present the results concerning the distribution of dust particles (of radius 1-2 $\mu$m) produced by 100 KBOs from both the pericenter and apocenter of each. Figs. 2a,b show the spatial structure of kuiperoidal dust up to 60 AU. Our simulations reveal a new dust component in the form of the gravitationally scattered kuiperoidal dust in the belts near Jupiter and Saturn. This [*scattered*]{} population is basically non-resonant, it is highly inclined and possesses large eccentricities. A major part of this component is ejected from the Solar system while passing by Saturn’s and Jupiter’s orbits.
The other, [*resonant*]{} component of kuiperoidal dust is responsible for the regions of elevated dust density. These resonant dust belts, especially near the Neptune’s orbit, can be seen in Fig. 1b. Our simulations used two particle sizes ($1-2~\mu$m and $5-10~\mu$m). The smaller the particle size, the smaller is
the contrast resonant structure / background, in accordance with Liou & Zook (1999). There is a remarkable density minimum between Mars and Jupiter. This minimum, which is seen more clearly in Fig. 2b, is due to the fact that Jupiter either ejects from the Solar system or transfers to more inclined and eccentric orbits an appreciable part of the dust drifting toward the Sun. An increase of dust number density in the region between Mars and Earth is explained by the role of the P-R drag, which results in the density run $\propto R^{-1}$. Finally, as a major result, we find that the number density of kuiperoidal dust increases with heliocentric distance between 4 and 10 AU but forms a plateau between 10 and 50 AU.
4.3. Which of Dust Components Prevails in the Solar System? {#which-of-dust-components-prevails-in-the-solar-system .unnumbered}
-----------------------------------------------------------
Our hypothesis that the kuiperoidal dust dominates in the Solar system (certainly, in its outer regions) has been verified by confronting it with available data on the dust distribution in the Solar system, both outer and inner. Two fundamental facts need to be mentioned: (i) at Earth, the dust density run $R^{-a}$ has the exponent $a=1.3$ (Divine 1993); and (ii) between Jupiter and Neptune, the dust density is almost constant (Humes 1980, Gurnett et al. 1997).
Fig. 3 shows, for various possible components of the IPD cloud, the number density of the simulated dust distribution in the ecliptic plane as a function of heliocentric distance. Within the distance interval of $0.5-1.5$ AU, an averaged density run $R^{-a}$ has the exponent $a=1.5-1.7$ for the cometary dust; $a=1.4$ for the kuiperoidal dust, and $a=1.0$ for the asteroidal component. We find that both the two-component and three-component models described above give us the exponent $a$ very close to $a=1.3$ observed at Earth. However, the difference between the two models, only marginal at Earth, becomes very significant at large heliocentric distances. Fig. 3b demonstrates the simulated dust density run for different components of the IPD cloud at heliocentric distances up to 60 AU. Both the cometary dust and small asteroidal dust decrease their number densities in the distance range of $10-30$ AU as $R^{-(2.4-3.3)}$, whereas that of the kuiperoidal dust changes insignificantly (as $R^{-0.05}$). Thus our hypothesis that the kuiperoidal dust dominates in the outer Solar system explains the Pioneer and Voyager data fairly well, while the traditional view that the bulk of the IPD cloud is produced by the Jupiter family comets turns out in contradiction with the available data of dust detectors, as is clearly seen in Fig. 3c.
5. Modelling the Zodiacal Light {#modelling-the-zodiacal-light .unnumbered}
===============================
5.1. Fitting the COBE DIRBE Data {#fitting-the-cobe-dirbe-data .unnumbered}
--------------------------------
We have computed the contribution to the zodiacal light from each of the components of the IPD cloud described in Secs. 3 and 4. Figs. 4a,b show the results from these computations. Also shown is a comparison with some representative DIRBE data on the zodiacal light, where the comparison is constructed by a weighted mixture of the various components. As can be seen from the ratio $I_{ecliptic}/I_{pole}$ shown in Fig. 4a, both the asteroidal and kuiperoidal components have a flatter shape than the actual IPD cloud. In contrast, the Jupiter family comets create a cloud thicker than the actual one.
\[!ht\]
As seen in Fig. 4b, a three-component (‘asteroidal-cometary-kuiperoidal’) physical model of the IPD cloud describes the COBE data with an average accuracy of 0.85%. Although this model employs three free physical parameters (which actually can be determined from observations), it offers an accuracy that is comparable with the best phenomenological model of the zodiacal cloud using about 50 free parameters (Kelsall et al. 1998).
5.2. Modelling the Zodiacal Light Far from Earth {#modelling-the-zodiacal-light-far-from-earth .unnumbered}
------------------------------------------------
Space observations far from Earth, [*e.g.*]{}, at 3 AU, would offer substantial improvements in the zodiacal light emission and scattering (Mather and Beichman 1996). Unfortunately, a rather accurate multi-parametric model of the zodiacal brightness derived by Kelsall et al. (1998) from the COBE data cannot be
reliably extrapolated to heliocentric distances as large as 3 AU. Meanwhile our physical modeling of the zodiacal cloud makes it possible to evaluate quantitatively the zodiacal light emission and scattering throughout the Solar system (GOMT 2000b).
Using the inferred distribution of the zodiacal dust, we have computed a variety of zodiacal light maps, both for thermal emission and scattered components, at different locations $(R,Z)$ of the observer. Fig. 5 gives a representative illustration of the brightness of asteroidal dust at 5 $\mu$m as a function of latitude $\varphi$ and longitude $\lambda$ (in the telescope’s frame) is given in logarithmic scale (the neighboring contour intensities differ by ${\sqrt e}$). The Sun’s position is (0,0) for $Z=0$ and is shifted to a negative $\varphi$ for $Z=0.25$ AU. At each location, there is a minimum in the zodiacal light which can be seen as a ‘dark spot’ (or several spots). The positions of those minima are explained by an interplay between dependencies of emissivity upon density and temperature.
5.3. Warp of the Zodiacal Cloud {#warp-of-the-zodiacal-cloud .unnumbered}
-------------------------------
The smooth zodiacal dust cloud is inclined to the ecliptic plane by $2.03\pm
0.017^\circ$ (Kelsall et al. 1998). Since during the annual Earth motion the bulk of the zodiacal dust is positioned either above the ecliptic plane or below it, the zodiacal light turns out to be variable.
The cause of warp of the zodiacal cloud is as yet a puzzle. Interestingly, a similar warp has been recently discovered in the circumstellar disk around Beta Pictoris (Heap et al. 2000 and refs. therein). We point out that the observed warp in both cases is caused by the presence of a massive planet such as Jupiter. To prove this, we have plotted the North pole emission using both the data of 41-week helium-cooled period and (virtually so far unused) the data of a more extended (about 3 years), ‘warm’ period of COBE mission (Kelsall et al. 1998). We find a well-pronounced 12-month period, which indicates that an inclined part of the zodiacal cloud is stationary in the inertial frame. Therefore, the warp of the cloud could be associated with the influence of Jupiter on a non-resonant part of the cloud, and the annual variations of the zodiacal emission are induced by the Earth orbital motion through the warp.
Our numerical simulation confirm the efficiency of this process. Since the warp of the disk around $\beta$ Pic can be visualized easier than that around the Sun, it would be instructive to address our numerical simulation of $\beta$ Pic dust disk to illustrate the warp induced by a Jupiter-like planet in an orbit inclined to the disk (see Gorkavyi et al. 2000c). As shown there, a comparison of that modeling with the STIS observations makes the proposed interpretation quite plausible.
5.4. Simulations of Zodiacal Dust in Circumstellar Disks {#simulations-of-zodiacal-dust-in-circumstellar-disks .unnumbered}
--------------------------------------------------------
While warping of a dusty disk, like $\beta$ Pic, seen edge-on might serve as a signature of an embedded planet, simulations of circumstellar disks seen face-on need a different approach. The major factors which cause the re-distribution of dust visible in the circumstellar disk in the presence of an embedded planet are: (i) gravitational scattering by the planet, which produces a central cavity (a ‘hole’) and (ii) resonances, which produce asymmetry in the dust distribution in the form of clumps, arcs, rings, etc. As the dust passes by the planets in its infall, it interacts with them by accumulating in the outer planetary resonances (Liou & Zook 1999, OGMT 2000). As we demonstrate in OGMT (2000), the resonant structure in the dusty circumstellar disks seen face-on, like Vega or Epsilon Eridani, could serve as an efficient tool of planet detection in those systems.
Thus, along with the ability to compute an inproved physical model of the IPD cloud in the Solar system and to explain some key aspects of the available data with it, our tools enable us to predict important details in the dust structure near other stars to be tested in future observations. For instance, using our modeling, we predict that the resonant asymmetric feature revolves around $\epsilon$ Eri and Vega with an angular velocity measurable within a few years (OGMT 2000, Gorkavyi et al. 2000c).
7. Future Work {#future-work .unnumbered}
==============
We plan to improve our physical modeling so as to fit the DIRBE data with a precision much better than 1%. To this end, one needs:
- to accurately compute the density distributions of all dust components incorporating particles of 5 to 10 different sizes and accounting for all known as well as simulated sources of dust;
- to account for evaporation and sputtering of dust as a function of heliocentric distance;
- to include the short-term (days to months) variability and small-scale phenomena in the zodiacal cloud.
The last point deserves a more detailed discussion. The time variations found in the zodiacal light are at a few % level for the short ($\lax 5~\mu$m) wavelength bands and of order of 0.5 to 1% for the longer wavelength bands (Kelsall et al. 1998), [*i.e.*]{}, appreciably greater than expected on the basis of known variability in the bolometric output from the Sun.
The revealed short-term variability might contain contributions of different origins, such as (i) variability of solar wind; (ii) inhomogeneities in the zodiacal cloud along the Earth orbit associated, e.g., with the Earth resonant ring and recent meteoroidal dust or cometary tails; and (iii) dust-source dependence and/or dust response to UV or particle change with time as the solar-wind and/or UV heating ’cook’ the particles.
Given that the IPD inhomogeneities are located at the distances from Earth not exceeding several AU, their parallaxes are as large as tens of degrees. This provides a unique opportunity to reconstruct, via a 3-D ‘computer tomography’, a 3-D map of small-scale inhomogeneities in the zodiacal cloud using the 41 week data.
Knowledge of the variability and local structure of the IPD cloud would allow us to make the next important step and to determine the absolute value of the zodiacal emission. Indeed, in a given direction on the sky, the Galactic emission is constant, whereas the zodiacal emission varies in time. The physical modeling of the emission amplitude in the above direction could enable us to disentangle the Galactic and zodiacal contributions. As a result, a very accurate fitting of the zodiacal light would be possible, able to solve two major problems: (i) to derive the basic physical parameters of the zodiacal cloud containing a valuable information concerning the structure, dynamics, and origin of this cloud, and (ii) the residuals would make it possible to constrain or even evaluate the contribution of the extragalactic infrared background.
8. Summary and Conclusions {#summary-and-conclusions .unnumbered}
==========================
We have developed a physical model of the zodiacal cloud incorporating the real dust sources of asteroidal, cometary, and kuiperoidal origin.
Using our codes, we have demonstrated their efficiency and power by performing the following simulations:
\(i) distribution of the scattered comets, which enables us to reveal the four ‘cometary belts’ associated with the orbits of four giant planets, which are expected to contain 20-30 million of cold comets;
\(ii) detailed analysis of a rich resonant structure found in these belts, which predicts the existence of gaps similar to the Kirkwood gaps;
\(iii) a 3-D physical model of the IPD cloud, which explains the available data of Pioneers and Voyagers dust detectors;
\(iv) zodiacal light distribution in the Solar system, which fits the COBE data with an average accuracy of 0.85%, and
\(v) resonant structure in dusty circumstellar disks of Vega and Epsilon Eridani and a warp in dusty disk of Beta Pictoris considered tobe a signature of embedded extrasolar planets.
Under a set of reasonable assumptions, it seems safe to conclude:
1\. The [*kuiperoidal*]{} dust plays a role more important than previously recognized. It appears to account for the space dust observations beyond 6 AU, while near Earth it could possibly contribute as much as 1/3 of total number density (1/4 of surface density) and 1/3 of the zodiacal emission near ecliptic.
2\. The two other components of the IPD cloud, the [*cometary*]{} and [*asteroidal*]{} dust contribute respectively 36% and 30% of the number density and the zodiacal emission (at ecliptic) near Earth. The cometary particles contribute 60% to the surface density of the IPD cloud near Earth. A solely two-component model (i.e. without the kuiperoidal dust) would give a worse fit of dust distribution at Earth and would fail entirely for the outer Solar system.
3\. Further simulations of resonances associated with the planets embedded in dusty circumstellar disks could enable a breakthrough in the understanding of the circumstellar disk structure and lead to possible planet detection long before direct imaging can find them.
Further development of a multicomponent, high-precision (at the level of a few $\times$0.1%) model of the IPD cloud would allow to solve at a new qualitative level a number of key astronomical problems:
- to get an important information about the major physical effects operating in the Solar system, such as the PR drag, resonant captures, gravitation scattering, role of interstellar particles (including their collisions with the IPD particles), evaporation of dust, efficiency of dust production at different distances from the Sun, etc.;
- to evaluate the parameters of yet undiscovered cometary and asteroidal populations contributing to the origin of the zodiacal cloud;
- to get a reliable template for exo-zodiacal clouds (circumstellar dusty disks) as a basis for revealing the embedded exo-planets;
- to help in interpreting and guiding a number of space missions with dust collectors, such as CASSINI and STARDUST;
- to help in planning the targets for space infrared telescopes, such as SIRFT and NGST;
- to improve evaluations of micro-meteoroid impacts for spacecraft;
- to subtract the zodiacal contribution from the COBE DIRBE data with a high precision to evaluate/constrain the Cosmic Infrared Background.
[**Acknowledgements**]{}. This work would be impossible without my collaboration with Nick Gorkavyi, John Mather, and Tanya Taidakova. I am thankful to Thomas Kelsall for many fruitful discussions. Support by NASA Grant NAG-7065 to George Mason University is acknowledged.
\[\] \[Backman, D.E., Dasgupta, A. & Stencel, R.E. 1995, ApJ 450, L35\] \[Dermott, S.F. et al. 1994, Nature 369, 719\] \[Dermott, S.F. et al. 1996, in [*Physics, Chemistry, and Dynamics of Interplanetary Dust*]{}, ed. $~$ B. Gustafson & M. Hanner, ASP Conf. Ser. 104, p. 143\] \[Divine, N. 1993, J. Geophys. Res. 98E, 17029\] \[Flynn, G.J. 1994, Lunar & Planetary Science XXV, 379\] \[Flynn, G.J. 1996, in [*Physics, Chemistry, and Dynamics of Interplanetary Dust*]{}, ed. $~$ B. Gustafson & M. Hanner, ASP Conf. Ser. 104, p. 171\] \[Gorkavyi, N.N., Ozernoy, L.M. & Mather, J.C. ($\equiv$ GOM) 1996, in [*Physics, Chemistry, and Dynamics of Interplanetary Dust*]{}, ed. B. Gustafson & M. Hanner, (San Francisco: ASP), ASP Conf. Ser. 104, p. 43\] \[Gorkavyi, N.N., Ozernoy, L.M. & Mather, J.C. ($\equiv$ GOM) 1997, ApJ 474, 496\] \[Gorkavyi, N.N., Ozernoy, L.M., Mather, J.C. & Taidakova, T. ($\equiv$ GOMT) 1997a, ApJ 488, 268\] \[Gorkavyi, N.N., Ozernoy, L.M., Mather, J.C. & Taidakova, T. 1997b, Bull. AAS 29, 782\] \[Gorkavyi, N.N., Ozernoy, L.M., Mather, J.C. & Taidakova, T. 1997c, Bull. AAS 29, 1310\] \[Gorkavyi, N.N. et al. ($\equiv$ GOMT) 1998a, Earth, Planets and Space, 50, 539\] \[Gorkavyi, N.N., Ozernoy, L.M., Mather, J.C. & Taidakova, T., 1998b, BAAS 30, 853\] \[Gorkavyi, N.N., Ozernoy, L.M., Mather, J.C. & Taidakova, T., 1998c, BAAS 30, 1143\] \[Gorkavyi, N.N. et al. ($\equiv$ GOMT) 2000a, [astro-ph/0006435]{}; Planetary Space. Sci. (submitted)\] \[Gorkavyi, N. et al. ($\equiv$ GOMT) 2000b, [*The NGST Science and Technology Exposition*]{} (E.P. Smith and K.S. Long, eds.) ASP Conf. Ser. 207, 462\] \[Gorkavyi, N., Ozernoy, L., Mather, J., & Heap, S. 2000c, In [*Disks, Planetesimals, and Planets*]{} (F. Garzon et al., eds.) ASP Conf. Ser. (in press); WWW e-print [astro-ph/0005347]{}\] \[Gurnett, D.A. et al. 1997, Geophys. Res. Lett. 24, p. 3125\] \[Haug, U. 1958, Zeitschrift für Astrophysik, 44, 71\] \[Hauser, M.G. et al. 1998, ApJ, 508, 25\] \[Heap, S. et al. 2000, Ap.J. (in press)\] \[Humes, D.H. 1980, J. Geophys. Res., 85, 5841\] \[Jackson, A.A. & Zook, H.A. 1989, Nature 337, 629\] \[Jewitt, D. 1999, Ann. Rev. Earth. Planet. Sci., 27, 287 \] \[Kelsall, T. et al. 1998, ApJ, 508, 44\] \[Kessler, D.J. 1981, Icarus 48, 39\] \[Leinert, L., Roser, S., & Buitrago, J. 1983, A&A, 118, 345\] \[Levison, H.F. & Duncan M.J. 1997, Icarus 127, 13\] \[Liou, J.-C. & Zook, H.A. 1999, Astron. J. 118, 580\] \[Liou, J.-C., Zook, H.A. & Dermott, S.F. 1996, in [*Physics, Chemistry, and Dynamics of Interplanetary Dust*]{}, ed. B. Gustafson & M. Hanner, (San Francisco: ASP), ASP Conf. Ser. 104, p. 163; Icarus 124, 429\] \[Malhotra, R., Duncan, M., & Levison, H. 1999. In [*Protostars and Planets IV*]{} (in press) $=$ astro-ph/9901155\] \[Mather, J.C. & Beichman, C.A. 1996, [*Unveiling the Cosmic Infrared Background*]{}, Ed. E. Dwek, AIP Conf. Proc. 348, 271\] \[Ozernoy, L.M., Gorkavyi, N.N., & Taidakova, T. 2000a, Planetary Space Science, 48, 993\] \[Ozernoy, L.M., Gorkavyi, N.N., & Taidakova, T. ($\equiv$ OGT) 2000b, Mon. Not. R.A.S. 2000 (submitted), an early version posted in [astro-ph/9812479]{}\] \[Ozernoy, L.M. et al. ($\equiv$ OGMT), 2000c, Astrophys. J. 537, L147\] \[Stern, A. 2000, in [*Highlights of Astronomy, JD 4*]{}, A. Lemaitre & H. Rickman, eds. ASP Conf. Ser. (in press)\] \[Taidakova, T. 1997, in [*Astronomical Data Analyses, Software and Systems VI*]{}, ed. G. Hunt & H.E. Payne, ASP Conf. Ser. 125, p. 174\] \[Taidakova, T. & Gorkavyi, N.N. 1999, [ *The Dynamics of Small Bodies in the Solar Systems: A Major Key to Solar Systems Studies*]{}, Eds. B.A. Steves and B.A. Roy, Kluwer Academic Publishers, 393 \] \[Valsecchi, G.B. & Manara, A. 1997, A&A, 323, 986\]
|
{
"pile_set_name": "ArXiv"
}
|
[Exact controllability in minimal time of the Navier-Stokes periodic flow in a 2D-channel]{}
[Gabriela Marinoschi]{}
Gheorghe Mihoc-Caius Iacob Institute of Mathematical Statistics and
Applied Mathematics of the Romanian Academy,
Calea 13 Septembrie 13, Bucharest, Romania
[email protected]
Abstract. This work is concerned with the necessary conditions of optimality for a minimal time control problem $(P)$ related to the linearized Navier-Stokes periodic flow in a 2D-channel, subject to a boundary input which acts on the transversal component of the velocity. The objective in this problem is the reaching of the laminar regime in a minimum time, as well as its preservation after this time. The determination of the necessary conditions of optimality relies on the analysis of intermediate minimal time control problems $(P_{k})$ for the Fourier modes $"k"$ associated to the Navier-Stokes equations and on the proof of the maximum principle for them. Also it is found that one can construct, on the basis of the optimal controllers of problems $(P_{k}),$ a small time called here *quasi minimal* and a boundary controller which realizes the required objective in $(P).$
**Key words**: minimal time controllability, boundary control, necessary conditions of optimality, Navier-Stokes equations.
**MSC2020**. 93B05, 93C20, 49K20, 35Q35
Introduction
============
In this paper, we focus on the determination of the necessary conditions of optimality for the linearized Navier-Stokes periodic flow in a channel, driven in minimal time towards a stationary laminar regime by a boundary control acting upon the transversal flow velocity.
Controllability of Navier-Stokes flow in finite small-time flow gave rise in the literature to a set of reference works. In the following, we briefly review some titles in the literature devoted to this subject. A maximum principle for the time optimal control of the 2D Navier-Stokes equations is presented in [@VB-97]. For some aspects concerning the Navier-Stokes controllability and stabilization we refer the reader to the monographs [VB-book-stab]{}, [@VB-control] and to the papers [@vb-stab-NA] and [@vb-stab-NS-SCL], investigating the stabilization of the Navier-Stokes flow in a channel by controllers with a vertical velocity observation which acts on the normal component of velocity, and by noise wall normal controllers, respectively. For techniques referring to minimal time controllability results we also mention the papers [@Wang-Wang-2003], [@Wang-Wang-2007], [@GM-ESAIM]. Moreover, we cite the more recent monograph [@Wang-book] where a detailed investigation of time optimal control problems is developed.
The small-time global exact null controllability problem for the Navier-Stokes equation was suggested by J.-L. Lions in [@Lions-61]. The control was a source term supported within a small subset of the domain, which is similar to controlling only a part of the boundary, with the Dirichlet boundary condition on the uncontrolled part of the boundary. An exhaustive presentation of various controllability problems, including also that of Navier-Stokes equations, is found in [@Coron-book]. Here, we shall indicate only a few titles related especially to the controllability by using a boundary control. In [@AF-OI-31], a small-time global exact null controllability is proved for a control supported on the whole boundary, while the paper [@Fursikov-Imanuvilov-1] is devoted to the proof of the local exact controllability of the 2D Navier-Stokes system in a bounded domain in the case when the control function is concentrated on the whole boundary or on some part of it. The small-time global exact null controllability for Navier-Stokes under an irrotational flow boundary condition on the uncontrolled boundaries in a 2D rectangular domain it is proved in [@Chapouly-13]. The exact boundary controllability of the Navier-Stokes system where the controls are supported in a given open subset of the boundary is provided in [@Rodrigues]. More recently, the paper [@Coron-2017] focuses on the small-time controllability presenting a new method, which takes into account the boundary layer for getting the control determination.
Let us consider the fluid flow in a $2$-$D$ infinitely long channel, governed by the incompressible Navier-Stokes equations:
$$\begin{tabular}{ll}
$u_{t}-\nu \Delta u+uu_{x}+vu_{y}=\theta _{x},$ & $v_{t}-\nu \Delta
v+uv_{x}+vv_{y}=\theta _{y},$ \\
$u_{x}+v_{y}=0,$ & \\
$u(t,x+2\pi ,y)=u(t,x,y),$ & $v(t,x+2\pi ,y)=v(t,x,y),$ \\
$u(t,x,0)=u(t,x,L)=0,$ & $v(t,x,0)=0,\mbox{ }v(t,x,L)=0,$ \\
$u(0,x,y)=u_{0},\mbox{ }v(0,x,y)=v_{0},$ & for $t\in \mathbb{R}_{+}=(0,\infty ),$ $x\in \mathbb{R},$ $y\in (0,L).$\end{tabular}
\label{0}$$
Here, $(u,v)$ is the fluid velocity, $\theta $ is the pressure, the subscripts $t,x,y$ represent the partial derivatives with respect to these variables.
We also consider the steady-state flow with zero vertical velocity, governed by (\[0\]). This flow velocity turns out to be of the form $(U(y),0),$ where $U(y)=-\frac{a}{2\nu }\left( \frac{y^{2}}{L}-y\right) ,$ $a\in \mathbb{R}_{+}$ (see e.g., [@Temam])).
The problem we are concerned with is to steer the flow (\[0\]) to the stationary regime $U(y)$, within a minimal finite time, by means of a boundary control $w$ acting at $y=L$ upon the transversal velocity component $v,$ namely by $v(t,x,L)=w(t,x).$ More precisely, the objective is to characterize the boundary control $w$ which could force the flow $(u,v)$ starting from $(u_{0},v_{0})\neq (U(y),0)$ to reach the laminar regime $(U(y),0)$ at a minimal time and, moreover, to preserve it at this value after that time. We stress that we are concerned with the determination of the necessary conditions of optimality and not with the proof of the controllability result. However, we succeed to prove that this action can be done within a *quasi-minimal* time, provided by optimal minimal times for the problems in modes of the Fourier transform of the Navier-Stokes linearized system. Further, we shall describe in detail the arguments.
We shall study this problem for the linearized flow around the laminar steady-state $(U(y),0)$. Also, since the flow is periodic along the longitudinal axis, we shall consider it on a period $(0,2\pi ).$ Thus, we linearize (\[0\]) around $(U(y),0)$ relying on the change of function $u\rightarrow u-U$ and continue to keep the same notation $u$ for the linearized longitudinal velocity. Then, the linearized controlled system reads$$\begin{tabular}{ll}
$u_{t}-\nu \Delta u+Uu_{x}+U_{y}v=\theta _{x},$ & $v_{t}-\nu \Delta
v+Uv_{x}=\theta _{y},$ \\
$u_{x}+v_{y}=0,$ & \\
$u(t,2\pi ,y)=u(t,0,y),$ & $v(t,2\pi ,y)=v(t,0,y),$ \\
$u(t,x,0)=u(t,x,L)=0,$ & $v(t,x,0)=0,\mbox{ }v(t,x,L)=w(t,x),$ \\
$u(0,x,y)=u_{0},\mbox{ }v(0,x,y)=v_{0},$ & for $t\in \mathbb{R}_{+},$ $x\in
(0,2\pi ),$ $y\in (0,L).$\end{tabular}
\label{1}$$
We express the flow controllability in minimal time by the problem $(P)$ $\mbox{\ \ \ \ \ \ \ \ \ Minimize }\Big\{ J(T,w)=T;\mbox{ }T>0,\mbox{ }w\in
H^{1}(0,T;L^{2}(0,2\pi )), \mbox{ }w(0,x)=0, $
$\begin{array}{l}
\displaystyle
\mbox{\ \ \ \ \ \ \ \ \ \ \ \ }\int_{0}^{T}\int_{0}^{2\pi }\left\vert w_{t}(t,x)\right\vert ^{2}dxdt\leq
\rho ^{2},\\
\mbox{\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }u(T,x,y)=0,\mbox{ }v(T,x,y)=0\mbox{ a.e. }(x,y)\in (0,2\pi )\times (0,L) \Big\} ,
\end{array}$
subject to system (\[1\]), with $u_{0}\neq 0,$ $v_{0}\neq 0.$ In addition, by resetting $w$ after $t=T$ one ensures that the null regime is preserved, as we shall see.
It is obvious that the minimal time should be positive. Indeed, by absurd, if $T=0,$ we would have $0=v(T)=v(0)=v_{0},$ and similarly for $u,$ which contradicts the hypothesis. The requirement $w(0,x)=0$ is done especially for technical purposes, but it is also in agreement with the fact that at the initial time the boundary condition at $y=L$ is no-slip. Finally, it is clear that in problem $(P)$ it is important to find information about the controller $w$ only on the interval $(0,T)$ within which the objective is reached. On the interval $(T,\infty )$ the function $w$ can take a whatever value, in particular $0,$ and this choice will have the effect of preserving the zero value for $v$ after the time $T$. That is why the property of $w$ of belonging to $H^{1}$ is required only for $t\in (0,T).$
Our purpose is to find the necessary conditions of optimality for $(P).$ To this end, the following controllability assumption, which will allow the derivation of an observability result for the adjoint system, will be in effect:
$(H)$ For each $(t_{0},T),$ $0\leq t_{0}<T<\infty ,$ and each $(u^{0},v^{0})\in (L^{2}(0,2\pi ;L^{2}(0,L)))^{2}$,
with $\left\Vert v^{0}\right\Vert _{L^{2}(0,2\pi ;(H^{2}(0,L))^{\ast })}\leq
1,$
there exists $w\in H^{1}(0,T;L^{2}(0,2\pi ))$ and $\gamma _{(t_{0},T)}>0,$ $\gamma _{(\cdot ,T)}\in L^{2}(0,T)$ with the properties$$w(\tau )=0\mbox{ for }0\leq \tau \leq t_{0}<T,\mbox{ }\left(
\int_{0}^{T}\int_{0}^{2\pi }\left\vert w_{\tau }(\tau ,x)\right\vert
^{2}dxd\tau \right) ^{1/2}\leq \gamma _{(t_{0},T)},$$such that $u^{t_{0},w}(T,x,y)=0,$ $v^{t_{0},w}(T,x,y)=0$ a.e. $(x,y)\in
(0,2\pi )\times (0,L).$
Here, $(u^{t_{0},w},v^{t_{0},w})$ is the solution to (\[1\]) starting from $(u^{0},v^{0})\neq (0,0)$ at time $t=t_{0},$ and controlled by $w$ and $(H^{2}(0,L))^{\ast }$ is the dual of the space $H^{2}(0,L).$ We note that $\gamma _{(t_{0},T)}$ depends on $(u^{0},v^{0}),$ on the interval $(t_{0},T),$ it is bounded on $(0,T-\delta )$ for all $\delta >0$ and $\gamma
_{(t_{0},T)}\rightarrow \infty $ as $t_{0}\rightarrow T.$ In fact $\gamma
_{(t_{0},T)}$ represents the controllability cost, which should be larger if the objective is expected to be reached in a smaller time, but its singularity is assumed to be square integrable (with respect to $t_{0}\in
(0,T)).$
In what concerns the controllability hypothesis $(H)$, its verification is beyond the purpose of this work.
Next, we are going to describe the organization of the paper. It is convenient to reduce problem $(P)$ to minimization problems for the Fourier coefficients of the velocity. To this end, we write $$f(t,x,y)=\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}f_{k}(t,y)e^{ikx},\mbox{ }f_{k}=\overline{f}_{-k},\mbox{ }f_{0}=0,\mbox{ } \label{0-3}$$(which ensures that $f$ is real), where $i=\sqrt{-1}\in \mathbb{C},$ the set of complex numbers and $\overline{f}$ is the complex conjugate. The notation $f$ stands for $u,v,\theta ,$ and $f_{k}$ stands for $u_{k},v_{k},\theta
_{k}.$ Obviously, $f_{k}(t,y)\in \mathbb{C}$. Similarly, $$\begin{aligned}
w(t,x) &=&\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}w_{k}(t)e^{ikx},\mbox{
}w_{k}=\overline{w}_{-k} \label{0-4} \\
u_{0}(x,y) &=&\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}u_{k0}(y)e^{ikx},\mbox{ }v_{0}(x,y)=\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq
0}v_{k0}(y)e^{ikx},\mbox{ }u_{k0}=\overline{u}_{-k0},\mbox{ }v_{k0}=\overline{v}_{-k0}. \notag\end{aligned}$$Replacing in (\[1\]) the functions by their Fourier series and identifying the coefficients we obtain the system$$\begin{aligned}
(u_{k})_{t}-\nu u_{k}^{\prime \prime }+(\nu k^{2}+ikU)u_{k}+U^{\prime }v_{k}
&=&ik\theta _{k},\mbox{ } \label{2} \\
(v_{k})_{t}-\nu v_{k}^{\prime \prime }+(\nu k^{2}+ikU)v_{k} &=&\theta
_{k}^{\prime }, \notag \\
iku_{k}+v_{k}^{\prime } &=&0,\mbox{ } \notag \\
u_{k}(t,0)=u_{k}(t,L)=v_{k}(t,0)=0,\mbox{ }v_{k}(t,L) &=&w_{k}(t), \notag \\
u_{k}(0,y)=u_{k0},\mbox{ }v_{k}(0,y) &=&v_{k0}. \notag\end{aligned}$$For simplicity, we denote by the superscripts $^{\prime },^{\prime \prime
},^{\prime \prime \prime },^{\mbox{iv}}$ the first four partial derivatives with respect to $y$ of the functions $u_{k}$ and $v_{k}.$
By the Parseval identity, in particular for$$\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\int_{0}^{T}\left\vert
(w_{k})_{t}(t)\right\vert ^{2}dt=\frac{1}{2\pi }\int_{0}^{2\pi }\left(
\int_{0}^{T}\left\vert w_{t}(t,x)\right\vert ^{2}dt\right) dx, \label{0-6}$$we can consider that $$\int_{0}^{T}\left\vert (w_{k})_{t}(t)\right\vert ^{2}dt\leq \rho _{k}^{2},\mbox{ where}\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\rho _{k}^{2}\leq
\rho ^{2}.$$Eliminating $\theta _{k}$ between the first equations in (\[2\]) by using $$v_{k}^{\prime }=-iku_{k} \label{3}$$we obtain the following equation in $v_{k}:$$$\begin{aligned}
(k^{2}v_{k}-v_{k}^{\prime \prime })_{t}+\nu v_{k}^{\mbox{iv}}-(2\nu
k^{2}+ikU)v_{k}^{\prime \prime }+(\nu k^{4}+ik^{3}U+ikU^{\prime \prime
})v_{k} &=&0,\mbox{ } \label{4} \\
v_{k}(t,0)=0,\mbox{ }v_{k}(t,L) &=&w_{k}(t), \label{4-1} \\
v_{k}^{\prime }(t,0)=v_{k}^{\prime }(t,L) &=&0, \label{4-2} \\
v_{k}(0,y) &=&v_{k0}(y), \label{4-3}\end{aligned}$$for $t\in \mathbb{R}_{+},$ $y\in (0,L).$
Consequently, for each $k\in \mathbb{Z},$ $k\neq 0$ we can consider the minimization problem for the mode "$k":$
$(P_k)$ $\qquad
\mbox{Minimize }\Big\{ J_{k}(T,w)=T;\mbox{ }T>0,\mbox{ }w\in
H^{1}(0,T),\mbox{ } w(0)=0, $
$\mbox{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ } \mbox{ }\int_{0}^{T}\left\vert
w_{t}(t)\right\vert ^{2}dt\leq \rho _{k}^{2},\mbox{
}v_{k}(T,y)=0\mbox{ a.e. }y\in (0,L)\Big\} , $\
subject to (\[4\])-(\[4-3\]), with $v_{k0}\neq 0.$
The controllability hypothesis $(H)$ and the Parseval identity provide for each mode $"k"$ the following consequence:
$(H_{k})$ For each $(t_{0},T),$ $0\leq t_{0}<T,$ and each initial datum $v^{0}\in L^{2}(0,L),$ $v^{0}\neq 0,$ $\left\Vert v^{0}\right\Vert
_{(H^{2}(0,L))^{\ast }}\leq 1,$ there exists $w\in H^{1}(0,T)$ and $\gamma
_{(t_{0},T)}\in L^{1}(0,T),$ satisfying $$w(\tau )=0\mbox{ for }0\leq \tau \leq t_{0},\mbox{ }\left(
\int_{0}^{T}\left\vert w_{\tau }(\tau )\right\vert ^{2}d\tau \right)
^{1/2}\leq \gamma _{(t_{0},T)},$$ such that $v^{t_{0},w}(T,y)=0,$ where $v^{t_{0},w}$ is the solution to ([4]{})-(\[4-3\]) starting from $v^{0}$ at time $t=t_{0}.$
Here, we included the constant $\frac{1}{2\pi }$ in (\[0-6\]) in $\gamma
_{(t_{0},T)}.$
As announced previously, the main part of the paper is directed to the determination of the necessary conditions of optimality for $(P)$, which will be deduced from those found for $(P_{k}).$
Here there is the structure of the paper. For technical reasons, by using an appropriate variable transformation, we shall study instead of $(P_{k})$, another problem $(\widehat{P_{k}})$ set on a fixed time interval. In Section 2, we shall prove, in Theorem 2.2, the well-posedness of the transformed state system, and conclude with the existence of a solution $(T_{k}^{\ast
},w_{k}^{\ast })$ to $(\widehat{P_{k}})$. A characterization of the optimality conditions cannot be done directly for problem $(\widehat{P_{k}}), $ so that we have to resort to an approximating problem $(P_{k,\varepsilon }),$ indexed along a small positive parameter $\varepsilon
,$ and prove the existence of a solution $(T_{k,\varepsilon }^{\ast
},w_{k,\varepsilon }^{\ast })$, in Theorem 3.1. A convergence result, formally expressed by $(P_{k,\varepsilon })\rightarrow (\widehat{P_{k}}),$ as $\varepsilon \rightarrow 0,$ will be proved in Theorem 3.2. The latter actually shows that if we fix an optimal pair $(T_{k}^{\ast },w_{k}^{\ast })$ in $(\widehat{P_{k}})$ we can recover it as a limit of a sequence of solutions $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast })$ to $(P_{k,\varepsilon }).$ Based on these results, we proceed to the calculation of the necessary conditions of optimality for the approximating minimization problem $(P_{k,\varepsilon })$ in Proposition 3.4. They can be established if $T_{k}^{\ast }$ is small enough and $\rho _{k}$ is chosen sufficiently large. Appropriate fine estimates following by an observability result allow to pass to the limit in the approximating optimality conditions to get those corresponding to problem $(\widehat{P_{k}})$ in Theorem 4.1
Finally, by relying on the Fourier characterization of $u,$ $v$ and $w$ we prove in Theorem 5.2 that, if $(P)$ has an admissible pair $(T_{\ast
},w_{\ast }),$ there exists $(T^{\ast },w^{\ast })$ which steers $(u_{0},v_{0})$ into $u(T^{\ast })=v(T^{\ast })=0$ and $T^{\ast }\leq T_{\ast
}.$ This pair is constructed on the basis of $(T_{k}^{\ast },w_{k}^{\ast })$ with $T_{k}^{\ast }$ minimal in problems $(P_{k}).$ For this reason we call it a *quasi minimal* time for $(P).$ However, it is not clear if it is precisely the minimal one.
**Notation.** Let $X_{\mathbb{R}}$ be a real Banach space and let $T>0$. We denote by $X$ the complexified space $X_{\mathbb{R}}+iX_{\mathbb{R}}$ and by $L^{p}(0,T;X),$ $W^{1,p}(0,T;X),$ $C^{l}([0,T];X)$ the complexified spaces containing functions of the form $f_{1}+if_{2},$ with $f_{1},f_{2}\in L^{2}(0,T;X_{\mathbb{R}}),$ $W^{k,p}(0,T;X_{\mathbb{R}})$ and $C^{l}([0,T];X_{\mathbb{R}}),$ respectively, for $p\in \lbrack 1,\infty ],$ $l\in \mathbb{N}.$
The space $W^{k,p}(0,T;X_{\mathbb{R}})=\{f\in L^{p}(0,T;X_{\mathbb{R}});$ $\frac{\partial ^{m}f}{\partial t^{m}}\in L^{p}(0,T;X_{\mathbb{R}}),$ $m=1,...,k\}.$
We shall use the standard Sobolev spaces $(H^{l}(0,1))_{\mathbb{R}},$ $(H_{0}^{1}(0,1))_{\mathbb{R}}$ and denote$$H_{\mathbb{R}}:=(L^{2}(0,L))_{\mathbb{R}},\mbox{ }(H_{0}^{2}(0,L))_{\mathbb{R}}=\{f\in (H^{2}(0,L))_{\mathbb{R}};\mbox{ }f(0)=f(L)=f^{\prime
}(0)=f^{\prime }(L)=0\}.$$We have $(H_{0}^{2}(0,L))_{\mathbb{R}}\subset (H_{0}^{1}(0,L))_{\mathbb{R}}\subset H_{\mathbb{R}}\subset (H_{0}^{1}(0,L))_{\mathbb{R}}^{\ast }\subset
((H_{0}^{2}(0,L))_{\mathbb{R}})^{\ast }$ with compact injections, where $(H_{0}^{1}(0,L))_{\mathbb{R}}^{\ast }$ and $((H_{0}^{2}(0,L))_{\mathbb{R}})^{\ast }$ are the duals of $(H_{0}^{1}(0,L))_{\mathbb{R}}$ and $(H_{0}^{2}(0,L))_{\mathbb{R}},$ respectively. Their corresponding complexified spaces $H_{0}^{2}(0,L),$ $H_{0}^{1}(0,L),$ $H,$ $(H_{0}^{1}(0,L))^{\ast },$ $(H_{0}^{2}(0,L))^{\ast }$ are defined as before, and satisfy $$H_{0}^{2}(0,L)\subset H_{0}^{1}(0,L)\subset H\subset (H_{0}^{1}(0,L))^{\ast
}\subset (H_{0}^{2}(0,L))^{\ast }$$with compact injections. Also, we define the spaces $$\begin{aligned}
V_{T} &:&=\{f\in H^{1}(0,T);\mbox{ }f(0)=0\},\mbox{ with the norm }\left\Vert w\right\Vert _{V_{T}}^{2}=\int_{0}^{T}\left\vert \dot{w}(t)\right\vert ^{2}dt, \label{VT} \\
V_{1} &:&=\{f\in H^{1}(0,1);\mbox{ }f(0)=0\},\mbox{ with the norm }\left\Vert w\right\Vert _{V_{1}}^{2}=\int_{0}^{1}\left\vert \dot{w}(t)\right\vert ^{2}dt. \notag\end{aligned}$$We denote by $\left\vert \zeta \right\vert $ the norm of $\zeta =\zeta
_{1}+i\zeta _{2}\in \mathbb{C}$, the space of complex numbers. The scalar product on $\mathbb{C}$ is defined as $$(a,b)_{\mathbb{C}}=a\overline{b},\mbox{ for }a,b\in \mathbb{C},\mbox{ with }\overline{b}\mbox{ the complex conjugate.} \label{0-5}$$The scalar product and norm in $H$ are defined by $$\left( \zeta ,z\right) _{H}=\int_{0}^{L}\zeta (y)\overline{z}(y)dy,\mbox{ }\left\Vert z\right\Vert _{H}=\left( \int_{0}^{L}\left\vert z(y)\right\vert
^{2}dy\right) ^{1/2},\mbox{ }\zeta ,z\in H. \label{5}$$The first and second derivatives of a function $w$ depending only on $t$ will be denoted by $\dot{w}$ and $\ddot{w}.$
Problems $(P_{k})$ and $(\protect\widehat{P_{k}})$
==================================================
In order to handle in a more convenient way the arguments in the proofs of the next results, and especially for calculating the approximating necessary conditions of optimality, we shall use a state system and a new minimization problem for the modes $k$, on a fixed time interval, by making an appropriate transformation in order to bring the interval $(0,T)$ into $(0,1).$ To this end, we set in the state system$$t=\widehat{t}T,\mbox{ }v_{k}(t,y)=\widehat{v_{k}}(\widehat{t}T,y):=\widehat{v_{k}}(\widehat{t},y),\mbox{ }\widehat{w_{k}}(t)=w_{k}(\widehat{t}),\mbox{ }
\label{t-cap}$$such that $\widehat{t}\in \lbrack 0,1]$ when $t\in \lbrack 0,T].$
Then, the restriction $\left\Vert w\right\Vert _{V_{T}}\leq \rho _{k}$ becomes $\left\Vert \widehat{w}\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T},$ with $V_{1}$ defined in (\[VT\]).
The state system (\[4\])-(\[4-3\]) is transformed into the appropriate system for $\widehat{v_{k}}$ $$\begin{aligned}
(k^{2}\widehat{v_{k}}-\widehat{v_{k}}^{\prime \prime })_{\widehat{t}}+T\left( \nu \widehat{v_{k}}^{\mbox{iv}}-(2\nu k^{2}+ikU)\widehat{v_{k}}^{\prime \prime }+(\nu k^{4}+ik^{3}U+ikU^{\prime \prime })\widehat{v_{k}}\right) &=&0,\mbox{ } \label{4'} \\
\widehat{v_{k}}(\widehat{t},0)=0,\mbox{ }\widehat{v_{k}}(\widehat{t},L) &=&\widehat{w_{k}}(\widehat{t}), \label{4''} \\
\widehat{v_{k}}^{\prime }(\widehat{t},0)=\widehat{v_{k}}^{\prime }(\widehat{t},L) &=&0, \label{4'''} \\
\widehat{v_{k}}(0,y) &=&v_{k0}(y), \label{4''''}\end{aligned}$$for $\widehat{t}\in (0,1),$ $y\in (0,L).$
In this way, problem $(P_{k})$ becomes $(\widehat{P_{k}})$ below: $(\widehat{P_{k}})$ $\quad\mbox{Minimize }\Big\{ J_{k}(T,\widehat w)=T;\ T>0,\ \widehat w\in V_{1}, \displaystyle\int_{0}^{1}\!\!\left\vert \dot{\widehat{w}} (t)\right\vert ^{2}dt\leq \rho _{k}^{2}T, \mbox{ } \widehat{v_{k}}(1,y)=0,\text{ a.e. }y\in (0,L)\Big\}$
subject to (\[4’\])-(\[4””\]), with $v_{k0}\neq 0,$ where $V_{1}$ was defined in (\[VT\]).
The controllability hypothesis $(H_{k})$ will be correspondingly written on the interval $(\widehat{t_{0}},1)$ and denoted by $(\widehat{H_{k}}).$
**Note.** *However, for not overloading the notation, we shall skip in sections 2-4 the notation with the decoration* $"\symbol{94}"$ *and will resume it in Theorem 5.2. Thus, in system* (\[4’\])-([4””]{}) *we shall* *write* $t,$* *$v_{k},$* *$w_{k}$*, instead of* $\widehat{t},\widehat{v_{k}},\widehat{w_{k}}.$
In this section we prove the well-posedness for the state system derived from (\[4’\])-(\[4””\]) and the existence of a solution to $(\widehat{P_{k}})$, which obviously imply the same results for (\[4\])-(\[4-3\]) and $(P_{k}).$
We begin with some definitions. For each $k\in \mathbb{Z}\backslash \{0\}$ let us define the operators $$E_{0k}:D(E_{0k})\subset H\rightarrow H,\mbox{ }D(E_{0k})=H^{2}(0,L)\cap
H_{0}^{1}(0,L),\mbox{ }E_{0k}z:=k^{2}z-z^{\prime \prime } \label{10-3}$$and$$\begin{aligned}
F_{0k} &:&D(F_{0k})\subset H\rightarrow H,\mbox{ }D(F_{0k})=H^{4}(0,L)\cap
H_{0}^{2}(0,L), \label{10-4} \\
F_{0k}z &=&\nu z^{\mbox{iv}}-(2\nu k^{2}+ikU)z^{\prime \prime }+(\nu
k^{4}+ik^{3}U+ikU^{\prime \prime })z. \notag\end{aligned}$$Since $E_{0k}$ is $m$-accretive, coercive, hence invertible, with the inverse continuous on $H,$ we can define the operator $$A_{k}:=F_{0k}E_{0k}^{-1},\mbox{ }A_{k}:D(A_{k})\subset H\rightarrow H,\mbox{
}D(A_{k})=\{v\in H;\mbox{ }E_{0k}^{-1}v\in D(F_{0k})\}. \label{10-2-4}$$We also observe that,$$v\in D(A_{k})\mbox{ iff }v=E_{0k}\varphi ,\mbox{ for }\varphi \in
H^{4}(0,L)\cap H_{0}^{2}(0,L). \label{iff}$$
By Lemma 1 in [@vb-stab-NS-SCL] we know that $A_{k}$ is closed and densely defined in $H,$ and $-A_{k}$ generates a $C_{0}$-analytic semigroup on $H,$ that is, its resolvent has the property $$\left\Vert (\sigma I+A_{k})^{-1}f\right\Vert _{H}\leq \frac{\left\Vert
f\right\Vert _{H}}{\left\vert \sigma \right\vert -\sigma _{0}},\mbox{ for
all }f\in H\mbox{ and }\left\vert \sigma \right\vert >\sigma _{0}.
\label{resolvent}$$
**Definition 2.1.** We call a solution to** **(\[4’\])-(\[4””\]) a function $$v_{k}\in C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap
L^{2}(0,1;H^{4}(0,L)),$$ which satisfies (\[4’\])-(\[4””\]) for a.a. $t>0.$
**Theorem 2.2.** *Let* $T>0$ *and* $v_{k0}\in
H^{4}(0,L)\cap H_{0}^{2}(0,L),$ $w_{k}\in V_{1}.$* Then, problem* (\[4’\])-(\[4””\]) *has a unique solution* $$v_{k}\in C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap
L^{2}(0,1;H^{4}(0,L)), \label{vk}$$*which satisfies the estimate*$$\begin{aligned}
&&\left\Vert v_{k}(t)\right\Vert _{H^{2}(0,L)}^{2}+\int_{0}^{1}\left\Vert
\frac{dv_{k}}{dt}(t)\right\Vert _{H^{2}(0,L)}^{2}dt+T\int_{0}^{1}\left\Vert
v_{k}(t)\right\Vert _{H^{4}(0,L)}^{2}dt \label{vk-est1} \\
&\leq &C\left( \left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap
H_{0}^{2}(0,L)}^{2}+T\int_{0}^{1}\left\vert w_{k}(t)\right\vert
^{2}dt+\int_{0}^{1}\left\vert \dot{w}_{k}(t)\right\vert ^{2}dt\right) ,\mbox{
\textit{for all} }t\geq 0. \notag\end{aligned}$$*The solution is continuous with respect to the data, that is, two solutions* $(v_{k}^{1},v_{k}^{2})$* corresponding to the data* $(v_{k0}^{1},w_{k}^{1})$* and* $(v_{k0}^{2},w_{k}^{2})$*satisfy the estimate* $$\begin{aligned}
&&\left\Vert (v_{k}^{1}-v_{k}^{2})(t)\right\Vert
_{H^{2}(0,L)}^{2}+\int_{0}^{1}\left\Vert \frac{d(v_{k}^{1}-v_{k}^{2})}{dt}(t)\right\Vert _{H^{2}(0,L)}^{2}dt \label{vk-est2} \\
&&+T\int_{0}^{1}\left\Vert (v_{k}^{1}-v_{k}^{2})(t)\right\Vert
_{H^{4}(0,L)}^{2}dt \notag \\
&\leq &C\left( \left\Vert (v_{k0}^{1}-v_{k0}^{2})\right\Vert
_{H^{4}(0,L)\cap H_{0}^{2}(0,L)}^{2}+\int_{0}^{1}\left( T\left\vert
(w_{k}^{1}-w_{k}^{2})(t)\right\vert ^{2}+\left\vert (\dot{w}_{k}^{1}-\dot{w}_{k}^{2})(t)\right\vert ^{2}\right) dt\right) ,\mbox{ } \notag\end{aligned}$$*for all* $t\geq 0.$
**Proof.** We recall that $V_{1}:=\{f\in H^{1}(0,1);$ $f(0)=0\}$. Let us introduce a function transformation in order to homogenize the boundary conditions, namely $$\widetilde{v}_{k}(t,y)=v_{k}(t,y)-\beta (y)w_{k}(t), \label{v-beta}$$where$$\beta (y)=-\frac{2}{L^{3}}y^{3}+\frac{3}{L^{2}}y^{2}. \label{beta}$$This transformation is chosen such that $\widetilde{v}_{k,\varepsilon }(t,0)=\widetilde{v}_{k,\varepsilon }(t,L)=\widetilde{v}_{k,\varepsilon }^{\prime
}(t,0)=\widetilde{v}_{k,\varepsilon }^{\prime }(t,L)=0$. Equation (\[4’\]) is transformed into $$\begin{aligned}
&&(k^{2}\widetilde{v}_{k}-\widetilde{v}_{k}^{\prime \prime })_{t}+T\left(
\nu \widetilde{v}_{k}^{\mbox{iv}}-(2\nu k^{2}+ikU)\widetilde{v}_{k}^{\prime
\prime }+(\nu k^{4}+ik^{3}U+ikU^{\prime \prime })\widetilde{v}_{k}\right)
\label{10-2} \\
&=&a_{k}(y)w_{k}(t)+Tb_{k}(y)\dot{w}_{k}(t), \notag\end{aligned}$$where$$a_{k}=-\left( \nu \beta ^{\mbox{iv}}-(2\nu k^{2}+ikU)\beta ^{\prime \prime
}+(\nu k^{4}+ik^{3}U+ikU^{\prime \prime })\beta \right) ,\mbox{ }b_{k}=-(k^{2}\beta -\beta ^{\prime \prime }) \label{ak-bk}$$and $a_{k},$ $b_{k}\in C^{\infty }(0,L).$ We denote $\widetilde{\widetilde{v}}_{k}(t)=k^{2}\widetilde{v}_{k}(t)-\widetilde{v}_{k}^{\prime \prime }(t),$ for $t\in (0,1),$ and note that since $\widetilde{v}_{k}$ vanishes at the boundaries, we have in fact $$\widetilde{\widetilde{v}}_{k}(t)=E_{0k}\widetilde{v}_{k}(t),\mbox{ for }t\in
(0,1). \label{vL0}$$Since $\widetilde{\widetilde{v}}_{k}(0)=(k^{2}v_{k0}-v_{k0}^{\prime \prime
})-(k^{2}\beta -\beta ^{\prime \prime })w_{k}(0)$ and $w_{k}(0)=0,$ equation (\[10-2\]) can be written as the equivalent Cauchy problem $$\begin{aligned}
\frac{d\widetilde{\widetilde{v}}_{k}}{dt}(t)+TA_{k}\widetilde{\widetilde{v}}_{k}(t) &=&Ta_{k}w_{k}(t)+b_{k}\dot{w}_{k}(t),\mbox{ a.e. }t\in (0,1),
\label{10-2-1} \\
\widetilde{\widetilde{v}}_{k}(0) &=&E_{0k}v_{k0}, \notag\end{aligned}$$where $A_{k}=F_{0k}E_{0k}^{-1},$ by (\[10-2-4\]). We recall that $-A_{k}$ generates a $C_{0}$-analytic semigroup and so the solution to (\[10-2-1\]) is given by $$\widetilde{\widetilde{v}}_{k}(t)=v_{1}(t)+v_{2}(t)+v_{3}(t),\mbox{ }$$where $$v_{1}(t)=e^{-tTA_{k}}\widetilde{\widetilde{v}}_{k}(0),\mbox{ }v_{2}(t)=T\int_{0}^{t}(e^{-(t-s)TA_{k}}a_{k})w_{k}(s)ds,\mbox{ }v_{3}(t)=\int_{0}^{t}(e^{-(t-s)TA_{k}}b_{k})\dot{w}_{k}(s)ds.$$Since $v_{k0}\in H^{4}(0,L)\cap H_{0}^{2}(0,L)$, it follows that $\widetilde{\widetilde{v}}_{k}(0)=E_{0k}v_{k0}\in D(A_{k}).$ Then, using the existence theorems for the solutions to equations with a $C_{0}$-analytic semigroup (see e.g. [@Pazy], Theorem 3.5, p. 114 and [@vbp-2012], Proposition 1.148, p. 60), it follows for the first term that$$v_{1}\in C([0,1];D(A_{k}))\cap C^{1}([0,1];H),$$$$\left\Vert v_{1}(t)\right\Vert _{D(A_{k})}+\left\Vert \frac{dv_{1}}{dt}(t)\right\Vert _{H}\leq C\left\Vert E_{0k}v_{k0}\right\Vert _{D(A_{k})}\leq
C\left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap H_{0}^{2}(0,L)},\mbox{ for
all }t\in \lbrack 0,1].$$The second term can be viewed as the solution to the Cauchy problem $$\begin{aligned}
\frac{dv_{2}}{dt}(t)+TA_{k}v_{2}(t) &=&Ta_{k}w_{k}(t),\mbox{ a.e. }t\in
(0,1), \\
v_{2}(0) &=&0,\end{aligned}$$where $w_{k}a_{k}\in L^{2}(0,1;H),$ whence $v_{2}\in C([0,1];H)\cap
W^{1,2}(0,1;H)\cap L^{2}(0,1;D(A_{k}))$ and $$\left\Vert v_{2}(t)\right\Vert _{H}^{2}+\int_{0}^{1}\left\Vert \frac{dv_{2}}{dt}(t)\right\Vert _{H}^{2}dt+T\int_{0}^{1}\left\Vert
A_{k}v_{2}(t)\right\Vert _{H}^{2}dt\leq CT\int_{0}^{1}\left\vert
w_{k}(t)\right\vert ^{2}dt.$$Similarly, $v_{3}\in C([0,1];H)\cap W^{1,2}(0,1;H)\cap L^{2}(0,1;D(A_{k}))$ and $$\left\Vert v_{3}(t)\right\Vert _{H}^{2}+\int_{0}^{1}\left\Vert \frac{dv_{3}}{dt}(t)\right\Vert _{H}^{2}dt+T\int_{0}^{1}\left\Vert
A_{k}v_{3}(t)\right\Vert _{H}^{2}dt\leq C\int_{0}^{1}\left\vert \dot{w}_{k}(t)\right\vert ^{2}dt.$$Gathering the results for $v_{1},v_{2},v_{3},$ we get$$\widetilde{\widetilde{v}}_{k}\in C([0,1];H)\cap W^{1,2}(0,1;H)\cap
L^{2}(0,1;D(A_{k})),$$$$\begin{aligned}
&&\left\Vert \widetilde{\widetilde{v}}_{k}(t)\right\Vert
_{H}^{2}+\int_{0}^{1}\left\Vert \frac{d\widetilde{\widetilde{v}}_{k}}{dt}(t)\right\Vert _{H}^{2}dt+T\int_{0}^{1}\left\Vert A_{k}\widetilde{\widetilde{v}}_{k}(t)\right\Vert _{H}^{2}dt \\
&\leq &C\left( \left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap
H_{0}^{2}(0,L)}^{2}+T\int_{0}^{1}\left\vert w_{k}(t)\right\vert
^{2}dt+\int_{0}^{1}\left\vert \dot{w}_{k}(t)\right\vert ^{2}dt\right) .\end{aligned}$$By (\[vL0\]) we obtain $$\widetilde{v}_{k}\in C([0,1];H^{2}(0,L)\cap H_{0}^{1}(0,L))\cap
W^{1,2}(0,1;H^{2}(0,L)\cap H_{0}^{1}(0,L))\cap L^{2}(0,1;H^{4}(0,L)\cap
H_{0}^{1}(0,L)),$$the last space being derived by $E_{0k}^{-1}(D(A_{k}))\subset H^{4}(0,L)\cap
H_{0}^{1}(0,L).$ Finally, by (\[v-beta\]) we obtain (\[vk\]). Estimate (\[vk-est1\]) follows by the estimate for $\widetilde{\widetilde{v}}_{k}$. Since the equation is linear we also get (\[vk-est2\]). This implies still the uniqueness. The proof is ended.$\square $
Even if we are only interested in obtaining the conditions of optimality, we also provide later, for the reader convenience, the proof of the existence of a solution to $(P_{k}).$
To this end, we define an admissible pair for $(P)$ a pair $(T_{\ast
},w_{\ast })$ with $w_{\ast }\in H^{1}(0,T;L^{2}(0,2\pi )),$ $w_{\ast
}(0,y)=0,$ $\int_{0}^{T}\int_{0}^{2\pi }\left\vert w_{\ast
t}(t,x)\right\vert ^{2}dxdt\leq \rho ^{2}$ and $u(T,x,y)=0,$ $v(T,x,y)=0$. By the Parseval identity this implies that $(T_{\ast },w_{\ast k})$ is an admissible pair for $(\widehat{P_{k}}),$ with $w_{\ast k}$ being the mode $k$ of $w_{\ast }$. This will be resumed in Theorem 5.2. We are not concerned here with the proof of the existence of an admissible pair. Related results can be found in the literature already cited referring to the controllability in small-time.
**Theorem 2.3.** *Let* $v_{k0}\in H^{4}(0,L)\cap
H_{0}^{2}(0,L),$ $v_{k0}\neq 0.$ *If* $(\widehat{P_{k}})$*has an admissible pair*, *then it has at least a solution* $(T_{k}^{\ast },w_{k}^{\ast })$ *with the corresponding optimal state* $v_{k}^{\ast }$*. Moreover, let us set* $\widetilde{w_{k}}(t):=w_{k}^{\ast }(t)$ *for* $t\in \lbrack 0,1],$ *and* $\widetilde{w_{k}}(t)=0$* for* $t\in (1,\infty ).$* Then,* $v_{k}^{\ast }(t)=0$ *for* $t>1.$
**Proof.** Following the assumption before, $(\widehat{P_{k}}) $ has an infimum denoted $T_{k}^{\ast }$ which is positive. We consider a minimizing sequence $(T_{k}^{n},w_{k}^{n})$ such that $T_{k}^{n}>0,$ $\left\Vert w_{k}^{n}\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T_{k}^{n}}$, with $v_{k}^{n}(1,y)=0$ and $$T_{k}^{\ast }\leq J(T_{k}^{n},w_{k}^{n})=T_{k}^{n}\leq T_{k}^{\ast }+\frac{1}{n},\mbox{ }n\geq 1. \label{J}$$This yields $T_{k}^{n}\rightarrow T_{k}^{\ast }$ as $n\rightarrow \infty .$ Also, there exists $w_{k}^{\ast }\in H^{1}(0,1)$ such that, on a subsequence, $w_{k}^{n}\rightarrow w_{k}^{\ast }$ weakly in $H^{1}(0,1),$ strongly in $C([0,1])$ by Arzelà theorem, and $\left\Vert w_{k}^{\ast
}\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T_{k}^{\ast }}.$ Thus, $w_{k}^{n}(0)\rightarrow w_{k}^{\ast }(0)=0.$ The solution to (\[4’\])-([4””]{}) corresponding to $(T_{k}^{n},w_{k}^{n})$ is denoted $v_{k}^{n},$ has the properties (\[vk\])-(\[vk-est2\]) with $T=T_{k}^{n}$ and $v_{k}^{n}(1,y)=0$. Thus, by a simple calculation, handling the property ([vk-est2]{}) we get$$v_{k}^{n}\rightarrow v_{k}^{\ast }\mbox{ strongly in }C[0,1];H^{2}(0,L))\cap
W^{1,2}(0,1;H^{2}(0,L))\cap L^{2}(0,1;H^{4}(0,L)).$$We can pass to the limit in (\[4’\])-(\[4””\]) written for $(T_{k}^{n},w_{k}^{n})$ to obtain that $v_{k}^{\ast }$ is the solution to (\[4’\])-(\[4””\]) corresponding to $(T_{k}^{\ast },w_{k}^{\ast }).$ Moreover, since $v_{k}^{n}\rightarrow v_{k}^{\ast }$ strongly in $C([0,1];H)$ we also have $v_{k}^{\ast }(1,y)=0,$ so that $(T_{k}^{\ast },w_{k}^{\ast })$ is optimal in $(\widehat{P_{k}}).$
Next, we prove the last assertion in the statement of the theorem. If $w_{k}^{\ast }$ is extended by $0$ on $(1,\infty ),$ the system for the variable $\chi _{k}$ starting from the initial datum at $t=1$ reads $$\begin{aligned}
(k^{2}\chi _{k}-\chi _{k}^{\prime \prime })_{t}+\nu \chi _{k}^{\mbox{iv}}-(2\nu k^{2}+ikU)\chi _{k}^{\prime \prime }+(\nu k^{4}+ik^{3}U+ikU^{\prime
\prime })\chi _{k} &=&0, \\
\chi _{k}(t,0)=0,\mbox{ }\chi _{k}(t,L)=0,\mbox{ }\chi _{k}^{\prime
}(t,0)=\chi _{k}^{\prime }(t,L) &=&0, \\
\chi _{k}(1,y)=v_{k}^{\ast }(1,y) &=&0,\end{aligned}$$for $(t,y)\in (1,\infty )\times (0,L)$. Obviously, it has the unique solution $0$, which extends the solution $v_{k}^{\ast }$ on $(0,1).$$\square $
We note that in problem $(\widehat{P_{k}}),$ the optimal state satisfies (\[4’\])-(\[4””\]) with $T=T_{k,}^{\ast }$ and $w_{k}^{\ast }.$
The approximating problem $(P_{k,\protect\varepsilon })$
========================================================
In this section, we introduce an approximating minimization problem $(P_{k,\varepsilon })$, prove the existence of a solution, its convergence to $(\widehat{P_{k}})$ and determine the approximating necessary conditions of optimality.
Let $(T_{k}^{\ast },w_{k}^{\ast })$ be a solution to $(\widehat{P_{k}})$ and let $\varepsilon >0.$ We introduce the following approximating problem:
$(P_{k,\varepsilon})$ $\mbox{Minimize}\Big\{ J_{k,\varepsilon }(T,w)=T+\displaystyle\frac{1}{2\varepsilon }\left\Vert (\sigma I+A_{k})^{-1}v_{k}(1)\right\Vert _{H}^{2}$
$\hspace*{40mm}+\displaystyle\frac{1}{2}\int_{0}^{1}\left\vert \int_{0}^{t}(w_{k}-w_{k}^{\ast })(\tau )d\tau
\right\vert ^{2}dt;\ T>0, \mbox{ }, w\in V_1,\ \left\Vert w\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T} \Big\} ,$\
subject to the approximating system (\[4’\])-(\[4””\]). We underline that $v_{k}(1)$ is $v_{k}(1,y).$
**Theorem 3.1.** *Let* $v_{k0}\in H^{4}(0,L)\cap
H_{0}^{2}(0,L),$ $v_{k0}\neq 0.$* Then, problem* $(P_{k,\varepsilon
})$ *has at least a solution* $(T_{k,\varepsilon }^{\ast
},w_{k,\varepsilon }^{\ast })$ *with the corresponding optimal state* $v_{k,\varepsilon }^{\ast }$*.*
**Proof.** For** **$(P_{k,\varepsilon })$ we see that there exists at least an admissible pair, which is $(T_{k}^{\ast
},w_{k}^{\ast }),$ the optimal pair in $(\widehat{P_{k}}).$ Then, $J_{k,\varepsilon }(T,w)$ is positive and so there exists $d_{\varepsilon
}=\inf J_{k,\varepsilon }(T,w)$ and it is positive. Indeed, by absurd if $J_{k,\varepsilon }(T,w)=0,$ then each term, including $T,$ should be equal with 0. This implies that in the second term of $J_{k,\varepsilon },$ $v_{k}(T=0,y)=0,$ which contradicts $v_{k0}\neq 0.$
We consider a minimizing sequence $(T_{k,\varepsilon
}^{n},w_{k,\varepsilon }^{n})$ with $T_{k,\varepsilon }^{n}>0,$ $\left\Vert
w_{k,\varepsilon }^{n}(t)\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T_{k,\varepsilon }^{n}},$ satisfying $$d_{\varepsilon }\leq J_{k,\varepsilon }(T_{k,\varepsilon
}^{n},w_{k,\varepsilon }^{n})\leq d_{\varepsilon }+\frac{1}{n},\mbox{ }n\geq
1. \label{14}$$Hence, there exists $T_{k,\varepsilon }^{\ast }>0$ such that $T_{k,\varepsilon }^{n}\rightarrow T_{k,\varepsilon }^{\ast },$ as $n\rightarrow \infty .$ On a subsequence, denoted still by $n,$ we have $w_{k,\varepsilon }^{n}\rightarrow w_{k,\varepsilon }^{\ast },$ weakly in $V_{1}$ and strongly in $C([0,1]),$ so that $w_{k,\varepsilon
}^{n}(0)\rightarrow w_{k,\varepsilon }^{\ast }(0)=0$ and $\left\Vert
w_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}.$ By (\[14\]) $$\int_{0}^{t}(w_{k,\varepsilon }^{n}-w_{k}^{\ast })(\tau )d\tau \rightarrow
\int_{0}^{t}(w_{k,\varepsilon }^{\ast }-w_{k}^{\ast })(\tau )d\tau ,\mbox{
uniformly for all }t\in \lbrack 0,1], \label{14-1}$$according to Arzelà theorem, because the sequence $\left(
\int_{0}^{t}(w_{k,\varepsilon }^{n}-w_{k}^{\ast })(\tau )d\tau \right) _{n}$ is bounded in $L^{2}(0,1)$ and its derivative is bounded in $L^{2}(0,1),$ too.
Then, the state system (\[4’\])-(\[4””\]) corresponding to $T_{k,\varepsilon }^{n}$ and $w_{k,\varepsilon }^{n}$ has, by Theorem 2.2, a unique solution continuous in time on $[0,1].$ This solution $v_{k,\varepsilon }^{n}$ has the properties (\[vk\])-(\[vk-est2\]), with $T=T_{k,\varepsilon }^{n}$ and $w_{k,\varepsilon }^{n}.$ By (\[vk-est2\]) we deduce that $v_{k,\varepsilon }^{n}\rightarrow v_{k,\varepsilon }^{\ast }$ strongly in $C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap
L^{2}(0,1;H^{4}(0,L))$ as $n\rightarrow \infty .$ Passing to the limit in (\[4’\])-(\[4””\]) written for $(T_{k,\varepsilon
}^{n},w_{k,\varepsilon }^{n})$ we get that $v_{k,\varepsilon }^{\ast }$ is the solution to (\[4’\])-(\[4””\]) corresponding to $T_{k,\varepsilon
}^{\ast }$ and $w_{k,\varepsilon }^{\ast }.$ Moreover, $(\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{n}(1)\rightarrow (\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)$ strongly in $H.$ Passing to the limit in (\[14\]), as $n\rightarrow \infty ,$ we get on the basis of the previous convergences and of the weakly lower semicontinuity of the norms, that $J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast
})=d_{\varepsilon },$ that is $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon
}^{\ast })$ is an optimal controller in $(P_{k,\varepsilon }).$ $\square $
We note that in problem $(P_{k,\varepsilon }),$ the optimal state satisfies (\[4’\])-(\[4””\]) with $T=T_{k,\varepsilon }^{\ast }$ and $w_{k,\varepsilon }^{\ast }.$
**Theorem 3.2.** *Let* $(T_{k,\varepsilon }^{\ast
},w_{k,\varepsilon }^{\ast },v_{k,\varepsilon }^{\ast })$ *be optimal in* $(P_{k,\varepsilon })$* and* $(T_{k}^{\ast },w_{k}^{\ast
},v_{k}^{\ast })$* be optimal in* $(\widehat{P_{k}}).$ *Then,* $$T_{k,\varepsilon }^{\ast }\rightarrow T_{k}^{\ast },\mbox{ }w_{k,\varepsilon
}^{\ast }\rightarrow w_{k}^{\ast }\mbox{ \textit{weakly in} }H^{1}(0,1)\mbox{
\textit{and strongly in} }C([0,1]), \label{50}$$$$v_{k,\varepsilon }^{\ast }\rightarrow v_{k}^{\ast }\mbox{ \textit{strongly in} }C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{4}(0,L))\cap L^{2}(0,1;H^{4}(0,L)).
\label{51}$$
**Proof.** Let $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon
}^{\ast },v_{k,\varepsilon }^{\ast })$* *be optimal in $(P_{k,\varepsilon })$ and denote $$h_{k,\varepsilon }^{\ast }(t)=\int_{0}^{t}(w_{k,\varepsilon }^{\ast
}-w_{k}^{\ast })(\tau )d\tau .$$The fact that $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast })$ is optimal in $(P_{k,\varepsilon })$ implies that $$\begin{aligned}
&&J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast
})=T_{k,\varepsilon }^{\ast }+\frac{1}{2\varepsilon }\left\Vert (\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)\right\Vert _{H}^{2}+\frac{1}{2}\int_{0}^{1}\left\vert h_{k,\varepsilon }^{\ast }(t)\right\vert ^{2}dt
\label{52-0} \\
&\leq &J_{k,\varepsilon }(T,w)=T+\frac{1}{2\varepsilon }\left\Vert (\sigma
I+A_{k})^{-1}v(1)\right\Vert _{H}^{2}+\frac{1}{2}\int_{0}^{1}\left\vert
\int_{0}^{t}(w-w_{k}^{\ast })(\tau )d\tau \right\vert ^{2}dt \notag\end{aligned}$$for any $T>0$ and $w\in V_{1},$ $\left\Vert w\right\Vert _{V_{1}}\leq \rho
_{k}\sqrt{T},$ where $v$ is the solution to the state system corresponding to $(T,w).$ Let us set in (\[52-0\]), $T=T_{k}^{\ast }$ and $w=w_{k}^{\ast
},$ the chosen optimal controller in $(\widehat{P_{k}}).$ Thus, the second and the last terms on the right-hand side of (\[52-0\]) vanish and $$J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast
})=T_{k,\varepsilon }^{\ast }+\frac{1}{2\varepsilon }\left\Vert (\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)\right\Vert _{H}^{2}+\frac{1}{2}\int_{0}^{1}\left\vert h_{k,\varepsilon }^{\ast }(t)\right\vert ^{2}dt\leq
T_{k}^{\ast }. \label{52-1}$$Then, $T_{k,\varepsilon }^{\ast }\rightarrow T_{k}^{\ast \ast },$ and on a subsequence denoted still by $\varepsilon ,$ we have $w_{k,\varepsilon
}^{\ast }\rightarrow w_{k}^{\ast \ast }$ weakly in $V_{1},$ strongly in $C([0,1]),$ and $w_{k,\varepsilon }^{\ast }(0)\rightarrow w_{k}^{\ast \ast
}(0)=0.$ Also, $\left\Vert w_{k}^{\ast \ast }\right\Vert _{V_{1}}\leq \rho
_{k}\sqrt{T_{k}^{\ast \ast }}.$
The solution $v_{k,\varepsilon }^{\ast \ast }$ corresponding to $(T_{k}^{\ast \ast },w_{k,\varepsilon }^{\ast \ast })$ exists, it is unique, according to Theorem 2.2 and has the properties (\[vk\])-(\[vk-est2\]) with $T=T_{k,\varepsilon }^{\ast }$ and $w_{k,\varepsilon }^{\ast }$. Therefore, by handling some calculations based on (\[vk-est2\]) we get $$v_{k,\varepsilon }^{\ast }\rightarrow v_{k}^{T_{k}^{\ast \ast },w_{k}^{\ast
\ast }}:=v_{k}^{\ast \ast }\mbox{ strongly in }W^{1,2}(0,1;H^{2}(0,L))\cap
L^{2}(0,1;H^{4}(0,L)) \label{52-2}$$where $v_{k}^{\ast \ast }$ is the solution to (\[4’\])-\[4””\]) corresponding to $(T_{k,\varepsilon }^{\ast \ast },w_{k,\varepsilon }^{\ast
\ast })$. By (\[52-1\]) we have$$\left\Vert (\sigma I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)\right\Vert
_{H}^{2}\leq 2\varepsilon T_{k}^{\ast }, \label{e1}$$so that $$\lim_{\varepsilon \rightarrow 0}\left\Vert (\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)\right\Vert _{H}^{2}=0 \label{e2}$$which implies that $v_{k,\varepsilon }^{\ast }(1,\cdot )\rightarrow 0$ strongly in $H.$ On the other hand, by (\[52-2\]), $v_{k,\varepsilon
}^{\ast }(1)\rightarrow v_{k}^{\ast \ast }(1),$ so that $v_{k}^{\ast \ast
}(1)=0.$ Again by (\[52-1\]), $T_{k,\varepsilon }^{\ast }\leq
J_{k,\varepsilon }(T_{\varepsilon }^{\ast },w_{\varepsilon }^{\ast })\leq
T_{k}^{\ast },$ implying at limit that $T_{k}^{\ast \ast }\leq T_{k}^{\ast
}. $ Since $T_{k}^{\ast \ast }$ and $w_{k}^{\ast \ast }$ satisfy the restrictions required in problem $(\widehat{P_{k}}),$ that is $T_{k}^{\ast
\ast }>0,$ $\left\Vert w_{k}^{\ast \ast }\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T_{k}^{\ast \ast }},$ and $v_{k}^{\ast \ast }(1)=0,$ recalling that $T_{k}^{\ast }$ is the infimum in $(\widehat{P_{k}})$ it follows that $T_{k}^{\ast \ast }=T_{k}^{\ast }.$
Again by (\[52-1\]) we see that $T_{k}^{\ast }+\limsup\limits_{\varepsilon
\rightarrow 0}\frac{1}{2\varepsilon }\left\Vert (\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)\right\Vert _{H}^{2}\leq
T_{k}^{\ast },$ which implies $$\limsup_{\varepsilon \rightarrow 0}\frac{1}{2\varepsilon }\left\Vert (\sigma
I+A_{k})^{-1}v_{k,\varepsilon }^{\ast }(1)\right\Vert _{H}^{2}=0.
\label{52-4}$$Also, $T_{k}^{\ast }+\limsup\limits_{\varepsilon \rightarrow 0}\frac{1}{2}\int_{0}^{1}\left\vert h_{k,\varepsilon }^{\ast }(t)\right\vert ^{2}dt\leq
T_{k}^{\ast },$ implying that $$\limsup_{\varepsilon \rightarrow 0}\int_{0}^{1}\left\vert h_{k,\varepsilon
}^{\ast }(t)\right\vert ^{2}dt=0. \label{52-5}$$Therefore, $$h_{k,\varepsilon }^{\ast }\rightarrow 0\mbox{ strongly in }L^{2}(0,1),\mbox{
as }\varepsilon \rightarrow 0 \label{52-6}$$and so it follows that $$h_{k,\varepsilon }^{\ast }(t)=\int_{0}^{t}(w_{k,\varepsilon }^{\ast
}-w_{k}^{\ast })(\tau )d\tau \rightarrow \int_{0}^{t}(w_{k}^{\ast \ast
}-w_{k}^{\ast })(\tau )d\tau =0,\mbox{ for all }t\in \lbrack 0,1],\mbox{ as }\varepsilon \rightarrow 0.$$Thus, we get $w_{k}^{\ast \ast }=w_{k}^{\ast }$ for all $[0,1]$ and so $v_{k}^{\ast \ast }(t)=v_{k}^{\ast \ast }(t)$ for all $t\in \lbrack 0,1].$
Based on the previous convergences, we pass to the limit in ([52-0]{}) and conclude that $\lim\limits_{\varepsilon \rightarrow
0}J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast \ast },w_{k,\varepsilon
}^{\ast \ast })=T_{k}^{\ast }=J_{k}(T_{k}^{\ast },w_{k}^{\ast })=T_{k}^{\ast
}$. $\square $
Systems in variations and the adjoint system for $(P_{k,\protect\varepsilon })$
-------------------------------------------------------------------------------
Since the minimization problem depends on two controllers, $T_{k,\varepsilon
}^{\ast }$ and $w_{k,\varepsilon }^{\ast },$ we shall study separate variations with respect to them. First, let us keep $T_{k,\varepsilon
}^{\ast }$ fixed and give variations to $w_{k,\varepsilon }^{\ast }.$ We shall obtain a first system in variations.
Let $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast })$ be an optimal controller in $(P_{k,\varepsilon }).$ For $\lambda >0,$ we set$$w_{k,\varepsilon }^{\lambda }=w_{k,\varepsilon }^{\ast }+\lambda \omega ,\mbox{ where }\omega =\widetilde{w}-w_{k,\varepsilon }^{\ast },\mbox{ }\widetilde{w}\in V_{1}. \label{53}$$We note that the state system satisfies (\[4’\])-(\[4””\]) with $T=T_{k,\varepsilon }^{\ast }$ and $w_{k,\varepsilon }^{\ast }.$ We define $Y_{\lambda }=\frac{v_{k,\varepsilon }^{T_{\varepsilon }^{\ast
},w_{\varepsilon }^{\lambda }}-v_{k,\varepsilon }^{\ast }}{\lambda },$ where $v_{k,\varepsilon }^{T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\lambda
}}$ is the solution to (\[4’\])-(\[4””\]) corresponding to $T_{k,\varepsilon }^{\ast }$ and $w_{k,\varepsilon }^{\lambda }.$ Taking into account the estimates of Theorem 2.2, we deduce by a straightforward calculation that $$Y_{\lambda }\rightarrow Y\mbox{ as }\lambda \rightarrow 0,\mbox{ strongly in
}C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap L^{2}(0,1;H^{4}(0,L)),$$and that $Y$ satisfies the equations$$\begin{aligned}
(k^{2}Y-Y^{\prime \prime })_{t}+T_{k,\varepsilon }^{\ast }\left( \nu Y^{\mbox{iv}}-(2\nu k^{2}+ikU)Y^{\prime \prime }+(\nu k^{4}+ik^{3}U+ikU^{\prime
\prime })Y\right) &=&0,\mbox{ } \label{54} \\
Y(t,0)=0,\mbox{ }Y(t,L)=\omega ,\mbox{ }Y^{\prime }(t,0)=Y^{\prime }(t,L)
&=&0, \label{54-1} \\
Y(0,y) &=&0, \label{54-3}\end{aligned}$$for $(t,y)\in (0,1)\times (0,L).$ Moreover, following Theorem 2.2, we state that (\[54\])-(\[54-3\]) has a unique solution $$Y\in C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap
L^{2}(0,1;H^{4}(0,L)). \label{55}$$
We introduce the dual system$$-(k^{2}p_{k,\varepsilon }-p_{k,\varepsilon }^{\prime \prime
})_{t}+T_{k,\varepsilon }^{\ast }\left( \nu p_{k,\varepsilon }^{\mbox{iv}}-(2\nu k^{2}+ikU)p_{k,\varepsilon }^{\prime \prime }-2ikU^{\prime
}p_{k,\varepsilon }^{\prime }+(\nu k^{4}+ik^{3}U)p_{k,\varepsilon }\right)
=0,\mbox{ } \label{56-1}$$$$p_{k,\varepsilon }(t,0)=p_{k,\varepsilon }(t,L)=p_{k,\varepsilon }^{\prime
}(t,0)=p_{k,\varepsilon }^{\prime }(t,L)=0, \label{56-2}$$$$k^{2}p_{k,\varepsilon }(1,y)-p_{k,\varepsilon }^{\prime \prime
}(1,y)=(\sigma I+A_{k})^{-2}\left( \frac{1}{\varepsilon }\overline{v_{k,\varepsilon }^{\ast }}(1)\right) , \label{56-3}$$for $(t,y)\in (0,1)\times (0,L).$
Now, in order to simplify the writing in the next calculations, we use a formal notation $$Ev:=k^{2}v-v^{\prime \prime },\mbox{ }Fz:=\nu v^{\mbox{iv}}-(2\nu
k^{2}+ikU)v^{\prime \prime }+(\nu k^{4}+ik^{3}U+ikU^{\prime \prime })v
\label{L-V}$$for $v\in H^{4}(0,L)$ and rewrite the state equation (\[4’\])-(\[4””\]) for the solution $v_{k,\varepsilon }^{\ast }$ corresponding to $w_{k,\varepsilon }^{\ast }$ as $$E(v_{k,\varepsilon }^{\ast })_{t}(t)+T_{k,\varepsilon }^{\ast
}Fv_{k,\varepsilon }^{\ast }(t)=0,\mbox{ a.e. }t\in (0,1), \label{75}$$$$v_{k,\varepsilon }^{\ast }(0)=v_{k0}. \label{76}$$
**Proposition 3.3.** *The adjoint system* (\[56-1\])-(\[56-3\]) *has a unique solution*** **$$p_{k,\varepsilon }\in C([0,1];H^{6}(0,L)\cap H_{0}^{2}(0,L))\cap
C^{1}([0,1];H^{4}(0,L)\cap H_{0}^{2}(0,L)). \label{57}$$
**Proof.** We recall** **the definition (\[10-3\]) and introduce, similarly to (\[10-4\]), the following operator:$$\begin{aligned}
F_{0k}^{\ast } &:&D(F_{0k}^{\ast })=D(F_{0k})\subset H\rightarrow H,
\label{57-2} \\
F_{0k}^{\ast }z &=&\nu z^{\mbox{iv}}-(2\nu k^{2}+ikU)z^{\prime \prime
}-2ikU^{\prime }z^{\prime }+(\nu k^{4}+ik^{3}U)z. \notag\end{aligned}$$We can interpret $F_{0k}^{\ast }$ as the dual of $F$ in the sense of distributions, that is $(Fv,\overline{\varphi })=(\overline{v},F_{0k}^{\ast
}\varphi ),$ for $\varphi \in C_{0}^{\infty }(0,L)$ and $v\in H^{4}(0,L)$. System (\[56-1\])-(\[56-3\])** **can be written$$-E_{0k}(p_{k,\varepsilon })_{t}(t)+T_{k,\varepsilon }^{\ast }F_{0k}^{\ast
}p_{k,\varepsilon }(t)=0,\mbox{ a.e. }t\in (0,1), \label{57-0}$$$$E_{0k}p_{k,\varepsilon }(1)=(\sigma I+A_{k})^{-2}\left( \frac{1}{\varepsilon
}\overline{v_{k,\varepsilon }^{\ast }}(1)\right) . \label{57-1}$$Also, we introduce $$B_{k}=F_{0k}^{\ast }E_{0k}^{-1},\mbox{ }B_{k}:D(B_{k})\subset H\rightarrow H,\mbox{ }D(B_{k})=\{v\in H^{2}(0,L);\mbox{ }E_{0k}^{-1}v\in
H_{0}^{2}(0,L)\}=D(A_{k}).$$By the same argument as for $A_{k}$ we infer that $-B_{k}$ generates a $C_{0} $-analytic semigroup.
We write the equivalent equation for $\widetilde{p_{k,\varepsilon }}(t):=E_{0k}p_{k,\varepsilon }(t),$ $$-\frac{d\widetilde{p_{k,\varepsilon }}}{dt}(t)+T_{k,\varepsilon }^{\ast
}B_{k}\widetilde{p_{k,\varepsilon }}(t)=0,\mbox{ a.e. }t\in (0,1),$$$$\widetilde{p_{k,\varepsilon }}(1)=(\sigma I+A_{k})^{-2}\left( \frac{1}{\varepsilon }\overline{v_{k,\varepsilon }^{\ast }}(1)\right) .$$The solution is $$\widetilde{p_{k,\varepsilon }}(t)=e^{T_{k,\varepsilon }^{\ast }B_{k}(1-t)}\widetilde{p_{k,\varepsilon }}(1),\mbox{ for all }t\in \lbrack 0,1]$$and since $\overline{v_{k,\varepsilon }^{\ast }}(1)\in H,$ by (\[vk\]), we have $\widetilde{p_{k,\varepsilon }}(1)\in D(A_{k}^{2})=D(B_{k}^{2}).$ Therefore, $$\widetilde{p_{k,\varepsilon }}\in C([0,1];D(B_{k}^{2}))\cap
C^{1}([0,1];D(B_{k})) \label{57-2-0}$$and $$p_{k,\varepsilon }(t)=E_{0k}^{-1}(\widetilde{p_{k,\varepsilon }}(t)),\mbox{
for all }t\in \lbrack 0,1] \label{57-3}$$turns out to be in the spaces (\[57\]), by the observation (\[iff\]) made before Theorem 2.2. The proof is ended. $\square $
Now, let us keep $w_{k,\varepsilon }^{\ast }$ fixed and give variations to $T_{k,\varepsilon }^{\ast }.$ For $\lambda >0,$ we define $Z^{\lambda }=\frac{v_{k,\varepsilon }^{T_{k,\varepsilon }^{\ast }+\lambda
,w_{k,\varepsilon }^{\ast }}-v_{k,\varepsilon }^{T_{k,\varepsilon }^{\ast
},w_{k,\varepsilon }^{\ast }}}{\lambda }$ and by a straightforward calculation we have $Z^{\lambda }\rightarrow Z$ as $\lambda \rightarrow 0,$ where $Z$ satisfies the system in variations$$\begin{aligned}
EZ_{t}(t)+T_{k,\varepsilon }^{\ast }FZ(t) &=&-T_{k,\varepsilon }^{\ast
}Fv_{k,\varepsilon }^{\ast }(t),\mbox{ a.e. }t\in (0,1), \label{Z} \\
Z(t,0) &=&Z(t,L)=Z^{\prime }(t,0)=Z^{\prime }(t,L)=0, \notag \\
Z(0,y) &=&0. \notag\end{aligned}$$Since the right hand side in the first equation in (\[Z\]) is in $L^{2}(0,T_{k,\varepsilon }^{\ast };H)$ it follows that (\[Z\]) has the unique solution$$Z\in C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap
L^{2}(0,1;H^{4}(0,L)). \label{Z1}$$
Necessary conditions of optimality for $(P_{k,\protect\varepsilon })$
---------------------------------------------------------------------
We recall that $V_{1}=\{f\in H^{1}(0,1);$ $f(0)=0\}.$ Let us introduce the set $$K_{T}=\{w\in V_{1};\mbox{ }\left\Vert w\right\Vert _{V_{1}}\leq \rho _{k}\sqrt{T}\} \label{K}$$and denote the normal cone to $K_{T}$ at $w$ by$$N_{K_{T}}(w)=\left\{ \chi \in V_{1}^{\ast };\mbox{ }{\rm Re}\left\langle
\chi (t),(w-w_{1})(t)\right\rangle _{V_{1}^{\ast },V_{1}}\geq 0,\mbox{ for
all }w_{1}\in K_{T}\right\} ,\mbox{ } \label{59}$$where $V_{1}^{\ast }$ is the dual of $V_{1}.$
We make a parenthesis for a discussion about this cone. It is known that $$N_{K_{T}}(w)=\left\{
\begin{array}{l}
\underset{\alpha >0}{\cup }\alpha \Lambda (w),\mbox{ if }\left\Vert
w\right\Vert _{V_{1}}=\rho _{k}\sqrt{T} \\
\{0\},\mbox{ \ \ \ \ \ \ \ \ if }\left\Vert w\right\Vert _{V_{1}}<\rho _{k}\sqrt{T} \\
\varnothing ,\mbox{ \ \ \ \ \ \ \ \ \ \ if }\left\Vert w\right\Vert
_{V_{1}}>\rho _{k}\sqrt{T},\end{array}\right.$$where $\alpha \in \mathbb{R}_{+},$ and $\Lambda :V_{1}\rightarrow
V_{1}^{\ast }$ is the canonical isomorphism from $V_{1}$ to $V_{1}^{\ast }$. Here, $\Lambda w=-\ddot{w}$ (see e.g., [@vb-springer-2010], pp. 2-4).
By abuse of notation, we denote still by $N_{K_{T}}(w)$ the restriction of $N_{K_{T}}(w)$ on $L^{2}(0,1).$ In this case $$\Lambda w=-\ddot{w},\mbox{ }\Lambda :D(\Lambda ):=\{w\in H^{2}(0,1);\mbox{ }w(0)=\dot{w}(1)=0\}\subset L^{2}(0,1)\rightarrow L^{2}(0,1).$$
Let $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast })$ be an optimal controller in $(P_{k,\varepsilon }),$ with $w_{k,\varepsilon }^{\ast }\in
K_{T_{k,\varepsilon }^{\ast }}.$ For $\lambda >0,$ we set $$w_{k,\varepsilon }^{\lambda }=w_{k,\varepsilon }^{\ast }+\lambda \omega ,\mbox{ where }\omega =\widetilde{w}-w_{k,\varepsilon }^{\ast },\mbox{ }\widetilde{w}\in K_{T_{k,\varepsilon }^{\ast }},$$that is, $\widetilde{w}\in V_{1},$ $\left\Vert \widetilde{w}\right\Vert
_{V_{1}}\leq \rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}.$ We recall the notation $$h_{k,\varepsilon }^{\ast }(t)=\int_{0}^{t}(w_{k,\varepsilon }^{\ast
}-w_{k}^{\ast })(\tau )d\tau . \label{51-0}$$
**Proposition 3.4.** *Let* $(T_{k,\varepsilon }^{\ast
},w_{k,\varepsilon }^{\ast })$* be an optimal control in* $(P_{k,\varepsilon })$ *with the optimal state* $v_{k,\varepsilon
}^{\ast }$*. Then,* $$\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon }^{\ast
}(t)=T_{k,\varepsilon }^{\ast }\nu \overline{p_{k,\varepsilon }^{\prime
\prime \prime }}(t,L)+\int_{t}^{1}h_{k,\varepsilon }^{\ast }(\tau )d\tau ,\mbox{ \textit{for all} }t\in \lbrack 0,1],\mbox{ } \label{eta}$$*where* $w_{k,\varepsilon }^{\ast }(0)=\dot{w}_{k,\varepsilon }(1)=0,$ $\ddot{w}_{k,\varepsilon }^{\ast }\in H^{2}(0,1),$ $\left\Vert
w_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }},$ $\alpha _{k,\varepsilon }\in \mathbb{R}_{+}$ *and* $$\begin{aligned}
&&\alpha _{k,\varepsilon }\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\left\Vert \ddot{w}_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}^{\ast
}}+T_{k,\varepsilon }^{\ast }{\rm Re}\int_{0}^{1}(\overline{v_{k,\varepsilon }^{\ast }}(t),F_{0k}^{\ast }p_{k,\varepsilon }(t))_{H}dt
\label{62} \\
&=&1-{\rm Re}\int_{0}^{1}\left( \int_{t}^{1}h_{\varepsilon }^{\ast }(\tau
)d\tau \right) \overline{w_{k,\varepsilon }^{\ast }}(t)dt,\mbox{ } \notag\end{aligned}$$*with* $p_{k,\varepsilon }$ *the solution to the dual backward equation* (\[56-1\])-(\[56-3\]).
**Proof.** The proof will be done in two steps.
**Step 1.** **The** **first condition of optimality**. Let $(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast })$ be an optimal controller in $(P_{k,\varepsilon }).$ As already said, we keep $T_{k,\varepsilon }^{\ast }$ fixed and give variations to $w_{k,\varepsilon
}^{\ast }$. Using the fact that $w_{k,\varepsilon }^{\ast }$ is optimal we can write $$J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast
})\leq J_{k,\varepsilon }(T,w),\mbox{ for all }w\in K_{T}.$$This holds true if we replace $T$ by $T_{k,\varepsilon }^{\ast }$ and $w$ by $w_{k,\varepsilon }^{\lambda }.$ By calculating $$\begin{aligned}
&&\lim_{\lambda \rightarrow 0}\frac{J_{k,\varepsilon }(T_{k,\varepsilon
}^{\ast },w_{k,\varepsilon }^{\lambda })-J_{k,\varepsilon }(T_{k,\varepsilon
}^{\ast },w_{k,\varepsilon }^{\ast })}{\lambda } \\
&=&\lim_{\lambda \rightarrow 0}\frac{1}{\lambda }\left( \frac{1}{2\varepsilon }\left( \left\Vert (\sigma I+A_{k})^{-1}v_{k,\varepsilon
}^{T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\lambda }}(1)\right\Vert
_{H}^{2}-\left\Vert (\sigma I+A_{k})^{-1}v_{k,\varepsilon
}^{T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast }}(1)\right\Vert
_{H}^{2}\right) \right. \\
&&\left. +\frac{1}{2}\left( \int_{0}^{1}\left\vert
\int_{0}^{t}(w_{k,\varepsilon }^{\lambda }-w_{k}^{\ast })(\tau )d\tau
\right\vert ^{2}dt-\int_{0}^{1}\left\vert \int_{0}^{t}(w_{k,\varepsilon
}^{\ast }-w_{k}^{\ast })(\tau )d\tau \right\vert ^{2}dt\right) \right)\end{aligned}$$we obtain $$\begin{aligned}
&&\lim_{\lambda \rightarrow 0}\frac{J_{k,\varepsilon }(T_{k,\varepsilon
}^{\ast },w_{k,\varepsilon }^{\lambda })-J_{k,\varepsilon }(T_{k,\varepsilon
}^{\ast },w_{k,\varepsilon }^{\ast })}{\lambda } \label{65-0} \\
&=&{\rm Re}\left\{ \int_{0}^{L}\frac{1}{\varepsilon }(\sigma I+A_{k})^{-2}\overline{v_{k,\varepsilon }^{\ast }}(1)\cdot
Y(1,y)dy+\int_{0}^{1}h_{k,\varepsilon }^{\ast }(t)\left( \int_{0}^{t}\overline{\omega }(s)ds\right) dt\right\} \geq 0. \notag\end{aligned}$$Here we used the definition of the scalar product in $\mathbb{C}$ and that ${\rm Re}(a\cdot \overline{b})={\rm Re}(\overline{a}\cdot b),$ for $a,b\in
\mathbb{C}$. Then, we calculate the last term in (\[65-0\])$${\rm Re}\int_{0}^{1}h_{k,\varepsilon }^{\ast }(t)\left( \int_{0}^{t}\overline{\omega }(s)ds\right) dt={\rm Re}\int_{0}^{1}\int_{s}^{1}h_{k,\varepsilon }^{\ast }(t)\overline{\omega }(s)dtds \label{g51}$$and replacing it in (\[65-0\]) we obtain$${\rm Re}\left\{ \int_{0}^{L}(\sigma I+A_{k})^{-2}\left( \frac{1}{\varepsilon }\overline{v_{k,\varepsilon }^{\ast }}(1)\right)
Y(1,y)dy+\int_{0}^{1}\left( \int_{t}^{1}h_{k,\varepsilon }^{\ast }(\tau
)d\tau \right) \overline{\omega }(t)dt\right\} \geq 0. \label{65}$$Using the notation (\[L-V\]) we can write eq. (\[54\]) in the equivalent form$$EY_{t}(t)+T_{k,\varepsilon }^{\ast }FY(t)=0 \label{65-1}$$with the boundary and initial conditions. Now, we multiply this equation scalarly in $H$ by $p_{k,\varepsilon }(t)$ and integrate with respect to $t$ over $(0,1).$ While performing all the integrals by parts, we put first into evidence the relation $$\int_{0}^{L}FY(t,y)\cdot p_{k,\varepsilon }(t,y)dy=\int_{0}^{L}F_{0k}^{\ast
}p_{k,\varepsilon }(t,y)\cdot Y(t,y)dy-\nu p_{k,\varepsilon }^{\prime \prime
\prime }(t,L)\omega (t), \label{F-F0}$$and obtain$$\begin{aligned}
&&\int_{0}^{1}\int_{0}^{L}\left\{ -(k^{2}p_{k,\varepsilon }-p_{k,\varepsilon
}^{\prime \prime })_{t}\right\} Ydydt \label{66} \\
&&+T_{k,\varepsilon }^{\ast }\int_{0}^{1}\int_{0}^{L}\{\nu p_{k,\varepsilon
}^{\mbox{iv}}-(2\nu k^{2}+ikU)p_{k,\varepsilon }^{\prime \prime
}-2ikU^{\prime }p_{k,\varepsilon }^{\prime }+(\nu
k^{4}+ik^{3}U)p_{k,\varepsilon }\}Ydydt \notag \\
&&+\int_{0}^{L}\left\{ k^{2}p_{k,\varepsilon }(1,y)-p_{k,\varepsilon
}^{\prime \prime }(1,y)\right\} Y(1,y)dy-T_{k,\varepsilon }^{\ast
}\int_{0}^{1}\omega (t)\nu p_{k,\varepsilon }^{\prime \prime \prime
}(t,L)dy=0. \notag\end{aligned}$$By (\[57\]), we see that $p_{k,\varepsilon }^{\prime \prime \prime }\in
C([0,1];H^{1}(0,L))$ and so the trace $p_{k,\varepsilon }^{\prime \prime
\prime }(t,L)$ is well defined and belongs to $C([0,1];\mathbb{C}).$ Recalling now the equations in the adjoint system (\[56-1\])-(\[56-3\]) and comparing with the integrands in (\[66\]) we get$$\int_{0}^{L}(\sigma I+A_{k})^{-2}\left( \frac{1}{\varepsilon }\overline{v_{k,\varepsilon }^{\ast }}(1)\right) \cdot Y(1,y)dy=T_{k,\varepsilon
}^{\ast }\int_{0}^{1}\omega (t)\cdot \nu p_{k,\varepsilon }^{\prime \prime
\prime }(t,L)dt. \label{67}$$Then plugging the latter into (\[65\]) and using again ${\rm Re}(a\cdot
b)={\rm Re}(\overline{a}\cdot \overline{b})$ we deduce the relation $$\begin{aligned}
&&{\rm Re}\left\{ \int_{0}^{1}T_{k,\varepsilon }^{\ast }\nu \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(t,L)\overline{\omega }(t)dt+\int_{0}^{1}\left( \int_{t}^{1}h_{k,\varepsilon }^{\ast }(\tau )d\tau
\right) \overline{\omega }(t)dt\right\} \label{68} \\
&=&{\rm Re}\int_{0}^{1}\left( -T_{k,\varepsilon }^{\ast }\nu \overline{p_{\varepsilon }^{\prime \prime \prime }}(t,L)-\int_{t}^{1}h_{\varepsilon
}^{\ast }(\tau )d\tau \right) (w_{k,\varepsilon }^{\ast }-\widetilde{w})(t)dt\geq 0, \notag\end{aligned}$$for all $\widetilde{w}\in K_{T_{k,\varepsilon }^{\ast }}$. The immediate result is that $$\eta _{k,\varepsilon }:=-T_{k,\varepsilon }^{\ast }\nu \overline{p_{\varepsilon }^{\prime \prime \prime }}(\cdot ,L)-\int_{\cdot
}^{1}h_{\varepsilon }^{\ast }(\tau )d\tau \in N_{K_{T_{k,\varepsilon }^{\ast
}}}(w_{k,\varepsilon }^{\ast }) \label{68-0}$$where $N_{K_{T_{k,\varepsilon }^{\ast }}}(w_{k,\varepsilon }^{\ast })$ is viewed as the cone from $V_{1}$ to $V_{1}^{\ast }.$ However, since $\eta
_{k,\varepsilon }\in L^{2}(0,1),$ we can consider it in the restriction of $N_{K_{T_{k,\varepsilon }^{\ast }}}(w_{k,\varepsilon }^{\ast })$ to $L^{2}(0,1),$ still denoted by $N_{K_{T_{k,\varepsilon }^{\ast
}}}(w_{k,\varepsilon }^{\ast }).$ Because an element of this cone is of the form $\eta _{k,\varepsilon }=-\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon }$ with $w_{k,\varepsilon }^{\ast }(0)=\dot{w}_{k,\varepsilon }^{\ast }(1)=0$ and $\alpha _{k,\varepsilon }>0$, we have $$-\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon }(t)=-T_{k,\varepsilon
}^{\ast }\nu \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(t,L)-\int_{t}^{1}h_{k,\varepsilon }^{\ast }(\tau )d\tau ,\mbox{ for all }t\in \lbrack 0,1], \label{70-0}$$which is just (\[eta\]). We note that $\ddot{w}_{k,\varepsilon }$ turns out to be in $H^{2}(0,1).$
The situation in which $\left\Vert w_{k,\varepsilon }^{\ast }\right\Vert
_{V_{1}}<\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}$ provides $\ddot{w}_{k,\varepsilon }(t)=0,$ which gives the solution $w_{k,\varepsilon }^{\ast
}(t)=0$ which is not relevant for our problem. In particular, it means that $w_{k}^{\ast }=0,$ that is the flow would be not controlled.
Thus, (\[eta\]) follows for the case when $\left\Vert w_{k,\varepsilon
}^{\ast }\right\Vert _{V_{1}}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}.$
**Step 2. The second condition of optimality.** Here** **we keep $w_{k,\varepsilon }^{\ast }$ fixed and give variations for $T_{k,\varepsilon
}^{\ast }.$ Since $T_{k,\varepsilon }^{\ast }$ realizes the minimum in $(P_{k,\varepsilon })$ we write $$J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast
})\leq J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast }+\lambda
,w_{k,\varepsilon }^{\ast }),\mbox{ }\lambda >0,$$that is, $$\begin{aligned}
&&J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast },w_{k,\varepsilon }^{\ast
})=T_{k,\varepsilon }^{\ast }+\frac{1}{2\varepsilon }\left\Vert (\sigma
I+A_{k})^{-2}v_{k,\varepsilon }^{T_{k,\varepsilon }^{\ast },w_{k,\varepsilon
}^{\ast }}(1)\right\Vert _{H}^{2}+\frac{1}{2}\int_{0}^{1}\left\vert
h_{k,\varepsilon }^{\ast }(t)\right\vert ^{2}dt \\
&\leq &J_{k,\varepsilon }(T_{k,\varepsilon }^{\ast }+\lambda
,w_{k,\varepsilon }^{\ast })=T_{k,\varepsilon }^{\ast }+\lambda +\frac{1}{2\varepsilon }\left\Vert (\sigma I+A_{k})^{-2}v_{k,\varepsilon
}^{T_{k,\varepsilon }^{\ast }+\lambda ,w_{k,\varepsilon }^{\ast
}}(1)\right\Vert _{H}^{2}+\frac{1}{2}\int_{0}^{1}\left\vert h_{k,\varepsilon
}^{\ast }(t)\right\vert ^{2}dt.\end{aligned}$$By performing the computations as before we obtain $$1+{\rm Re}\int_{0}^{1}\frac{1}{\varepsilon }(\sigma
I+A_{k})^{-2}v_{k,\varepsilon }^{\ast }(1)Z(1,y)dy=0, \label{cond2}$$where $Z$ is the solution to the system in variations (\[Z\]). We consider the same adjoint system (\[57-0\])-(\[57-1\]). By multiplying the first equation in (\[Z\]) by $p_{k,\varepsilon }(t)$ scalarly in $H,$ integrating along $t\in (0,1)$ and taking the real part, we obtain $${\rm Re}\int_{0}^{L}E_{0k}p_{k,\varepsilon }(1)Z(1,y)dy=-{\rm Re}\int_{0}^{1}\int_{0}^{L}T_{k,\varepsilon }^{\ast }Fv_{k,\varepsilon }^{\ast
}p_{k,\varepsilon }dydt. \label{cond2-1}$$By using the final condition in the adjoint system and comparing (\[cond2\]) and (\[cond2-1\]) we get$${\rm Re}\int_{0}^{1}\int_{0}^{L}T_{k,\varepsilon }^{\ast }Fv_{k,\varepsilon
}^{\ast }p_{k,\varepsilon }dydt=1. \label{cond2-3}$$Finally, applying a relation similar to (\[F-F0\]) in the left-hand side of (\[cond2-3\]) we obtain$${\rm Re}\int_{0}^{1}\int_{0}^{L}T_{k,\varepsilon }^{\ast }F_{0k}^{\ast
}p_{k,\varepsilon }\cdot v_{k,\varepsilon }^{\ast }dydt-{\rm Re}\int_{0}^{1}T_{k,\varepsilon }^{\ast }\nu p_{k,\varepsilon }^{\prime \prime
\prime }(t,L)w_{k,\varepsilon }^{\ast }dt=1$$and recalling again that ${\rm Re}(a\cdot b)={\rm Re}(\overline{a}\cdot
\overline{b})$ we get$$-{\rm Re}\int_{0}^{1}T_{k,\varepsilon }^{\ast }\nu \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(t,L)\overline{w_{k,\varepsilon
}^{\ast }}dt+{\rm Re}\int_{0}^{1}T_{k,\varepsilon }^{\ast }(\overline{v_{k,\varepsilon }^{\ast }}(t),F_{0k}^{\ast }p_{k,\varepsilon }(t))_{H}dt=1.
\label{cond2-2}$$Using (\[eta\]) we still can write $$\begin{aligned}
&&{\rm Re}\int_{0}^{1}\left( -T_{k,\varepsilon }^{\ast }\nu
p_{k,\varepsilon }^{\prime \prime \prime }(t,L)-\int_{t}^{1}h_{\varepsilon
}^{\ast }(\tau )d\tau \right) \overline{w_{k,\varepsilon }^{\ast }}(t)dt+{\rm Re}\int_{0}^{1}T_{k,\varepsilon }^{\ast }(\overline{v_{k,\varepsilon
}^{\ast }}(t),F_{0k}^{\ast }p_{k,\varepsilon }(t))_{H}dt \notag \\
&=&1-{\rm Re}\int_{0}^{1}\left( \int_{t}^{1}h_{\varepsilon }^{\ast }(\tau
)d\tau \right) \overline{w_{k,\varepsilon }^{\ast }}dt. \label{cond3}\end{aligned}$$But $\eta _{k,\varepsilon }(t)\in N_{K_{T_{k,\varepsilon }^{\ast
}}}(w_{k,\varepsilon }^{\ast })$ and so we have $${\rm Re}\int_{0}^{1}\eta _{k,\varepsilon }(t)\overline{w_{k,\varepsilon
}^{\ast }}(t)dt={\rm Re}\left\langle \eta _{k,\varepsilon
}(t),w_{k,\varepsilon }^{\ast }(t)\right\rangle _{V_{1}^{\ast },V_{1}}=\rho
_{k}\sqrt{T_{k,\varepsilon }^{\ast }}\left\Vert \eta _{k,\varepsilon
}\right\Vert _{V_{1}^{\ast }}. \label{cond4}$$Let $\left\Vert w_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}.$ Taking into account that in this case there is $\alpha _{k,\varepsilon }>0$ such that $\eta _{k,\varepsilon
}(t)=-\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon }^{\ast },$ we get by (\[cond3\]) and (\[cond4\]), relation (\[62\]), as claimed.
Moreover, on the one hand, $\left\langle -\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon }^{\ast },\overline{w_{k,\varepsilon }^{\ast }}(t)\right\rangle _{V_{1}^{\ast },V_{1}}=\alpha _{k,\varepsilon
}\int_{0}^{1}\left\vert \dot{w}_{k,\varepsilon }^{\ast }(t)\right\vert
^{2}dt=\alpha _{k,\varepsilon }\rho _{k}^{2}T_{k,\varepsilon }^{\ast }.$ On the other hand by (\[cond4\]) we get $${\rm Re}\left\langle -\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon
}^{\ast },\overline{w_{k,\varepsilon }^{\ast }}(t)\right\rangle
_{V_{1}^{\ast },V_{1}}=\alpha _{k,\varepsilon }\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\left\Vert \ddot{w}_{k,\varepsilon }^{\ast
}\right\Vert _{V_{1}^{\ast }}, \label{cond5}$$hence $\left\Vert \ddot{w}_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}^{\ast
}}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}$ is verified.
The situation $\left\Vert \ddot{w}_{k,\varepsilon }^{\ast }\right\Vert
_{V_{1}^{\ast }}<\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}$ was excluded, so that (\[eta\]) and (\[62\]) follow for the case when $\left\Vert
\ddot{w}_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}^{\ast }}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}.$ This ends the proof. $\square $
The maximum principle for $(P_{k})$
===================================
We recall that by $(\widehat{H_{k}}),$ which is $(H_{k})$ translated in the new variable (\[t-cap\]), the system (\[4’\])-(\[4””\]) is controllable. Namely, for each initial datum $v^{t_{0}}\in L^{2}(0,L),$ $\left\Vert v^{t_{0}}\right\Vert _{(H^{2}(0,L))^{\ast }}\leq 1,$ there exists $w\in H^{1}(0,1),$ and $\gamma _{(\cdot ,1)}\in L^{2}(0,1),$ satisfying $w(\tau )=0$ for $0\leq \tau \leq t_{0},$ $\left( \int_{0}^{1}\left\vert \dot{w}(\tau )\right\vert ^{2}d\tau \right) ^{1/2}\leq \gamma _{(t_{0},1)},$ such that $v^{t_{0},w}(1,y)=0$ a.e. $y\in (0,L).$ Here, $v^{t_{0},w}$ is the solution to (\[4’\])-(\[4””\]) starting at time $t_{0}$. (
Here, all functions are in fact those denoted by $"$ $\widehat{}$ $"$ depending on $\widehat{t},$ but this decoration is skipped in this section).
We denote by $C_{\nu ,U,k}:=Ck^{4}\left( \nu +\left\Vert U\right\Vert
_{W^{1,\infty }(0,L)}\right) ,$ where $C$ is a positive number$.$
**Theorem 4.1.** *Let* $(T_{k}^{\ast },w_{k}^{\ast
},v_{k}^{\ast })$* be optimal in* $(\widehat{P_{k}})$*. If* $$\begin{aligned}
&&C_{\nu ,U,k}(1+k^{2})\left\Vert \gamma \right\Vert _{L^{1}(0,1)}\sqrt{T_{k}^{\ast }}\left( T_{k}^{\ast }+\sqrt{T_{k}^{\ast }}\right) <1,\mbox{ }
\label{cond-rok} \\
&&\rho _{k}\left( 1-C_{\nu ,U,k}(1+k^{2})\left\Vert \gamma \right\Vert
_{L^{1}(0,1)}\sqrt{T_{k}^{\ast }}(T_{k}^{\ast }+\sqrt{T_{k}^{\ast }})\right)
\notag \\
&>&C_{\nu ,U,k}(1+k^{2})\sqrt{T_{k,\varepsilon }^{\ast }}\left\Vert \gamma
\right\Vert _{L^{1}(0,1)}\left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap
H_{0}^{2}(0,L)}, \notag\end{aligned}$$*then, there exists* $\alpha _{k}^{\ast }>0$ *such that* $$\alpha _{k}^{\ast }\ddot{w}_{k}^{\ast }(t)=T_{k}^{\ast }\nu \overline{p_{k}^{\prime \prime \prime }}(t,L),\mbox{ \textit{a.e.} }t\in (0,1),\mbox{ }w_{k}^{\ast }(0)=\dot{w}_{k}^{\ast }(1)=0,\mbox{ }\left\Vert w_{k}^{\ast
}\right\Vert _{V_{1}}=\rho _{k}\sqrt{T_{k}^{\ast }}, \label{101}$$* * $$\alpha _{k}^{\ast }\rho _{k}^{2}T_{k}^{\ast }+T_{k}^{\ast }{\rm Re}\int_{0}^{1}(\overline{v_{k}^{\ast }}(t),F_{0k}^{\ast }p_{k}(t))_{H}dt=1,
\label{102}$$*where* $p_{k}$ *is the solution to the adjoint equation*$$-E_{0k}(p_{k})_{t}(t)+T_{k}^{\ast }F_{0k}^{\ast }p_{k}(t)=0,\mbox{ \textit{a.e.} }t\in (0,1). \label{103}$$
**Proof.** Using the notation (\[L-V\]) let us consider the generic system $$\begin{aligned}
Ev_{t}(t)+T_{k,\varepsilon }^{\ast }Fv(t) &=&0, \notag \\
v(t,0) &=&0,\mbox{ }v(t,L)=w(t),\mbox{ }v^{\prime }(t,0)=v^{\prime }(t,L)=0,
\label{91} \\
v(0) &=&v_{0k}\in L^{2}(0,L), \notag\end{aligned}$$which has a unique solution, according to Theorem 2.2 and which is controllable, according to $(\widehat{H_{k}}).$ Then, we repeat a similar calculus providing (\[66\]). Thus, we multiply the first equation in ([91]{}) scalarly by $p_{k,\varepsilon }(t),$ integrate along $(t,1)$ and get$$\begin{aligned}
&&\int_{0}^{L}v(1,y)E_{0k}p_{k,\varepsilon
}(1)dy-\int_{0}^{L}v(t,y)E_{0k}p_{k,\varepsilon }(t)dy \label{91-0} \\
&&-\int_{t}^{1}T_{k,\varepsilon }^{\ast }\nu p_{k,\varepsilon }^{\prime
\prime \prime }(\tau ,L)w(\tau )d\tau =0,\mbox{ for }t\in (0,1). \notag\end{aligned}$$Now, we apply $(\widehat{H_{k}})$ for a particular choice, for $t_{0}=t,$ $v^{t_{0}}=v^{t}:=v(t),$ $\left\Vert v(t)\right\Vert _{(H^{2}(0,L))^{\ast
}}\leq 1,$ $w\in V_{1},$ $w(\tau )=0$ for $0\leq \tau \leq t,$ $\gamma
_{(t,1)}:=\gamma (t)$ with $\gamma \in L^{2}(0,1)$ and $\left\Vert
w\right\Vert _{V_{1}}\leq \gamma (t).$ Then, the solution $v^{t,w}$ to ([91]{}), corresponding to the initial datum $v^{t,w}(t,y)=v(t,y),$ satisfies $v^{t,w}(1,y)=0.$ By taking the real part in (\[91-0\]) we get$${\rm Re}(E_{0k}p_{k,\varepsilon }(t),\overline{v}(t))_{H}=-\nu
T_{k,\varepsilon }^{\ast }{\rm Re}\int_{t}^{1}\overline{p_{k,\varepsilon
}^{\prime \prime \prime }}(\tau ,L)\overline{w}(\tau )d\tau . \label{91-00}$$
We choose now $\overline{v}(t)=\frac{E_{0k}p_{k,\varepsilon }(t)}{\left\Vert
E_{0k}p_{k,\varepsilon }(t)\right\Vert _{H}}$ which ensures that $\left\Vert
\overline{v}(t)\right\Vert _{H}=1,$ so that $\left\Vert \overline{v}(t)\right\Vert _{(H^{2}(0,L))^{\ast }}\leq 1$ and compute $$\left\Vert E_{0k}p_{k,\varepsilon }(t)\right\Vert _{H}^{2}=k^{4}\left\Vert
p_{k,\varepsilon }^{\prime \prime }(t)\right\Vert _{H}^{2}+2k^{2}\left\Vert
p_{k,\varepsilon }^{\prime }(t)\right\Vert _{H}^{2}+\left\Vert
p_{k,\varepsilon }(t)\right\Vert _{H}^{2}\geq 2\left\Vert p_{k,\varepsilon
}(t)\right\Vert _{H^{2}(0,L)}^{2},$$for $k\geq 1$.
Since $\overline{w}(\tau )=0$ for $\tau \in \lbrack 0,t],$ $p_{k,\varepsilon
}^{\prime \prime \prime }(\cdot ,L)\in L^{2}(0,1)$ and $V_{1}\subset
H^{1}(0,1)\subset L^{2}(0,1)\subset (H^{1}(0,1))^{\ast }\subset V_{1}^{\ast
},$ we have by (\[91-00\])$$\begin{aligned}
\left\Vert p_{k,\varepsilon }(t)\right\Vert _{H^{2}(0,L)} &\leq &\frac{1}{\sqrt{2}}\left\Vert E_{0k}p_{k,\varepsilon }(t)\right\Vert _{H}\leq \frac{1}{\sqrt{2}}\nu T_{k,\varepsilon }^{\ast }\left\vert \int_{0}^{1}\overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\tau ,L)\overline{w}(\tau )d\tau
\right\vert \label{201} \\
&\leq &\frac{1}{\sqrt{2}}\nu T_{k,\varepsilon }^{\ast }\left\vert
\left\langle \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\tau ,L),\overline{w}(\tau )\right\rangle _{V_{1}^{\ast },V_{1}}\right\vert \leq
\frac{1}{\sqrt{2}}\nu T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot ,L)\right\Vert
_{V_{1}^{\ast }}\left\Vert \overline{w}\right\Vert _{V_{1}} \notag \\
&\leq &\frac{1}{\sqrt{2}}\nu T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot ,L)\right\Vert
_{V_{1}^{\ast }}\gamma (t), \notag\end{aligned}$$where we took into account that $\left\Vert w\right\Vert _{V_{1}}\leq \gamma
(t),$ by $(\widehat{H_{k}}).$ Net, we note that $$\left\Vert \psi \right\Vert _{H^{2}(0,L)}=\sup_{\left\Vert \omega
\right\Vert _{(H^{2}(0,L))^{\ast }}\leq 1}\left\vert \left\langle \psi
,\omega \right\rangle _{H^{2}(0,L),(H^{2}(0,L))^{\ast }}\right\vert .
\label{201-0}$$Going back to (\[91-00\]) $${\rm Re}\left\langle E_{0k}p_{k,\varepsilon }(t),\overline{v}(t)\right\rangle _{H^{2}(0,L),(H^{2}(0,L))^{\ast }}={\rm Re}(E_{0k}p_{k,\varepsilon }(t),\overline{v}(t))_{H}=-\nu T_{k,\varepsilon
}^{\ast }{\rm Re}\int_{t}^{1}\overline{p_{k,\varepsilon }^{\prime \prime
\prime }}(\tau ,L)\overline{w}(\tau )d\tau$$and taking $\sup $, we deduce by (\[201-0\]) that $$\begin{aligned}
\left\Vert E_{0k}p_{k,\varepsilon }(t)\right\Vert _{H^{2}(0,L)}
&=&\sup_{\left\Vert v(t)\right\Vert _{(H^{2}(0,L)^{\ast }}\leq 1}\left\vert
\left\langle E_{0k}p_{k,\varepsilon }(t),\overline{v}(t)\right\rangle
_{H^{2}(0,L),(H^{2}(0,L))^{\ast }}\right\vert \label{200} \\
&\leq &\nu T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon
}^{\prime \prime \prime }}(\cdot ,L)\right\Vert _{V_{1}^{\ast }}\left\Vert
\overline{w}\right\Vert _{V_{1}}\leq \nu T_{k,\varepsilon }^{\ast
}\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot
,L)\right\Vert _{V_{1}^{\ast }}\gamma (t), \notag\end{aligned}$$where we used again the part of $(\widehat{H_{k}})$ written before. Now, by a straightforward calculation we deduce that $$\left\Vert p_{k,\varepsilon }(t)\right\Vert _{H^{4}(0,L)}\leq \sqrt{2}\left\Vert E_{0k}p_{k,\varepsilon }(t)\right\Vert _{H^{2}(0,L)}+\sqrt{2}k^{2}\left\Vert p_{k,\varepsilon }^{\prime \prime }(t)\right\Vert
_{L^{2}(0,L)}. \label{94-0}$$Using now (\[200\]) and (\[201\]) we write $$\left\Vert p_{k,\varepsilon }(t)\right\Vert _{H^{4}(0,L)}\leq \sqrt{2}\nu
T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon }^{\prime
\prime \prime }}(\cdot ,L)\right\Vert _{V_{1}^{\ast }}\gamma (t)+k^{2}\nu
T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon }^{\prime
\prime \prime }}(\cdot ,L)\right\Vert _{V_{1}^{\ast }}\gamma (t)$$and obtain the observability relation$$\left\Vert p_{k,\varepsilon }(t)\right\Vert _{H^{4}(0,L)}\leq (\sqrt{2}+k^{2})\nu T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon
}^{\prime \prime \prime }}(\cdot ,L)\right\Vert _{V_{1}^{\ast }}\gamma (t).
\label{94}$$By (\[57-2\]) we have the relation $\left\Vert F_{0k}^{\ast
}p_{k,\varepsilon }(t)\right\Vert _{H}\leq C_{\nu ,U,k}\left\Vert
p_{k,\varepsilon }(t)\right\Vert _{H^{4}(0,L)}.$
Now, we rewrite (\[eta\]) (multiplied by $\rho _{k}\sqrt{T_{k,\varepsilon
}^{\ast }})$ and replace $\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\alpha
_{k,\varepsilon }\left\Vert \ddot{w}_{k,\varepsilon }^{\ast }\right\Vert
_{V_{1}^{\ast }}$ using (\[62\]), as follows:$$\begin{aligned}
&&\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\left\Vert T_{k,\varepsilon
}^{\ast }\nu \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot
,L)\right\Vert _{V_{1}^{\ast }}\leq \rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\alpha _{k,\varepsilon }\left\Vert \ddot{w}_{k,\varepsilon }^{\ast
}\right\Vert _{V_{1}^{\ast }}+\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\left\Vert \int_{\cdot }^{1}h_{k,\varepsilon }^{\ast }(\tau )d\tau
\right\Vert _{V_{1}^{\ast }} \\
&\leq &T_{k,\varepsilon }^{\ast }\int_{0}^{1}\left\Vert \overline{v_{k,\varepsilon }^{\ast }}(t)\right\Vert _{H}\left\Vert F_{0k}^{\ast
}p_{k,\varepsilon }(t)\right\Vert _{H}dt+1+\int_{0}^{1}\left\vert
\int_{t}^{1}h_{k,\varepsilon }^{\ast }(\tau )d\tau \right\vert \left\vert
\overline{w_{k,\varepsilon }^{\ast }}(t)\right\vert dt+C_{\varepsilon }(\rho
_{k},T_{k,\varepsilon }^{\ast }) \\
&\leq &T_{k,\varepsilon }^{\ast }C_{v}\int_{0}^{1}C_{\nu ,U,k}\left\Vert
p_{k,\varepsilon }(t)\right\Vert _{H^{4}(0,L)}dt+1+C_{\varepsilon }(\rho
_{k},T_{k,\varepsilon }^{\ast }) \\
&\leq &C_{v}C_{\nu ,U,k}T_{k,\varepsilon }^{\ast }(\sqrt{2}+k^{2})\int_{0}^{1}\nu T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot ,L)\right\Vert
_{V_{1}^{\ast }}\gamma (t)dt+1+C_{\varepsilon }(\rho _{k},T_{k}^{\ast }) \\
&\leq &C_{v}C_{\nu ,U,k}T_{k}^{\ast }(\sqrt{2}+k^{2})\nu T_{k}^{\ast
}\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot
,L)\right\Vert _{V_{1}^{\ast }}\left\Vert \gamma \right\Vert
_{L^{1}(0,1)}+1+C_{\varepsilon }(\rho _{k},T_{k}^{\ast }),\end{aligned}$$where $C_{\varepsilon }(\rho _{k},T_{k}^{\ast })$ goes to zero, as $\varepsilon \rightarrow 0,$ by (\[52-6\]).
Here, we took into account the observability relation (\[94\]), the boundedness of $v_{k,\varepsilon }^{\ast }$ in $C([0,1];H^{2}(0,L))$ by (\[vk-est1\]) (with the constant in (\[vk-est1\]) called here $C_{v})$ and the fact that $T_{k,\varepsilon }^{\ast }\leq T_{k}^{\ast }$ by ([52-1]{}). Finally, we obtain$$\begin{aligned}
&&\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}\nu T_{k,\varepsilon }^{\ast
}\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot
,L)\right\Vert _{V_{1}^{\ast }}\leq 1+C_{\varepsilon }(\rho
_{k},T_{k,\varepsilon }^{\ast })+\sqrt{2}C_{\nu ,U,k}T_{k}^{\ast
}(1+k^{2})\left\Vert \gamma \right\Vert _{L^{1}(0,1)} \label{95} \\
&&\times \sqrt{C}\left( \left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap
H_{0}^{2}(0,L)}+\rho _{k}(T_{k,\varepsilon }^{\ast }+\sqrt{T_{k,\varepsilon
}^{\ast }})\right) \nu T_{k,\varepsilon }^{\ast }\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot ,L)\right\Vert
_{V_{1}^{\ast }}. \notag\end{aligned}$$The constants $\sqrt{2}$ and $\sqrt{C}$ coming from (\[vk-est1\]) will be included in $C_{\nu ,U,k}.$ Therefore, by denoting $$\begin{aligned}
D_{k}(T) &:&=1-C_{\nu ,U,k}(1+k^{2})\left\Vert \gamma \right\Vert
_{L^{1}(0,1)}\sqrt{T}(T+\sqrt{T}),\mbox{ } \\
G_{k}(T) &:&=C_{\nu ,U,k}(1+k^{2})\sqrt{T}\left\Vert \gamma \right\Vert
_{L^{1}(0,1)}\left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap H_{0}^{2}(0,L)},\end{aligned}$$we get$$T_{k,\varepsilon }^{\ast }\nu \left\Vert \overline{p_{k,\varepsilon
}^{\prime \prime \prime }}(\cdot ,L)\right\Vert _{V_{1}^{\ast }}\sqrt{T_{k,\varepsilon }^{\ast }}(\rho _{k}D_{k}(T_{k,\varepsilon }^{\ast
})-G_{k}(T_{k,\varepsilon }^{\ast }))\leq 1+C_{\varepsilon }(\rho
_{k},T_{k}^{\ast }). \label{96}$$On the one hand, on the basis of (\[cond-rok\]) we have that $$1>C_{\nu ,U,k}(1+k^{2})\sqrt{T_{k}^{\ast }}\left( T_{k}^{\ast }+\sqrt{T_{k}^{\ast }}\right) \left\Vert \gamma \right\Vert _{L^{1}(0,1)}\geq C_{\nu
,U,k}(1+k^{2})\sqrt{T_{k,\varepsilon }^{\ast }}\left( T_{k,\varepsilon
}^{\ast }+\sqrt{T_{k,\varepsilon }^{\ast }}\right) \left\Vert \gamma
\right\Vert _{L^{1}(0,1)},$$since $T_{k,\varepsilon }^{\ast }<T_{k}^{\ast }$ by Theorem 3.2, so $D_{k}(T_{k,\varepsilon }^{\ast })>0$. On the other hand, $G_{k}(T_{k}^{\ast
})>G_{k}(T_{k,\varepsilon }^{\ast })$ and $D(T_{k}^{\ast
})<D(T_{k,\varepsilon }^{\ast }),$ so that $\rho _{k}D_{k}(T_{k,\varepsilon
}^{\ast })-G_{k}(T_{k,\varepsilon }^{\ast })>\rho _{k}D_{k}(T_{k}^{\ast
})-G_{k}(T_{k}^{\ast }).$ All these imply$$\begin{aligned}
\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot
,L)\right\Vert _{V_{1}^{\ast }} &\leq &\frac{1+C_{\varepsilon }(\rho
_{k},T_{k}^{\ast })}{\nu T_{k,\varepsilon }^{\ast }\sqrt{T_{k,\varepsilon
}^{\ast }}(\rho _{k}D_{k}(T_{k,\varepsilon }^{\ast })-G_{k}(T_{k,\varepsilon
}^{\ast }))} \label{97} \\
&<&\frac{1+C_{\varepsilon }(\rho _{k},T_{k}^{\ast })}{\nu T_{k,\varepsilon
}^{\ast }\sqrt{T_{k,\varepsilon }^{\ast }}(\rho _{k}D_{k}(T_{k}^{\ast
})-G_{k}(T_{k}^{\ast }))}<\infty , \notag\end{aligned}$$independent of $\varepsilon ,$ since $T_{k}^{\ast },T_{k,\varepsilon }^{\ast
}>0$ and $\frac{1}{T_{k,\varepsilon }^{\ast }}\rightarrow \frac{1}{T_{k}^{\ast }}.$
Going back to (\[94\]), we deduce that$$\int_{0}^{1}\left\Vert p_{k,\varepsilon }(t)\right\Vert
_{H^{4}(0,L)}^{2}dt\leq \left( \sqrt{2}(1+k^{2})\nu T_{k,\varepsilon }^{\ast
}\left\Vert \overline{p_{k,\varepsilon }^{\prime \prime \prime }}(\cdot
,L)\right\Vert _{V_{1}^{\ast }}\right) ^{2}\int_{0}^{1}\gamma ^{2}(t)dt\leq
C, \label{98}$$independent on $\varepsilon ,$ since $\gamma \in L^{2}(0,1).$ Here, $C$ denotes several constants depending on the parameters and $k.$
Therefore, on a subsequence we have $$p_{k,\varepsilon }\rightarrow p_{k}\mbox{ weakly in }L^{2}(0,1;H^{4}(0,L)),\mbox{ as }\varepsilon \rightarrow 0. \label{99-1}$$By the adjoint equation $$E_{0k}(p_{k,\varepsilon })_{t}=T_{k,\varepsilon }^{\ast }F_{0k}^{\ast
}p_{k,\varepsilon }\rightarrow T_{k}^{\ast }F_{0k}^{\ast }p_{k}\mbox{ weakly
in }L^{2}(0,1;H),$$hence $$(p_{k,\varepsilon })_{t}\rightarrow T_{k}^{\ast }E_{0k}^{-1}F_{0k}^{\ast
}p_{k}:=(p_{k})_{t}\mbox{ weakly in }L^{2}(0,1;H^{2}(0,L)\cap
H_{0}^{1}(0,L)). \label{99-2}$$Thus, we get at limit the adjoint equation $$-E_{0k}(p_{k})_{t}(t)+T_{k}^{\ast }F_{0k}^{\ast }p_{k}(t)=0\mbox{ a.e. }t\in
(0,1), \label{99-6}$$which is (\[103\]).
Since $H^{4}(0,1)$ is compact in $H^{4-\varepsilon ^{\prime }}(0,1)$ for any $\varepsilon ^{\prime }>0,$ it follows by (\[99-1\]), (\[99-2\]) and the Aubin-Lions lemma that $$p_{k,\varepsilon }\rightarrow p_{k}\mbox{ strongly in }L^{\infty
}(0,1;H^{4-\varepsilon ^{\prime }}(0,L)). \label{99-3}$$By Ascoli-Arzelà theorem we have $$p_{k,\varepsilon }\rightarrow p_{k}\mbox{ strongly in }C([0,1];H^{2}(0,L)).$$Also, (\[99-3\]) implies $$p_{k,\varepsilon }^{\prime \prime \prime }\rightarrow p_{k}^{\prime \prime
\prime }\mbox{ strongly in }L^{2}(0,1;H^{1-\varepsilon ^{\prime }}(0,L))\mbox{ for all }\varepsilon ^{\prime }>0$$and also, by the trace convergence $$p_{k,\varepsilon }^{\prime \prime \prime }(\cdot ,L)\rightarrow
p_{k}^{\prime \prime \prime }(\cdot ,L)\mbox{ strongly in }L^{2}(0,1).$$On the other hand, by (\[97\]) $$p_{k,\varepsilon }^{\prime \prime \prime }(\cdot ,L)\rightarrow \chi \mbox{
weakly in }V_{1}^{\ast },\mbox{ } \label{99-4}$$which combined with the previous convergence yields $$\chi (t)=p_{k}^{\prime \prime \prime }(t,L)\mbox{ a.e. }t\in (0,1).
\label{99-5}$$The convergence of $(v_{k,\varepsilon }^{\ast })_{\varepsilon }$ is ensured by Theorem 2.2, namely $$v_{k,\varepsilon }^{\ast }\rightarrow v_{k}^{\ast }\mbox{ strongly in }C([0,1];H^{2}(0,L))\cap W^{1,2}(0,1;H^{2}(0,L))\cap L^{2}(0,1;H^{4}(0,L))\mbox{.}$$
We recall that $w_{k,\varepsilon }^{\ast }\rightarrow w_{k}^{\ast }$ weakly in $H^{1}(0,1)$ and uniformly in $C([0,1]).$
We have already viewed that the case $\left\Vert w_{k,\varepsilon }^{\ast
}\right\Vert _{V_{1}}<\rho _{k}$ is not acceptable.
Thus, we have to discuss only the case $\left\Vert w_{k,\varepsilon }^{\ast
}\right\Vert _{V_{1}}=\left\Vert \ddot{w}_{k,\varepsilon }^{\ast
}\right\Vert _{V_{1}^{\ast }}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}.$
We assert that there exists $\alpha _{k}^{\ast }>0$ such that $\alpha
_{k,\varepsilon }\rightarrow \alpha _{k}^{\ast }.$ Indeed, from (\[eta\]) and (\[97\]) we derive that $(\alpha _{k,\varepsilon })_{\varepsilon }$ is bounded. By absurd, if it is not, it means that $\ddot{w}_{k,\varepsilon
}^{\ast }\rightarrow 0,$ that is $\ddot{w}_{k}^{\ast }=0,$ which we have already seen that it is not compatible with our problem.
Since $\left\Vert \ddot{w}_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}^{\ast
}}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}$, on the one hand, $\ddot{w}_{k,\varepsilon }^{\ast }\rightarrow \ddot{w}_{k}^{\ast }$ weakly in $V_{1}^{\ast },$ as $\varepsilon \rightarrow 0.$
On the other hand, by (\[eta\]) we have that $$\alpha _{k,\varepsilon }\ddot{w}_{k,\varepsilon }^{\ast }\rightarrow
T_{k}^{\ast }\nu \overline{p_{k}^{\prime \prime \prime }}(\cdot ,L)\mbox{
weakly in }V_{1}^{\ast }$$because $h_{k,\varepsilon }^{\ast }\rightarrow 0$ strongly in $L^{2}(0,1)$ and so in $V_{1}^{\ast },$ by (\[52-6\]). We immediately infer that ([101]{}) takes place.
Then, by passing to the limit in (\[62\]), where $\left\Vert \ddot{w}_{k,\varepsilon }^{\ast }\right\Vert _{V_{1}^{\ast }}=\rho _{k}\sqrt{T_{k,\varepsilon }^{\ast }}$ we obtain (\[102\]). This ends the proof.$\square $
Problem $(P)$
=============
**Definition 5.1.** We call a *quasi minimal* solution to problem $(P)$ a pair $(T^{\ast },w^{\ast })$ given by $$T^{\ast }:=\sup\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\{T_{k}^{\ast };\mbox{ }v_{k}^{\ast }(T_{k}^{\ast })=0\},\mbox{ }w^{\ast
}(t,x)=\sum\limits_{k\in \mathbb{Z}_{k},\mbox{ }k\neq 0}w_{k}^{\ast
}(t)e^{ikx} \label{T*}$$where $(T_{k}^{\ast },w_{k}^{\ast })$ is optimal in $(P_{k}),$ for each $k\in \mathbb{Z}\backslash \{0\}$.
We recall that an admissible pair for $(P)$ is a pair $(T_{\ast },w_{\ast })$ with $w_{\ast }\in H^{1}(0,T;L^{2}(0,2\pi )),$ $w_{\ast }(0,y)=0,$ $\int_{0}^{T}\int_{0}^{2\pi }\left\vert w_{\ast t}(t,x)\right\vert
^{2}dxdt\leq \rho ^{2}$ and $u(T,x,y)=0,$ $v(T,x,y)=0$.
Further, we shall resume the notation $"\symbol{94}"$ corresponding to the functions in $(\widehat{P_{k}})$ by the transformation (\[t-cap\]).
**Theorem 5.2.** *Let* $(u_{0},v_{0})\in L^{2}(0,2\pi
;H^{3}(0,L)\cap H_{0}^{1}(0,L))\times L^{2}(0,2\pi ;H^{4}(0,L)\cap
H_{0}^{2}(0,L))$ *and* $(u_{0},v_{0})\neq (0,0).$ *If* $(P)$ *has an admissible pair* $(T_{\ast },w_{\ast }),$ *there exists a quasi minimal solution* $(T^{\ast },w^{\ast })$ *to* $(P)$* given by* (\[T\*\]), *with the corresponding state* * *$$\begin{aligned}
u^{\ast } &\in &W^{1,2}(0,T^{\ast };L^{2}(0,2\pi ;H^{1}(0,L)))\cap
L^{2}(0,T^{\ast };L^{2}(0,2\pi ;H^{3}(0,L))) \label{sol-u} \\
v^{\ast } &\in &W^{1,2}(0,T^{\ast };L^{2}(0,2\pi ;H^{2}(0,L)))\cap
L^{2}(0,T^{\ast };L^{2}(0,2\pi ;H^{4}(0,L))). \notag\end{aligned}$$*Moreover, for the modes for which* $\rho _{k},$ $T_{k}^{\ast }$ *and* $k$ *satisfy* $$\begin{aligned}
&&C_{\nu ,U,k}(1+k^{2})\left\Vert \gamma \right\Vert _{L^{1}(0,T_{k}^{\ast
})}\left( \sqrt{T_{k}^{\ast }}+1\right) <1,\mbox{ } \label{94-6} \\
&&\rho _{k}\sqrt{T_{k}^{\ast }}\left( 1-C_{\nu ,U,k}(1+k^{2})\left\Vert
\gamma \right\Vert _{L^{1}(0,T_{k}^{\ast })}(\sqrt{T_{k}^{\ast }}+1)\right)
\notag \\
&>&C_{\nu ,U,k}(1+k^{2})\left\Vert \gamma \right\Vert _{L^{1}(0,T_{k}^{\ast
})}\left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap H_{0}^{2}(0,L)}, \notag\end{aligned}$$*the optimal pairs* $(T_{k}^{\ast },w_{k}^{\ast })$ *in* $(P_{k})$ *are given by* * *$$\alpha _{k}^{\ast }T_{k}^{\ast }\ddot{w}_{k}^{\ast }(t)=\nu \overline{p_{k}^{\prime \prime \prime }}(t,L),\mbox{ \textit{a.e. }}t\in (0,1),\mbox{ }w_{k}^{\ast }(0)=\dot{w}_{k}^{\ast }(T_{k}^{\ast })=0,\mbox{ }\left\Vert
w_{k}^{\ast }\right\Vert _{V_{T_{k}^{\ast }}}=\rho _{k}, \label{94-7}$$$$\alpha _{k}^{\ast }\rho _{k}^{2}T_{k}^{\ast }+{\rm Re}\int_{0}^{T_{k}^{\ast
}}(\overline{v}_{k}^{\ast }(t),F_{0k}^{\ast }p_{k}(t))_{H}dt=1,\mbox{ }
\label{94-8}$$*where* $\alpha _{k}^{\ast }>0,$ $v_{k}^{\ast }$ *is the solution to* (\[4\])-(\[4-3\]) *and* $p_{k}$ *is the solution to the adjoint equation*$$-E_{0k}(p_{k})_{t}(t)+F_{0k}^{\ast }p_{k}(t)=0,\mbox{ \textit{a.e.} }t\in
(0,T_{k}^{\ast }). \label{94-9}$$
**Proof.** By the hypothesis, there exists $T_{\ast }$ such that $u(T_{\ast },x,y)=v(T_{\ast },x,y)=0.$ Then, by the Parseval identity, it follows for the modes $k$ that $w_{\ast k}\in H^{1}(0,T_{\ast })$ and $u_{k}(T_{\ast },y)=v_{k}(T_{\ast },y)=0.$ Moreover, one can choose $\rho
_{k}:=\left( \int_{0}^{T}\left\vert (w_{\ast k})_{t}(t)\right\vert
^{2}dt\right) ^{1/2},$ which ensures that$\sum\limits_{k\in \mathbb{Z},\mbox{
}k\neq 0}\rho _{k}^{2}\leq \rho ^{2}.$ Thus, $(T_{\ast },w_{\ast k})$ turns out to be an admissible pair for $(P_{k})$ and correspondingly, $(T_{\ast },\widehat{w_{\ast k}})$ is an admissible pair in $(\widehat{P_{k}}).$ Then, by Theorem 2.3, it follows that $(\widehat{P_{k}})$ has a solution $(T_{k}^{\ast },\widehat{w_{k}^{\ast }}),$ and since $T_{k}^{\ast }$ is minimal it follows that $T_{k}^{\ast }<T_{\ast }.$ The state $\widehat{v_{k}^{\ast }}$ has the properties (\[vk\]) and $\widehat{v_{k}^{\ast }}(1,y)=0$. Moreover, by extending $\widehat{w_{k}^{\ast }}$ by $0$ after $\widehat{t}=1$ we get $\widehat{v_{k}^{\ast }}(t)=0$ for $\widehat{t}>1$.
Now, we resume the transformation (\[t-cap\]) and set $\widehat{t}=\frac{t}{T_{k}^{\ast }},$ where $t\in \lbrack 0,T_{k}^{\ast }]$ if $\widehat{t}\in
\lbrack 0,1].$ All the results obtained for $(\widehat{P_{k}})$ will be correspondingly transported to $(P_{k})$ on $(0,T_{k}^{\ast }).$ Thus, $(P_{k})$ has an optimal pair $(T_{k}^{\ast },w_{k}^{\ast }),$ $v_{k}^{\ast
}(T_{k}^{\ast })=0$ and $v_{k}^{\ast }(t)=0$ for $t>T_{k}^{\ast }.$
By the third equation in (\[2\]) we have $u_{k}^{\ast }(t,y):=\frac{i}{k}(v_{k}^{\ast })^{\prime }(t,y)$ and so $u_{k}^{\ast }(T_{k}^{\ast })=0$ and $u_{k}^{\ast }(t)=0$ for $t>T_{k}^{\ast }.$
We set $T^{\ast }=\sup\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\{T_{k}^{\ast
};$ $v_{k}^{\ast }(T_{k}^{\ast })=0\}$ and $w^{\ast }$ as in (\[T\*\]). It follows that $(T^{\ast },w^{\ast })$ is a quasi minimal solution to $(P)$ and since $T_{k}^{\ast }<T_{\ast }$ it follows that $T^{\ast }<T_{\ast }.$ By the Parseval identity, $u^{\ast }$ and $v^{\ast }$ constructed by ([0-3]{}) satisfy $u^{\ast }(T^{\ast })=v(T^{\ast })=0,$ as required in $(P),$ and $u^{\ast }(t)=v^{\ast }(t)=0$ for $t>T^{\ast }.$
Also, the solution $v_{k}^{\ast }$ to (\[4\])-(\[4-3\]) satisfies$$\begin{aligned}
&&\left\Vert v_{k}(t)\right\Vert _{H^{2}(0,L)}^{2}+\int_{0}^{T_{k}^{\ast
}}\left\Vert \frac{dv_{k}}{dt}(t)\right\Vert
_{H^{2}(0,L)}^{2}dt+\int_{0}^{T_{k}^{\ast }}\left\Vert v_{k}(t)\right\Vert
_{H^{4}(0,L)}^{2}dt \\
&\leq &C\left( \left\Vert v_{k0}\right\Vert _{H^{4}(0,L)\cap
H_{0}^{2}(0,L)}^{2}+\int_{0}^{T_{k}^{\ast }}\left\vert w_{k}(t)\right\vert
^{2}dt+\int_{0}^{T_{k}^{\ast }}\left\vert \dot{w}_{k}(t)\right\vert
^{2}dt\right) ,\mbox{ for all }t\geq 0.\end{aligned}$$Then, $$\begin{aligned}
&&\frac{1}{2\pi }\int_{0}^{2\pi }\left\Vert v^{\ast }(t,x)\right\Vert
_{H^{2}(0,L)}^{2}dx=\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\left\Vert
v_{k}^{\ast }(t)\right\Vert _{H^{2}(0,L)}^{2} \\
&\leq &C\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\left( \left\Vert
v_{k0}\right\Vert _{H^{4}(0,L)\cap H_{0}^{2}(0,L)}^{2}+\rho
_{k}^{2}(T_{k}^{\ast })^{2}+\rho _{k}^{2}T_{k}^{\ast }\right) \\
&\leq &C\left( \frac{1}{2\pi }\int_{0}^{2\pi }\left\Vert v_{0}(x)\right\Vert
_{H^{2}(0,L)}^{2}dx+\rho ^{2}\max \{(T^{\ast })^{2},T^{\ast }\}\right)
:=I_{1},\mbox{ for all }t\in \lbrack 0,T^{\ast }],\end{aligned}$$since $\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\rho _{k}^{2}\leq \rho
^{2}.$ Similarly, we proceed for showing that $v^{\ast }$ belongs to the other two spaces.
For $u^{\ast }$ we calculate e.g., $$\begin{aligned}
&&\frac{1}{2\pi }\int_{0}^{2\pi }\left\Vert u^{\ast }(t,x)\right\Vert
_{H^{1}(0,L)}^{2}dx=\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\left\Vert
u_{k}^{\ast }(t)\right\Vert _{H^{1}(0,L)}^{2}=\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\left\Vert \frac{i}{k}(v_{k}^{\ast })^{\prime
}(t)\right\Vert _{H^{1}(0,L)}^{2} \\
&=&\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\frac{1}{k^{2}}\left\Vert
(v_{k}^{\ast })^{\prime }(t)\right\Vert _{H^{1}(0,L)}^{2}\leq
\sum\limits_{k\in \mathbb{Z},\mbox{ }k\neq 0}\left\Vert v_{k}^{\ast
}(t)\right\Vert _{H^{2}(0,L)}^{2}\leq I_{1},\end{aligned}$$and proceed similarly for the other norms, so that $u^{\ast }$ and $v^{\ast
} $ belong to the spaces (\[sol-u\]).
The second part of the statement follows immediately by making the transformation $\widehat{t}=tT_{k}^{\ast }$ in (\[cond-rok\]), (\[101\])-(\[103\]). This ends the proof.$\square $
**Remark 5.3.** We conclude that the study of the controllability problem $(P)$ returns the fact that, if $(P)$ has an admissible pair $(T_{\ast },w_{\ast })$ one can prove that there exists $(T^{\ast },w^{\ast })$ which ensures the flow stabilization towards the stationary laminar regime, with $T^{\ast }\leq T_{\ast }.$ The pair $(T^{\ast },w^{\ast })$ has an important property, namely it is constructed via the solutions of minimal time controllability problems for the modes of the Fourier transforms of the Navier-Stokes linearized system.
Finally, let us comment the conditions (\[94-6\]) that enhance the determination of (\[94-7\])-(\[94-9\]) which are the necessary conditions to be satisfied by the optimal pairs in $(P_{k}).$ One can note that by (\[94-6\]) the determination of the optimality conditions can be done for a finite number of modes $k,$ if $\rho _{k}$, that is, if $\rho $ is large enough.
By strengthening a little bit the first condition in (\[94-6\]), we have the possibility of calculating a rough estimate of the number of modes allowing the determination of the optimality conditions, on the basis of a relation involving the problem parameters and $T_{\ast }$. Let us assume$$C_{\nu ,U,k}(1+k^{2})\left\Vert \gamma \right\Vert _{L^{1}(0,T_{\ast
})}\left( \sqrt{T_{\ast }}+1\right) <1,\mbox{ } \label{cond-1}$$which obviously implies the first condition in (\[94-6\]), since $T_{\ast
}>T_{k}^{\ast }$. However, it provides a smaller number of $k$ than that obtained by (\[94-6\]). By calculating the norm $\left\Vert U\right\Vert
_{W^{1,\infty }(0,L)}$ we get $C_{\nu ,U,k}=Ck^{4}\left( \nu +\frac{a}{8\nu }(4+L)\right) ,$ with $C$ a positive constant. Thus, we see that a larger number of $k$ can be obtained function of a smaller controllability cost reflected by $\left\Vert \gamma \right\Vert _{L^{1}(0,T_{\ast })}$, or a smaller $T_{\ast }.$ Also, this number depends on $a$ (fixing the steady-state flow $U(y)),$ $L$ and $\nu ,$ whose values can lead to a smaller $C_{\nu ,U,k}.$
We also add that one can construct an approximating solution, by restraining the solution using this finite number of modes. Denote $S_{\rho }:=\{k\in
\mathbb{Z}\backslash \{0\};$ $k$ satisfies (\[cond-1\])$\}$ and express the initial data with respect to these modes only: $u_{0}(x,y)=\sum\limits_{k\in S_{\rho }}u_{k0}(y)e^{ikx},$ $v_{0}(x,y)=\sum\limits_{k\in
S_{\rho }}v_{k0}(y)e^{ikx}.$ Then we construct an approximating optimal state defined by $u^{\ast }(t,x,y)=\sum\limits_{k\in S_{\rho }}u_{k}^{\ast
}(t,y)e^{ikx},$ $v^{\ast }(t,x,y)=\sum\limits_{k\in S_{\rho }}v_{k}^{\ast
}(t,y)e^{ikx},$ corresponding to the optimal pair $(T_{k}^{\ast
},w_{k}^{\ast })$ given by (\[94-7\])-(\[94-9\]) with $T^{\ast
}=\sup\limits_{k\in S_{\rho }}\{T_{k}^{\ast };$ $v_{k}^{\ast }(T_{k}^{\ast
})=0\},$ $w^{\ast }(t,x)=\sum\limits_{k\in S_{\rho }}w_{k}^{\ast
}(t)e^{ikx}. $ Such a solution constructed on the basis of a finite number of modes can be used in numerical computations.
[99]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently it has been demonstrated that an ensemble of trapped ions may serve as a quantum annealer for the number-partitioning problem \[Nature Comm. DOI: 10.1038/ncomms11524\]. This hard computational problem may be addressed employing a tunable spin glass architecture. Following the proposal of the trapped ions annealer, we study here its robustness against thermal effects, that is, we investigate the role played by thermal phonons. For the efficient description of the system, we use a semiclassical approach, and benchmark it against the exact quantum evolution. The aim is to understand better and characterize how the quantum device approaches a solution of, an otherwise, difficult to solve NP-hard problem.'
author:
- David Raventós
- 'Tobias Gra[ß]{}'
- 'Bruno Juliá-Díaz'
- Maciej Lewenstein
bibliography:
- 'bib.bib'
title: Semiclassical approach to finite temperature quantum annealing with trapped ions
---
Introduction {#sec:intro}
============
Quantum computers and quantum simulators are nowadays becoming a reality thanks to the advances in ion trapping and integrated superconducting technology [@2017arXiv171203773A; @supercon; @Linke3305; @Monroe1164; @Devoret1169]. A possible device which is quickly being developed are quantum annealers. Annealing, as opposed to quenching, is a method to produce the ground state of a target Hamiltonian by slowly deforming/adjusting a well-known ground state of a different Hamiltonian. Annealing is in fact a concept originating from classical metallurgy, extended in the 1980s to classical optimization problems, and known as simulated annealing [@Kirkpatrick671; @cerny1985]. In the current quantum versions, quantum annealing is very much analogous to quantum adiabatic computing, but is typically targeted towards the classical optimization problems. The idea is to add a simple, noninteracting, but noncommuting term to the original classical Hamiltonian. This simple additional term should dominate the system at the initial time, so that the ground state will be easy to find, since it will correspond to a non-interacting system. The non-commuting nature of the additional term ensures that the initial and target ground states are not symmetry protected. Then, the additional term is adiabatically removed and the ground state is expected to go slowly from the initial one to the one of the Hamiltonian of interest [@Brooke779; @PhysRevE.58.5355; @Farhi472], see also the recent review [@RevModPhys.90.015002]. This scheme is nowadays plausible with a large number of possible platforms, including trapped ions, cavity QED, circuit QED, superconducting junctions [@nature10012] and atoms in nanostructures. The first commercially accessible quantum annealers are in the market [@Rnnow420; @Boixo2014; @youtube].
Since the original proposals [@PhysRevLett.87.257904; @PhysRevLett.92.207901] trapped ions quantum simulators are the subject of intensive theoretical and experimental research. Starting from realization of the simple instances of quantum magnetism [@friedenauer08], they have reached quite a maturity in the recent experimental developments (cf. [@2015arXiv151203559M; @PhysRevA.94.023401; @2016arXiv160906429L; @PhysRevLett.119.080501; @PhysRevLett.118.053001; @timecrystal; @Zhang2017]). The recent paper by Bollinger’s group [@2017arXiv171107392S], in addition to the excellent experimental work, contains also an outstanding analysis of quantum dynamics of the relevant Dicke model, in which the ions interact essentially with one phononic mode.
Quantum dynamics in general, and in particular for the Dicke-like ion-phonon models, are very challenging for numerical simulations. Exact treatments are possible for small systems only, so that various approximate methods have to be used. One of them is the truncated Wigner approximation, in which both ionic and phononic operators are replaced by complex numbers, the dynamics becomes “classical”, and only the initial data mimic the “quantumness” of the problem [@TWA]. This approach was used in Ref. [@PhysRevA.96.033607] to study the quantum non-equilibrium dynamics of spin-boson models. More sophisticated “mean-field” approaches decorrelate ions from phonons, but treat at least either ions or phonons fully quantum mechanically – this approach is in particular analyzed in the present paper. Quantum aspects of the models in question were studied in the series of papers [@PhysRevA.95.013602; @Garttner2017; @PhysRevA.94.053637].
We have considered recently the exact quantum dynamics of few ion systems to demonstrate the robustness of chiral spin currents in a trapped-ion quantum simulator using Floquet engineering [@PhysRevA.97.010302]. Our earlier works include studies of dual trapped-ion quantum simulators as an alternative route towards exotic quantum magnets [@1367-2630-18-3-033011], and studies of ion chains with long range interactions forming “magnetic loops” [@PhysRevA.91.063612]. Topological edge states in periodically-driven trapped-ion chains [@PhysRevLett.119.210401], trapped ion quantum simulators of Rabi lattice models with discrete gauge symmetry [@2015PhRvA..92a3624N], and hidden frustrated interactions and quantum annealing in trapped ion spin-phonon chains [@PhysRevA.93.013625] were also considered recently. Novel ideas for spin-boson models simulated with trapped ions can be found in Ref. [@2017arXiv170400629L].
In Ref. [@HBHL], it has been proposed to use trapped ions for solving difficult optimization problems via quantum annealing. Such scheme, applied to the concrete example of number-partitioning, has come under scrutiny in Ref. [@ncomms11524]. The idea is to profit from the known mapping between the number partitioning problem and the ground state of spin Hamiltonians [@0305-4470-15-10-028]. In the interesting domain, where the number partitioning problem is notably difficult, the system is actually in the spin-glass-like phase, which renders finding the actual ground state an involved task for classical methods. The annealing method proposed was found to work well for small number of ions at zero temperature. In this work we explore in detail a semiclassical approximation to the original problem, where the quantum correlations between the spins and the phonon bath are neglected. This, however, allows us to solve the Heisenberg equations of motion in an efficient way for much larger ion systems. Notably, the approach allows us to explore finite temperature effects on the annealing protocol.
Our present work is structured as follows. In Sec. \[sec:start\] we explain the specific technical details of the studied model and the calculations. In particular, the Hamiltonian is explained in \[sec:Ham\], the reduction methods in \[sec:SCdecoupling\] and the annealing protocol \[sec:annprot\] along with the basic definitions \[sec:fide\]. In Sec. \[secIII\], we introduce our semiclassical approximation, and benchmark it with a full quantum treatment. In Sec. \[sec:CalcsAtTfin\] we show the results of the semiclassical approach applied to finite temperature. Here, we set the initial phonon population to non-zero thermal values. A summary and our main conclusions are provided in Sect. \[sec:concl\]. Finally, in the appendix we provide some further tests to our numerical integration method (appendix \[app:Integrationtests\]), and a brief study of the optimal bias for the annealing protocol (appendix \[sec:optim\]).
System {#sec:start}
======
Hamiltonian {#sec:Ham}
-----------
We study a chain of $N$ trapped ions interacting by effective spin-spin interactions subjected to a transverse time-dependent magnetic field. Interactions are generated by Raman coupling the pseudo-spin degrees of freedom to the phonon modes which are obtained expanding the Coulomb force between the ions around their equilibrium positions [@PhysRevLett.92.207901]. The phonon spectrum is defined through its natural frequencies $\omega_k$ and modes $\xi^{i}_{k}$. The dynamics of the system is described by a time-dependent Hamiltonian that in the Schrödinger picture reads, \_[S]{} ( t )/&= & \_[k]{}\^[M]{} [\_k \^\_k \_k]{} +\_[i,k]{}\^[N,M]{} [\^[(i)]{}\_[k]{} (\^\_k + \_k) \_[x]{}\^[( i )]{}]{}\
&+ & \_[i]{}\^[N]{} [B ( t ) \_[z]{}\^[( i )]{}]{} + \_[x]{}\^[( p )]{}, \[eq:Hamiltonian\] where $\hat{a}_k$($ \hat{a}^{\dagger}_k$) is the annihilation (creation) operator of one phonon in the $k$-th mode and $\omega_k$ is the frequency of that mode. The operators $\sigma_{x}^{\left( i \right)}$, $\sigma_{y}^{\left( i \right)}$, and $\sigma_{z}^{\left( i \right)}$ are the spin operators in the $i$-th position. The frequencies $\Omega$ and $\omega_{\rm L}$ are the Rabi frequency and the beatnote frequency of the laser, respectively. The dimensionless parameters $\eta^{(i)}_{k} $ are the Lamb-Dicke parameters proportional to the displacement of an ion $i$ in the vibrational mode $k$, see Ref. [@PhysRevLett.92.207901]. As usual, $t$ is time, and a time-dependent magnetic field $B(t)$ allows us to perform the quantum annealing. A small bias term, proportional to $\varepsilon$, has been added in the $p$-th position to remove the ${\cal Z}_2$ degeneracy. The upper limit of the sum over the ions $i$ is $N$, the number of ions. The upper limit of the sum in modes $k$ is $M$, the number of modes. The total number of phonon modes is $3N$, but the Raman beam couples to only $N$ modes, selected by the wave vector difference of the lasers. At this point, we may keep our analysis general by making no assumption about the number $M$ of modes. However, all phonons which are considered are assumed to be coupled to the spin in the same way. Hereinafter, the upper limits of the sums will be omitted for brevity.
Equations of motion {#sec:SCdecoupling}
-------------------
We compute the Heisenberg equations of motion for the quantum average of every operator in the Hamiltonian, $\hat{a}_k$, $\hat{a}^{\dagger}_k$, $\sigma_{x}^{\left( i \right)}$, $\sigma_{y}^{\left( i \right)}$, and $\sigma_{z}^{\left( i \right)}$. Given that these are time independent operators, the calculation reduces to commutators. Additionally, we replace $\hat{a}^{\dagger}_k+\hat{a}_k$ by $2 \Re \left[ \hat{a}_k\right]$ and $\hat{a}_k-\hat{a}^{\dagger}_k$ by $2 {\rm i} \Im \left[ \hat{a}_k\right]$. The equations of motion (all with real coefficients) read, &=& \_k \[eem\]\
&=& -\_k - \_[j]{} \^[(j)]{}\_[k]{} \_[x]{}\^[( j )]{} ,\
&=& -2 B( t ) \_[y]{}\^[( i )]{},\
&= & -4 \_[l]{} \^[(i)]{}\_[l]{} \_[z]{}\^[( i )]{}\
&&+2 B( t ) \_[x]{}\^[( i )]{} -2 \_[z]{}\^[( i )]{} \_[p,i]{},\
&=&4 \_[l]{} \^[(i)]{}\_[l]{} \_[y]{}\^[( i )]{}+2\_[y]{}\^[( i )]{} \_[p,i]{}.
Annealing protocol {#sec:annprot}
------------------
The functional form and value of $\Omega$, $B \left( t \right)$ and $\varepsilon$ determine the annealing protocol. In these annealing schemes, the initial value of the transverse magnetic field $B \left( t=0 \right)$ must be sufficiently strong to initialize the system in the paramagnetic phase, that is, $B$ must be larger than the effective spin-spin interactions $J \sim \Omega^2 \omega_{\rm rec}/ (\delta \omega_{\rm rad} )$, where $\omega_{\rm rec}$ is the recoil energy of the photon-ion coupling, $\omega_{\rm rad}$ is the radial trap frequency, and $\delta$ the detuning from the nearest phonon mode. For typical values, e.g. $\Omega \sim \delta \sim 100$ kHz, $\omega_{\rm rec}\sim 25$ kHz, $\omega_{\rm rad} \sim 5$ MHz, we obtain effective interactions $J\sim 1$ kHz, so we need an initial field strength $B(0)\sim 10$ kHz. The annealing scheme proceeds by turning down the magnetic field according to some functional form in order to adiabatically achieve the ground state of the Hamiltonian of interest. Given the adiabatic theorem, for a closed system initiallized in the ground state, the final system is guaranteed to be in the ground state as long as the system is gapped along the annealing path, and the variation is slow enough. Generalization to open systems has been proposed in Ref. [@PhysRevA.93.032118].
We have used a decreasing exponential form for the transverse magnetic field, $B \left( t\right)=B(0)\,e^{-\frac{t}{\tau}}$ with a decay rate $\tau$. The other parameters, $\Omega$ and $\varepsilon$, remain constant. An example of the evolution of the system under this protocol is shown in Fig. \[fig:sigma\_x\]. Initially, ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}=0$ for all $i$, and the total phononic population is set to $0$. Within tens of microseconds the phononic modes are populated. Not surprisingly, the mode next to the resonance becomes the most populated one, with a population being orders of magnitude larger than the population of the other modes. In contrast to these rapid changes of the phonon state, the spin dynamics is much slower. The spin expectations ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ remain mostly clustered around zero for hundreds of microseconds. When $B \left( t\right) \simeq \varepsilon$, the values ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ start to deviate from zero, and some acquire positive values, while others become negative. Thus, the spin curves separate from each other, and we call the time at which this happens the [*separation time*]{}. At some point after the separation time, the spin curves saturate, that is, from then on ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ remain constant in time. We define the [*waiting time*]{} as the time when all ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ have stopped varying. At the waiting time, the phononic populations stabilize around certain values, although their oscillations do never vanish.
The quantum annealer produces final values of ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ which are not fully polarized , that is, $|{\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}|<1$. Thus, the final state differs from the classical ground state of the target Hamiltonian, that is, the Hamiltonian in the absence of a transverse field. Thus, we take as readout of the annealing protocol the average spin values [@PhysRevE.94.032105; @PhysRevA.96.042310; @srep22318]. This is not a problem, as long as for all spins the sign matches with the one in the classical state. As explained in detail in Ref. [@ncomms11524], the spin configuration of the target Hamiltonian is determined by the dominant mode, defined as the one with frequency just below the beatnote frequency $\omega_{\rm L}$. There are different reasons why the final ground state might show a different spin pattern: Either, the annealing was too fast, that is, the value of $\tau$ was chosen too small, or the effective spin model is not valid. This is the case when $\omega_{\rm L}$ is too close to a resonance $\omega_k$.
![(Upper panel) Evolution of ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ in a system with 4 spins, with initial populations of phononic modes set to $0$. The biased spin is the pink one, see text for details. (Lower panel) Evolution of the populations of the phononic modes in a system with 4 spins, with initial populations of phononic modes set to $0$. $\delta = 1$ MHz.[]{data-label="fig:sigma_x"}](Ex_Ev_SigsandOccs.ps){width="50.00000%"}
Fidelity of the annealing protocol {#sec:fide}
----------------------------------
In order to quantify the success of the annealing protocol, that is, the ability of the method to identify the target ground state of the spin system, we will define the following fidelity, $$F=\begin{cases}
\min\limits_{i} {\left| {\langle\sigma_{x}^{(i)}\rangle}\right|} &{\rm if\;}
{\rm sign}{\left[{\langle\sigma_{x}^{(i)}\rangle}\right]}={\rm sign}{\left[ \eta_{k_d}^{(i)} \right]}, \forall i\\
0 & {\rm Otherwise}
\end{cases},
\label{eq:Fide_def}$$ where $\eta_{i j}$ is the $i$th component of the dominant mode $k_d$ for a fixed value of the beatnote frequency $\omega_{\rm L}$. That is, the fidelity is zero if the signs of $\langle \sigma_x^i\rangle$ do not match the signs of $\eta_i$ of the dominant mode. If the signs are reproduced, the value of the fidelity is defined as the smallest expectation value of the spins of the ions. Note that with this definition, any non-zero fidelity is good enough for correctly identifying the ground state pattern, assuming the absence of noise in the system.
Semiclassical approximation {#secIII}
===========================
Semiclassical equations of motion
---------------------------------
Now we will develop a semiclassical approximation to the exact equations of motion, Eq. (\[eem\]), that will allow us to study larger systems of ions and the effects of temperature on the annealing protocols. We make the following approximations: \_k \_\^[( i )]{} && \_k \_\^[( i )]{}\
\^\_k \_\^[( i )]{} && \^\_k \_\^[( i )]{} with $\mu = \left\{ x,y,z\right\}$. These approximations ignore the quantum correlations in the coupling between bosonic and spin modes.
Additionally defining the auxiliary variables $S_k \left( t \right) \equiv \sin{\left(\omega_{\rm L} t\right)}
\sum_{j} \Omega \eta^{j}_{k} \langle \sigma_{x}^{\left( j \right)} \rangle$ and $J^{\left( i \right)} \left( t \right) \equiv \sin{\left(\omega_{\rm L} t\right)}
\Omega \left(2\sum_{l} \eta^{i}_{l} \langle \Re{ \left[\hat{a}_l\right]}
\rangle \right) +\varepsilon \delta_{p,i}$, we obtain the approximate equations of motion, &=& \_k , \[eq:Eqs\_motion\_simpl\]\
&=& -\_k -S\_k ( t ),\
&=& -2
0 & B( t ) & 0\
-B( t ) & 0 & J\^[( i )]{} ( t )\
0 & -J\^[( i )]{} ( t ) & 0
\^[( i )]{} where a spin vector notation, $\langle \vec{\sigma}^{\left( i \right)} \rangle =\left( \langle \sigma_{x}^{\left( i \right)} \rangle,
\langle \sigma_{y}^{\left( i \right)} \rangle, \langle \sigma_{z}^{\left( i \right)} \rangle \right)$, has been used. This is a system of $2\times M+ 3\times N$ non-linear first-order ordinary differential equations. Hence, it is numerically solved with a first order, ordinary differential equation solver that uses the Gragg–Bulirsch–Stoer method, stepsize control and order selection, called ODEX [@ODEXcite].
Comparison of semiclassical approximation to full quantum evolution
-------------------------------------------------------------------
![\[fig:fig4\] Quantum (a,c) and semiclassical (b,d) evolution of ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$. The figure is computed with $N=4$ ions with a detuning $\delta =2900$ kHz. Panels (a) and (b) correspond to $\tau=0.35$ ms and panels (c) and (d) to $\tau=0.7$ ms. ](figura.eps){width="50.00000%"}
To benchmark the semiclassical method, we have compared it against a full quantum evolution of the system using Krylov subspaces. The latter is a method to study the dynamical evolution under time-dependent Hamiltonians that computes a reduced evolution operator omitting contributions smaller than a certain threshold, that is, transitions to irrelevant states. Despite this neglection, the Krylov evolution can be considered an *exact* numerical simulation, as it iteratively determines which part of the Hilbert space is irrelevant at a given accuracy.
In both cases, semiclassical and full quantum, the evolution of a given initial state under a time-depending Hamiltonian is calculated with time steps in a recurrent way. In the fully quantum calculation, the time steps are of the order of $1$ ns, while in the semiclassical description they are variable, but can be orders of magnitude larger. In the quantum case, we have to specify a quantum state —a complex vector in the joint Fock basis of phononic and spin modes—, containing the amplitudes of every state of the basis. In the semiclassical case, we only have to supply the initial mean values of every operator.
It should be noted that the exact quantum evolution requires truncating the maximum phonon number which in our case was set to two phonons per mode. Such truncation of the Hilbert space requires sufficiently cool systems. And even with this truncation, the quantum evolution is restricted to a small number of ions. Considering only one transverse phonon branch, i.e. $N$ phonon modes, with a maximum population of two phonons per mode, the Hilbert space dimension is $2^N \times 3^N$, that is, a dimension of 46656 for $N=6$ ions. The semiclassical approach, in contrast, allows us to explore larger systems.
![Evolution of ${\langle\hat{n}_3 \hat{\sigma}_{x}^{\left( i \right)}\rangle}$ (a) and of the relative difference $({\langle\hat{n}_3 \hat{\sigma}_{x}^{\left( i \right)}\rangle}-{\langle\hat{n}_3\rangle} {\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle})/{\langle\hat{n}_3 \hat{\sigma}_{x}^{\left( i \right)}\rangle} $ for a system of four spins with initial populations of phonons set to zero for a detuning $\delta=1000$ kHz. []{data-label="fig:diffs"}](Ex_Ev_ComparExRelErr){width="50.00000%"}
### Time evolution of ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$
The semiclassical model captures well the qualitative behavior of the evolution of ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$, as exemplified in Fig. \[fig:fig4\]. In the figure we compare the semiclassical evolution with the exact dynamics for $N=4$ ions for two different decay times. The discrepancy between semiclassical and exact evolution is smallest for shorter times ($t\lesssim \tau$), where the semiclassical model is able to correctly capture the details of the dynamical evolution, most notably little wiggles in the evolution of the biased ion, $i=1$. Importantly, the semiclassical model also agrees with the exact evolution regarding general features such as the separation time, and the sign of each ${\langle\hat{\sigma}_{x}^{\left( i \right)}\rangle}$ in the long-time limit. As discussed in more detail in the next paragraph, this enables a quite accurate prediction of annealing fidelities, despite the fact that the approximation disregards some quantum properties. Thus it provides a computationally efficient way to study the behavior of larger systems.
For a better understanding of the errors in the semiclassical approach, we have exactly calculated the evolution of $\langle \hat n_3 \sigma_x^{(i)} \rangle$, see Fig. \[fig:diffs\](a), and of $\langle \hat n_3 \rangle \langle \sigma_x^{(i)} \rangle$. As our semiclassical approximation is based on substituting the former correlator by the latter one, the discrepancy between both expressions is an indicator for the quality of the semiclassical approach. In Fig. \[fig:diffs\](b), we plot the relative difference as a function of time: Initially, the phonon and spin degrees of freedom are taken as uncorrelated, thus, the semiclassical and the exact description coincide at $t=0$. On short time scales, both correlators have small absolute values, but their relative difference becomes large. For times larger than the separation time, the absolute values of the correlators increase, and the relative errors decrease. On long time scales, the errors oscillate around mean values of the order 0.1. This observation suggests that the main errors made in the semiclassical approximation are introduced at short times, where the transverse magnetic field and its temporal derivative takes large values.
![\[fig:QFide\_na4\] Fidelity obtained with the full quantum evolution $F_{\rm E}$ (a), the semiclassical model $F_{\rm SC}$ (b), and the difference between the semiclassical model and the full quantum (c). In both cases, we consider a system of four ions, and we plot the magnitudes as a function of $\omega_{\rm L}$ and $\tau$. The fidelity is readout after a time $20 \,\tau$.](Fides_na4_diff.eps){width="50.00000%"}
### Fidelity {#secfide}
As we have seen the semiclassical approximation provides a reasonable description of the dynamics in many configurations. Let us now explore in more detail in which parameter regions it predicts the correct fidelity for the annealing protocol. In Fig. \[fig:QFide\_na4\] we present a comparison of the fidelities obtained from the exact time evolution and from the semiclassical approach for a system of four ions. We tune through a broad range of beatnote frequencies $\omega_L$, and vary the decay time $\tau$ of the magnetic field. The overall agreement is very good: Both methods predict a small fidelity when the field decays too fast (small $\tau$), or when the system is too close to one of the phonon resonances. The semiclassical evolution, however, slightly overestimates the fidelity for small $\tau$, and also slightly below each phonon resonance, that is, on the ferromagnetic side of the resonance. Notably, the semiclassical approach works quite well in the glassy regimes above the resonances, where it estimates correctly the regions in which the annealer fails for any annealing time.
As discussed earlier, the failure of the annealing protocol for small $\tau $ is due to non-adiabatic behavior in the fast varying field. The failure near the resonance, though, cannot be fixed by increasing $\tau$, and has its origin in the deviation from of the Dicke dynamics from the effective spin model. Although such deviations are expected on both sides of a phonon resonance, the region of zero fidelity is seen only on the glassy side of each resonance. From that perspective, the size of the spin gap seems to play a role as well, although in this regime we should not compare it to $\hbar/\tau$, but to those spin-phonon energy scales which are neglected in the effective spin model, that is, the first order term in a Magnus expansion, see Ref. [@PhysRevLett.103.120502].
The main advantage of the semiclassical model is that it can easily be applied to larger systems. In Fig. \[fig:SCFide\_na6\], we consider systems of six and eight ions. Notably, a broad region of zero fidelity occurs for eight ions between $\omega_5$ and $\omega_6$. Its origin is unclear to us, and further calculations on the fully quantum evolution would be needed in order to discriminate whether they are true effects or merely calculation artifacts.
![\[fig:SCFide\_na6\] Fidelity of a system of six ions (a) and a system of eight ions (b) as a function of $\omega_{\rm L}$ and $\tau$ using the semiclassical approximation. Final times are $20 \tau$.](Fide_na6i8){width="50.00000%"}
We finish this section by discussing the factor which limits the scalability of the quantum annealer. As seen above, below a critical detuning from the resonance the fidelity drops to zero. This sets a limit to the scalability of the quantum annealer, because, with the number of modes being proportional to the number of ions, the mode spacing decreases when the system size is increased. However, we note that phonon spectra are not equidistant, and within the transverse branch, the phonon spacing is largest at the lower energetic end of the phonon spectrum. Thus, to achieve finite fidelity in an up-scaled system, one may need to operate in the regime of low-energy phonons. The semiclassical estimates in Ref. [@ncomms11524] suggest that the quantum annealing still works in systems with more than twenty ions, presumably large enough to detect quantum speed-up.
Finite temperature effects on the annealing protocol {#sec:CalcsAtTfin}
====================================================
Finite temperature effects are expected to reduce the quality of the annealing protocol, recently, however, thermal effects are shown to aid the annealing protocol for a 16-qubit problem in a superconducting setup [@ncomms2920]. In this section we study the robustness of the annealing protocol when the phonons are initially at finite temperature. To do so, we consider an initial state with phonon mode populations set as follows: we fix the temperature $T$ of the phonons, then, the mean values of the number of phonons for each mode are sampled according to the a bosonic thermal bath probability distribution at $T$. With these initial conditions, the system is then evolved semiclassically according to Eqs. (\[eq:Eqs\_motion\_simpl\]). This process is repeated with different initial values of the population of phonons, sampled appropriately. After the evolution, the statistical moments are calculated in order to infer the thermal properties of the system at the final time.
It should be noted that our dynamical model only captures the coherent Hamiltonian evolution, but no decoherence processes due to interactions with the environment. Thus, in order to account for all thermal effects, heating events, as they occur for instance due to trap inhomogeneties, should be taken into account by considering an increased initial temperature.
Classical thermal phonons
-------------------------
We assume that the initial populations of the phononic modes are determined by a phonon temperature. In the canonical ensemble, the expected value of the number operator of the phonons in the $k$-th mode is, $$\langle \hat{n}_{k} \rangle = \frac{1}{e^{\beta \hbar \omega_k}-1}\,.
\label{eq:n_k}$$ The corresponding Hamiltonian of the symmetrized phonon field is $\hat{\cal H}_{\rm ph} = \hbar \sum_{k} \omega_k\frac{ \hat{a}^{\dagger}_k \hat{a}_k
+ \hat{a}_k \hat{a}^{\dagger}_k}{2}$. The expected value of the annihilation operator $\alpha_k \equiv {\langle\hat{a}_k\rangle}$ is sampled as, $$P\left(\alpha_k \right) = \frac{1}{\pi \langle \hat{n}_{k} \rangle} e^{-\frac{|\alpha_k|^2}{\langle \hat{n}_{k} \rangle}} \,.
\label{eq:p_alpha_k}$$ The complex-valued Gaussian probability distribution function (PDF) for the random variable $\alpha_k$ is a product of two Normal PDFs —one real the other purely imaginary—for the random variables $\Re\left[\alpha_k\right]$ and $\Im\left[\alpha_k\right]$. Both distributions have a mean $\mu=0$ and variance $\sigma^2=\hat{n}_{k}/2$. We thus use for convenience, $$P\left(\alpha_k \right) =
N \left(\Re\left[\alpha_k\right];0,\langle \hat{n}_{k} \rangle/2\right)
N \left(\Im\left[\alpha_k\right];0,\langle \hat{n}_{k} \rangle/2\right)\,,
\label{eq:p_reimalpha_k}$$ being $N \left(x;\mu,\sigma^2\right)$ the Normal PDF of the random variable $x$ with mean $\mu$ and variance $\sigma^2$.
Effects of temperature on the protocol
--------------------------------------
![\[fig:ThermalFideandTotPop\] We plot the fidelity (a,b) and the square of the phonon coherences (c,d), as a function of the initial phonon temperature, as obtained from the semiclassical calculation for a system of four ions. Different lines correspond to different detunings. Panels (a) and (c) consider cases where the second phonon resonance is approached from below (where the magnetic order is ferromagnetic), while panels (b) and (d) consider cases where the same resonance is approached from above (where a glassy regime occurs near the resonance). In our system, the second resonance occurs at a frequency $\omega_2=14332.7$ kHz. All calculations were done for $\tau = 10$ ms. Each point is obtained by sampling over 1,000 runs. The adscribed error is 2 $\sigma$.](output_wL_compar_panel.eps){width="50.00000%"}
To evaluate the effects of temperature on the proposed annealing protocol we will consider different initial temperatures and detunings. In all cases we will fix the decay time $\tau=10$ ms. Fig. \[fig:ThermalFideandTotPop\] shows the fidelity and the total phononic population per mode in the system. In the figure we compare results obtained with several values of $\omega_L$ and a broad range of temperatures of the phonons. Quite generally, panels (a) and (b) show that up to a certain temperature the fidelity is not affected by thermal phonons, but above this temperature the fidelity drops to zero. The value of this temperature strongly depends on the detuning, and decreases by several orders of magnitudes when we change from a far-detuned configuration to a near-resonance scenario. For instance, in the far-detuned regime at $\omega_{\rm L}=\omega_2-902.41$ kHz (solid squares) the critical temperature is of the order $0.1$ K, while close to the resonance at $\omega_{\rm L}=\omega_2-35.33$ kHz (small triangles), the fidelity drop occurs at a temperature of the order $10^{-4}$ K. Such behavior is seen both when we approach the phonon resonance from above (panel (b))or below (panel (a)), but we remind that above the resonance there is a finite region in which the fidelity is zero even at $T=0$, cf. the discussion of Fig.( \[fig:QFide\_na4\]) in Sec. \[secfide\].
Regarding the phonon population, we assume that a reasonable estimate of $\langle \hat{n}_k \rangle$ is given by $|\alpha_k|^2$, as would be the case if the phonon field remains coherent. In Fig. \[fig:ThermalFideandTotPop\](c,d), we may distinguish two different regimes: At low temperatures, i.e. for $T\lesssim 10^{-5}$ K, the final phonon population is dominated by those phonons which are produced by the spin-phonon coupling. The number of those phonons is independent from the temperature, and as shown earlier in Fig. 1(b), such phonons are generated also at $T=0$. The proximity to the resonance induces a larger population of the dominant mode, resulting in a larger value for the population as we approach the resonance. In contrast, for high temperatures, the phonon population is dominated by thermal phonons present already in the beginning of the evolution. In this case, the phonon population is more or less constant during the evolution, and the population number strongly depends on the temperature, according to the initial values from the Boltzmann distribution. As in this case, the time evolution does not noticeably change the phonon distribution, initial and final distribution are very close, and so are initial and final temperature. Between the low- and high- temperature phase, there is a narrow crossover regime, where the number of dynamically generated phonons is similar to the number of thermal phonons. The temperature at which this happens generally depends on the detuning, i.e. on $\omega_{\rm L}$. In all cases, the fidelity drop occurs only in the high-temperature regime, that is, the number of thermal photons must be large compared to the dynamically generated phonons in order to negatively affect the spin evolution.
Thermal tolerance of the Lamb-Dicke regime
------------------------------------------
The Hamiltonian Eq. (\[eq:Hamiltonian\]) describes the trapped ion system when it is in the Lamb-Dicke regime, that is, for $k x \ll 1$. In a harmonic oscillator with frequency $\omega$, we have$\langle x^2 \rangle = \frac{\hbar}{m \omega} \left(\langle n \rangle + \frac{1}{2} \right)$. With $k=\sqrt{2m\omega_{\rm rec}/\hbar}$, and introducing the Lamb-Dicke parameter $\eta=\sqrt{\omega_{\rm rec}/\omega}$, we re-write the Lamb-Dicke condition as $$\begin{aligned}
\eta \sqrt{ 2\langle n \rangle +1 } \ll 1 \,.\end{aligned}$$ All phonon frequencies $\omega$ are of the order of the radial trap frequency, $\omega_{\rm rad}=2\pi \times 2,655$ kHz. In our simulation, the recoil frequency is taken as $\omega_{\rm rec}=2\pi \times 15$ kHz, so we obtain a Lamb-Dicke parameter $\eta \approx 0.075$. Thus, the Lamb-Dicke regime requires $\sqrt{\langle n \rangle +0.5} \ll 10$, which is fulfilled by a phonon occupation $\langle n \rangle \lesssim 1$. From that perspective, we have to disregard those calculations where the phonon population exceeds this number. From Fig. \[fig:ThermalFideandTotPop\](c,d) we find that, for the Rabi frequency we have used, phonon numbers above 1 occur in the thermally dominated regime, independent from the detuning. This regime is characterized by temperatures $>10^{-4}$ K.
Summary and conclusions {#sec:concl}
=======================
We have considered a chain of trapped ions with an internal state (“spin”) coupled to vibrational modes via Raman lasers. The couplings are such that the effective model describing the ions is a long range spin model with tunable, pseudo-random couplings, leading to a spin-glass-like phase. The goal of our approach is the adiabatic distillation of the ground state in the glassy phase starting from a completely paramagnetic state. To this aim we consider the addition of a time-dependent transverse magnetic field. Our procedure goes as follows: At the initial time, the magnetic field is strong enough to ensure the ground state of the spins is a ferromagnetic state, with all spins aligned in the transverse direction. As time evolves we slowly, ideally adiabatically, remove the magnetic field such that the final Hamiltonian is our effective long-range spin model in the spin-glass-like phase.
We have simulated our annealing protocol using the exact evolution by means of a Krylov subspace method which is feasible for a small number of ions. In order to consider larger systems as well as to study the effect of temperature on the time evolution we have developed a semiclassical formalism which ignores the quantum correlations between the ions and the phonons. The quality of this method has been benchmarked by comparing its predictions with the exact evolution for four ions. The semiclassical model is found to provide a very accurate qualitative picture of our proposed method, and allows us to correctly identify the parameter region where the annealing protocol works well. By means of the semiclassical model we have thus extended our study to larger number of ions, providing an accurate picture of the ability of the annealing protocol to find the correct ground state depending on the annealing time.
Finally, the semiclassical model has allowed us to study the robustness of the scheme for initial phonon states at finite temperature. We find that the effect of temperature strongly depends on the detuning from a phonon resonance. While in most configurations, the quantum annealing does not break down within the Lamb-Dicke regime, close to a resonance the situation is different. Here, the fidelity of the annealing may drop even before the Lamb-Dicke limit is reached, see Fig. \[fig:ThermalFideandTotPop\]. Thus, while state-of-art spin model simulations which are carried out far off any phonon resonance (e.g. Refs. [@PhysRevLett.119.080501; @Zhang2017]) require only cooling to the Lamb-Dicke limit, quantum annealing in the interesting glassy regime requires more cooling. Accordingly, our finding motivates the development of new, more efficient cooling techniques, as for instance cooling based on electromagnetically-induced transparency [@PhysRevLett.85.5547; @PhysRevLett.110.153002; @PhysRevA.93.053401], which is very well suited to simultaneously achieve low populations in all radial modes.
This work has been funded by EU grants (EQuaM (FP7-ICT-2013-C No. 323714), OSYRIS (ERC-2013-AdG No. 339106), SIQS (FP7-ICT-2011-9 No. 600645), and QUIC (H2020-FETPROACT-2014 No. 641122)), Spanish Ministerio de Economía y Competitividad grants (Severo Ochoa (SEV-2015-0522), FOQUS (FIS2013-46768-P), FIS2014-54672-P and FISICATEAMO (FIS2016-79508-P)), Generalitat de Catalunya (2014 SGR 401, 2014 SGR 874, and CERCA program), and Fundació Cellex. TG acknowledges support of the NSF through the PFC@JQI.
:
![\[fig:optim\_bias\] Separation time (a)) and Fidelity (b)) of systems with $6$ spins for several values of $\varepsilon$ and $\tau$, with initial populations of phononic modes set to $0$. Waiting time is fixed to $20 \tau$ and $\omega_L = \omega_{\rm N}-1500$ kHz.](optim_bias){width="50.00000%"}
Robustness of the semiclassical calculation {#app:Integrationtests}
===========================================
One can test the quality of the integration method by checking whether the constants of motion are conserved. In time-independent systems, the total energy of the system is ususally the conserved quantity which can be computs most easily. Unfortunately, since the Hamiltonian in Eq. \[eq:Hamiltonian\] does not commute with itself at different times due to the transverse magnetic field in the annealing term, the total energy of the system is not conserved. Furthermore, it is challenging to find an analytical expression for an alternative conserved quantity given the infinitely non-commutativity of the Pauli matrices algebra. We have tested the integration method checking its time reversibility. The latter test has given relative differences below $10^-7$, which is within the range of the computation precision.
Optimal bias for the exponential annealing function {#sec:optim}
===================================================
Although the only function of the bias potential $\epsilon$ is to break $\mathcal{Z}_2$ symmetry in the target Hamiltonian, it turns out that the value and position of the bias have a non-neglible effect on the outcome of the annealing process. Here, we investigate which values of the bias $\epsilon$ and the annealing parameter $\tau$ minimize the separation time. As seen in Fig. \[fig:optim\_bias\] a), the separation time is minimized for smaller values of $\tau$, at any value of the bias larger than Hz. However, small $\tau$ are known to affect negatively the fidelity. As we see from Fig. \[fig:optim\_bias\] b), there is, even for decay times as short as a few ms, a range of bias potentials (roughly between 1 kHz and 10 kHz), where the fidelity gets large. Thus, this range defines the optimal choice for $\epsilon$, which we have also used in our calculations.
We also note that the magnitude of the bias provides a bound for the maximum absolute value of $\langle \sigma_{x} \rangle$, that is, the spin expectation on the biased site at the end of the annealing depends on the strength of the bias potential. For a weak bias, this spin will deviate only weakly from zero, limiting the overall fidelity.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Naoki Kobayashi\
The University of Tokyo
- |
Ugo Dal Lago\
University of Bologna
- |
Charles Grellois\
Aix-Marseille University
- Naoki Kobayashi
- Ugo Dal Lago
- Charles Grellois
title:
- |
On the Termination Problem\
for Probabilistic Higher-Order Recursive Programs
- |
On the Termination Problem\
for Probabilistic Higher-Order Recursive Programs
---
Introduction {#sec:intro}
============
\[SEC:INTRO\]
Probabilistic Higher-Order Recursion Schemes () and Termination Probabilities {#sec:problem}
=============================================================================
\[SEC:PROBLEM\]
Undecidability of Almost Sure Termination of Order-2 {#sec:undecidability}
=====================================================
\[SEC:UNDECIDABILITY\]
Fixpoint Characterization of Termination Probability {#sec:fixpoint}
====================================================
\[SEC:FIXPOINT\]
Computing Upper-Bounds of Termination Probability {#sec:upperbound}
=================================================
\[SEC:UPPERBOUND\]
Preliminary Experiments {#sec:exp}
=======================
Related Work {#sec:related}
============
Conclusion {#sec:conc}
==========
Acknowledgments {#acknowledgments .unnumbered}
---------------
We would like to thank Kazuyuki Asada and Takeshi Tsukada for discussions on the topic. This work was supported by JSPS <span style="font-variant:small-caps;">KAKENHI</span> Grant Number JP15H05706 and by ANR <span style="font-variant:small-caps;">Elica</span> Grant Number ANR-14-CE25-0005.
[10]{}
M. Avanzini, U. D. Lago, and A. Yamada. On probabilistic term rewriting. In J. P. Gallagher and M. Sulzmann, editors, [*Functional and Logic Programming - 14th International Symposium, [FLOPS]{} 2018, Nagoya, Japan, May 9-11, 2018, Proceedings*]{}, volume 10818 of [*Lecture Notes in Computer Science*]{}, pages 132–148. Springer, 2018.
G. Bacci, R. Furber, D. Kozen, R. Mardare, P. Panangaden, and D. Scott. Boolean-valued semantics for the stochastic [$\lambda$]{}-calculus. In [*Proceedings of the 33rd Annual [ACM/IEEE]{} Symposium on Logic in Computer Science, [LICS]{} 2018, Oxford, UK, July 09-12, 2018*]{}, pages 669–678, 2018.
C. Baier and J.-P. Katoen. . The MIT Press, 2008.
G. Barthe, M. J. Frade, E. Gim[é]{}nez, L. Pinto, and T. Uustalu. Type-based termination of recursive definitions. , 14(1):97–141, 2004.
O. Bournez and F. Garnier. Proving positive almost-sure termination. In J. Giesl, editor, [*Term Rewriting and Applications, 16th International Conference, [RTA]{} 2005, Nara, Japan, April 19-21, 2005, Proceedings*]{}, volume 3467 of [*Lecture Notes in Computer Science*]{}, pages 323–337. Springer, 2005.
T. Br[á]{}zdil, V. Brozek, V. Forejt, and A. Kucera. Branching-time model-checking of probabilistic pushdown automata. , 80(1):139–156, 2014.
T. Br[á]{}zdil, J. Esparza, S. Kiefer, and A. Kucera. Analyzing probabilistic pushdown automata. , 43(2):124–163, 2013.
F. Breuvart and U. Dal Lago. On intersection types and probabilistic lambda calculi. In [*Proceedings of PPDP 2018*]{}. ACM, 2018. To appear.
C. H. Broadbent and N. Kobayashi. Saturation-based model checking of higher-order recursion schemes. In [*Proceedings of CSL 2013*]{}, volume 23 of [*LIPIcs*]{}, pages 129–148, 2013.
C. Calude. . Springer-Verlag, Berlin, Heidelberg, 2nd edition, 2002.
A. Chakarov and S. Sankaranarayanan. Probabilistic program analysis with martingales. In N. Sharygina and H. Veith, editors, [*Computer Aided Verification*]{}, pages 511–526, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg.
K. Chatterjee, P. Novotn[ý]{}, and D. Zikelic. Stochastic invariants for probabilistic termination. In G. Castagna and A. D. Gordon, editors, [*Proceedings of the 44th [ACM]{} [SIGPLAN]{} Symposium on Principles of Programming Languages, [POPL]{} 2017, Paris, France, January 18-20, 2017*]{}, pages 145–160. [ACM]{}, 2017.
E. M. Clarke, T. A. Henzinger, H. Veith, and R. Bloem, editors. . Springer, 2018.
P. Cousot. Types as abstract interpretations. In P. Lee, F. Henglein, and N. D. Jones, editors, [*Conference Record of POPL’97: The 24th [ACM]{} [SIGPLAN-SIGACT]{} Symposium on Principles of Programming Languages, Papers Presented at the Symposium, Paris, France, 15-17 January 1997*]{}, pages 316–331. [ACM]{} Press, 1997.
R. Crubill[é]{} and U. Dal Lago. On probabilistic applicative bisimulation and call-by-value [$\lambda$]{}-calculi. In Z. Shao, editor, [*Programming Languages and Systems - 23rd European Symposium on Programming, [ESOP]{} 2014, Held as Part of the European Joint Conferences on Theory and Practice of Software, [ETAPS]{} 2014, Grenoble, France, April 5-13, 2014, Proceedings*]{}, volume 8410 of [*Lecture Notes in Computer Science*]{}, pages 209–228. Springer, 2014.
U. Dal Lago and C. Grellois. Probabilistic termination by monadic affine sized typing. In [*Programming Languages and Systems - 26th European Symposium on Programming, [ESOP]{} 2017, Held as Part of the European Joint Conferences on Theory and Practice of Software, [ETAPS]{} 2017, Uppsala, Sweden, April 22-29, 2017, Proceedings*]{}, pages 393–419, 2017.
U. Dal Lago, D. Sangiorgi, and M. Alberti. On coinductive equivalences for higher-order probabilistic functional programs. In Jagannathan and Sewell [@DBLP:conf/popl/2014], pages 297–308.
U. Dal Lago and M. Zorzi. Probabilistic operational semantics for the lambda calculus. , 46(3):413–450, 2012.
V. Danos and R. Harmer. Probabilistic game semantics. , 3(3):359–382, 2002.
K. de Leeuw, E. F. Moore, C. E. Shannon, and N. Shapiro. Computability by probabilistic machines. , 34:183–212, 1956.
T. Ehrhard, C. Tasson, and M. Pagani. Probabilistic coherence spaces are fully abstract for probabilistic [PCF]{}. In Jagannathan and Sewell [@DBLP:conf/popl/2014], pages 309–320.
J. Esparza, A. Gaiser, and S. Kiefer. Proving termination of probabilistic programs using patterns. In P. Madhusudan and S. A. Seshia, editors, [*Computer Aided Verification*]{}, pages 123–138, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg.
K. Etessami and M. Yannakakis. Recursive markov chains, stochastic grammars, and monotone systems of nonlinear equations. , 56(1):1:1–1:66, 2009.
K. Etessami and M. Yannakakis. Model checking of recursive probabilistic systems. , 13(2):12:1–12:40, 2012.
K. Etessami and M. Yannakakis. Recursive markov decision processes and recursive stochastic games. , 62(2):11:1–11:69, 2015.
L. M. F. Fioriti and H. Hermanns. Probabilistic termination: Soundness, completeness, and compositionality. In S. K. Rajamani and D. Walker, editors, [*Proceedings of the 42nd Annual [ACM]{} [SIGPLAN-SIGACT]{} Symposium on Principles of Programming Languages, [POPL]{} 2015, Mumbai, India, January 15-17, 2015*]{}, pages 489–501. [ACM]{}, 2015.
H. Gimbert and Y. Oualhadj. Probabilistic automata on finite words: Decidable and undecidable problems. In S. Abramsky, C. Gavoille, C. Kirchner, F. M. auf der Heide, and P. G. Spirakis, editors, [*Automata, Languages and Programming, 37th International Colloquium, [ICALP]{} 2010, Bordeaux, France, July 6-10, 2010, Proceedings, Part [II]{}*]{}, volume 6199 of [*Lecture Notes in Computer Science*]{}, pages 527–538. Springer, 2010.
S. Goldwasser and S. Micali. Probabilistic encryption. , 28(2):270–299, 1984.
N. D. Goodman, V. K. Mansinghka, D. M. Roy, K. Bonawitz, and J. B. Tenenbaum. Church: A language for generative models. In [*In UAI*]{}, pages 220–229, 2008.
C. Grellois and P. Melli[è]{}s. Finitary semantics of linear logic and higher-order model-checking. In Italiano et al. [@DBLP:conf/mfcs/2015-1], pages 256–268.
C. Grellois and P. Melli[è]{}s. Relational semantics of linear logic and higher-order model checking. In [*24th [EACSL]{} Annual Conference on Computer Science Logic, [CSL]{} 2015, September 7-10, 2015, Berlin, Germany*]{}, volume 41 of [ *LIPIcs*]{}, pages 260–276. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015.
M. Hague, A. Murawski, C.-H. L. Ong, and O. Serre. Collapsible pushdown automata and recursion schemes. In [*Proceedings of 23rd Annual IEEE Symposium on Logic in Computer Science*]{}, pages 452–461. IEEE Computer Society, 2008.
C. Heunen, O. Kammar, S. Staton, and H. Yang. A convenient category for higher-order probability theory. In [*32nd Annual [ACM/IEEE]{} Symposium on Logic in Computer Science, [LICS]{} 2017, Reykjavik, Iceland, June 20-23, 2017*]{}, pages 1–12. [IEEE]{} Computer Society, 2017.
J. Hromkovic. . Texts in Theoretical Computer Science. An [EATCS]{} Series. Springer, 2005.
J. Hughes, L. Pareto, and A. Sabry. Proving the correctness of reactive systems using sized types. In H. Boehm and G. L. S. Jr., editors, [*[POPL]{}’96*]{}, pages 410–423. [ACM]{} Press, 1996.
G. F. Italiano, G. Pighizzini, and D. Sannella, editors. , volume 9234 of [*Lecture Notes in Computer Science*]{}. Springer, 2015.
S. Jagannathan and P. Sewell, editors. . [ACM]{}, 2014.
C. Jones and G. D. Plotkin. A probabilistic powerdomain of evaluations. In [*Proceedings of the Fourth Annual Symposium on Logic in Computer Science [(LICS]{} ’89), Pacific Grove, California, USA, June 5-8, 1989*]{}, pages 186–195. [IEEE]{} Computer Society, 1989.
A. Jung and R. Tix. The troublesome probabilistic powerdomain. , 13:70 – 91, 1998. Comprox III, Third Workshop on Computation and Approximation.
B. L. Kaminski and J. Katoen. On the hardness of almost-sure termination. In Italiano et al. [@DBLP:conf/mfcs/2015-1], pages 307–318.
T. Knapik, D. Niwinski, and P. Urzyczyn. Higher-order pushdown trees are easy. In [*FoSSaCS 2002*]{}, volume 2303 of [*LNCS*]{}, pages 205–222. Springer, 2002.
N. Kobayashi. Model-checking higher-order functions. In [*Proceedings of PPDP 2009*]{}, pages 25–36. ACM Press, 2009.
N. Kobayashi. Types and higher-order recursion schemes for verification of higher-order programs. In [*Proceedings of POPL*]{}, pages 416–428. ACM Press, 2009.
N. Kobayashi. Model checking higher-order programs. , 60(3), 2013.
N. Kobayashi. Inclusion between the frontier language of a non-deterministic recursive program scheme and the dyck language is undecidable. submitted to a TCS special issue, 2018.
N. Kobayashi and C.-H. L. Ong. A type system equivalent to the modal mu-calculus model checking of higher-order recursion schemes. In [*Proceedings of LICS 2009*]{}, pages 179–188. IEEE Computer Society Press, 2009.
N. Kobayashi and C.-H. L. Ong. Complexity of model checking recursion schemes for fragments of the modal mu-calculus. , 7(4), 2011.
N. Kobayashi, R. Sato, and H. Unno. Predicate abstraction and [CEGAR]{} for higher-order model checking. In [*Proceedings of [[ACM]{} SIGPLAN Conference on Programming Language Design and Implementation ([PLDI]{})]{}*]{}, pages 222–233. ACM Press, 2011.
D. Koller and N. Friedman. . The MIT Press, 2009.
Y. V. Matiyasevich. . The MIT Press, 1993.
A. McIver, C. Morgan, B. L. Kaminski, and J. Katoen. A new proof rule for almost-sure termination. , 2([POPL]{}):33:1–33:28, 2018.
A. Mista, A. Russo, and J. Hughes. Branching processes for quickcheck generators. , abs/1808.01520, 2018.
R. Motwani and P. Raghavan. . Cambridge University Press, New York, NY, USA, 1995.
A. S. Murawski and J. Ouaknine. On probabilistic program equivalence and refinement. In [*[CONCUR]{} 2005 - Concurrency Theory, 16th International Conference, [CONCUR]{} 2005, San Francisco, CA, USA, August 23-26, 2005, Proceedings*]{}, pages 156–170, 2005.
C.-H. L. Ong. On model-checking trees generated by higher-order recursion schemes. In [*LICS 2006*]{}, pages 81–90. IEEE Computer Society Press, 2006.
S. M. Poulding and R. Feldt. Automated random testing in multiple dispatch languages. In [*2017 [IEEE]{} International Conference on Software Testing, Verification and Validation, [ICST]{} 2017, Tokyo, Japan, March 13-17, 2017*]{}, pages 333–344. [IEEE]{} Computer Society, 2017.
S. Ramsay, R. Neatherway, and C.-H. L. Ong. An abstraction refinement approach to higher-order model checking. In [*Proceedings of POPL 2014*]{}, pages 61–72. ACM, 2014.
N. Saheb[-]{}Djahromi. Probabilistic [LCF]{}. In J. Winkowski, editor, [*Mathematical Foundations of Computer Science 1978, Proceedings, 7th Symposium, Zakopane, Poland, September 4-8, 1978*]{}, volume 64 of [*Lecture Notes in Computer Science*]{}, pages 442–451. Springer, 1978.
S. Salvati and I. Walukiewicz. Krivine machines and higher-order schemes. In [*Proceedings of ICALP 2011*]{}, volume 6756 of [*Lecture Notes in Computer Science*]{}, pages 162–173. Springer, 2011.
D. Sangiorgi and V. Vignudelli. Environmental bisimulations for probabilistic higher-order languages. In [*Proceedings of the 43rd Annual [ACM]{} [SIGPLAN-SIGACT]{} Symposium on Principles of Programming Languages, [POPL]{} 2016, St. Petersburg, FL, USA, January 20 - 22, 2016*]{}, pages 595–607, 2016.
E. S. Santos. Probabilistic turing machines and computability. , 22(3):704–710, 1969.
R. Statman. On the lambda*Y* calculus. , 130(1-3):325–337, 2004.
S. Staton, H. Yang, F. D. Wood, C. Heunen, and O. Kammar. Semantics for probabilistic programming: higher-order functions, continuous distributions, and soft constraints. In [*Proceedings of the 31st Annual [ACM/IEEE]{} Symposium on Logic in Computer Science, [LICS]{} ’16, New York, NY, USA, July 5-8, 2016*]{}, pages 525–534, 2016.
T. Tsukada and C. L. Ong. Compositional higher-order model checking via *[$\omega$]{}*-regular games over b[ö]{}hm trees. In [*Proceedings of [CSL-LICS]{} ’14*]{}, pages 78:1–78:10. [ACM]{}, 2014.
F. Wood, J. W. Meent, and V. Mansinghka. . In S. Kaski and J. Corander, editors, [*Proceedings of the Seventeenth International Conference on Artificial Intelligence and Statistics*]{}, volume 33 of [*Proceedings of Machine Learning Research*]{}, pages 1024–1032, Reykjavik, Iceland, 22–25 Apr 2014. PMLR.
M. Yannakakis and K. Etessami. Checking [LTL]{} properties of recursive markov chains. In [*Second International Conference on the Quantitative Evaluaiton of Systems [(QEST]{} 2005), 19-22 September 2005, Torino, Italy*]{}, pages 155–165. [IEEE]{} Computer Society, 2005.
Appendix {#appendix .unnumbered}
========
Additional Materials for Section \[sec:undecidability\] {#sec:app-section3}
=======================================================
Additional materials for Section \[sec:upperbound\] {#sec:app-sec5}
===================================================
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We derive the low-energy expansion of $\left( Z\alpha\right) ^{2}$ and $\left( Z\alpha\right) ^{4}$ terms of the polarization operator in the Coulomb field. Physical applications such as the low-energy Delbrück scattering and “magnetic loop” contribution to the $g$ factor of the bound electron are considered.'
address:
- 'Budker Institute of Nuclear Physics, 630090, Novosibirsk, Russia'
- 'Novosibirsk State University, 630090, Novosibirsk, Russia'
author:
- 'G.G.Kirilin'
- 'R.N. Lee'
title: '$(Z\alpha)^{4}$ order of the polarization operator in Coulomb field at low energy'
---
polarization operator ,multiloop calculations ,quantum electrodynamics ,$g$ factor 31.30.jf ,12.20.Ds ,31.15.xp ,31.30.js
Introduction
============
One of the predictions of the quantum field theory is a vacuum polarization by an external field. An important case thoroughly studied both experimentally and theoretically is the vacuum polarization effects in atomic field. Methods used for the study of this effect essentially depend on the nuclear charge $Z\left\vert e\right\vert $. At low $Z$, the perturbation theory with respect to $Z\alpha$ is applicable ($\alpha
=e^{2}=1/137$ is the fine structure constant, $\hbar=c=1$). At high $Z$, the interaction with the external field should be taken into account exactly, which can be done with the help of the electron Green function in this field. This approach often requires quite involved numerical calculations, which usually fail to give the results for low $Z$. Thus, the two approaches tend to be complementary. Usually, the perturbative calculations of vacuum polarization effect are limited by the leading order, since the first nonvanishing correction involves two more loops. Nowadays, the modern methods of calculation of the multiloop integrals are sufficiently powerful for the calculation of higher orders in $Z\alpha$. It provides a possibility to compare the results of these approaches.
One of the basic nonlinear QED processes in the atomic field is the Delbrück scattering [@Delbruck1933], the scattering of the photon in the Coulomb field due to the vacuum polarization. The amplitude of this process in the Born approximation has been obtained long ago for arbitrary energies in Ref.[@Costantini:1971cj]. At high energies and small scattering angles, when the quasiclassical approximation is valid, the amplitude is known exactly in $Z\alpha$, see Refs. [@Cheng1969a; @Cheng1970; @Cheng1972; @Milstein1983; @Milstein1983a; @Lee1999]. Recently in Ref.[@Kirilin2008], the Delbrück amplitude has been calculated numerically exactly in the parameter $Z\alpha$ at low energies. It was shown that the contribution of the higher orders (Coulomb corrections) to the amplitude can be well fitted by the polynomial $C_{4}(Z\alpha)^{4}%
+C_{6}(Z\alpha)^{6}$. The calculation of $(Z\alpha)^{4}$ term in perturbation theory would provide the independent check of the result of Ref.[@Kirilin2008].
In present paper, we consider the polarization operator $\Pi^{\mu\nu}\left(
\omega,\mathbf{k},\mathbf{q}\right) $ in the Coulomb field for small external momenta$$\omega\sim\left\vert \mathbf{k}\right\vert \sim\left\vert \mathbf{q}%
\right\vert \sim\lambda m,$$ where $\lambda$ is a dimensionless small parameter,$$\lambda\ll1\,. \label{eq:lambda}%$$ We calculate the expansion of the polarization operator in the Coulomb field in $\lambda$ and $Z\alpha$ up to the order $\lambda^{4}\left( Z\alpha\right)
^{4}$. The low-energy Delbrück scattering amplitude is readily expressed in terms of this operator. This polarization operator is also an essential ingredient of calculations of different physical observables in atoms, like Lamb shift and magnetic moment of the bound particle.
The polarization operator in the external Coulomb field is determined as follows: $$\Pi^{\mu\nu}\left( \omega,\mathbf{k},\mathbf{q}\right) =4\pi ie^{2}\int
d\mathbf{x}d\mathbf{y\,}dt e^{-i\omega t+i\mathbf{kx}-i\mathbf{q y}%
}\left\langle \mathrm{vac}\right\vert \mathrm{T}J^{\mu}\left( t,\mathbf{x}%
\right) J^{\nu}\left( 0,\mathbf{y}\right) \left\vert \mathrm{vac}%
\right\rangle \,,\label{eq:PmunuDefinition}%$$ where $J^{\mu}=\bar{\psi}\gamma^\mu\psi$ is the electron current and the state $\left\vert \mathrm{vac}\right\rangle $ corresponds to the vacuum state in the presence of the Coulomb potential $Z\left\vert e\right\vert /r$. In $e^{2}$ order, we have $$\Pi^{\mu\nu}\left( \omega,\mathbf{k},\mathbf{q}\right) =4\pi ie^{2}\int
\frac{d\varepsilon}{2\pi}d\mathbf{x}d\mathbf{y\,} e^{i\mathbf{kx}%
-i\mathbf{qy}}\mathrm{Tr}\left[ \gamma^{\mu}G\left( \mathbf{x}%
,\mathbf{y}|\varepsilon-\omega\right) \gamma^{\nu}G\left( \mathbf{y,x}%
|\varepsilon\right) \right] \,,\label{eq:PiviaG}%$$ where $G\left( \mathbf{y,x}|\varepsilon\right) $ is the Green function of the electron in the Coulomb field. Due to the gauge invariance, the polarization operator obeys the constraints$$k_{\mu}\Pi^{\mu\nu}\left( \omega,\mathbf{k},\mathbf{q}\right) =q_{\nu}%
\Pi^{\mu\nu}\left( \omega,\mathbf{k},\mathbf{q}\right) =0,\label{e1:10}%$$ where $k^{0}=q^{0}=\omega$. Using these constraints, we can express $\Pi
^{\mu\nu}$ via five independent tensor structures, which we choose as follows $$\begin{aligned}
m^{3}\Pi^{\mu\nu} & =f_{1}\left( g^{\mu\nu}\,k\cdot q-q^{\mu}k^{\nu}\right)
-f_{3}\epsilon^{\mu\alpha\beta\gamma}n_{\alpha}\epsilon^{\nu\rho\sigma\tau
}n_{\rho}\,\frac{k_{\beta}\,\left( k-q\right) _{\gamma}\left( k-q\right)
_{\sigma}q_{\tau}}{\left( k-q\right) ^{2}}\nonumber\\
& +\left( n^{\mu}k^{\alpha}-\omega g^{\mu\alpha}\right) \left( q^{\beta
}n^{\nu}-\omega g^{\beta\nu}\right) \left[ f_{2}g_{\alpha\beta}+f_{4}%
\,\frac{k_{\alpha}q_{\beta}}{\omega^{2}}-f_{5}\frac{\left( k-q\right)
_{\alpha}\left( k-q\right)_{\beta}}{(k-q)^{2}}\right]
,\label{eq:Parametrization}%\end{aligned}$$ where $n=\left( 1,\mathbf{0}\right) $ and $f_{i}$ are some scalar functions of $\omega$, $\mathbf{k}$ and $\mathbf{q}$.
It is important to note that the low-energy expansion of $\Pi^{\mu\nu}$ itself and the functions $f_{1-5}$ is not reduced to the multiple Taylor expansion in $k^{i},q^{i}$ and $\omega$, i.e., $f_{1-5}$ are nonanalytic functions of the external momenta. Nevertheless, the expansion in powers of the parameter $\lambda$ from Eq.(\[eq:lambda\]) is still possible. Different terms of the expansion come from one of two different regions of integration. We separate the contributions of these regions, using the dimensional regularization, in spirit of Refs. [@Beneke:1997zp; @Smirnov:1998vk]. The methods of separation and calculation of the contributions of these two regions are given in Sections \[Sec:Separation\] and \[Sec:Calculation\], respectively. The results and conclusions are presented in Section \[Sec:Results\].
Separation of the contributions of soft and hard regions\[Sec:Separation\]
==========================================================================
In order to demonstrate our method, let us consider the behaviour of the integral $$\begin{gathered}
J^{\left( \mathcal{D}\right) }\left( \mathbf{q}^{2}\right) =\int
\frac{d^{\mathcal{D}}\Delta d^{\mathcal{D}}p}{\pi^{\mathcal{D}}}\,j\left(
\mathbf{p},\boldsymbol{\Delta},\mathbf{q}\right) \\
=\int\frac{\pi
^{-\mathcal{D}}d^{\mathcal{D}}\Delta d^{\mathcal{D}}p}{\mathbf{\Delta}%
^{2}\left( \mathbf{q}+\boldsymbol{\boldsymbol{\Delta}}\right) ^{2}\left(
\mathbf{p}^{2}+1\right) \left[ \left( \mathbf{p}+\mathbf{q}\right)
^{2}+1\right] \left[ \left( \mathbf{p}-\boldsymbol{\boldsymbol{\Delta}%
}\right) ^{2}+1\right] },\end{gathered}$$ at $\mathbf{q}^{2}\ll1$. The corresponding diagram is depicted in Fig.\[Fig:Example\].
\[ptb\]
[Example.eps]{}
The small-$\mathbf{q}^{2}$ expansion of this integral, obtained in Ref.[@Broadhurst:1993mw], has the form:$$J^{\left( \mathcal{D}\right) }\left( \mathbf{q}^{2}\right) =\sum
_{n=0}^{\infty}C_{n}\left( \mathcal{D}\right) \,\left( -\mathbf{q}%
^{2}\right) ^{n}+\left( \mathbf{q}^{2}\right) ^{\frac{\mathcal{D}}{2}%
-2}\sum_{n=0}^{\infty}D_{n}\left( \mathcal{D}\right) \,\left(
-\mathbf{q}^{2}\right) ^{n}\,, \label{e2:3}%$$$$\begin{aligned}
C_{n}\left( \mathcal{D}\right) & =\frac{\Gamma\left( n+3-\mathcal{D}%
/2\right) }{\left( n+3-\mathcal{D}/2\right) \left( \mathcal{D}-3\right)
}\left[ \Gamma\left( \mathcal{D}/2-1\right) \frac{\Gamma\left(
n+5-\mathcal{D}\right) \Gamma\left( n+3-\mathcal{D}/2\right) }%
{\Gamma\left( n+\mathcal{D}/2\right) \Gamma\left( 2n+6-\mathcal{D}\right)
}\right. \nonumber\\
& \left. -\frac{\Gamma\left( n+2\right) \Gamma\left( 2-\mathcal{D}%
/2\right) }{\Gamma\left( 2n+3\right) }\right] ,\end{aligned}$$$$D_{n}\left( \mathcal{D}\right) =\Gamma\left( 2-\mathcal{D}/2\right)
\Gamma\left( \mathcal{D}/2-1\right) ^{2}\frac{\Gamma\left( n+1\right)
\Gamma\left( n+3-\mathcal{D}/2\right) }{\Gamma(\mathcal{D}-2)\Gamma\left(
2n+3\right) }\,.$$ The representation (\[e2:3\]) determines $J^{\left( \mathcal{D}\right)
}\left( \mathbf{q}^{2}\right) $ at $0<\mathbf{q}^{2}<4.$ The limit $\mathbf{q}^{2}\rightarrow0$ essentially depends on $\mathcal{D}$. For $\mathcal{D}>4$, this limit is equal to $C_{0}\left( \mathcal{D}\right) $ and can be considered as the value of the function $J^{\left( \mathcal{D}%
\right) }$ at $\mathbf{q}^{2}=0.$ For $\mathcal{D}<4$, there is no finite $\mathbf{q}^{2}\rightarrow0$ limit of the representation (\[e2:3\]). However, the value $J^{\left( \mathcal{D}\right) }\left( 0\right) $ defined via the analytic continuation with respect to $\mathcal{D}$ from the region $\mathcal{D}>4$ is still $C_{0}\left( \mathcal{D}\right) .$ In other words, the limit $\mathbf{q}^{2}\rightarrow0$ is not commuting with the analytic continuation with respect to $\mathcal{D}$. The same claim is valid for the derivatives of $J^{\left( \mathcal{D}\right) }\left( \mathbf{q}%
^{2}\right) $ with respect to $\mathbf{q}^{2}$. As a consequence, the integrand expanded in $\mathbf{q}$ gives only the terms $\propto\left(
\mathbf{q}^{2}\right) ^{n}$ after the integration within the dimensional regularization. These terms correspond to the first sum in the right-hand side of Eq.(\[e2:3\]): $$\int\frac{d^{\mathcal{D}}\Delta\,d^{\mathcal{D}}p}{\pi^{\mathcal{D}}}\left[
\sum_{n=0}^{\infty}\frac{q^{i_{1}}\ldots q^{i_{n}}}{n!}\left. \frac
{\partial^{n}\,j\left( \mathbf{p},\boldsymbol{\Delta},\mathbf{q}\right)
}{\partial q^{i_{1}}\ldots\partial q^{i_{n}}}\right\vert _{\mathbf{q}%
=0}\right] =\sum_{n=0}^{\infty}C_{n}\left( \mathcal{D}\right) \,\left(
-\frac{\mathbf{q}^{2}}{4}\right) ^{n}.$$ Note that the expansion of the massless propagators is valid only in the region $\Delta\gg q$ (hard region). In order to obtain the rest terms of the expansion (\[e2:3\]) one has to determine the contribution of the region $\Delta\sim q$ (soft region). To separate these contributions, we use the following trick. In the soft ** region the massive propagators can be expanded in both $\mathbf{q}$ and $\boldsymbol{\Delta}$. Let us truncate this expansion at some fixed order $N$ in $\mathbf{q}$ and $\boldsymbol{\Delta}$: $$\begin{aligned}
j_{\mathrm{soft}}\left( N,\mathbf{p},\boldsymbol{\Delta},\mathbf{q}\right)
& =\sum_{n=0}^{N}j_{\mathrm{soft}}^{(n)}\left( \mathbf{p},\boldsymbol{\Delta
},\mathbf{q}\right) ,\\
j_{\mathrm{soft}}^{(n)}\left( \mathbf{p},\boldsymbol{\Delta},\mathbf{q}%
\right) & =\left. \frac{\partial^{n}\tau^{4}j\left( \mathbf{p}%
,\tau\boldsymbol{\Delta},\tau\mathbf{q}\right) }{n!\partial\tau^{n}%
}\right\vert _{\tau=0}\,,\text{ so that }j_{\mathrm{soft}}^{(n)}=O\left(
q^{n-4}\right) .\nonumber\end{aligned}$$ Now we can identically rewrite $J^{\left( \mathcal{D}\right) }\left(
\mathbf{q}^{2}\right) $ as $$\begin{aligned}
& J^{\left( \mathcal{D}\right) }\left( \mathbf{q}^{2}\right)
=J_{\mathrm{hard}}^{\left( \mathcal{D}\right) }\left( \mathbf{q}%
^{2}\right) +J_{\mathrm{soft}}^{\left( \mathcal{D}\right) }\left(
\mathbf{q}^{2}\right) \nonumber\\
& =\int\frac{d^{\mathcal{D}}\Delta\,d^{\mathcal{D}}p}{\pi^{\mathcal{D}}%
}\,\left[ j\left( \mathbf{p},\boldsymbol{\Delta},\mathbf{q}\right)
-\,j_{\mathrm{soft}}\left( N,\mathbf{p},\boldsymbol{\Delta},\mathbf{q}%
\right) \right] +\int\frac{d^{\mathcal{D}}\Delta\,d^{\mathcal{D}}p}%
{\pi^{\mathcal{D}}}\,j_{\mathrm{soft}}\left( N,\mathbf{p},\boldsymbol{\Delta
},\mathbf{q}\right) . \label{eq:add and subtract}%\end{aligned}$$ The contribution of the soft region in the first term in Eq.(\[eq:add and subtract\]) is suppressed as $q^{\mathcal{D+}N-3}$, thus the integral of the difference $j\left( \mathbf{p},\boldsymbol{\Delta}%
,\mathbf{q}\right) -\,j_{\mathrm{soft}}\left( N,\mathbf{p}%
,\boldsymbol{\Delta},\mathbf{q}\right) $ is determined by the region $\Delta\sim1\gg q$ up to $O\left( q^{2\left\lfloor \left( \mathcal{D+}%
N-3\right) /2\right\rfloor }\right) $ term. In fact, the expansion of $\,j_{\mathrm{soft}}\left( N,\mathbf{p},\boldsymbol{\Delta},\mathbf{q}%
\right) $ in $\mathbf{q}$ gives scaleless functions of $\boldsymbol{\Delta}$, which vanish after the integration in the dimensional regularization. Finally, we have $$\begin{aligned}
J^{\left( \mathcal{D}\right) }\left( \mathbf{q}^{2}\right) &
=J_{\mathrm{hard}}^{\left( \mathcal{D}\right) }\left( \mathbf{q}%
^{2}\right) +J_{\mathrm{soft}}^{\left( \mathcal{D}\right) }\left(
\mathbf{q}^{2}\right) \nonumber\\
& =\sum_{n=0}^{\infty}\int\frac{d^{\mathcal{D}}\Delta\,d^{\mathcal{D}}p}%
{\pi^{\mathcal{D}}}\,j_{\mathrm{hard}}^{(n)}\left( \mathbf{p}%
,\boldsymbol{\Delta},\mathbf{q}\right) +\sum_{n=0}^{\infty}\int
\frac{d^{\mathcal{D}}\Delta\,d^{\mathcal{D}}p}{\pi^{\mathcal{D}}%
}\,j_{\mathrm{soft}}^{(n)}\left( \mathbf{p},\boldsymbol{\Delta}%
,\mathbf{q}\right) ,\\
j_{\mathrm{hard}}^{(n)}\left( \mathbf{p},\boldsymbol{\Delta},\mathbf{q}%
\right) & =\left. \frac{\partial^{n}j\left( \mathbf{p},\boldsymbol{\Delta
},\tau\mathbf{q}\right) }{n!\partial\tau^{n}}\right\vert _{\tau
=0},~j_{\mathrm{soft}}^{(n)}\left( \mathbf{p},\boldsymbol{\Delta}%
,\mathbf{q}\right) =\left. \frac{\partial^{n}\tau^{4}j\left( \mathbf{p}%
,\tau\boldsymbol{\Delta},\tau\mathbf{q}\right) }{n!\partial\tau^{n}%
}\right\vert _{\tau=0}.\nonumber\end{aligned}$$
Let us now use this approach for the calculation of the low-energy expansion of the polarization operator in the Coulomb field.
\[ptb\]
[DelbruckN.eps]{}
The $(Z\alpha)^{N}$ contribution ($N$ is even) is determined by the $N$-loop diagrams depicted in Fig.\[fig:DelbruckN\]. In the dimensional regularization, it can be represented in the form$$\Pi_{(Z\alpha)^{N}}^{\mu\nu}\left( \omega,\mathbf{k},\mathbf{q}\right)
=\int\frac{d\varepsilon\,d^{\mathcal{D}}p}{\left( 2\pi\right) ^{\mathcal{D}%
+1}}\frac{\prod_{i=1}^{N-1}d^{\mathcal{D}}\Delta_{i}}{\left( 2\pi\right)
^{\left( N-1\right) \mathcal{D}}}\,\mathfrak{P}^{\mu\nu}\left(
\varepsilon,\mathbf{p},\boldsymbol{\Delta}_{1},\ldots\boldsymbol{\Delta}%
_{N-1},\omega,\mathbf{k},\mathbf{q}\right) . \label{eq:PolarizationOperatorN}%$$ Similar to the previous example, there are two different region of integration$$\begin{aligned}
\text{\textit{hard} region, when } & \varepsilon\sim p^{i}\sim
\Delta_{1..N-1}^{i}\sim m,\label{e1:2}\\
\text{\textit{soft} region, when } & \left\{
\begin{array}
[c]{rl}%
\varepsilon\sim p^{i} & \sim m,\\
\Delta_{1..N-1}^{i} & \sim\lambda\,m.
\end{array}
\right. \label{e1:3}%\end{aligned}$$ Again, the expansion of the polarization operator is the sum of the integrals of the expansion of the integrand in hard and soft regions. In the coordinate representation, these regions have a simple physical meaning. The characteristic size of the electron field fluctuations (the size of the electron loop) is of the order $1/m$. Hard region corresponds to the configurations where the distance between the Coulomb source and the electron loop is of the order of $1/m$. Soft region corresponds to the creation of the virtual electron-positron pair far from the Coulomb source. Obviously, there is no contribution from the region where only some of the momenta $\boldsymbol{\Delta}_{n}$ are hard while the rest are soft. In the momentum representation, these regions correspond to massless tadpole diagrams which are zero in the dimensional regularization. Thus, the expansion of the polarization operator (\[eq:PolarizationOperatorN\]) has the form$$\begin{gathered}
\Pi_{(Z\alpha)^{N}}^{\mu\nu}\left( \omega,\mathbf{k},\mathbf{q}\right)
=\Pi_{(Z\alpha)^{N},\text{hard}}^{\mu\nu}\left( \omega,\mathbf{k}%
,\mathbf{q}\right) +\Pi_{(Z\alpha)^{N},\text{soft}}^{\mu\nu}\left(
\omega,\mathbf{k},\mathbf{q}\right) \\
=\sum_{n}\int\frac{d\varepsilon\,d^{\mathcal{D}}p}{\left( 2\pi\right)
^{\mathcal{D}+1}}\frac{\prod_{i=1}^{N-1}d^{\mathcal{D}}\Delta_{i}}{\left(
2\pi\right) ^{\left( N-1\right) \mathcal{D}}}\,\mathfrak{P}_{\text{hard}%
}^{\mu\nu\left( n\right) }+\sum_{n}\int\frac{d\varepsilon\,d^{\mathcal{D}}%
p}{\left( 2\pi\right) ^{\mathcal{D}+1}}\frac{\prod_{i=1}^{N-1}%
d^{\mathcal{D}}\Delta_{i}}{\left( 2\pi\right) ^{\left( N-1\right)
\mathcal{D}}}\,\mathfrak{P}_{\text{soft}}^{\mu\nu\left( n\right)
},\label{eq:POHardSoft}\end{gathered}$$ $$\begin{aligned}
\mathfrak{P}_{\text{hard}}^{\mu\nu\left( n\right) } & =\left.
\frac{\partial^{n}\mathfrak{P}^{\mu\nu}\left( \varepsilon,\mathbf{p}%
,\boldsymbol{\Delta}_{1},\ldots\boldsymbol{\Delta}_{N-1},\tau\omega
,\tau\mathbf{k},\tau\mathbf{q}\right) }{n!\partial\tau^{n}}\right\vert
_{\tau=0},\\
\mathfrak{P}_{\text{soft}}^{\mu\nu\left( n\right) } & =\left.
\frac{\partial^{n}\tau^{N-2}\mathfrak{P}^{\mu\nu}\left( \varepsilon
,\mathbf{p},\tau\boldsymbol{\Delta}_{1},\ldots\tau\boldsymbol{\Delta}%
_{N-1},\tau\omega,\tau\mathbf{k},\tau\mathbf{q}\right) }{n!\partial\tau^{n}%
}\right\vert _{\tau=0}.\end{aligned}$$ Note that the simple power counting allows one to estimate the leading terms of the hard and soft contribution as$$\begin{aligned}
\Pi_{(Z\alpha)^{N},\text{hard}}^{\mu\nu}\left( \omega,\mathbf{k}%
,\mathbf{q}\right) & \sim\lambda^{2},\nonumber\\
\Pi_{(Z\alpha)^{N},\text{soft}}^{\mu\nu}\left( \omega,\mathbf{k}%
,\mathbf{q}\right) & \sim\lambda^{\left( N-1\right) \left(
\mathcal{D}-1\right) +1}.\end{aligned}$$ Using Eq.(\[eq:POHardSoft\]), one can calculate the contributions of the hard and soft regions separately.
Method of calculation\[Sec:Calculation\]
========================================
\[ptb\]
[SoftTree.eps]{}
The contribution of the soft region can be graphically represented as the tree diagram shown in Fig.\[fig:HELVertex\]. The local multiphoton vertex depicted as a thick dot corresponds to the expansion of the fermionic loop with respect to the soft momenta $\boldsymbol{\Delta}_{i},k,q$. The expansion is expressed via the integrals of the following form$$\int\frac{d\varepsilon\,d^{\mathcal{D}}p}{\left( 2\pi\right) ^{\mathcal{D}%
+1}}\left( \varepsilon^{2}-\mathbf{p}^{2}-m^{2}+i0\right) ^{-n}%
=\frac{i\left( -1\right) ^{n}\Gamma\left( n-\mathcal{D}/2-1/2\right)
}{\left( 4\pi\right) ^{\left( \mathcal{D}+1\right) /2}\Gamma\left(
n\right) m^{2n-\mathcal{D}-1}}\,.$$ The remaining integrals over $\boldsymbol{\Delta}_{i}$ can be easily evaluated in the coordinate representation. Naturally, the contribution of the soft region can be calculated also with the help of the derivative expansion of the one-loop effective QED action.
The contribution of the hard region determined by Eq.(\[eq:POHardSoft\]) is expressed in terms of the $N$-loop tadpoles. In particular, in $\left( Z\alpha\right) ^{2}$ order, the basic integral has the following form$$%I=\int\frac{d\varepsilon\,d^{\mathcal{D}}p\,d^{\mathcal{D}}\Delta}{\left(
%2\pi\right) ^{2\mathcal{D}+1}}\left( \varepsilon^{2}-\mathbf{p}^{2}%
%-m^{2}+i0\right) ^{-n_{1}}\left( \varepsilon^{2}-\left( \mathbf{p-}%
%\boldsymbol{\Delta}\right) ^{2}-m^{2}+i0\right) ^{-n_{2}}\left(
%\boldsymbol{\Delta}^{2}\right) ^{-n_{3}}.
I=\int\frac{d\varepsilon\,d^{\mathcal{D}}p\,d^{\mathcal{D}}\Delta}{\left(
2\pi\right) ^{2\mathcal{D}+1}}\frac{1}{\left( \varepsilon^{2}-\mathbf{p}^{2}%
-m^{2}+i0\right)^{n_{1}}\left( \varepsilon^{2}-\left( \mathbf{p-}%
\boldsymbol{\Delta}\right)^{2}-m^{2}+i0\right)^{n_{2}}\left(
\boldsymbol{\Delta}^{2}\right)^{n_{3}}}.$$ After the Wick rotation $\varepsilon\rightarrow i\varepsilon$ and rescaling $\mathbf{p}\rightarrow\sqrt{\varepsilon^{2}+m^{2}}\mathbf{p},\mathbf{~}%
\boldsymbol{\Delta}\rightarrow\sqrt{\varepsilon^{2}+m^{2}}\boldsymbol{\Delta}%
$, we integrate over $\varepsilon$ and obtain$$\begin{aligned}
%I & =i\left( -1\right) ^{n_{1}+n_{2}}\frac{m^{1-2\gamma}\Gamma\left(
%\gamma-1/2\right) }{2\sqrt{\pi}\,\Gamma\left( \gamma\right) }\int
%\frac{d^{\mathcal{D}}p\,d^{\mathcal{D}}\Delta}{\left( 2\pi\right)
%^{2\mathcal{D}}}\left( \mathbf{p}^{2}+1\right) ^{-n_{1}}\left( \left(
%\mathbf{p-}\boldsymbol{\Delta}\right) ^{2}+1\right) ^{-n_{2}}\left(
%\boldsymbol{\Delta}^{2}\right) ^{-n_{3}},\\
I & =i\frac{m^{1-2\gamma}\Gamma\left(
\gamma-1/2\right) }{2\sqrt{\pi}\,\Gamma\left( \gamma\right) }\int
\frac{d^{\mathcal{D}}p\,d^{\mathcal{D}}\Delta}{\left( 2\pi\right)
^{2\mathcal{D}}}\frac{\left( -1\right) ^{n_{1}+n_{2}}}{\left(\mathbf{p}^{2}+1\right)^{n_{1}}\left( \left(
\mathbf{p-}\boldsymbol{\Delta}\right) ^{2}+1\right)^{n_{2}}\left(
\boldsymbol{\Delta}^{2}\right)^{n_{3}}},\\
\gamma & =n_{1}+n_{2}+n_{3}-\mathcal{D}.\nonumber\end{aligned}$$ The remaining integral is the two-loop tadpole in $\mathcal{D}=3+\epsilon$ dimensions which can be easily expressed in terms of $\Gamma$-functions.
Performing the similar integration over $\varepsilon$ in $\left(
Z\alpha\right) ^{4}$ order, we express the contribution of the hard region in terms of the integrals of the topology (and its subtopologies) depicted in Fig.\[Fig:Topology\].
\[ptb\]
[CakeDecorated.eps]{}
The general form of such integral is$$J_{n_{1}\ldots n_{10}}=\int\frac{d^{\mathcal{D}}k_{1}d^{\mathcal{D}}%
k_{2}d^{\mathcal{D}}k_{3}d^{\mathcal{D}}k_{4}D_{9}^{-n_{9}}D_{10}^{-n_{10}}%
}{\pi^{2\mathcal{D}}D_{1}^{n_{1}}D_{2}^{n_{2}}D_{3}^{n_{3}}D_{4}^{n_{4}}%
D_{5}^{n_{5}}D_{6}^{n_{6}}D_{7}^{n_{7}}D_{8}^{n_{8}}},$$ where$$\begin{gathered}
D_{1,\ldots,4}=k_{1,\ldots,4}^{2}+1,\qquad\ \\
D_{5}=\left( k_{1}-k_{2}\right) ^{2},\quad D_{6}=\left( k_{2}-k_{3}\right)
^{2},\nonumber\\
D_{7}=\left( k_{3}-k_{4}\right) ^{2},\quad D_{8}=\left( k_{4}-k_{1}\right)
^{2},\nonumber\\
D_{9}=\left( k_{1}-k_{3}\right) ^{2},\quad D_{10}=\left( k_{2}%
-k_{4}\right) ^{2}.\nonumber\end{gathered}$$
The IBP reduction procedure [[@Chetyrkin1980; @Chetyrkin1981]]{} allows one to express any vacuum integral of the considered topology via the five master integrals shown in Fig.\[fig:MIs\].
On these diagrams the solid and dashed lines denote the massive $\left( k^{2}+1\right) ^{-1}$ and massless $\left( k^{2}\right) ^{-1}$ propagators. For each loop momentum the integration measure is taken as $$\frac{d^{\mathcal{D}}k}{\pi^{\mathcal{D}/2}}.$$ Four of these integrals are trivially expressed in terms of $\Gamma
$-functions. Their explicit forms are presented in Appendix. The only nontrivial master integral is $J_{\text{Cake}}$. After the IBP reduction, the master integral $J_{\text{Cake}}$ enters the polarization operator with the coefficient, having the first-order pole in the point $\mathcal{D}=3$. Therefore, we need to determine the $O\left( \epsilon^{0}\right) $ and $O\left( \epsilon^{1}\right) $ terms of the expansion of $J_{\text{Cake}}$.
We find it convenient to use the recurrence relation with respect to space-time dimension, see Ref.. First, we use the Feynman parameterization to obtain the relation$$\begin{aligned}
J_{\text{Cake}}^{(\mathcal{D}-2)}\overset{def}{\equiv}J_{1111111100}%
^{(\mathcal{D}-2)} & =8J_{1112122200}^{(\mathcal{D})}+8J_{1112221200}%
^{(\mathcal{D})}+8J_{1122211200}^{(\mathcal{D})}+8J_{1222111200}%
^{(\mathcal{D})}\nonumber\\
& +4J_{1122121200}^{(\mathcal{D})}+4J_{1212112200}^{(\mathcal{D}%
)}+4J_{1212121200}^{(\mathcal{D})}+J_{2222111100}^{(\mathcal{D})}.
\label{eq:recurrence}%\end{aligned}$$ Then, using the IBP identities, we express the integrals in the right-hand side of this relation via the five master integrals from Fig.\[fig:MIs\]. In particular, the last term in Eq.(\[eq:recurrence\]) can be expressed via the master integrals as follows$$\begin{aligned}
J_{2}^{(5+\epsilon)}\overset{def}{\equiv}J_{2222111100}^{(5+\epsilon)} &
=a_{\text{Cake}}J_{\text{Cake}}^{(5+\epsilon)}+\frac{a_{\text{Clover}}%
}{\epsilon^{3}}J_{\text{Clover}}^{(5+\epsilon)}\nonumber\\
& +a_{\text{Infinity}}J_{\text{Infinity}}^{(5+\epsilon)}+\frac
{a_{\text{Tumbler}}}{\epsilon^{2}}J_{\text{Tumbler}}^{(5+\epsilon)}%
+\frac{a_{\text{Melon}}}{\epsilon^{2}}J_{\text{Melon}}^{(5+\epsilon)}\,.
\label{eq:j2222}%\end{aligned}$$ The coefficients $a_{i}$ are presented in the Appendix. They are chosen to be finite in the limit $\epsilon\rightarrow0$. After the reduction, we have the following recurrence relation:$$\begin{gathered}
J_{\text{Cake}}^{(3+\epsilon)}=\epsilon b_{\text{Cake}}J_{\text{Cake}%
}^{(5+\epsilon)}+\frac{b_{\text{Clover}}}{\epsilon^{2}}\,J_{\text{Clover}%
}^{(5+\epsilon)}\\+b_{\text{Infinity}}J_{\text{Infinity}}^{(5+\epsilon)}%
+\frac{b_{\text{Tumbler}}}{\epsilon}\,J_{\text{Tumbler}}^{(5+\epsilon)}%
+\frac{b_{\text{Melon}}}{\epsilon}\,J_{\text{Melon}}^{(5+\epsilon)}\,.\end{gathered}$$ Again, the coefficients $b_{i}$ are chosen to be finite in the limit $\epsilon\rightarrow0$ and are presented in the Appendix. Now we use the following trick. Let us express $J_{\text{Cake}}^{(5+\epsilon)}$ from Eq.(\[eq:j2222\]) and substitute into Eq. (\[eq:recurrence\]). We obtain$$\begin{gathered}
J_{\text{Cake}}^{(3+\epsilon)} =\frac{\epsilon b_{\text{Cake}}%
}{a_{\text{Cake}}}\,J_{2}^{(5+\epsilon)}+\frac{b_{\text{Cake}}}{\epsilon^{2}%
}\left( \frac{b_{\text{Clover}}}{b_{\text{Cake}}}-\frac{a_{\text{Clover}}%
}{a_{\text{Cake}}}\right) J_{\text{Clover}}^{(5+\epsilon)}\\
+b_{\text{Cake}%
}\left( \frac{b_{\text{Infinity}}}{b_{\text{Cake}}}-\frac{\epsilon
\,a_{\text{Infinity}}}{a_{\text{Cake}}}\right) J_{\text{Infinity}%
}^{(5+\epsilon)}+\frac{b_{\text{Cake}}}{\epsilon}\left( \frac{b_{\text{Tumbler}}%
}{b_{\text{Cake}}}\,-\frac{a_{\text{Tumbler}}}{a_{\text{Cake}}}\right)
J_{\text{Tumbler}}^{(5+\epsilon)}\\
+\frac{b_{\text{Cake}}}{\epsilon}\left(
\frac{b_{\text{Melon}}}{b_{\text{Cake}}}-\frac{a_{\text{Melon}}}%
{a_{\text{Cake}}}\right) J_{\text{Melon}}^{(5+\epsilon)}\,.
\label{eq:jCakeviaj2D}%\end{gathered}$$ Since the integral $J_{2}^{\left( \mathcal{D}\right) }$ is finite in $\mathcal{D}=5$ and the coefficient in front of this integral in Eq.(\[eq:jCakeviaj2D\]) contains $\epsilon$ factor, the first term in the right-hand side of Eq.(\[eq:jCakeviaj2D\]) does not contribute in $\epsilon^{0}$ order. Expanding the coefficient $a_{i},b_{i}$ and the four simple master integrals, we obtain$$J_{\text{Cake}}^{(3+\epsilon)}=\frac{\pi^{4}}{6}+\epsilon\left[ \frac{\pi
^{4}}{3}\left( C-\ln2-\frac{11}{8}\right) -\pi^{2}-\frac{3}{4}\,J_{2}%
^{(5)}\right] +O\left( \epsilon^{2}\right) \,, \label{eq:jCakeviaj2}%$$ where $C=0.577...$ is the Euler constant. Note, that using this trick we have obtained the $O\left( \epsilon^{0}\right) $ term of $J_{\text{Cake}%
}^{(3+\epsilon)}$ for free.
In order to calculate the $O\left( \epsilon\right) $ term, let us consider the general solution of the recurrence (\[eq:recurrence\]). Taking into account the explicit form of the coefficient $b_{\text{Cake}}$, we obtain$$J_{\text{Cake}}^{(\mathcal{D})}=J_{0}\left( \mathcal{D}\right) \left[
P\left( \mathcal{D}\right) +\sum_{i=1}^{4}\sum_{n=1}^{\infty}\frac
{c_{i}^{(\mathcal{D}+2n)}}{J_{0}(\mathcal{D}+2n)}J_{i}^{(\mathcal{D}%
+2n)}\right] , \label{eq:solutionDrelation}%$$ where$$\begin{aligned}
J_{0}\left( \mathcal{D}\right) & =\frac{2^{3\mathcal{D}}\Gamma\left(
1-\mathcal{D}/2\right) \Gamma\left( 3\mathcal{D}/2-11/2\right) }%
{\Gamma(\mathcal{D}-2)\Gamma(\mathcal{D}-3)^{2}},\\
c_{\text{Clover}} & =\frac{b_{\text{Clover}}}{\epsilon^{3}b_{\text{Cake}}%
}\ ,\quad c_{\text{Infinity}}=\frac{b_{\text{Infinity}}}{\epsilon
b_{\text{Cake}}},\nonumber\\
c_{\text{Tumbler}} & =\frac{b_{\text{Tumbler}}}{\epsilon^{2}b_{\text{Cake}}%
}\ ,\quad c_{\text{Melon}}=\frac{b_{\text{Melon}}}{\epsilon^{2}b_{\text{Cake}%
}},\nonumber\end{aligned}$$ and $P\left( \mathcal{D}\right) =P\left( \mathcal{D}+2\right) $ is a periodic function of $\mathcal{D}$. Note that, using the explicit form of the simple master integrals, Eq.(\[eq:MIsimple\]), the sums in Eq.(\[eq:solutionDrelation\]) can be checked to converge rapidly. In order to fix the function $P\left( \mathcal{D}\right) $, we have to calculate the leading asymptotic of $J_{\text{Cake}}^{(\mathcal{D})}$ at $\mathcal{D}\rightarrow+\infty$. However, the calculation of this asymptotic is not a simple problem. Instead, we may proceed in the alternative way by applying the method of difference equations described in Ref.. According to this method, we derive the recurrence relation in $x$ for $J_{x111111100}^{(\mathcal{D})}$:$$\begin{aligned}
J_{x111111100}^{(\mathcal{D})} & =C(x)J_{x+1,111111100}^{(\mathcal{D}%
)}+F(x),\label{eq:Xrecurrencerelation}\\
C(x) & =\frac{x\left( x+11-3\mathcal{D}\right) }{\left( x+5-3\mathcal{D}%
/2\right) \left( x+3-\mathcal{D}\right) }\,,\end{aligned}$$ where $F(x)$ is expressed in terms of finite sums of $\Gamma$- functions. The solution of this relation is$$\begin{aligned}
J_{x111111100}^{(\mathcal{D})} & =G^{(\mathcal{D})}\left( x\right) \left[
R^{(\mathcal{D})}\left( x\right) +\sum_{y=x}^{\infty}\frac{F(y)}{G\left(
y\right) }\right] ,\label{eq:solutionXrelation}\\
G^{(\mathcal{D})}\left( x\right) & =\frac{\Gamma\left( x+5-3\mathcal{D}%
/2\right) \Gamma\left( x+3-\mathcal{D}\right) }{\Gamma\left( x\right)
\Gamma\left( x+11-3\mathcal{D}\right) },\end{aligned}$$ where $R^{(\mathcal{D})}\left( x\right) $ is a periodic function of $x$, which can be determined through the $J_{x111111100}^{(\mathcal{D})}$ asymptotic behaviour in $x\rightarrow\infty$ limit. At large $x$, the $x$-dependence of $J_{x111111100}^{(\mathcal{D})}$ factorizes into $\int
d^{\mathcal{D}}k_{1}\,\left( k_{1}^{2}+1\right) ^{-x}$. Using the large-$x$ behaviour of the function $G^{(\mathcal{D})}\left( x\right) $, we find$$\left. \frac{J_{x111111100}^{(\mathcal{D})}}{G^{(\mathcal{D})}\left(
x\right) }\right\vert _{x\rightarrow\infty}\sim x^{3-\mathcal{D}}.$$ Therefore, for $\mathcal{D}>3$ we have $R\left( x\right) =0$ and we can use Eq.(\[eq:solutionXrelation\]) to estimate numerically the function $P\left( \mathcal{D}\right) $ in Eq.(\[eq:solutionDrelation\]). It should be noted that the recurrence relation (\[eq:Xrecurrencerelation\]) can be hardly applied near $\mathcal{D}=3$ due to the slow convergence of the sum in the right-hand side of Eq.(\[eq:solutionXrelation\]). Performing the estimation of $P\left( \mathcal{D}\right) $ for several non-integer values of $\mathcal{D}$, we find that in all cases the value of the function $P\left( \mathcal{D}\right) $ is compatible with zero up to $10^{-10}$. Thus, our ansatz is $P\left( \mathcal{D}\right) =0$, and we have$$J_{\text{Cake}}^{(3+\epsilon)}=\frac{8^{\epsilon+3}\Gamma\left(
-\frac{\epsilon}{2}-\frac{1}{2}\right) \Gamma\left( \frac{3\epsilon}%
{2}-1\right) }{\epsilon\Gamma(\epsilon)^{3}}\sum_{n=0}^{\infty}T\left(
2n+\epsilon\right) , \label{eq:JD}%$$$$\begin{aligned}
T\left( \nu\right) & =\frac{8^{-\nu-5}(\nu+1)\left( 14\nu^{3}+40\nu
^{2}+35\nu+10\right) \Gamma\left( -\frac{\nu}{2}-\frac{1}{2}\right)
^{3}\Gamma(\nu)^{3}}{\Gamma\left( \frac{3}{2}\nu+2\right) }\nonumber\\
& -\frac{\pi^{7/2}\left( 6417\nu^{5}+13266\nu^{4}+8368\nu^{3}+980\nu
^{2}-585\nu-126\right) \Gamma(\nu)}{6144\nu(3\nu-1)(3\nu+1)\Gamma
(-2\nu)\Gamma(3\nu+4)\cos\left( \frac{\pi\nu}{2}\right) \sin^{2}(\pi\nu
)\cos\left( \frac{3\pi\nu}{2}\right) }\nonumber\\
& +\frac{\pi^{5/2}\left( 585\nu^{3}+372\nu^{2}+7\nu-12\right) \Gamma
(\nu)\Gamma\left( \frac{\nu+1}{2}\right) ^{2}\tan\left( \frac{3\pi\nu}%
{2}\right) }{12288(3\nu-1)(3\nu+1)\Gamma(2\nu+2)\cos\left( \frac{\pi\nu}%
{2}\right) \sin\left( 2\pi\nu\right) }\nonumber\\
& -\frac{\pi^{3/2}\left( 3897\nu^{4}+3870\nu^{3}+797\nu^{2}-198\nu
-54\right) \Gamma(\nu)^{2}\Gamma\left( \frac{\nu+1}{2}\right) ^{3}%
\tan\left( \frac{3\pi\nu}{2}\right) }{36864(3\nu-1)(3\nu+1)\Gamma
(2\nu+2)\Gamma\left( \frac{3}{2}\nu+\frac{3}{2}\right) \sin\left( 2\pi
\nu\right) \sin\left( \frac{\pi\nu}{2}\right) }\,.\end{aligned}$$ The series in this representation are converging rapidly. The first term in $T\left( \nu\right) $, the most slowly decreasing one, behaves as $3^{-3n}$. Thus, roughly speaking, each two consecutive terms in the sum give three more decimal digits of precision. We claim this expression to be valid for arbitrary $\epsilon$. For example, one can easily reproduce terms of the expansion of $J_{\text{Cake}}$ near $\mathcal{D}=4$ obtained in [[@Schroder:2005va]]{}.
Using Eqs.(\[eq:j2222\]), (\[eq:JD\]), we obtain $$\begin{aligned}
J_{2}^{(5)} & =-\frac{1}{81}\pi^{2}\left( 63\pi^{2}-488-72\zeta_{3}\right)
+\sum_{n=1}^{\infty}T_{2}(n)\,,\label{eq:j2222res}\\
T_{2}(n) & =\frac{(-1)^{n}\left( 56n^{3}+80n^{2}+35n+5\right) \pi
^{2}((n-1)!)^{3}}{18(2n+1)^{2}(3n+1)!}\nonumber\\
& -\frac{(-1)^{n}2^{-4n-1}\left( 2340n^{3}+744n^{2}+7n-6\right) \pi
^{4}((2n)!)^{3}}{9n\left( 36n^{2}-1\right) (n!)^{2}(4n+1)!}\nonumber\\
& -\frac{16\pi^{2}(2n-1)!(4n-1)!}{27\left( 36n^{2}-1\right) ^{2}%
(6n+3)!}\nonumber\\
& \times\left( 11088576n^{6}+7641216n^{5}+691632n^{4}-424512n^{3}%
-79356n^{2}+616n+585\right) \nonumber\\
& +\frac{2(-1)^{n}\pi^{2}((2n)!)^{5}(3n)!}{27n^{2}\left( 36n^{2}-1\right)
^{2}(n!)^{3}(4n+2)!(6n+1)!}\nonumber\\
& \times\left( 4489344n^{6}+2794176n^{5}+206352n^{4}-146880n^{3}%
-26032n^{2}-28n+153\right) \nonumber\\
& -\frac{4\pi^{2}(2n)!(4n)!\left( H_{2n-1}+2H_{4n-1}-3H_{6n+3}\right)
}{27n^{2}\left( 36n^{2}-1\right) (6n+3)!}\nonumber\\
& \times\left( 102672n^{5}+106128n^{4}+33472n^{3}+1960n^{2}-585n-63\right)
\nonumber\\
& -\frac{(-1)^{n}\pi^{2}((2n)!)^{5}(3n)!\left( 3H_{n-1}-10H_{2n-1}%
-3H_{3n+1}+4H_{4n+2}+6H_{6n+3}\right) }{54n^{2}\left( 36n^{2}-1\right)
(n!)^{3}(4n+1)!(6n+1)!}\nonumber\\
& \times\left( 31176n^{4}+15480n^{3}+1594n^{2}-198n-27\right) \,,\nonumber\end{aligned}$$$$J_{2}^{(5)}=0.0516516357945\ldots.$$ Here $H_{n}=\sum_{k=1}^{n}k^{-1}$ is a harmonic number. Unfortunately, we have not been able to express the sums in Eq.(\[eq:j2222res\]) in terms of $\zeta$-functions and alike. However, the numerical convergence of the above series is perfect and we obtain from Eqs.(\[eq:jCakeviaj2\]), (\[eq:j2222res\])$$J_{\text{Cake}}^{(3+\epsilon)}=\frac{\pi^{4}}{6}-\epsilon\times
58.3184377060\ldots+O\left( \epsilon^{2}\right) \,.$$
Methods of calculation and values of multiloop vacuum integrals in arbitrary space-time dimension $\mathcal{D}$ are of independent interest, because these integrals appear as parts of the amplitudes for various physical processes: from QCD and QED radiative corrections [@Steinhauser:2002rq] to the thermodynamics of finite temperature QCD-like theories [@Kajantie:2003ax]. In particular, the master integrals in Fig.\[fig:MIs\] enter the basis intensively used in modern QCD calculations [@Chetyrkin:2006xg; @Chetyrkin:2005ia; @Faisst:2006sr; @Chetyrkin:2006dh].
Results and Conclusion\[Sec:Results\]
=====================================
The perturbative expansion of the form factors in Eq.(\[eq:Parametrization\]) has the form$$f_{i}=\sum_{n=2,4,\ldots}\alpha\left( Z\alpha\right) ^{n}f_{i}^{\left(
n\right) }.$$ We have calculated the low-energy expansion of $f_{i}^{\left( 2\right) }$ and $f_{i}^{\left( 4\right) }$ up to $O\left( \lambda^{2}\right) $, which corresponds to the expansion of the polarization operator up to $O\left(
\lambda^{4}\right) $. Using Eq.(\[eq:POHardSoft\]), we obtain in $(Z\alpha)^{2}$ order$$\begin{aligned}
f_{1}^{(2)} & =\frac{7}{2(4!)^{2}}+\frac{1}{6!}\frac{|\mathbf{k}-\mathbf{q}%
|}{m}+\frac{1}{2^{2}(5!)^{2}m^{2}}\left[ \frac{117}{4}\,\omega^{2}%
+49\,\mathbf{k}\cdot\mathbf{q}-\frac{57}{2}\left( \mathbf{k}^{2}%
+\mathbf{q}^{2}\right) \right] ,\label{eq:f12}\\
f_{2}^{(2)} & =-\frac{73\,}{2^{2}(4!)^{2}}-\frac{11}{6!}\frac{|\mathbf{k}%
-\mathbf{q}|}{m}\nonumber\\
& +\frac{1}{2^{5}(5!)^{2}m^{2}}\left[ -\frac{10923}{2}%
\,\omega^{2}+3095\,\mathbf{k}\cdot\mathbf{q}+1111\left( \mathbf{k}%
^{2}+\mathbf{q}^{2}\right) \right] ,\label{eq:f22}\\
f_{3}^{(2)} & =-\frac{7}{6!}\frac{|\mathbf{k}-\mathbf{q}|}{m}+\frac
{587}{2^{3}(5!)^{2}}\frac{\left( \mathbf{k}-\mathbf{q}\right) ^{2}}{m^{2}%
},\label{eq:f32}\\
f_{4}^{(2)} & =\frac{573}{2^{4}(5!)^{2}}\frac{\omega^{2}}{m^{2}%
},\label{eq:f42}\\
f_{5}^{(2)} & =-\frac{4}{6!}\frac{|\mathbf{k}-\mathbf{q}|}{m}+\frac
{1369}{2^{5}(5!)^{2}}\,\frac{\left( \mathbf{k}-\mathbf{q}\right) ^{2}}%
{m^{2}}.\label{eq:f52}%\end{aligned}$$ The $O\left( \lambda^{0}\right) $ terms of $f_{1,2}^{(2)}$ have been obtained in Ref.[@Costantini:1971cj]. When $\mathbf{k}=\omega=0,$ the expansion of $f_{1}^{\left( 2\right) }\left( q\right) $ agrees with the exact result obtained in Ref.[@Lee:2007]. The next-to-leading terms, proportional to $|\mathbf{k}-\mathbf{q}|$, come from the soft region and exhibit the above-mentioned nonanalytic behavior.
In $(Z\alpha)^{4}$ order we obtain$$\begin{aligned}
f_{1}^{(4)}= & \frac{1}{2^{7}}\left( \frac{2}{\pi^{2}}J_{2}^{\left(
5\right) }-\frac{4}{3}\,\zeta_{3}+\frac{5}{9}\,\zeta_{2}+\frac{13}%
{18}\right)\label{eq:f14}\\
& +\frac{1}{2^{10}5\,m^{2}}\left[ \left( -\frac{109}{2\pi^{2}%
}J_{2}^{\left( 5\right) }+\frac{475}{3}\,\zeta_{3}-\frac{4111}{30}%
\,\zeta_{2}+\frac{2575}{72}\right) \omega^{2}\right. \nonumber\\
& +\left( \mathbf{k}^{2}+\mathbf{q}^{2}\right) \left( -\frac{6}{\pi^{2}%
}J_{2}^{\left( 5\right) }-\frac{25}{18}\,\zeta_{2}+\frac{29}{8}\right)
+\left. \mathbf{k}\cdot\mathbf{q}\left( -\frac{51}{\pi^{2}}J_{2}^{\left(
5\right) }+\frac{481}{30}\,\zeta_{2}-\frac{3053}{108}\right) \right]
,\nonumber\\
f_{2}^{(4)}= & \frac{1}{2^{7}}\left( \frac{7}{2\pi^{2}}J_{2}^{(5)}-\frac
{10}{3}\,\zeta_{3}+\frac{1267}{72}\,\zeta_{2}-\frac{3661}{144}\right)\label{eq:f24}
\\
&+\frac{1}{2^{10}5\,m^{2}}\left[ \left( \frac{473}{4\pi^{2}}J_{2}^{\left(
5\right) }-145\zeta_{3}+\frac{279829}{720}\,\zeta_{2}-\frac{1233655}%
{2592}\right) \omega^{2}\right. \nonumber\\
& +\left( \mathbf{k}^{2}+\mathbf{q}^{2}\right) \left( \frac{2}{\pi^{2}%
}J_{2}^{\left( 5\right) }-\frac{95}{2}\,\zeta_{3}-\frac{359}{360}\,\zeta
_{2}+\frac{18769}{288}\right) \nonumber \\
&+\left. \mathbf{k}\cdot\mathbf{q}\left(
-\frac{235}{2\pi^{2}}J_{2}^{\left( 5\right) }-\frac{32999}{180}\,\zeta
_{2}+\frac{610}{3}\,\zeta_{3}+\frac{7571}{144}\right) \right]\nonumber,\end{aligned}$$ $$\begin{aligned}
f_{3}^{(4)}= & \frac{1}{2^{10}5\,}\frac{(\mathbf{k}-\mathbf{q})^{2}}{m^{2}%
}\left( \frac{87}{2\pi^{2}}J_{2}^{\left( 5\right) }-\frac{12641}%
{180}\,\zeta_{2}+\frac{24551}{216}\right) ,\\
f_{4}^{(4)}= & \frac{1}{2^{10}5\,}\frac{\omega^{2}}{m^{2}}\left( -\frac
{71}{2\pi^{2}}J_{2}^{\left( 5\right) }+\frac{25}{3}\,\zeta_{3}+\frac
{2857}{60}\,\zeta_{2}-\frac{111685}{1296}\right) ,\\
f_{5}^{(4)}= & \frac{1}{2^{10}5}\frac{(\mathbf{k}-\mathbf{q})^{2}}{m^{2}%
}\left( \frac{31}{2\pi^{2}}J_{2}^{\left( 5\right) }-30\zeta_{3}%
+\frac{11639}{180}\,\zeta_{2}-\frac{31319}{432}\right) .\end{aligned}$$ In $(Z\alpha)^4$ order, the nonanalytic contribution to the form factors is suppressed as $O(\lambda^5)$ and thus is far beyond the accuracy chosen. Note, that the technique used in this paper can be applied without modification to the calculation of the higher terms of the low-energy expansion in $(Z\alpha)^2$ and $(Z\alpha)^4$ orders.
The Coulomb corrections to the form factors $f_{1}$ and $f_{2}$ in order $\lambda^{0}$ were calculated in Ref.[@Kirilin2008] numerically. Although the interaction with the Coulomb field was taken into account exactly, it turned out that the results can be well fitted by the polynomial function of $Z\alpha$:$$\begin{aligned}
f_{1} & =\frac{7}{2(4!)^{2}}(Z\alpha)^{2}+3.35\cdot10^{-4}(Z\alpha
)^{4}+1.6\cdot10^{-4}(Z\alpha)^{6},\label{eq:fitF1}\\
f_{2} & =-\frac{73\,}{2^{2}(4!)^{2}}(Z\alpha)^{2}-3.55\cdot10^{-3}%
(Z\alpha)^{4}-2.1\cdot10^{-3}(Z\alpha)^{6}.\label{eq:fitF2}%\end{aligned}$$ In $O\left( \lambda^{0}\right) $ order, our results (\[eq:f14\]), (\[eq:f24\]) for the $(Z\alpha)^{4}$-order corrections numerically coincide with those of Eqs.(\[eq:fitF1\]),(\[eq:fitF2\]) with an accuracy of a few percent.
\[ptb\]
[gBound.eps]{}
As the demonstration of possible applications of our result, let us calculate the $O\left( \left( Z\alpha\right) ^{7}\right) $ contribution of the “magnetic loop” to the $g$ factor of the bound electron, see Fig.\[fig:gBound\]. The corresponding correction to the $g$ factor of the electron in $nL_{J}$ state has the form [@Lee:2004vb] $$\begin{aligned}
\frac{\Delta g}{g_{0}} & =\frac{16}{\pi\left( 2\kappa-1\right) }\int
\frac{dq}{q}\,f_{1}\left( q\right) \,G\left( q\right) \,,\\
f_{1}\left( q\right) & =\left. f_{1}\right\vert _{\mathbf{k}=\omega
=0}\,,\quad G\left( q\right) =\int dr\,a\left( r\right) b\left( r\right)
\left( \sin qr-qr\cos qr\right) ,\\
g_{0} & =\frac{2\kappa}{2\kappa+1},\quad\kappa=\left( J+1/2\right)
\,\mathrm{sign}\left( L-J\right) \,,\end{aligned}$$ where $a$ and $b$ are determined by the form of the bound electron wave function$$\psi\left( \mathbf{r}\right) =\left(
\begin{array}
[c]{c}%
a\left( r\right) \Omega_{JLM}\left( \mathbf{n}\right) \\
ib\left( r\right) \tilde{\Omega}_{JLM}\left( \mathbf{n}\right)
\end{array}
\right) .$$ The characteristic scale of the function $G\left( q\right) $ is $mZ\alpha/n\ll m$, so we have two regions of integration: $q\sim mZ\alpha/n$ and $q\sim m$. In the leading order, only the first region is essential, the contribution of this region is of the order $O\left( \left( Z\alpha\right)
^{5}\right) $. The leading correction $\sim O\left( \left( Z\alpha\right)
^{6}\right) $ for $L\neq0$ states also comes from the region $q\sim
mZ\alpha/n,$while for $L=0$ the whole interval $mZ\alpha/n\lesssim q\lesssim
m$ is essential. This correction has been found in Ref.[@Lee:2004vb]. Note that the integrals in Eq.(18) of Ref.[@Lee:2004vb] can be taken analytically, and the correction to $g$ factor up to the order $O\left( \left( Z\alpha\right) ^{6}\right)
$ can be represented as $$\begin{aligned}
\left( \frac{\Delta g}{g_{0}}\right) _{\left( Z\alpha\right) ^{5}+\left(
Z\alpha\right) ^{6}} & =\frac{7\alpha\left( Z\alpha\right) ^{5}}%
{288n^{3}J\left( J+1\right) \left( 2J+1\right) }\nonumber\\
& +\delta_{L=0}\frac{4\alpha\left( Z\alpha\right) ^{6}}{135\pi}\frac
{1}{n^{3}}\left[ \log\frac{n}{2Z\alpha}-\frac{641}{240}-H_{n}+\frac{\left(
n+1\right) \left( 4n-1\right) }{6n^{2}}\right] \nonumber\\
& +\delta_{L\neq0}\frac{2\alpha(Z\alpha)^{6}}{45\pi n^{3}(2L+1)(2\kappa
-1)^{2}}\left( \frac{3}{L\left( L+1\right) }-\frac{1}{n^{2}}\right) \,.\end{aligned}$$ In order to find the next-to-leading correction, we separate the contributions of the two regions similar to what has been described above. The details of this calculation will be presented elsewhere. It turns out that the complete $O\left( \left( Z\alpha\right) ^{7}\right) $ result for the correction to $g$ factor can be expressed via several first term of expansion of the function $f_{1}$ near $q=0$, namely$$\begin{gathered}
\left( \frac{\Delta g}{g_{0}}\right) _{\left( Z\alpha\right) ^{7}}
=\frac{4\alpha\left( Z\alpha\right) ^{7}}{n^{5}J(J+1)(2J+1)}\\
\times\left[
n^{2}\frac{6J\left( J+1\right) +1}{(2J+1)^{2}J\left( J+1\right) }%
+n\frac{3}{(2J+1)}-2-\frac{1}{2(2\kappa-1)}\right] f_{1}^{\left( 2\right)
}\left( 0\right) \\
-\delta_{L=0}\frac{8\alpha\left( Z\alpha\right) ^{7}}{3n^{5}}\left(
1+5n^{2}\right) m^{2}f_{1}^{\left( 2\right) \prime\prime}\left( 0\right)
+\frac{4\alpha\left( Z\alpha\right) ^{7}}{n^{3}J\left( J+1\right) \left(
2J+1\right) }f_{1}^{\left( 4\right) }\left( 0\right) .\end{gathered}$$ Now, owing to Eq.(\[eq:f14\]), we have the last essential ingredient to obtain the correction. Using Eqs.(\[eq:f12\]),(\[eq:f14\]), we obtain$$\begin{gathered}
\left( \frac{\Delta g}{g_{0}}\right) _{\left( Z\alpha\right) ^{7}}
=\frac{7\alpha\left( Z\alpha\right) ^{7}}{288n^{5}J(J+1)(2J+1)}\\
\times\left[n^{2}\frac{6J\left( J+1\right) +1}{(2J+1)^{2}J\left( J+1\right) }%
+n\frac{3}{(2J+1)}-2-\frac{1}{2(2\kappa-1)}\right]
+\delta_{L=0}\frac{19\alpha\left( Z\alpha\right) ^{7}}{7200n^{5}}\left(
1+5n^{2}\right) \\ +\frac{\alpha\left( Z\alpha\right) ^{7}}{32n^{3}J\left(
J+1\right) \left( 2J+1\right) }\left( \frac{2}{\pi^{2}}J_{2}^{\left(
5\right) }-\frac{4}{3}\,\zeta_{3}+\frac{5}{9}\,\zeta_{2}+\frac{13}%
{18}\right) .\label{eq:delta_g7}%\end{gathered}$$ In particular, for $1S_{1/2}$ and $2P_{1/2}$ states we have$$\begin{aligned}
\left( \frac{\Delta g}{g_{0}}\right) _{1S} & =1.62\times10^{-2}%
\alpha\left( Z\alpha\right) ^{5}+9.431\,4\times10^{-3}\alpha\left(
Z\alpha\right) ^{6}\left( \ln\frac{1}{2Z\alpha}-2.67\right)\nonumber \\
& +4.1\times
10^{-2}\alpha\left( Z\alpha\right) ^{7}\,,\nonumber\\
8\left( \frac{\Delta g}{g_{0}}\right) _{2P} & =1.62\times10^{-2}%
\alpha\left( Z\alpha\right) ^{5}+5.894\,6\times10^{-3}\alpha\left(
Z\alpha\right) ^{6}\nonumber\\
&+3.26\times10^{-2}\alpha\left( Z\alpha\right) ^{7}.\end{aligned}$$ The contribution of the $O\left( \left( Z\alpha\right) ^{7}\right) $ term is rather essential, e.g., for $Z=6$ (carbon) the ratio of this term to $O\left( \left( Z\alpha\right) ^{6}\right) $ term for $1S_{1/2}$ state is $-0.81$. The last term in Eq.(\[eq:delta\_g7\]) corresponds to the contribution of the electron loop with four Coulomb exchanges. It is interesting to compare the magnitude of this term with that of the first two terms. As it was claimed in Ref.[@Lee:2004vb] this term appears to be numerically small. E.g., for the ground state, the contribution of the last term is only $2.2$ percent.
Appendix
========
The explicit form of the four simple master integrals from Fig.\[fig:MIs\] reads:$$\begin{aligned}
J_{\text{Infinity}}^{\left( \mathcal{D}\right) } & \overset{def}{\equiv
}J_{0101111100}^{(\mathcal{D})} \nonumber \\
& =\frac{\Gamma(6-2\mathcal{D})\Gamma\left(
5-3\mathcal{D}/2\right) ^{2}\Gamma\left( 2-\mathcal{D}/2\right) ^{2}%
\Gamma\left( \mathcal{D}/2-1\right) ^{4}\Gamma\left( 3\mathcal{D}%
/2-4\right) }{\Gamma(10-3\mathcal{D})\Gamma(\mathcal{D}-2)^{2}\Gamma\left(
\mathcal{D}/2\right) },\nonumber\\
J_{\text{Melon}}^{\left( \mathcal{D}\right) } & \overset{def}{\equiv
}J_{0011110100}^{(\mathcal{D})} =\frac{\Gamma(5-2\mathcal{D})\Gamma\left(
4-3\mathcal{D}/2\right) ^{2}\Gamma(3-\mathcal{D})\Gamma\left( \mathcal{D}%
/2-1\right) ^{3}}{\Gamma(8-3\mathcal{D})\Gamma\left( \mathcal{D}/2\right)
},\nonumber\\
J_{\text{Tumbler}}^{\left( \mathcal{D}\right) } & \overset{def}{\equiv
}J_{0111100100}^{(\mathcal{D})} =\frac{\Gamma\left( 4-3\mathcal{D}%
/2\right) \Gamma(3-\mathcal{D})^{2}\Gamma\left( 1-\mathcal{D}/2\right)
\Gamma\left( 2-\mathcal{D}/2\right) \Gamma\left( \mathcal{D}/2-1\right)
^{2}}{\Gamma(6-2\mathcal{D})\Gamma\left( \mathcal{D}/2\right) },\nonumber\\
J_{\text{Clover}}^{\left( \mathcal{D}\right) } & \overset{def}{\equiv
}J_{1111000000}^{(\mathcal{D})} =\Gamma\left( 1-\mathcal{D}/2\right)
^{4}. \label{eq:MIsimple}%\end{aligned}$$ The coefficients in Eq.(\[eq:j2222\]) are$$\begin{aligned}
a_{\text{Cake}} & =\frac{(\epsilon+1)^{3}\left( 9\epsilon^{2}%
+3\epsilon-4\right) }{48(3\epsilon+2)},\nonumber\\
\frac{a_{\text{Clover}}}{\epsilon^{3}} & =\frac{(\epsilon+1)(\epsilon+3)^{3}\left(
6\epsilon^{3}+5\epsilon^{2}+3\epsilon+2\right) }{384\epsilon^{3}%
(3\epsilon+2)},\nonumber\\
a_{\text{Infinity}} & =-\left( \frac{9\epsilon^{3}}{16}-\frac{3\epsilon
^{2}}{8}-\frac{51\epsilon}{16}-\frac{5}{3}\right) \frac{(\epsilon
+1)_{2}(2\epsilon+2)_{3}}{(3\epsilon+1)_{4}},\nonumber\\
\frac{a_{\text{Tumbler}}}{\epsilon^{2}} & =-\left( 4\epsilon^{6}+\frac
{73\epsilon^{5}}{6}+16\epsilon^{4}+\frac{967\epsilon^{3}}{54}+\frac
{397\epsilon^{2}}{27}+\frac{17\epsilon}{3}+\frac{2}{3}\right) \frac
{(\epsilon+1)\left( 3\epsilon/2+3/2\right) _{4}}{(2\epsilon)_{4}%
(3\epsilon)_{3}},\nonumber\\
\frac{a_{\text{Melon}}}{\epsilon^{2}} & =\left( \frac{9\epsilon^{5}}{16}%
-\frac{3\epsilon^{4}}{4}-\frac{117\epsilon^{3}}{16}-\frac{305\epsilon^{2}}%
{24}-\frac{51\epsilon}{8}-\frac{3}{4}\right) \frac{\left( 3\epsilon
/2+5/2\right) _{2}(2\epsilon+2)_{4}}{\epsilon(\epsilon+2)(3\epsilon)_{5}},\end{aligned}$$ Here $x_{n}=\Gamma\left( x+n\right) /\Gamma\left( x\right) =x\left(
x+1\right) \ldots\left( x+n-1\right) $. The coefficients in Eq.(\[eq:recurrence\]) are$$\begin{aligned}
\epsilon b_{\text{Cake}} & =-\frac{\epsilon(\epsilon+1)^{3}(\epsilon
+2)(\epsilon+3)}{48(3\epsilon-2)(3\epsilon+2)},\nonumber\\
\epsilon^{-2}b_{\text{Clover}} & =\frac{(\epsilon+1)(\epsilon+3)^{4}\left(
14\epsilon^{3}+40\epsilon^{2}+35\epsilon+10\right) }{384\epsilon
^{2}(3\epsilon-2)(3\epsilon+2)},\nonumber\\
b_{\text{Infinity}} & =-\left( 3-\frac{7\epsilon}{4}-93\epsilon^{2}%
-\frac{585\epsilon^{3}}{4}\right) \frac{(\epsilon+2)\epsilon_{4}%
(2\epsilon+2)_{3}}{4\,(3\epsilon-2)_{7}},\nonumber\\
\epsilon^{-1}b_{\text{Tumbler}} & =\left( 7+\frac{65\epsilon}{2}%
-\frac{490\epsilon^{2}}{9}-\frac{4184\epsilon^{3}}{9}-737\epsilon^{4}%
-\frac{713\epsilon^{5}}{2}\right) \frac{(\epsilon+1)_{3}\left(
3\epsilon/2+5/2\right) _{3}}{64\,\left( \epsilon+1/2\right) _{2}%
(3\epsilon-2)_{5}},\nonumber\\
\epsilon^{-1}b_{\text{Melon}} & =\left( 3+11\epsilon-\frac{797\epsilon^{2}%
}{18}-215\epsilon^{3}-\frac{433\epsilon^{4}}{2}\right) \frac{3\,\left(
3\epsilon/2+5/2\right) _{2}(2\epsilon+2)_{5}}{16\,(3\epsilon-2)_{7}}.\end{aligned}$$
[10]{}
L. Meitner, H. Kösters, and M. Delbrück. . , 84:137–144, 1933.
V. Costantini, B. De Tollis, and G. Pistoni. . , A2:733–787, 1971.
Hung Cheng and Tai Tsun Wu. . , 182(5):1873–1898, Jun 1969.
Hung Cheng and Tai Tsun Wu. . , 2(10):2444–2457, Nov 1970.
Hung Cheng and Tai Tsun Wu. . , 5(12):3077–3087, Jun 1972.
A. I. Milshtein and V. M. Strakhovenko. . , 58:8–13, 1983.
V.M. Strakhovenko A.I. Mil’shtein. . , 95:135–138, 1983.
R. N. Lee, A. I. Milsten, and V. M. Strakhovenko. . , 89:41, 1999.
G. G. Kirilin and I. S. Terekhov. . , 77(3):032118, 2008.
M. Beneke and Vladimir A. Smirnov. . , B522:321–344, 1998.
Vladimir A. Smirnov and E. R. Rakhmetov. . , 120:870–875, 1999.
David J. Broadhurst, J. Fleischer, and O. V. Tarasov. . , C60:287–302, 1993.
K.G. Chetyrkin, A.L. Kataev, and F.T. Tkachev. . , 174:345, 1980.
K.G. Chetyrkin and F.T. Tkachev. Integration by parts: [T]{}he algorithm to calculate $\beta$-functions in 4 loops. , 192:159, 1981.
O. V. Tarasov. . , 54:6479, 1996.
S. Laporta. , 15:5087, 2000.
Y. Schroder and A. Vuorinen. . , 06:051, 2005.
Matthias Steinhauser. . , 364:247–357, 2002.
K. Kajantie, M. Laine, K. Rummukainen, and Y. Schroder. . , 04:036, 2003.
K. G. Chetyrkin, J. H. Kuhn, and C. Sturm. . , C48:107–110, 2006.
K. G. Chetyrkin, J. H. Kuhn, and C. Sturm. . , B744:121–135, 2006.
M. Faisst, P. Maierhoefer, and C. Sturm. . , B766:246–268, 2007.
K. G. Chetyrkin, M. Faisst, C. Sturm, and M. Tentyukov. . , B742:208–229, 2006.
R. N. Lee, A. I. Milstein, I. S. Terekhov, and Savely G. Karshenboim. . , 85:541, 2007.
R. N. Lee, A. I. Milstein, I. S. Terekhov, and Savely G. Karshenboim. . , A71:052501, 2005.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present high-resolution optical [*HST*]{} Space Telescope Imaging Spectrograph (STIS) observations made with two slits crossing four of the optically brightest starburst clumps in the vicinity of the nucleus of M82. These provide H$\alpha$ kinematics, extinction, electron density and emission measures. From the radial velocity curves derived from both slits we confirm the presence of a stellar bar. We also find that the super star cluster M82-A1 has a position and radial velocity consistent with it being at the end of one of the unique $x_{2}$ bar orbits formed by an inner Lindblad resonance. We derive a new model for the orientation of the bar and disc with respect to the main starburst clumps, and propose that clump A has formed within the bar region as a result of gas interactions between the bar orbits, whereas region C lies at the edge of the bar and regions D and E are located further out from the nucleus but heavily obscured. We derive extremely high interstellar densities of 500–900 , corresponding to ISM pressures of $P/k \approx 0.5$–$1.0\times 10^{7}$ K, and discuss the implications of the measured gas properties surrounding the nuclear star clusters on the production and evolution of the galactic wind. Despite varying pressures, the ionization parameter is uniform down to parsec-scales, and we discuss why this might be so. Where the signal-to-noise (S/N) of our spectra are high enough, we identify multiple emission-line components. Through detailed Gaussian line-fitting, we identify a ubiquitous broad (200–300 ) underlying component to the bright H$\alpha$ line, and discuss the physical mechanism(s) that could be responsible for such widths. We conclude that the evaporation and/or ablation of material from interstellar gas clouds caused by the impact of the high-energy photons and fast-flowing cluster winds produces a highly turbulent layer on the surface of the clouds from which the emission arises.'
author:
- 'M. S. Westmoquette'
- 'L. J. Smith'
- 'J. S. Gallagher III'
- 'R. W. O’Connell'
- 'D. J. Rosario'
- 'R. de Grijs'
bibliography:
- '/Users/msw/Documents/work/Thesis/thesis/references.bib'
nocite:
- '[@westm07a]'
- '[@westm07a; @westm07b]'
title: '*HST*/STIS spectroscopy of the environment in the starburst core of M82'
---
Introduction {#sect:intro}
============
Galaxies undergoing starburst events, triggered by interactions or mergers, are important objects to study in the local universe because of the insights they provide into the violent star formation processes that occurred in galaxies at much earlier epochs. M82 is the best-studied nearby [3.6 Mpc; @freedman94] example of such a galaxy. The current starburst is thought to have been triggered by a tidal interaction with M81 a few $\times$ $10^8$ yrs ago [@yun_ho_lo94], but investigations of this starburst are hampered by the galaxy’s almost edge-on inclination [$i \sim 80^{\circ}$; @lynds63; @mckeith95]. Much of the starburst core therefore suffers from heavy extinction along the line-of-sight, so most observations have concentrated on wavelengths beyond the visible domain [e.g. @rieke80; @achtermann95; @satyapal95; @satyapal97; @forster01].
@oconnell78 identified a number of high surface brightness clumps or regions in M82, denoted A, C, D and E that define the optical starburst core, and cover an area of $\sim$500 pc in diameter. However, the study of starbursts at resolutions attainable with the [*Hubble Space Telescope (HST)*]{} has revealed the existence of hyper-luminous compact star clusters, so-called super star clusters (SSCs), within sites of intense star-formation. With [*HST*]{} Planetary Camera imaging, @oconnell95 identified over one hundred candidate SSCs within the visible starburst, while more recently @melo05 catalogued a total of 197 young massive clusters in the starburst core with [*HST*]{}/WFPC2 observations, and they associated 86 with region A alone. This incredible density of star clusters gives rise to a very unusual, highly energetic environment, and is very different to anything we find in our own Galaxy or even in other less intense starbursts.
In a companion paper [@smith06 hereafter ], we describe observations of an isolated SSC in region A, designated M82-A1, and find it to be surrounded by a surprisingly compact H region ($r = 4.5\pm 0.5$ pc compared to the cluster radius, $r = 3.5\pm 0.5$ pc) at an unusually high pressure ($P/k=1$–$2\times 10^{7}$ cm$^{-3}$ K). The cluster H region was found to have an ionization parameter of $\log U = -2.24\pm 0.18$, in excellent agreement with the IR study of @forster01. These authors found, despite the seemingly chaotic and rapidly varying nature of the ISM, a uniform $\log U \approx -2.3$ for scales ranging all the way from a few tens to 500 pc, leading them to suggest that a similar star-formation efficiency and evolutionary stage exists across the whole starburst region. Our observation of a similar ionization parameter extends the uniformity of the starburst conditions down to scales of only a few parsecs. An alternative model presented by @dopita02 associates a constant observed $U$ with situations where dust effectively competes with gas in absorbing ionizing photons. We briefly discuss this issue in Section \[sect:emission\_meas\]. In , we discussed how such a compact, high density H region should theoretically not exist at the estimated age of the cluster ($6.4\pm 0.5$ Myr) if it had evolved according to the standard model for a pressure-driven bubble. We found we could explain its existence through the high ambient pressures resulting in non-standard evolution.
High ambient interstellar pressures have also been inferred from measurements of the size of supernova remnants (SNR) in the core of M82. Radio studies have shown that they are unusually compact [radii $<$4 pc; @muxlow94; @pedlar99]. This suggests that, together with the results mentioned above, the densities and pressures are the dominant drivers of the H region evolution and that pressure in the starburst core is a major factor influencing the starburst evolution [see also @rigby04].
M82 hosts a stellar bar that is thought to have formed during its last encounter with M81 [@telesco91; @wills00; @greve02a]. @wills00 developed a model of the M82 bar system by comparing the predictions of their analytical model to neutral (H, CO) and ionized (\[Ne\]12.8$\mu$m) gas observations. They found that the ionized gas is dominated by a faster rotating component that they identify with so-called $x_{2}$ bar orbit family [thus corresponding to the ionized ring identified by @achtermann95], whilst the neutral and molecular gas appears more consistent with the slower rotating cusped $x_{1}$ orbit family [$x_{2}$-orbits are associated with an inner Lindblad resonance; @athanassoula92a]. @ab99 perform extensive hydrodynamical simulations of bars, concentrating specifically on the gaseous component, and find the gas in fact follows stream-lines that correspond only qualitatively to the $x_{1}$- and $x_{2}$-orbits. Detailed inspection shows that there are differences arising from the fact that in reality the gas stream-lines only loosely resemble the stellar orbit families. In @greve02a, K. Wills and her collaborators used the near-IR Ca$\lambda8542$ stellar photospheric absorption line observations from @mckeith93 to compare their bar model directly to stellar velocities rather than the previously used gas measurements. By making this comparison, @greve02a were able to estimate that the mass of the stars following $x_{2}$-orbits is $\sim$15 per cent of the total mass of the bar.
The densely packed star clusters, high densities and the presence of a bar give rise to unusual, highly energetic ISM conditions. @lord96 developed a model for the ISM in the central 700 pc of M82 based on far-IR spectroscopic observations of forbidden-lines [see also @forster01; @forster03]. They proposed that the observed emission can be explained as originating in multiple H regions and PDRs (photo-dissociation regions) mixed uniformly with the ionizing stars and clusters. They found that the clouds have a characteristic size, $r_{\rm cl} \sim 0.4$–1 pc (with a molecular core and extended ionized envelope), density, $\sim$ $10^{4}$ , and mass range, $M_{\rm cl} \sim 200$–3000 , and are highly pressured ($P/k \sim 3\times 10^{6}$ K). These states are maintained by a hot ($\sim$$10^{6}$ K), diffuse surrounding medium of equal pressure, supported by the influence of a large number of SN shocks. In order to maintain the observed level of ionization, @lord96 found the average separation of the clouds and ionizing sources had to be of order their size (1–7 pc). This agrees well with the average cluster separations catalogued by @melo05, and with high-resolution CO observations [@mao00; @weis01] that also show that the clouds are partly disrupted and dissociated. By modelling the observed forbidden-line profiles, @lord96 argued that the clouds are clustered in two ‘hot-spots’ of size $125\times 125$ pc located at $\sim$$\pm 15''$ either side of the nucleus, representing a cross-section through the molecular and ionized gas torus found to surround the nuclear bar [@larkin94; @shen95; @achtermann95; @weis01].
M82 exhibits one of the largest optically visible outflows or ‘superwinds’ in the local Universe. The outflow is centred on regions A and C [@shopbell98; @ohyama02], and is composed of a complex morphology of loops and filaments. Recent ground- and space-based narrow-band optical imaging have shown that, contrary to previous thought, the M82 outflow is comprised of many channels which seem to point back to the individual starburst clumps [@wills99 Gallagher et al., in prep.], suggesting that the wind energy is injected by multiple discrete sources (i.e. SSCs), rather than a monolithic starburst region [see e.g. @t-t03]. The study of the ISM environment in which SSCs exist (the starburst core) and from which the superwind is driven, can hope to give insights into the initial conditions for galactic wind formation.
Observations of the ionized gas in the central region of M82 at visible wavelengths with [*HST*]{} offer the potential of probing the starburst region on sub-arcsecond scales ($0\farcs1 \approx 1.8$ pc) and further quantifying all of the above issues. In this paper, we present spectroscopy obtained with the Space Telescope Imaging Spectrograph (STIS) of the ionized gas near the nucleus, including regions A and C. We measure the kinematics, densities and extinction of the ionized gas on parsec-scales in order to characterise the nature of the ionized gas in the M82 starburst.
The remainder of this paper is organised as follows: in Section \[sect:data\], we describe the data and the reduction methods employed, including a description of how we extracted spectra at regular intervals along both slits and fitted the resulting emission-line profiles. The results of our line fitting are described in Section \[sect:results\], together with a discussion of the kinematics and line dynamics. In Section \[sect:prop\_gas\] we present our analysis of the properties of the ionized gas, including our flux, extinction and electron density results; in Section \[sect:state\_ISM\], we go on to describe the physical state of the gas and compare our results to previous models of the ISM. In Section \[sect:SB\_struct\] we analyse the structure of the starburst and derive a new model for the orientation of the bar and disc with respect to the starburst clumps. In this section we also discuss what implications our results have for our understanding of the superwind. Our findings are summarised in Section \[sect:summary\].
Observations and Analysis {#sect:data}
=========================
{width="7.5cm"}
{width="7.5cm"}
{width="7.5cm"}
{width="7.5cm"}
{width="7.5cm"}
{width="7.5cm"}
{width="7.5cm"}
{width="7.5cm"}
As part of a larger programme devoted to spectroscopy of SSCs in M82 (GO 9117; P.I. O’Connell), we obtained *HST*/STIS spectra for two slit positions crossing the galaxy’s starburst core. Fig. \[fig:finder\] shows the position of the two STIS slits overlaid on an *HST*/ACS WFC F658N image (GO 9788; P.I. Ho). The slits ($52\times 0.1$ arcsecs) pass close to the “nucleus” (2.2 $\mu$m peak) and just above and below the dust lane that perpendicularly bisects the disc of M82. The first slit, which we will refer to as slit A1, intersects the core of region A and the southern edge of region C, and was centred on the brightest isolated cluster in region A (which we have designated M82-A1) with position angle PA = $229^\circ$. The second slit, slit B2, was positioned on a bright cluster towards the north-east of the starburst core with a position angle PA = $235^\circ$, and passes through portions of regions D and E. The coordinates and exposure times for each slit position are given in .
Description of 2D spectra {#sect:desc_spectra}
-------------------------
Spectra were taken with two gratings giving coverage from 2900–5700Å with the G430L and 6295–6865Å with the G750M setting with a spatial resolution of $0.05$ arcsec pixel$^{-1}$. A detailed description of the data reduction procedure is given in . The spectral resolution of the data was measured for the G430L data by taking the average FWHM of Gaussian fits to the unresolved H$\beta$ line at various points along the length of the slit where extended H$\beta$ emission is observed. We found the resolution to be 2.5 pixels (in good agreement with the quoted value of 2–3 pixels given in the STIS Data Handbook for extended objects), corresponding to 6.7 Å in the G430L grating and 1.4 Å in the G750M grating. Since variations in slit illumination and of the source structure at each point could affect the resolution slightly, it is possible that some of the narrowest lines we measure (uncorrected FWHM $\le$ 75 ) may be unresolved. The material covered by the majority of the north-easternmost portion of both slits is very faint (see Fig. \[fig:finder\]), and not detected in our observations. The extinction in this region as measured by @satyapal95 and @alonso03 may be very high, A$_{V} \sim 10$–15 mag. In spectra where we do detect emission, we observe the nebular emission lines of H$\beta$, \[O\]$\lambda 5007$, \[N\]$\lambda\lambda 6548,6583$, H$\alpha$ and \[S\]$\lambda\lambda 6716,6731$.
Fig. \[fig:2d\_slit\] shows the H$\alpha$ and \[N\] spectral region of the G750M 2D spectral images for slits A1 and B2. Both 2D slit images are able to share the same $x$-axis since their position angles (PAs) are sufficiently similar (scale is given in arcseconds offset from the position of M82-A1 – marked with a dashed line on Fig. \[fig:finder\]). The spatial extent of the slit images is also restricted to show only the nuclear regions where emission is actually detected (represented by the white shaded section of the slits in Fig. \[fig:finder\]). We see emission arising from both discrete sources and holes in the foreground dust screen. Bright emission from the ambient interstellar gas is also present.
Fig. \[fig:2d\_slit\] clearly shows how the emission line profiles vary continuously along the length of the slit. It would be ideal to track these variations on a pixel-to-pixel basis, but in order to make accurate measurements of the gas conditions, the spectra need to have a sufficiently high signal-to-noise (S/N) ratio. Therefore we summed the spectra using multiple-pixel apertures. To measure the kinematics of the gas, observations of a strong emission line (e.g. H$\alpha$) are all that is required. However, to derive gas properties that require information from additional, fainter lines, a higher S/N was needed. To satisfy these two criteria, we extracted two sets of spectra from each of the 2D images. A high S/N set for each slit was made by stepping along in 20 pixel intervals ($\equiv$$1\farcs0$), extracting apertures of width equal to this interval (hereafter referred to as set A20), and resulted in spectra from slit A1 with an average S/N of 10 in H$\beta$ (G430L grating) and 20 in H$\alpha$ (G750M grating). In order to maximise spatial-resolution for the kinematics measurements whilst maintaining sufficient S/N for accurate line fitting, we decreased the extraction widths to 10 pixels ($\equiv$$0\farcs5$) separated by intervals of 10 pixels to create the second data-set (hereafter referred to as set A10). Extracted spectra were given sequential IDs starting from the north-east of each slit to identify them in later analysis. No background subtraction was attempted for either of the slit A1 or B2 datasets due to the bright and variable nature of the diffuse emission in the starburst core.
Fig. \[fig:spec\_plots\] shows four representative spectra from regions A, C, D and E illustrating the variation between these clumps. The arrows on Fig. \[fig:2d\_slit\] mark the position that each spectrum shown in Fig. \[fig:spec\_plots\] was extracted from, where the numbers indicate the spectrum IDs referred to above, and correspond to the individual plot titles in Fig. \[fig:spec\_plots\].
Cluster distributions and approximate ages within the clumps {#sect:ews}
------------------------------------------------------------
The clumps of emission seen in Figs \[fig:finder\] and \[fig:2d\_slit\] represent a densely packed mixture of individual clusters and their surrounding H regions, ionized gas clumps and bright diffuse ISM seen through holes in the foreground dust screen, all with sizes and separations on the order of a few to a few tens of pc. The extent to which these phenomena dominate the physical nature of the spectra along the length of the slits varies dramatically.
In the cores of the starburst clumps, the slits cross a number of individual clusters (see e.g. , figure 1 inset). We have experimented with extracting spectra close to individual sources catalogued by @melo05, but unfortunately the slits to do not pass directly through the centre of any isolated clusters, making the extraction of unique spectra difficult. Those that we did extract were of too low S/N to allow accurate model fitting. Measuring the equivalent width (EW) of the H$\alpha$ line in these core regions can, however, give an approximate indication of age trends in the stellar population by comparing to evolutionary synthesis models. It must be borne in mind, however, that since there is no way that we can measure the light from individual clusters, effects such as crowding, slit losses, and the presence of bright ambient diffuse emission will strongly affect the meaning of our EW measurements.
By measuring the H$\alpha$ EW along the central core regions of both slits and using the EW predictions of a [starburst99]{} [[sb99]{}; @leitherer99], $10^6$ model with a Kroupa IMF formulation, we find that the cluster ages in regions A, C, D and E lie between 5.5 and 6.5 Myr (log(EW) $\approx$ 1.5–2.5). If, however, a significant fraction of the Lyman continuum photons are absorbed by dust [@dopita02; @dopita06a and discussed in Section \[sect:emission\_meas\]], then cluster ages determined in this way will be overestimated. Nevertheless, these values are consistent with previous age estimates of clusters in the starburst core [@forster03; @smith06].
Line Profile Fitting {#sect:fitting}
--------------------
The S/N and spectral resolution of the 10-pixel wide spectra are high enough to resolve multiple components in H$\alpha$ in many regions along the slit. In general, a bright, narrow component is superposed on an underlying broad, fainter component, but obviously the identification of the broad component is dependent on the S/N of the particular spectrum. The component profiles are best approximated by a Gaussian function since the line broadening mechanism is dominated by Doppler effects caused by the gas temperature or turbulent state.
Examples of H$\alpha$ line profiles from different parts of regions A, C, D and E are shown in Fig. \[fig:ap10\_hafits\] together with the Gaussian profiles needed to model the integrated shapes (see below). Again, the position along the slit that each H$\alpha$ profile has been extracted from is marked by numbered arrows on Fig. \[fig:2d\_slit\]. These were chosen to represent a range in the variety of line shapes seen: bright, narrow emission lines with underlying broad components; double-peaked lines with no detectable underlying broad emission; and mixtures of the two. Where a broad component is present, it is often, but not always, redshifted compared to the narrower component. Double-peaked lines with velocity differences of $\sim$50–100 are seen in both the cores of region A and C.
We have used the general-purpose IDL-based curve-fitting utility, <span style="font-variant:small-caps;">pan</span> [Peak ANalysis; @dimeo], to automate the Gaussian fitting process of each of the extracted spectral lines. Briefly, <span style="font-variant:small-caps;">pan</span> embodies an interactive environment for specification of a profile’s initial guess parameters, and uses a $\chi^{2}$ minimisation algorithm to optimise the fit to the data. A more detailed description of the program and how we have customised it for our use is given in Westmoquette et al. (2007a).
Each line profile was fit with the appropriate number of Gaussian components, giving a measurement of the radial velocity, flux, and FWHM for each component. The lowest S/N lines could only be fit with a single Gaussian. However along most of the slit where emission is detected, the line profile shape distinctly contains more than one component. In cases where a double-component fit was most appropriate (determined from a comparison of the $\chi^{2}$ fit value to the single component result, and by visual inspection), the initial guess was always made setting the first Gaussian as the narrow component (hereafter referred to as *c1*), and the second Gaussian as the broader component (hereafter *c2*); in the case of a double peak, the bluest component was assigned as *c1* and the redder one as *c2*. For a few cases a triple-component fit was necessary, and the additional component (*c3*) was always specified with a initial guess assigning it to a supplemental narrow line at the same wavelength as the main narrow line. This consistent approach helped limit the confusion that might arise during analysis as to which Gaussian fit belonged to which component of the line, as well as aiding the $\chi^{2}$ minimisation process employed by <span style="font-variant:small-caps;">pan</span>. This assignment convention means that wherever an underlying broad component exists, we always refer to it as *c2*.
Kinematics of the ionized gas {#sect:results}
=============================
There have been a number of previous studies aimed at measuring the characteristics of the gaseous component in the disc and wind of M82 [e.g. @gotz90; @heckman90; @mckeith93; @mckeith95; @shopbell98]. From these studies, it is apparent that the wind outflow, as defined by two well-separated ($\sim$300 ) velocity components, is observed out to distances greater than 1 kpc from the nucleus along the minor axis. The disc of M82 is inclined at an angle of $\sim$80$^{\circ}$ [@lynds63] such that the near side of the disc is projected on to the north-west side of the nucleus and the nuclear regions, including clumps A, C, D and E, are observed from underneath through the southern side [although this has been disputed in the past; @wills00]. We therefore directly view the roots of the wind near regions A and C. The two STIS slits are inside the 300 pc wind injection zone [@shopbell98] where disc material is being entrained into the flow, and likely sample much of the inter-cluster material within the main starburst clumps.
Radial velocity variations {#sect:results_radvel}
--------------------------
The measured radial velocities of the H$\alpha$ emission-line components from Gaussian fits to the A10 spectrum-sets are shown in Fig. \[fig:ap10vel\], and are plotted with respect to the systemic velocity of M82 [$v_{\rm sys} = +200$ ; @vaucouleurs91; @mckeith93]. We have again plotted the datasets from both slits together. As described in the above section, in general the H$\alpha$ line profile in the brightest parts of all four regions is composed of a bright, narrow component superimposed on a broad, fainter component with a FWHM $\sim$ 100 . At a few positions, we see a third component (*c3*). Error bars are shown for points to the north-east of M82-A1 and for all *c3* detections. For most of the other points, the error bars are smaller than the size of the plotting symbol.
The brightest component, *c1*, is expected to trace the velocity of the densest (i.e. bulk of the) ionized gas in M82. From Fig. \[fig:ap10vel\], the slit A1 velocity curve for *c1* reaches a maximum velocity of $\approx$+100 at an offset of $\sim$$+2''$, flattens out between $-2''$ to $-7''$ coincident with the conspicuous dust-lane bisecting the galaxy disc [see Fig. \[fig:finder\]; also @oconnell95], then begins to rise again beyond $-7''$. The velocity of M82-A1 , together with two points either side of it, are redshifted by $\sim$40 from the general trend. On the opposite side, the rotation curve shown by *c1* reaches a maximum blueshifted velocity of $\approx$$-130$ then begins to decrease again beyond $+20''$. At these offsets, however, the A2 slit is farthest from the major-axis and begins to sample the inner regions of the wind – we return to this point below. We measure a *c1* H$\alpha$ velocity gradient for the region nearest to the nucleus of 12 arcsec$^{-1}$, in excellent agreement with the measurements of @mckeith93 and @shopbell98 of 11 arcsec$^{-1}$. The radial velocities of the *c1* gas in slit B2 follow the same velocity trend as the gas sampled along slit A1, but with a shallower gradient of 7 arcsec$^{-1}$. We discuss what these measurements tell us about the rotation curve in Section \[sect:SB\_struct\].
The broad underlying component, *c2*, largely follows the *c1* velocity curve although there are some deviations. In the core of region C, *c2* is redshifted by up to 60 compared to *c1*. Here we detect a third component (*c3*), which is redshifted by a further 30–40 , and may represent the front side of an expanding structure and possibly the roots of the wind flow. Further along the slit, to the south-west of region C, *c2* becomes blueshifted by $\sim$40 with respect to *c1*, and there are two detections of a 40–80 blueshifted third component. Again, these could be signatures of expanding structures and/or ordered flows. An unusual situation occurs at $\sim$$+4''$ offset (right in the core of region A), where *c1* is redshifted compared to its expected velocity, and it is *c2* that instead conforms to the trend. Here we also detect a third component, blueshifted by $\sim$80 compared to *c1*, which might explain the unexpected velocities if they are both two halves of an expanding shell where the faintest component is the nearest. In slit B2, we can only detect *c2* in the brightest parts of regions D and E, and find that their velocities are similar in all cases to that of *c1*, except in the core of region E where there is evidence of a 20–40 offset. Critically, however, none of the radial velocity differences measured equal the galactic wind velocity shifts of $\sim$300 . This evidence suggests that we are observing inside the injection zone where organized flow velocities are low [consistent with @shopbell98] and/or that the flows at this point are oriented in the plane of the sky (i.e. in a poleward direction).
Line widths and components {#sect:results_fwhm}
--------------------------
Fig. \[fig:ap10fw\] shows the FWHM of the individual H$\alpha$ line components for both slits, corrected for the instrumental resolution (1.4 Å) but not thermal broadening. As mentioned in Section \[sect:data\], small variations in the spectral resolution due to changes in the slit illumination/source structure could mean that the narrowest lines are only marginally resolved. The large scatter in *c1* FWHM values (left-hand panel) across the two slits can be partly explained by a contribution from multiple unresolved kinematic components. We have attempted to show this by highlighting *c1* profiles from both slits that only required a single Gaussian to fit with circled symbols: the FWHMs of circled points are, in almost every case, broader than the narrow component (*c1*) of a multiple-Gaussian fit. This is particularly obvious in the $-9''$ to $0''$ region of slit A1 where a second component cannot be fit at all due to the low S/N of the extracted spectra. It is very unlikely that these differences in width reflect real changes, so we conclude that multiple components are present everywhere, but the S/N is not sufficient in all cases to permit a more accurate fit to be made. This argument is strengthened by the observation that in every spectrum extracted from either slit where the H$\alpha$ line has a high enough S/N, we see evidence of a *c2*- or *c3*-type component.
To measure the following averages, we ignore all single-Gaussian-only fits, as these are assumed to be biassed by unresolved components. For slit A1, we find an average FWHM of $\sim$$60\pm 40$ for *c1* (whereas the average width of single-Gaussian-only fits is $\sim$$160\pm 40$ ). For *c2*, we find an average FWHM of $\sim$$210\pm 60$ , but in a few cases the width of *c2* approaches 300 . Where detected, the mean FWHM of *c3* is $\sim$50 , which is consistent with the mean line width of *c1*. In general, the FWHM of the gas in slit B2 is narrower: in region D the width of *c1* is as low as 20–30 , with a corresponding *c2* FWHM of 150–200 . In parts of region D and the area between regions D and E, only one component can be fit (shown by circled symbols), with a width of $\sim$100–140 . The line widths in region E are similar to region D, and in both regions *c2* is never $>$200 .
### Line broadening mechanisms
Fig. \[fig:ap10fw\] shows that all but a few line components are fit with Gaussians with widths $>$ 30 . Interestingly, the widths are systematically narrower in regions D and E (even including those that could only be fit with a single component), possibly indicating a more quiescent gas state. We will now explore the possible explanations for the observed widths of the narrow (*c1*) and broad (*c2*) lines in turn.
Supersonic line widths have been observed in many intense star-formation sites both in nearby galaxies (30 Dor: @chuken94 [@melnick99]; NGC 604: @yang96; NGC 2363: @roy92 [@g-d94]) and in more distant dwarf galaxies [@izotov96; @homeier99; @marlowe95; @mendez97; @sidoli06; @rozas06b]. In these studies, gravitational broadening through virial motions of ionized gas clouds [@melnick77; @terlevich81], and multiple unresolved kinematical components along the line-of-sight [@chuken94] have been proposed as explanations for the broadening of the brightest component (what we refer to as *c1*). A number of other mechanisms proposed, such as dust or electron scattering [@roy92], can be immediately disregarded on physical grounds.
The fact that we rarely measure FWHMs $<$ 30 across both slits indicates a global origin for a large fraction of the observed widths. A clear lower-limit to the width of *c1* observed in some studies of similar environments [@m-t96; @martinez07; @westm07a; @westm07b] supports this explanation. The only two broadening mechanisms that can work on a global scale are gravitationally induced virial motions and the stirring effects of wide-scale, intense star-formation on the ambient ISM. We therefore propose that the turbulent motions induced by these two mechanisms provide the observed base level of line-broadening present over the whole of the region observed. However, neither of these globally determined values can account for the large local variations in *c1* line-width above 30 up to 100 . The bright background and variable extinction lends support to the fact that an inevitable additional contribution must result from including multiple, superimposed, small-scale kinematical components along every sight-line, some of which may be of fairly high velocity, within the small-scale emitting regions in the core of M82 (Section \[sect:emission\_meas\]). We already alluded to this effect above, where we invoked the unresolved nature of the line profile to explain the increased width of the profiles where only a single-Gaussian was required for a satisfactory fit. In fact observations presented by @rozas06b show that multiple emission line components are very common in H regions, and can be interpreted as expanding shells driven by the mechanical energy input of SNe.
A ubiquitous broad underlying component (FWHM $\sim$ 150–250 ) is observed in both slits where the H$\alpha$ line is of sufficient S/N, which cannot be broken down into individual subcomponents. Like the phenomenon of the supersonic narrow component, broad underlying profiles have also been observed in many intense star-formation sites [see in particular @melnick99], but to our knowledge have never been studied in detail in the central regions of M82. Observations of the broad component in other systems are less well constrained, particularly for more distant galaxies, so the proposed explanations cited in the literature are more varied. These include: rotational effects, direct observations of stellar/cluster winds [@g-d94]; the effects of SN remnants [SNRs; @roy92; @izotov96]; large-scale superbubble expansion and/or blow-out [@roy92; @marlowe95; @t-t97]; multiple resolved/unresolved expanding shells [@homeier99; @rozas06b]; champagne flows [@yorke84; @t-t00; @melnick99]; and ablation of gas from molecular clumps resulting from the impact of stellar/cluster winds [@melnick99].
Only one of these proposed explanations, that of galactic rotation, is not due to the effects of massive stars, and may be discounted by considering the following: the maximum radial velocity relative to the systemic velocity of M82, $v_{\rm rad}^{\rm max}$, is of the order 100 . If this were purely caused by rotation, then, when looking through one side of the galaxy, the line width would also be $\sim$100 . However, we see line widths of up to 300 , thus proving that the broad line component cannot result from simple rotation. In reality, the $v_{\rm rad}$ observed is due to large-scale orbital motion, not local gas rotation, and rotational broadening is likely close to negligible.
We can also discount direct observations of the stellar winds themselves from the spatial distribution of the line emission. Similarly we are not directly seeing supernova remnants (SNRs) as we do not find extensive evidence for shock spectra. The effects of large-scale superbubble blow-out are highly position dependent and would be expected to produce strong, clearly observable line-splitting, thus not making this a suitable explanation for the observed line broadening. Although multiple unresolved shells are very likely to contribute to the line broadening of *c1*, it is difficult to understand how multiple discrete kinematical components could produce a smooth, Gaussian shaped, 200–300 wide profile with no evidence of the discrete components themselves [see also @melnick99]. Champagne flows are not expected to produce outflow speeds much higher than the sound speed in the photoionized gas (10 for gas of $T\sim 10^{4}$ K), and are therefore not capable of producing the level of broadening we observe. Furthermore, many of the previous studies for which these mechanisms were proposed are of much larger physical regions than what we are observing, meaning that they may not be that appropriate to our $\sim$10 pc scale observations. This leaves only the final explanation: that of hydrodynamical evaporation/ablation of gas from the surface of illuminated gas clouds.
These arguments are strongly supported by results from an analysis of spatially-resolved spectroscopy of a similar, high-energy environment in the starburst galaxy NGC 1569, presented in Westmoquette et al. (2007a, 2007b). Here, we find striking spatial correlations between the width of *c2* and the intensity of *c1*, which strongly supports the hydrodynamic interaction mechanism.
To explain this mechanism, we can envision a situation whereby the energetics and dynamics of the brightest clumps are driven by an interaction with the high-energy, ionizing photons and fast-flowing wind from the surrounding star clusters, and that all the observed emission originates from a thin interface layer at the clump surfaces. As the gas clouds are impacted by hot, fast, cluster winds, thermal evaporation and/or mechanical ablation of gas from their surfaces produces a highly turbulent velocity field that pervades the whole region as the gas becomes entrained into the flow [@charnley90; @begelman90; @slavin93; @pittard06]. We explore how this conclusion relates to the galactic wind as a whole in Section \[sect:gal\_wind\].
@lord96 predict that the ionizing star clusters are well mixed with molecular matter in the form of clouds of characteristic sizes, $r_{\rm cl} \lesssim 1$ pc (Section \[sect:intro\]), thus providing copious sites from which gas can be evaporated and/or ablated. Thus, the fact that multiple PDR regions are predicted from far-IR spectroscopy, together with the observed close packing of the individual sources, the uniformity of the starburst conditions throughout the central regions, and the identification of a ubiquitous broad-line component, all point towards gas evaporation/ablation by cluster winds causing a highly turbulent velocity field being the root cause of the observed broad line widths.
Properties of the ionized gas {#sect:prop_gas}
=============================
Flux Variations {#sect:results_flux}
---------------
Fig. \[fig:ap10flux\] shows the distribution of the integrated H$\alpha$ line-flux in *c1* and *c2* along the section of the slits highlighted in white in Fig. \[fig:finder\]. Also plotted are the continuum flux levels (measured in the range 6620–6700 Å) for each slit, multiplied and offset for ease of comparison (factors are listed in the figure caption). The relative continuum flux levels are consistent with the *HST* F814W image, and clearly show the bright clumps A, D and E. These regions can also be seen in both the flux levels of *c1* and *c2*. In region A, the two peaks in *c1* correspond well to peaks in the continuum brightness, but at the location of M82-A1 (offset = $0''$), the H$\alpha$ intensity does not peak so strongly. This could be due to a spatial sampling effect if the continuum and H$\alpha$ intensities are offset by less than one spatial resolution element, but if the levels shown are correct, it would suggest that the ionizing flux emerging from the core of region A is higher than at M82-A1, possibly indicating that the age of region A is even younger than that of M82-A1 .
The flux in *c2* relative to *c1* indicates the significance of the *c2* emitting gas. *c2* is less than or equal to half the intensity of *c1* in the north-eastern half of region A and most of region C. *c2* has an equal intensity to *c1* in the south-western half of region A and in region D (but here the peak is spatially offset by $2''$). In region E, *c2* peaks at twice the brightness of *c1* (but again is spatially offset by $\sim$$2''$). A close examination of the line profiles at this position shows that *c2* has likely been misidentified, and should really be assigned to *c3*. However, the flux peaks are real, and are definitely offset from the continuum peak.
Interstellar Extinction {#sect:extinc}
-----------------------
The starburst core of M82 suffers from a high and non-uniform extinction. For example, several studies have found values for the visual extinction, $A_{V}$, of 2–12 mag from near-IR nebular line ratios (assuming a foreground screen model) [@satyapal95; @forster01]. In this analysis we adopt a foreground screen model for simplicity, although we realise that this is likely to be unrealistic considering the level of dust present in the M82 disc [values of $A_{V}\approx 20$–40 mag have been derived using mixed gas and dust models; @forster01]. One consolation is that at optical wavelengths, we can only see the least obscured part of the starburst, thus the effect of mixed dust will be fairly minimal. Using slits oriented along the minor-axis, @heckman90 also measure the extinction in the central regions of M82, and find an excess to the north of the nucleus up to a radius of $\sim$300 pc, and that in the south, the extinction falls rapidly to a small and constant value beyond $R\approx 300$ pc. These findings are consistent with the assumed inclination of the galaxy.
To calculate the reddening distribution of the gas we compared the observed flux ratio of H$\beta$/H$\alpha$ (assuming no underlying stellar absorption) to the theoretical case B ratio given by @humstor87, using the Galactic extinction law of @howarth83. Shown in Fig. \[fig:ap20ex\] is the extinction in magnitudes measured at 20 pixel ($1''$) intervals along the slits. The uncertainty in the measurement of the H$\beta$ line flux dominates the error bars shown, and indeed only four extracted spectra from slit B2 had detectable H$\beta$ emission. For the few extracted spectra from slit A1 where we detect H$\gamma$, we also derived $A_{V}$ from the H$\gamma$/H$\beta$ ratio. The results were consistent with the extinction values obtained using the H$\alpha$/H$\beta$ ratio, although the errors were large.
The average extinction in region A is $A_{V}=5.5$ mag, which falls to an average of 4 mag in region C, and the data are consistent with a further decrease past region C towards the south-west. The extinction towards M82-A1 is less than that towards the majority of region A, perhaps explaining why this cluster stands out so clearly in optical images. We can only measure the reddening for one point to the north-east of M82-A1, and although the errors are large, it is consistent with being at a much higher value than the surroundings. This coincides with the beginning of a region of complete obscuration at optical wavelengths, extending towards the north-east of the starburst core (the dust-lane). Extinction measurements of the diffuse gas in this region ($-10$ to $-2$ arcsecs) are therefore made impossible due to the very low S/N or absence of the H$\beta$ lines.
The extinction in slit B2 could only be measured for four points representing the brightest emission in the cores of regions D and E. In region D we find $A_{V}\approx 4$ mag, and is lower than that in region A. In region E, however, the extinction appears to be much higher at $A_{V}\approx 6$–6.5 mag.
Electron Density {#sect:dens}
----------------
We have measured the \[S\]$\lambda 6717$/$\lambda 6731$ line ratio by fitting Gaussian profiles to the emission lines in each of the A20 spectra for both slits. The distribution of derived electron densities, , (assuming an electron temperature, = $10^{4}$ K) is plotted in Fig. \[fig:ap20dens\].
In slit A1, we find the electron density to be highest around M82-A1 . \[S\] emission is not detected at all in spectra extracted from the north-east of M82-A1 (negative offsets), so in an attempt to derive an approximate average density for this region, we extracted a 60 pixel wide spectrum. However even summing over this many pixels did not yield a \[S\] detection, and therefore no density could be derived. From M82-A1, the density falls rapidly to $\sim$1000 in the centre of region A and to 600–800 cm$^{-3}$ in region C. Our observations provide constraints on the steepness of the density gradient from region A to A1: an enhancement of $\sim$ 1000 takes place over a projected distance of only 2.5 arcsecs (45 pc). In the area located between the two cluster complexes A and C, falls below the low-density limit ($<$100 ) for the \[S\] ratio method, and is indicated by upper-limit arrows in Fig. \[fig:ap20dens\].
Our measurements are lower than those of @oconnell78, who also measured $n_{\rm e}$ from the \[S\] doublet ratio for regions A and C, and found an *average* density of $n_{\rm e}=1800$ . The discrepancy may arise because at lower spatial-resolution, their data inevitably suffered from a higher degree of contamination and luminosity weighting than our high-resolution data, and were therefore biassed towards the densest gas. @heckman90 also measured the electron density from their minor-axis slits, and found an of $\sim$1000 at the nucleus, falling to $\sim$100–300 at lateral distances of 1–2 kpc, which is in good agreement with our measurements.
We measure a very different distribution of densities in slit B2: close to the centre of region D[^1], the electron density is as low as 300 , but quickly rises to over 2000 in the area between regions D and E. Here slit B2 passes just to the north of a smaller cluster complex (2–4 arcsecs north of the 2.2 $\mu$m nucleus; Fig. \[fig:finder\]), which if really surrounded by gas at this density would be a very stifled region indeed. Alternatively, this density peak could result from scattered light from the starburst core/nucleus, under the assumption that higher pressures are associated with more intense zones of star formation. In region E, the electron density is also high with values of $\sim$1300–1700 , similar to that measured near M82-A1.
Physical state of the ionized ISM {#sect:state_ISM}
=================================
Emission Measure {#sect:emission_meas}
----------------
In we discussed the surprising uniformity of the starburst over a large scale range by comparing our calculation of the ionization parameter of the M82-A1 H region to the results of the IR study by @forster01. We now use the observed H$\alpha$ fluxes to derive an estimate of the scale-length of the emitting regions to compare to and extend this discussion.
A characteristic scale-length can be derived from the emission-measure, which is defined by: $$F_{{\rm H}\alpha}^{\rm int} = \int n_{\rm e}^{2} \: \alpha_{\rm eff}^{{\rm H}\alpha} \: \frac{h \nu({\rm H}\alpha)}{4\pi D^{2}} \: \epsilon \: {\rm d}V,$$ where $\alpha_{\rm eff}^{{\rm H}\alpha}$ is the case B recombination coefficient for H$\alpha$ [$8.64\times 10^{-14}$ cm$^{3}$ s$^{-1}$; @osterbrock89], $h \nu({\rm H}\alpha)$ is the energy of an H$\alpha$ photon, $D$ is the distance to the source, and $\epsilon$ is the gas volume filling factor. To remove the distance dependence, and to re-arrange in terms of a scale-length, d$V$ can be substituted with d$A$d$\ell$, where d$A = A / 4\pi D^{2}$ $\equiv$ angular area on the sky ($0.1\times 1$ arcsec in this case). Integrating over the whole emitting volume gives the mean intensity of H$\alpha$, $$\big< \, F_{{\rm H}\alpha}^{\rm int} \, \big> = \big< \, n_{\rm e}^{2}\, \big> \: \alpha_{\rm eff}^{{\rm H}\alpha} \: \left<\epsilon\right> \: \ell \qquad \rm{erg~s^{-1}~cm^{-2}~arcsec^{-2}}.$$ Hence, in reality what we can measure is $\ell(\epsilon=1)$ for all three H$\alpha$ line components in slit A1. Since $\epsilon \leq 1.0$, this is a lower-limit to the size-scales of the emitting H columns.
We find $\ell(\epsilon=1)$ to be less than 100 pc at all points, and is as low as $\sim$few parsecs to a few tens of parsecs in the north-east half of region A and the majority of region C. The consistency between line components shows that the gas from which each is emitted has similar characteristic sizes. Within the south-western half of region A and the A–C inter-clump region, the scale-lengths are larger, rising up to $\sim$50 pc.
The roughly constant values of the ionization parameter, log $U \approx -2.3$, within the main M82 starburst zone suggests that dust could be successfully competing with gas for ionizing photons on the surfaces of clouds [see @dopita02]. In this case the @dopita02 model also suggests radiation pressure could be a significant factor. The luminosity densities in the M82 starburst zone, however, are well below those of the narrow line regions of AGN [e.g. @cecil02a] or found in ULIRGs [e.g. @thompson05] where models show that radiation acting on dust becomes important. Consistent with this view, we also find that the energy density in turbulent ISM motions exceeds the thermal pressure, and so even if radiation and thermal pressure are comparable, they are unlikely to be strongly influencing the M82 gas dynamics.
Comparison to ISM models {#sect_disc_ism}
------------------------
We can now discuss how our findings relate to the ISM models introduced in Section \[sect:intro\], and what that implies for the M82 starburst. In the above section we found that the characteristic scale-size of the emitting regions remains below $\sim$50 pc ($3''$) throughout regions A and C. The uniformity of the starburst conditions must therefore stem from the fact that the scale of the emitting regions remains very small [as suggested by @lord96], hence the conditions are able to respond rapidly in order for the starburst to maintain such a near-constant state. However, within this overall uniformity there are variations. The smallest characteristic sizes ($\sim$few parsecs) are found in or near the dense cores of the clumps, whereas the largest sizes are found in the inter-clump region. These differences imply that the most compact clouds are found in the clump cores where the star-formation is most intense and the gas pressures are high.
{width="7cm"}
{width="7cm"}
Diagnostic flux ratios derived from the most commonly observed nebular lines can give a further indication of the ionization parameter and the mean level of ionization within the M82 starburst. The ratios of \[S\]$\lambda$6717+$\lambda$6731/H$\alpha$ and \[N\]$\lambda$6548+$\lambda$6583/H$\alpha$ are primarily tracers of the ionization parameter of the gas [@dopita00; @dopita06b], and are shown in Fig. \[fig:diagnostics\] as a function of position along slit A1. We find that the \[S\]/H$\alpha$ ratio changes very slowly across regions A and C, from $\sim$$-1.8$ dex at M82-A1, to $\sim$$-1$ dex to the south-west of region C. For \[N\]/H$\alpha$, we were able to measure the ratio independently for both *c1* and *c2* in many places, but find no significant difference between the two (except for perhaps in the inter-clump region where *c2* is consistently higher than *c1* for seven consecutive points). Again, the ratio stays roughly constant across regions A and C at an average of $-0.2\pm 0.5$ dex. As a comparison, @shopbell98 measure equally high \[N\]/H$\alpha$ ratios ($-0.5$ to $-0.2$ dex) in the inner collimated zone for their low velocity component. Having access to the bluer nebular lines of \[O\] and H$\beta$ means that we can use their ratio as an indicator of the mean temperature ($T_{\rm eff}$) of the ionizing sources [@veilleux87; @dopita00]. Unfortunately though, our ability to measure this ratio, like that of measuring the extinction, is limited by the low S/N of the H$\beta$ line. Where this line is detected, we find the flux ratio to remain constant within the uncertainties at log(\[O\]/H$\beta$) $\approx$ $-0.5$ over both regions A and C.
@rigby04 discuss the effectiveness of certain optical and IR nebular line ratios, including \[O\]/H$\beta$, at tracing the mean temperature of the ionization source by comparing the consistency of the indicators between themselves and to models. They suggest that finding ratios that imply a low mean $T_{\rm eff}$ in starburst galaxies [e.g. @thornley00 this work] could result from either a lack of massive stars (M $\gtrsim$ 40 ), or more likely, from the fact that the ultracompact H region phase is prolonged by the high densities and pressures in starburst ISMs. Under the influence of the high pressures, the most massive stars within the clusters may spend a significant proportion of their lifetimes embedded within the dense, highly-extincted [$A_{V} \lesssim 50$ mag; @hanson02] regions in which they form, and they would remain undetectable, even through re-processed nebular optical or IR emission. Although their bolometric output may contribute to the surrounding gas for their entire main-sequence lifetime, it may influence nebular excitation indicators for a smaller fraction of this time [$\sim$85 per cent; @rigby04].
This situation agrees with what we know about the environment in M82, including its remarkably homogeneous ionization parameter [@forster01] and our detection of the dense, compact H region surrounding M82-A1 . Clearly not all clusters can be deeply embedded, since in many cases we can actually see the starlight from the cluster directly. However, the consistency in the line ratios (tracing nebular excitation and $T_{\rm eff}$) across the starburst regions must result from both the highly fractious nature of the ISM and the uniformity of the escaping radiation field. This radiation field must be attenuated by the high density gas to an extent to which it no longer reflects the presumably very inhomogeneous distribution of high-energy photons from the most massive stars within the clusters.
Structure of the M82 starburst {#sect:SB_struct}
==============================
Evidence of the bar {#sect:bar}
-------------------
We have used the high spectral- and spatial-resolution of our observations to track the kinematics of the H$\alpha$ line components across the two slits in unprecedented detail. We now use these measurements to provide constraints on the orbital parameters of the ionized gas, and the location of the individual starburst clumps within the galaxy.
As discussed in Section \[sect:intro\], @wills00 developed a model for the M82 bar system by analysing velocities derived from a number of neutral and ionized emission lines. They found the best-fitting bar model to have a peak radial velocity projected onto the sky of 140 , an angular velocity of 217 arcsec$^{-1}$, a total length of 1 kpc, a core radius of 25 pc, but assume an opposite inclination to that which is commonly accepted (i.e. $-80^{\circ}$). We now compare the results of their model to our observations to determine if we see evidence for any of the orbit families. In Fig. \[fig:ap10vel\_mckeith\], we plot the observed major-axis Ca stellar absorption and \[S\]$\lambda$9069 and Pa(10) nebular emission-line velocities from @mckeith93 and our H$\alpha$ *c1* and *c3* velocities [see also @greve02a]. The $x_{2}$-orbits, as traced by the stellar Ca measurements, can clearly be seen departing from the velocities determined from the gas emission-lines at offsets of $\sim$$+13''$ and $\sim$$0''$. Interestingly, M82-A1 appears to be located near the very top of the $x_{2}$-orbit pattern, unambiguously separated from the $x_{1}$-orbits. In other parts of the plot, the close match between the IR and optical data-sets indicates that the H$\alpha$-emitting gas follows the stellar rotation closely, although the rapidly varying extinction introduces inevitable complications. The inflection points at $\sim$0 and $\sim$+17$''$ represent the intersection between the $x_{2}$- and $x_{1}$-like orbits, where the gas is expected to shock and begin its journey towards the nucleus. Region C is clearly not associated with the $x_{2}$-like orbits, implying that it may be located within the molecular ring of material beyond the bar.
Locating the starburst zones {#sect: }
----------------------------
Bearing in mind that the @wills00 bar model assumed the opposite inclination to that which is generally accepted, we can now re-asses how the bar might be oriented with respect to the line-of-sight, and to the individual clumps. The @lord96 ISM model predicts a concentration of molecular clouds (PDR regions) at either end of the stellar bar, which may be attributed to the interaction of bar $x_{1}$- and $x_{2}$-orbits causing the build-up of dust along the leading $x_{1}$-orbits [a phenomenon sometimes termed ‘spraying’; @athanassoula92b; @wills00]. Fig. \[fig:degrijs\_CO\] shows contours of integrated CO(J=1$\to$0) line intensity from @weis01 overlaid on an *HST* broad-band colour-composite, clearly showing how the peak of the CO emission corresponds to the location of the dust clouds either side of the nucleus and the proposed position of the two PDR hot-spots [@lord96].
We can therefore associate this pile-up of molecular material/dust at either ends of the bar with the highly obscured regions to the north-east of clumps A and D and between clumps C and E extending out to the south-west. Our measurement of a rapid increase in extinction [also found by @satyapal95; @alonso03] and electron density to the north-east of regions A and D (see also Fig. \[fig:degrijs\_CO\]), and the high extinction and density in region E and the D–E inter-clump region, is consistent with this conclusion. Evidence for a rapid fall-off in extinction to the south-west of region C may indicate the torus edge in this direction (Fig. \[fig:ap20ex\]). The size and location of these dust clouds is consistent with the predicted bar length ($\sim$1 kpc), and the comparative levels of extinction and prominence of the north-eastern dust cloud (again see Fig. \[fig:degrijs\_CO\]) would imply that this is on the near-side of the starburst. This dust cloud also appears to wrap around region D and extend up into the northern inner-wind region – whether or not this dust is being entrained into the wind flow at this point would require detailed kinematical observations.
In the region between clumps A and C, we do not measure a discontinuity in the extinction, but we do find a sudden drop in the electron density (Fig. \[fig:ap20dens\]). This suggests that there is a physical ‘hole’ in the gas distribution between these two clumps and that the two regions are physically distinct, rather than the decrease in surface-brightness being due to an obscuring dust-lane. The remaining gas in this ‘void’ must still be kinematically associated with the two clumps [confirming the suggestion of @oconnell78] since there is no velocity discontinuity here. We can therefore now associate region C, located to the south of the bar region, as being outside the $x_{2}$-orbit region, and part of (or a result of star-formation within) the molecular torus surrounding the bar.
The fact that we can see evidence for both ends of the bar indicates that we are able to sample light from a considerable fraction of the M82 system. Contrary to previous models of the M82 system that assume large fractions of the M82 system are optically invisible [e.g. @rieke93; @satyapal97], the main hidden region appears to be only in the middle of the bar near the nucleus. The most recent and perhaps most intense star-forming sites are located inside compact clouds spread throughout the system but concentrated in the torus surrounding the bar.
The fact that M82-A1 appears to be located at one end of the $x_{2}$-orbits places a strong constraint on the orientation of the bar within M82, and indicates that region A may have formed as a result of intense star-formation in the intersection between the $x_{1}$- and $x_{2}$-orbit families. The shallower velocity gradient between regions D and E indicate that they are located at a larger radius than that of regions A and C, but the distinction between these two regions is less clear than that between clumps A and C. Comparing Fig. \[fig:finder\] to Fig. \[fig:degrijs\_CO\] shows that there is a great deal of extended emission and variable extinction in this inter-clump region, and the confused and variable velocities in this region indicate that we are seeing glimpses of star clusters through holes in a foreground dust screen. However, the fact that we find narrower line widths in regions D and E compared to A and C (Section \[sect:results\_fwhm\]) supports the scenario of them being located at larger radii, since narrower line widths would indicate a more quiescent gas state, consistent with material further away from the nucleus.
Fig. \[fig:bar\_orientation\] shows a schematic of our proposed spatial configuration of the starburst clumps and dust lanes in relation to the orientation of the bar orbits. This model implies that the eastern side of the bar is more distant, and that it extends from behind the dense dust cloud to the east of region A to just past region E. This orientation also implies that the eastern side of the inner $x_{2}$-orbits, where M82-A1 is located, is nearest.
Implications for the production of the galactic wind {#sect:gal_wind}
----------------------------------------------------
In this section, we will use our observations together with the discussion given in the previous two sections to consider what implications our findings may have on how the superwind is produced in the starburst core. We will first consider how our spatially-resolved measurements of the gas density may affect the individual cluster winds, then go on to discuss possible explanations for the width of the observed line profiles, and finally how the dynamics of the broad component may indicate that we are seeing the roots of the large-scale galactic superwind.
### Densities and pressures
In , we explored how the unusually high density of the gas surrounding M82-A1 has stifled the growth of its H region, and commented how it is hard to understand the development of the superwind if this cluster is representative of the entire starburst. We are now in a better position to know what the representative conditions are, and how these might vary over the face of the starburst. It is clear from high-resolution *HST* images of the galactic wind (Gallagher et al., in prep.) that the outflow *is* being driven directly from the central clumps (region A in particular), but our measured densities of the gas within these cluster complexes are still very high (500–900 , or $P/k \approx 0.5$–1.0$\times 10^{7}$ K; ) compared to standard Galactic ISM values. However, they are less than half of that found near M82-A1, and this highlights a possibly significant difference between a relatively isolated cluster such as M82-A1 and the densely-packed clusters in the clump cores. It appears that instead of having collimated winds from individual SSCs, the cluster complexes (clumps) produce a high pressure zone, or an ’energy injection centre’, and the winds result from hot gas expanding out of these zones, upwards into the halo. This scenario will be explored in more detail by Gallagher et al. (in prep.).
### Dynamics of the broad line component
We can now turn our attention to whether, by observing this broad, underlying component to the H$\alpha$ line, we are indirectly detecting the roots of the superwind flow. In our proposed hypothesis, *c2* represents gas that has been stripped off the surface of gas clouds embedded in the starburst, so in principle the radial velocity offset between this component and the brighter (hence presumably denser) *c1* gas can be used to trace the flow of the hot superwind into which it is being entrained.
Slit A1 passes through the very core of region A, and even though here we see some of the broadest line widths (up to 300 ), we do not see any velocity offset between the components, indicating that at this point there is still no preferred outflow direction. In region C, the slit passes through the southern side of this clump, rather than its core. Already at this radius, some evidence of an ordered wind flow is detected in the kinematical studies of @mckeith95 and @shopbell98. These authors find that within 200 pc south of the nucleus the H$\alpha$ components are separated by $\sim$50 . However, this is still within the chaotic zone of the outflow, and certainly well within the energy injection zone, where it is thought the wind flow is still subsonic but rapidly accelerating [@shopbell98]. In this region we find that *c2* is redshifted with respect to *c1* by $\sim$40 and we also detect an $\sim$80 redshifted *c3* profile (see Fig. \[fig:ap10vel\]). These values are consistent with the aforementioned studies, and show that here we are beginning to see an ordered flow with a preferred direction, and that *c2* is tracing the wind roots.
The region beyond clump C is interesting in a number of ways: here the *c1* velocities rapidly decrease back to $v_{\rm sys}$, departing from the major-axis velocity measurements of @mckeith93, and the offset between *c1* and *c2* reverses direction, becoming blueshifted by $\sim$40 relative to *c1*. This far end of the slit is coincident with lateral ionized streamers pointing towards the south-west (see Fig. \[fig:finder\]), indicating that this region may be part of a distinct flow, separate to the main wind. If clump C is located within the molecular torus surrounding the bar, then this region may be less influenced by the bar’s collimating effects, explaining the lateral flow.
Summary {#sect:summary}
=======
The use of high spatial resolution [*HST*]{} optical images and STIS spectra have enabled us to fill in a few more pieces of the large and complex jigsaw puzzle of M82. The data-set has allowed us to probe the ionized gas environment across the nuclear starburst in unprecedented detail. We have concentrated on mapping the ionized gas properties in the starburst core using two STIS slit pointings. One slit passes through the core of region A, and the southern edge of region C, whilst the other slit passes through part of region D and ends within region E.
- By performing a comparison of our radial velocity measurements to previous observations and models, we confirm the presence of a stellar bar with an inner Lindblad resonance (ILR), resulting in a set of unique family of $x_{2}$-orbits perpendicular to the bar major-axis. The radial velocity of M82-A1 is consistent with being on an $x_{2}$-orbit at one end of the perpendicular ellipse, implying that region A may have formed as a result of intense star-formation at the intersection of the $x_{1}$- and $x_{2}$-orbits [as predicted by models; @athanassoula92a].
- Since the interaction of $x_{1}$-orbits with the perpendicular $x_{2}$-orbits is predicted to lead to a build-up of gas and dust along the leading $x_{1}$-orbits [@ab99], we conclude that the prominent dust clouds to the north-east of regions A and D, and between regions C and E extending towards the south-west, must have been formed in this way.
- By comparing CO observations to an *HST* broad-band colour composite of M82, we identify that the peak of the molecular emission corresponds to the molecular torus surrounding the ionized ring [@achtermann95; @weis01], the proposed positions of the PDR hotspots from @lord96, and the prominent dust clouds located on either side of the nucleus. Since region C is clearly not associated with the $x_{2}$-orbits, we assume it to be located within this molecular torus.
- Our data are consistent with a physical distinction between clumps A and C, but gas between the clumps follows the same radial velocity pattern as the clumps, suggesting that all of this material is in the rotating disk of M82.
- A shallower velocity gradient between regions D and E leads us to believe that they are located at a larger radius to clumps A and C, supported by the finding of narrower line widths in these regions. However, the strong, patchy foreground extinction indicates that we may be seeing these cluster complexes through gaps in a thick foreground dust screen.
- By measuring the \[S\] doublet ratio along the slits, we find the gas densities (pressures) in the individual clumps to be high (500–900 ; $P/k \approx 0.5$–$1.0\times 10^{7}$ K), but significantly lower than that measured for the gas immediately surrounding M82-A1. We find that instead of seeing collimated winds from individual SSCs, the cluster complexes (clumps) each appear to produce a high pressure zone, or an ’energy injection centre’, and that the winds result from hot gas expanding out of these zones. We will explore this scenario in more detail in a forthcoming contribution (Gallagher et al. in prep.).
- We extend the uniformity of the ISM conditions found by @forster01 down to parsec-scales by deriving the emitting scale-sizes and by measuring the ionization parameter sensitive line-ratios of \[S\]/H$\alpha$ and \[N\]/H$\alpha$. We find the size-scale of the emitting regions to be always less than 50 pc along slit A1, and that the smallest regions (few pc) are in or near the dense cores of clumps A and C, implying that the most compact clouds are found where the star-formation is most intense. Our results confirm the prediction of uniform conditions throughout the starburst resulting from the highly fragmented nature of the ISM and/or the importance of dust in absorbing a significant fraction of the Lyman continuum luminosity. Our data also support the well-mixed nature of the embedded ionizing sources and ionized gas.
- Our line-profile analysis has shown that the widths of all but a few line components are $>$30 . We have explored the possible mechanisms for producing the observed broad-line widths in both the narrow component (*c1*) and the broad component (*c2*). The most likely explanation for broadening of *c1* is a combination of gravitationally induced virial effects and the stirring of the ISM through the intense star-formation activity. A base level of turbulent broadening ($\sim$30 ) exists over the whole starburst region, whilst the scatter to broader widths (up to 100 ) is likely to result from the presence of multiple unresolved kinematical components (e.g. expanding shells) along the line-of-sight. The contribution from unresolved components is inevitable due to the presence of large numbers of densely packed SSCs.
- A ubiquitous broad component (150–250 ) is seen in all nebular lines with a high enough S/N. We argue that the broad line widths are produced by the interaction of high-energy, ionizing photons and fast-flowing winds from the star clusters with the cool gas clumps found throughout the starburst zone. As the wind impacts the surface of the clouds, evaporation and ablation of cloud material results in a highly turbulent velocity field which manifests as this broad component.
- In our model, the broad component represents turbulent gas stripped from the surface of gas clouds through the action of stellar winds and SN from the surrounding SSCs, so therefore could be used to track the hot superwind as the cooler material becomes entrained into the wind flow. We find no velocity offset between *c1* and *c2* in of region A, implying that here in the clump cores, the wind has yet to possess a preferred outflow direction. Where the slit passes through the southern edge of region C, we begin to see an ordered flow: *c2* has a consistent redshifted offset, and a further redshifted narrow component (*c3*) can be identified.
- Beyond clump C, the offset between the *c1* and *c2* velocities is reversed, and *c1* departs from the radial velocity pattern measured along the major-axis. Since these measurements are coincident with a number of lateral south-west pointing streamers, this has lead us to believe that this region may not be associated with the main collimated flow. If clump C is located within the molecular torus surrounding the bar, then this region may be less influenced by the bar’s collimating effects, explaining the lateral flow.
In conclusion, it seems a number of key ingredients are needed for the M82 starburst to exist. A central concentration of gas is required, together with clumps of dense star formation and high interstellar pressures. Of course, these three essential components do not exist in isolation, they link together and are to a larger or lesser extent the result of external influences. Large-scale galaxy encounters (which we know M82 has experienced) very often result in the formation of a bar, and a bar is a very efficient method to funnel gas into the nuclear regions in order to build up the required central gas concentration. Furthermore, both clumps and bars can be the products of disc instabilities that occur when discs are too dynamically cool i.e. gas rich. We can therefore reason that the properties of the M82 starburst are all inextricably linked to one another. The study of one will always require a study (or at the least a good understanding) of the others.
Furthermore, we can now explain both the highly fragmented nature of the ISM and the near-uniform starburst conditions through the high densities and pressures within the starburst core, since high ambient densities result in compact, small, high density interstellar gas clouds. In high-pressured galaxies such as M82, the characteristic scale-sizes are therefore smaller than less-pressured systems, and this means that the gas conditions are able to respond quickly to any changes. This ability to respond rapidly results in the observed near-constant state of the starburst (seen though indicators such as ionization parameter and line diagnostics).
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the referee, Mike Dopita, for his very insightful comments that led to some interesting discussion. We also thank David Stys and Linda Dressel at STScI for their help in reducing the STIS observations. The Image Reduction and Analysis Facility ([iraf]{}) is distributed by the National Optical Astronomy Observatories which is operated by the Association of Universities for Research in Astronomy, Inc. under cooperative agreement with the National Science Foundation. [stsdas]{} is the Space Telescope Science Data Analysis System; its tasks are complementary to those in [iraf]{}. Support for program \#9117 was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555.
[^1]: Note: slit B2 is not aligned with the major axis of region D and misses most of the eastern extent.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Extreme UltraViolet (EUV) images of the optically-thin solar corona in multiple spectral channels give information on the emission as a function of temperature through differential emission measure (DEM) inversions. The aim of this paper is to describe, test, and apply a new DEM method named the Solar Iterative Temperature Emission Solver (SITES). The method creates an initial DEM estimate through a direct redistribution of observed intensities across temperatures according to the temperature response function of the measurement, and iteratively improves on this estimate through calculation of intensity residuals. It is simple in concept and implementation, is non-subjective in the sense that no prior constraints are placed on the solutions other than positivity and smoothness, and can process a thousand DEMs per second on a standard desktop computer. The resulting DEMs replicate model DEMs well in tests on Atmospheric Imaging Assembly (AIA) synthetic data. The same tests show that SITES performs less well on very narrow DEM peaks, and should not be used for temperature diagnostics below $\sim$0.5MK in the case of AIA observations. The SITES accuracy of inversion compares well with two other established methods. A simple yet powerful new method to visualise DEM maps is introduced, based on a fractional emission measure (FEM). Applied to a set of AIA full-disk images, the SITES method and FEM visualisation show very effectively the dominance of certain temperature regimes in different large-scale coronal structures. The method can easily be adapted for any multi-channel observations of optically-thin plasma and, given its simplicity and efficiency, will facilitate the processing of large existing and future datasets.'
author:
- Huw
- James
bibliography:
- './biblio.bib'
title: 'SITES: Solar Iterative Temperature Emission Solver for differential emission measure inversion of EUV observations'
---
Introduction {#intro}
============
Understanding the physics of the Sun’s atmosphere demands increasingly detailed and accurate observations. The development of new analysis methods to gain physical observables from remote sensing observations is an ongoing and critically important effort. As part of this effort, this paper presents a new Differential Emission Measure (DEM) method for the temperature/density analysis of solar coronal optically-thin emission lines. The Extreme UltraViolet (EUV) spectrum from the solar atmosphere contains several strong emission lines from highly-ionised species above a relatively low background. These lines are emitted from the hot corona only, thus narrowband EUV observations are an excellent probe of the low corona, with little contamination from the underlying photosphere and lower atmosphere.
The concept of using EUV line intensities to estimate the temperature of the emitting plasma is based on the temperature of formation of the line: a range of temperatures at which a certain ion can exist, and the relative population of that ion as a function of temperature. Thus calibrated observations of two lines with different formation temperatures can give a constraint on the dominant plasma temperature. Based on this concept, the simplest approach to estimating a dominant coronal temperature is the line ratio method, which assumes an isothermal plasma (see, for example, the description and criticism of ).
In the general case, imaging instruments provide an observed intensity integrated across a narrow bandpass that spans one or more spectral line - this is the case for an EUV imaging instrument such as AIA. Thus the temperature response of each channel may be computed based on the wavelength response of that channel, and modelled line intensities from an established atomic database (such as Chianti, [@dere1997]) using certain assumptions (e.g. Maxwell-Boltzmann distributions and thermal equilibrium). The measured intensity of multiple bandpasses, or channels, with different temperature responses, allow the estimation of emission as a function of temperature, or a DEM. A DEM is a powerful characterisation of the coronal plasma - it is an estimate of the total number of electrons squared along the observed line of sight (similar to a column mass) at a given temperature. The DEM method has revealed the general temperature characteristics of the main structures seen in the corona: for example, closed-field active regions are hot and multithermal ($>$2MK), open-field regions are colder ($<$1.1MK), and in between is the quiet corona ($\sim$1.4MK) [@delzanna2013; @hahn2011; @mackovjak2014; @hahn2014]. Changes in DEM over time are related to heating or cooling, and can be applied over large datasets to reveal solar cycle trends [@morgan2017].
For an imaging instrument such as AIA, the DEM method inverts measured intensities in a small number of bandpasses to give the emission as a function of temperature across a large number of temperature bins. This is an underdetermined problem that requires additional constraints on the solution, such as positivity and smoothness. There are several types of DEM methods in use, well summarized in the introduction to . One method is that of , which uses Tikhonov regularization to find an optimal weighting between fitting the data and satisfying additional constraints of positivity of the DEM (negative emission is unphysical), minimising the integrated emission, and smoothness of the result. To our knowledge, the most computationally fast method is that of , based on Simplex optimization of a set of smooth basis functions, or a sparse matrix. use a parametric functional form for the DEM, solved with a regularized inversion combined with an iterative scheme for removal of negative DEM values. A similar parametric form is also used by in the context of coronal tomography and a localised DEM.
This work presents a new DEM inversion method in section \[method\]. The method is introduced in the context of the type of imaging observations made by an instrument such as AIA, but can easily be generalised to any observation where the measurement temperature response is known. Tests of the method on synthetic observations made from model DEMs are made in section \[synthetic\], along with a non-rigorous test on computation time. Section \[aiadata\] discusses uncertainty in AIA measurements, and applies the method to data. An effective method to visualise DEM maps is also presented in section \[aiadata\]. A brief summary is given in section \[summary\].
The DEM method {#method}
==============
A set of intensities $I_0, I_1, ...I_{n-1}$ are measured by $i=0,1,...,n-1$ AIA channels, with associated errors $\sigma_i$. Each channel’s response as a function of temperature, $R_{ij}$, is known for a set of temperature bins indexed $j=0,1,...,n_t-1$. This work uses the response functions as given by the standard AIA Solarsoft routines, calculated from the Chianti atomic database [@dere1997; @landi2012], cross-calibrated over time with EVE observations and including a correction to the 94Å channel calibration [@boerner2014]. An example of these functions are shown in figure \[response\]a. We assume that each response function has a relative error $\varepsilon_i$, constant over all temperature. These uncertainties are discussed in a following section.
Before considering weightings associated with the relative noise in each channel, we first introduce the simple concept of relative temperature responses. The relative response for a given channel (indexed $i$) and temperature (indexed $j$), $S_{ij}$, is calculated as $$\label{relres}
S_{ij}=\frac{R_{ij}}{\sum_{i=0}^{n-1}R_{ij} },$$ so that, at a given temperature bin, the relative responses sum to unity over all channels. The relative responses are an useful value, and are shown in figure \[response\]b. For example, at very high temperatures ($>$10MK), the relative response of the 193 channel is almost 1, showing that any DEM method using solely AIA data will be very uncertain at these flare temperatures, since only one channel is giving information at this temperature. A similar argument holds for low ($<$0.1MK) temperatures, where the relative response of the 304 channel increases to 0.7. The relative response (further weighted by the relative error in each channel, to be introduced later) is used in the DEM method to combine the information from each channel. Thus, as can be seen from figure \[response\]b, at temperatures near 0.9MK, the resulting DEM will be dominated by the 171 channel.
It is convenient to include information on the relative measurement error of each channel, and the estimated errors of the response functions in the relative response. Thus equation \[relres\] becomes $$\label{relres2}
S_{ij}=\frac{w_i R_{ij}}{\sum_{i=0}^{n-1}w_i R_{ij} },$$ where $w$ is a weighting based on the measurement and calibration errors, $$\label{w}
w_i=\frac{1}{\sqrt{(\frac{\sigma_i}{I_i})^2 + \varepsilon_i^2}}.$$ Thus channels with smaller relative errors (i.e. higher signal to noise, and/or less uncertainty in response function) will have greater weighting in the DEM estimate.
An initial DEM estimate, $D_j$ is given by $$\label{dem1}
D_j=\left[ \sum_i^{n-1} S_{ij} \left( I_i \frac{R_{ij} \Delta T_j}{\sum_j^{n_t-1} (R_{ij} \Delta T_j)^2 } \right) \right] \otimes K,$$ where $K$ is a smoothing kernel. In words, a set of $n$ DEM profiles is calculated, one for each channel, based directly on the response function of each channel (the expression within the round brackets). Since the observed intensity is distributed over the DEM temperature range according to the response function of that channel, integrating these individual DEMs over temperature would result in exactly the observed intensities. These individual DEMs are combined into a single DEM through a weighted mean, using the weighting of the relative responses (i.e. product with the relative responses, $S_{ij}$, and summation over $i$). This DEM is convolved with a narrow Gaussian kernel over temperature $K$, to ensure a smooth DEM. The kernel $K$ is a Gaussian profile in logarithmic temperature, with a width (standard deviation) of 3.2 bins in logarithmic temperature, for 43 temperature bins over a temperature range of 0.07-20MK. These values are found through trial and error, with the criteria that the smoothing width is kept at a minimum value whilst still resulting in smooth DEMs. The width of the smoothing kernel is, in fact, the only subjective choice in this procedure. If more temperature bins are set, then the width of the kernel should be increased in proportion.
From this initial DEM, a set of modelled intensities $M_i$ is computed for each channel by $$\label{iexp}
M_i = \sum_j^{n_t-1}D_j R_{ij} \Delta T_j.$$ The residual, or difference between the observed and modelled intensities, is calculated as $I_i^\prime=I_i - M_i$. This residual intensity is fed back into equation \[dem1\] (taking the place of $I_i$ in the equation), and the resulting residual DEM added to the previous DEM. At this step, the main DEM is thresholded to a minimum value of zero since the residuals may result in a negative DEM at certain temperatures - thus a positivity constraint is applied. This process is iterated until convergence is reached, defined as when the weighted mean of the absolute ratios between the measurement residuals at the current iteration and the initial measurement, drops below an appropriately small threshold, for example 1%. The weights for this mean are those given by equation \[w\]. This is a sensible criteria for convergence - the process stops when the mean changes to the output DEM become small, with weighting towards the higher certainty measurements. There are similarities in this iterative approach to that of , which computes residual data intensities at several iterations in order to adjust the estimated DEM and eliminate negative intensities. However, the core DEM estimation at each iteration, given by equation \[dem1\], is quite different to their method.
An estimate for the DEM error $d$ at each temperature bin $j$ is $$\label{err}
d_{j} = \sqrt{ \sum_i^{n-1} S_{ij} \left[ \left( \sigma_i/I_i \right) ^2 + \varepsilon_i^2 \right] }.$$ The relative measurement error, $\sigma_i/I_i$, and the response function relative uncertainty $\varepsilon_i$, are summed in quadrature, giving the total squared measurement error for each channel. These are multiplied by the relative response $S_{ij}$ in order to distribute over temperatures, and summed over all channels, corresponding to the equivalent steps in the DEM estimate of equation \[dem1\]. The square root of this value gives the final DEM uncertainty. A complete error propagation treatment should consider the smoothing kernel and multiple iterations, but these steps would defeat the aim of implementing an efficient method. The uncertainties given by the simple calculation of equation \[err\] give values that agree well with tests involving varying the input measurements according to measurement noise, as is shown in subsection \[robustnoise\].
Demonstration using synthetic data {#synthetic}
==================================
A simple test {#simpletest}
-------------
A model DEM is produced by the Gaussian $$\label{gauss}
D^\prime = A \exp \left( \left[ \frac{t-t_c}{w_t} \right]^2 \right),$$ with peak maximum $A=1.4\times 10^{21}$[cm$^{-5}$ K$^{-1}$]{}, central temperature $t_c =$1.4MK, and width $w_t=0.9$MK. Using the AIA response functions (as shown in figure \[response\]), synthetic observations are created for the 7 channels, in units of [DN cm$^{5}$ $s^{-1}$ pix$^{-1}$]{} by integrating the product of the DEM with the response functions over temperature. Measurement uncertainties are given by the AIA Solarsoft routine aia\_bp\_estimate\_error. The synthetic observations are input into SITES, using 43 temperature bins within a temperature range of 0.07-20MK, with a regular bin size in logarithmic temperature. The method terminates at 101 iterations when convergence, as defined in the method, reaches 1%. To avoid edge effects caused by the smoothing truncation, the first and last DEM bins are discarded, leaving 41 temperature bins in the results. The maximum absolute relative difference between input target intensities and the method’s derived intensities (the $M_i$ of equation \[iexp\]) is 3.5% for the 335 channel. The mean absolute measurement difference $T_I$ over all channels is defined as $$\label{measdev}
T_I = \frac{1}{n} \sum_i^{n-1} \frac{| I_i - M_i | }{I_i},$$ and is 1.2% for this simple test.
The resulting DEM is compared to the target input DEM in figure \[simpledem\]. The median absolute relative DEM deviation $T_D$, between the input DEM $D^\prime$ and output estimated DEM $D$, is defined as $$\label{demdev}
T_D = \mathrm{median} \frac{| D_j^\prime - D_j | }{D_j^\prime},$$ and is 26% for this simple test. The median is calculated over temperature bins $j$ and is used here rather than the mean to avoid the influence of very small, or zero, values of input DEM at some temperatures bins.
The correlation $c$ between the input and output DEM curves is defined as $$\label{corr}
c= \frac{ \sum_j^{n_t-1} ( D_j - \bar{D_j} ) ( D_j^\prime - \bar{D_j^\prime} ) } {\sqrt{ \sum_j^{n_t-1} ( D_j - \bar{D_j} )^2 \sum_j^{n_t-1} ( D_j^\prime - \bar{D_j^\prime} )^2 }}$$ and is 98% for this simple test. Thus the position and width of the main peak is well fitted.
One concern is the range of temperatures used for the calculation. The examples here have temperatures limited to between 0.07 and 20MK. Some channels have significant values in their response functions outside this range, leading to an inherent uncertainty that can be included in the estimate of each response function’s uncertainty. Thus an estimate of the relative uncertainty in the response of each channel is given by $$\varepsilon_i = \sqrt {e_i^2 +\left( \frac{\sum_{j0} R_{ij}\Delta T_j}{\sum_{j1} R_{ij}\Delta T_j}\right)^2},$$ where $e_i$ is the calibration uncertainty for each channel, and the subscript $j1$ are the indices of temperature bins included within the temperature range, and $j0$ otherwise. The $e_i$ are given by AIA Solarsoft routines, and is 50% for the 94, 131 and 304 channels, 25% otherwise. $\varepsilon_i$ is a large uncertainty, ranging from 27% for the 171 channel, to 103% for the 131 channel. Channels with large contributions to their response functions outside of the temperature range of interest have a lesser weighting in calculating the final DEM.
The single-Gaussian DEM is used as a test of SITES across a broad range of Gaussian central temperature and Gaussian widths. The central temperature is increased from $\log T$ 5.3 to 7.05 in 160 increments, and the widths from $\log T$ 0.1 to 0.35 in 160 increments (note this differs from the example of figure \[simpledem\], which is formed from Gaussians in linear temperature). For each input DEM, synthetic measurements are calculated and given as input to SITES, as above for figure \[simpledem\]. The correlation between input and output DEM, $c$, as given by equation \[corr\] is calculated, giving a measure of the similarities of the profiles. This is shown in figure \[simpleparams\]a. A broad range of central temperatures and widths bounded by the dotted line give correlations above 95%. Poor correlations, below 80%, are found for low temperatures below $\log T$ 5.7, and for very narrow profiles at all central temperatures. Figure \[simpleparams\]b shows the mean absolute relative deviation of the input and output measurements, $T_I$, as given by equation \[measdev\]. The worst match, at close to 10% deviation, is found for low temperatures or higher temperatures at narrow widths. The deviation otherwise is good, with the majority of the parameter space at values of 4% or lower. This is to be expected, given that the iterative scheme is designed to reduce this deviation. Figure \[simpleparams\]c shows the median absolute relative deviation between the input and output DEMs, $T_D$, as given by equation \[demdev\]. For the broad region dominated by very high correlations, the deviation is around 15-50%. This deteriorates to over 50% for low temperatures below $\log T$ 5.7, or for narrow profiles at all temperatures.
In summary, SITES performs poorly for narrow DEM profiles at all temperatures. This is inherent to estimating DEMs from an instrument such as AIA, regardless of the method, given the broad multiple-peaked temperature profiles in most channels. SITES performs very poorly for DEMs peaked at cool temperatures below $\log T$ 5.7 ($\sim$0.5MK). At higher temperatures, and broader peaks, SITES performs very well, with 95% correlation with the target input DEMs.
A complex test {#complextest}
--------------
A more complex model DEM sums 2 Gaussian peaks over a constant background. The background emission has a value of $10^{20}$[cm$^{-5}$ K$^{-1}$]{}, and the Gaussian peaks have amplitude $A= [1.6,1.6] \times 10^{21}$[cm$^{-5}$ K$^{-1}$]{}, centered at temperatures $t_c=[0.8,4.5]$ MK, with widths $w_t=[0.35,3.0]$ MK. Synthetic observations are created from this model DEM as for the simple case above. The comparison between input and output DEM is shown in figure \[complexdem\]. The number of iterations is 114, and $T_I$ is 1.3%, with a maximum deviation of 3.4% for the 94 channel. $T_D$ is 12%. The position of the 2 peaks in the resulting DEM estimate agree well with the input DEM - the method is effective at finding these peaks, reflected in the $c=97.7$% correlation between input and output DEM.
The performance of SITES is tested for various combinations of parameters for the two-Gaussian plus background DEM profile through varying the central temperature of each peak. This experiment is repeated for two cases of wide and narrow Gaussians. The Gaussians are formed in logarithmic temperature (note this differs from the example of figure \[complexdem\], which is formed from Gaussians in linear temperature). Figure \[cp\_examples\] shows 4 characteristic examples of the two-Gaussian parameter space. Figure \[cp\_examples\]a is for a cool peak at $\log T = 5.5$ and a hot peak at $\log T = 6.55$. For the wide Gaussian profiles (solid lines), the hot peak is well-fitted by SITES, but the fit for the cool peak is poor. The same holds for the narrow Gaussians (dashed lines). The position of the hot narrow peak is found by SITES, although the method struggles to fit the profile properly, with regions next to the peak at zero emission. Figure \[cp\_examples\]b is for a cool peak at $\log T = 5.5$ and a hot peak at $\log T = 7.0$, with similar results to \[cp\_examples\]a. Figure \[cp\_examples\]c is for a cool peak at $\log T = 6.2$ and a hot peak at $\log T = 6.55$, thus the wide Gaussians are blended. SITES fits this profile very well. There are two closely-placed yet distinct peaks in the narrow Gaussian DEM. The SITES DEM also shows two peaks, but is far smoother than the target DEM. Note also the tendency for regions close to the two peaks to have zero emission. Figure \[cp\_examples\]d is for a cool peak at $\log T = 6.2$ and a hot peak at $\log T = 7.0$. Similar to the previous case, the wide DEMs are fitted very well. The narrow peaks are found by SITES, but are smoother, and tend to zero in nearby regions.
Figure \[cp\_wide\] shows the performance of SITES for a range of central temperatures for both Gaussian peaks, for the case of the wide Gaussians. The correlation between the input and SITES DEMs, shown in figure \[cp\_wide\]a, is excellent ($c>80$%) for all peak hot temperatures, and cool temperatures above 0.5MK. Below this cool temperature, the performance of SITES is poor, despite the close fit to the input measurement as shown in figure \[cp\_wide\]b. The same poor fit for low temperatures is seen in the median absolute relative deviation of the input and output DEMs in figure \[cp\_wide\]c.
Figure \[cp\_narrow\] shows the same parameter test for narrow Gaussian profiles. Overall, the correlation, fit to measurement, and fit to DEM have deteriorated throughout the parameter space. The very poor fit at low cold peak temperatures remains. In summary, the conclusions for a complex double-Gaussian DEM profile are similar to the case of a single Gaussian in the previous section. SITES performs poorly for narrow DEM profiles at all temperatures, and performs very poorly for DEMs which contain peaks at cool temperatures below $\log T$ 5.7 ($\sim$0.5MK). At higher temperatures, and broader peaks, SITES performs very well, with $c=95$% correlation with the target input DEMs.
Computational speed and convergence threshold
---------------------------------------------
To test computational efficiency, SITES is applied 1000 times to the complex 3-Gaussian plus background DEM distribution, imposing a variation on input channel intensities based on their randomisation according to the measurement uncertainty estimates at each run. This experiment is repeated for convergence thresholds of 1, 2, 4, 8, 16 and 32%. Figure \[convergence\]a summarises the performance of SITES as a function of the increasing convergence thresholds through the median absolute residuals of the measurements (goodness of fit), and the median absolute deviation of the resulting DEMs compared to the target model DEM. There is no significant deterioration of achieving the target DEM up to the 8% convergence threshold. The measurement residuals similarly remain small up to the 8% convergence threshold. On a Linux desktop Intel Core i7-4790 CPU with 16Gb memory the 1000 runs are timed, with the number of DEMs calculated per second shown in figure \[convergence\]b. Based on these results, for real data we set a convergence threshold of 4%, which can process around 1000 DEMs per second. This speed is similar to regularized matrix inversion-based methods such as or .
Robustness to noise {#robustnoise}
-------------------
This section tests the reliability of SITES in the presence of noise. A complex DEM formed from 3 Gaussian peaks and a constant background is used to create synthetic measurements. This DEM is shown as the solid black line in figure \[robust\]a. For a thousand repetitions, the measurements are varied randomly according to a noise amplitude given by the measurement and calibration errors, and the resulting DEMs recorded. The convergence factor is set at 4%, at a value that will typically be used for practical use on real data.
Figure \[robust\]a shows the mean DEM, calculated over the thousand repetitions, as a dotted line. This can be compared to the input model DEM which is shown as a bold solid line. The grey shaded region shows the standard deviation of DEMs over the thousand repetitions. The error bars show the mean DEM errors as calculated by equation \[err\]. Figure \[robust\]b shows the input measurements in each channel, in the absence of noise, as triangle points, with the error bars showing the noise level. The cross symbols and error bars show the mean and standard deviation of the fitted measurements (i.e. gained from the output DEM through equation \[iexp\]). Despite the large variations in the DEM values, the 3-peak profile is well replicated. The presence of noise does not lead to DEMs that deviate significantly beyond that expected given the uncertainties. The uncertainty estimate of equation \[err\] reflects well the true variation of the output DEMs. Integrating the product of the DEMs with the response functions (equation \[iexp\]) shows that the method is fitting the input data correctly. As can be seen in figure \[robust\]b, the only systematic discrepancy is seen for the low-signal 131 channel, where the method gives a small positive residual.
Figure \[robust2\]a shows the distribution of DEMs resulting from running the experiment for a signal 10 times lower than the previous example. In this very noisy case, SITES performs reasonably well, although the third DEM peak at high temperature is overestimated. The estimated error bars have increased correctly given the increase in noise across temperatures up to $\sim$2MK. Above this temperature, the uncertainty is underestimated. From figure \[robust2\]b, the measurement residuals are systematically too high for the lower-signal 94, 131 and 335 channels.
Comparison with other methods
-----------------------------
SITES is compared here with the method of , hereafter called Sparse Matrix Inversion (SMI), and with the method of , hereafter called Tikhonov Regularization (TR). Both the simple single Gaussian DEM of section \[simpletest\] and the multiple Gaussian plus constant background DEM of section \[complextest\] are used to create synthetic measurements that are given as input to SITES, SMI and TR. All three methods use identical temperature response functions, measurements and measurement errors for inversion. The TR method is called with the default order equal to zero, and we show the positive-constrained solution.
The resulting emissions as functions of temperature for the single Gaussian case is shown in figure \[cheung\]a. The result for the default choice of the SMI Gaussian basis functions is shown as a solid green line. It is obvious that this choice of basis functions gives an EM result which is too wide. Halving the width of the basis functions (dashed green line) gives a decent fit to the input EM curve, although emission is too high towards the high-temperature wing of the distribution. TR gives a good fit except at the highest range of temperature, where a steep increase is seen. SITES also has a small increase at the highest temperature bin. SITES outperforms both SMI and TR for this example, in closely fitting the Gaussian peak and giving zero DEM at higher temperatures.
Figure \[cheung\]b shows the result for a double-Gaussian input DEM. In the case of using the broad (default) SMI basis functions (solid green line), the estimated EM broadly covers the correct temperature region, but fails to identify the individual peaks. The narrow basis functions (dashed green line) successfully identifies the EM peak near $T=1$MK, but fails to invert the other peak, and gives an overall profile which is too narrow across temperature. TR is effective in finding the cooler $T=1$MK peak but fails to identify the main peak near 4MK. SITES outperforms both SMI and TR for the two-Gaussian DEM profiles, in successfully finding all three Gaussian peaks plus the constant background.
The comparison of SITES to the TR method is extended to a parameter search for the case of a single-Gaussian plus background input DEM profile. The parameter space is the same as in section \[simpletest\], but with a reduced number of bins (30 bins in Gaussian central temperature and 20 bins in Gaussian width). Results are shown in figure \[comp\], with the top row showing the SITES performance (almost identical to figure \[simpleparams\], with a different color scale range), and the bottom row showing the TR method performance. The DEM input-output correlation of figures \[comp\]a and b show a poor inversion for both methods at low temperatures ($\log T < 5.7$). Above this temperature, SITES outperforms TR for almost all central temperatures and widths. SITES also more closely fits the input data by a considerable margin, as shown in figures \[comp\]b and e. Figures \[comp\]c and e show the median absolute relative deviation of input and output DEMs for both methods. SITES has a larger region of small deviation ($<40$%), and more profiles that have a very small deviation ($<20$%), but also has some regions of higher deviation than TR. In summary, SITES generally gives better perofrmance than TR in this noiseless comparison.
As suggested by figure \[comp\]e, the TR method may be underfitting the data, therefore the comparison with SITES may be unfair since the input data has no randomness associated with noise. This is addressed by repeating the test 15 times, allowing the input data to vary randomly according to a Poisson distribution, comparing the output DEM at each repetition to the input DEM, and taking the mean correlation and measurement/DEM deviations over the 15 cases. To give an idea of the noise amplitude, at a central DEM temperature of $\log T$=6.4 and $\log T$ width 0.26, the relative Poisson noise is 19% for the lowest signal 94 channel, and 1.4% for the 193 channel. Results are shown in figure \[comp2\]. The input/output DEM correlation is generally better for SITES compared to TR (figures \[comp2\]a and d), whilst the DEM deviation (figures \[comp2\]c and f) is worse for SITES. In summary, both methods perform similarly for noisy data, with SITES giving an overall better match to the general DEM profiles (a broader region of higher correlation), and TR giving closer absolute values of DEM (a broader region of lower deviation).
We note that we have not investigated with any rigour the various parameters of SMI. We have, for example, only used two choices of the basis function widths. We further note that SMI is extremely fast compared to SITES, around a factor of 100 faster depending on the choice of SITES convergence factor. For the TR method, we have experimented with changing the choice of order (which sets the regularization constraints), with similar results to those shown for order equal to zero. At a convergence threshold of 4%, SITES is of comparable speed to TR.
Application to AIA data {#aiadata}
=======================
Data processing and error estimates {#aiaprocessing}
-----------------------------------
The standard SDO procedure read\_sdo.pro is used to open a set of full-resolution images in the 7 EUV channels of AIA. An example from 2015/01/01 03:00 is used here. Figure \[aiaimage\] shows a colour composite processed using Multiscale Gaussian Normalization to provide context [@morgan2014]. Each channel’s image is shifted in the $x$ and $y$ dimensions so that the central pixel corresponds to the solar disk center, as given by the header image geometry information. A secondary sub-pixel fine alignment is achieved through aligning each image to the 193 channel image, using a phase correlation method to estimate the required shift [@druckmuller2009; @fisher2008], and cubic interpolation to apply the shift. For the example set of images, these pixel shifts are listed in table \[tabledata\]. The mean signal calculated over all pixels on the solar disk is listed for each channel in the table.
[Channel]{} [$T (s)$]{} [$x_s$]{} [$y_s$]{} [$\bar{I}$ (DN pix$^{-1}$)]{}
------------- ------------- ----------- ----------- -------------------------------
[94]{} [2.9]{} [0.86]{} [-1.47]{} [3]{}
[131]{} [2.9]{} [1.59]{} [-1.10]{} [13]{}
[171]{} [2.0]{} [-0.37]{} [-0.62]{} [271]{}
[193]{} [2.0]{} [- ]{} [- ]{} [411]{}
[211]{} [2.9]{} [-0.10]{} [0.36]{} [207]{}
[304]{} [2.9]{} [0.91]{} [-0.91]{} [28]{}
[335]{} [2.9]{} [0.81]{} [-0.64]{} [5]{}
: Some characteristics of an AIA observation set, with columns showing channel, exposure time, $x$-shift (fine alignment relative to the 193 channel), $y$-shift, and mean intensity (on the disk).[]{data-label="tabledata"}
The uncertainty of the measurements given by the AIA Solarsoft routine aia\_bp\_estimate\_error includes the Poisson photon count, dark subtraction, read noise, count quantization and image compression uncertainties. Figures \[aiaerrors\]b and c shows the range of intensities enclosed by the estimated errors for the 193 and 94 channel respectively, for a horizontal cut across the images shown by the dashed red line in figure \[aiaerrors\]a. In high-signal regions/channels, the measurement error is small and the dominant uncertainty is in the response functions (calibration uncertainty). In low-signal regions/channels the method is influenced by both the response function and measurement uncertainties. At the expense of spatial and temporal resolution, rebinning images to smaller size through neighbourhood averaging, and combining two or more consecutive observations over time, will decrease measurement noise in the low signal channels to a more acceptable level.
The dataset of 2015/01/01 is rebinned to $512 \times 512$ pixels, from the original $4096 \times 4096$ pixels. Since 64 original measurements are combined (averaged) for each pixel, the measurement noise decreases by a factor of $1/8$. The DEM method is applied to all pixels at heights below 1.15[$ R_{\odot}$]{}, and DEMs converted to EM by product with the width of the temperature bins. Emission is shown for four example temperatures in figure \[aiadem0\].
Effective visualisation of DEMs is challenging, since the output result from an imaging instrument is a datacube, thus one can show emission at a given temperature yet the context of emission at other temperatures is absent. Such direct DEM images are also dominated by the high emission, at all temperatures, of active regions. One effective method is the emission-weighted-mean or median temperature displayed with a colour/hue table that can show temperatures and emission, as shown for example in figure 15 of . For visually comparing DEM maps in the context of dominance of different regions by certain temperature ranges, we introduce the simple concept of Fractional emission measure (FEM). FEM in a temperature bin (indexed $j$) is calculated from a DEM by $$\label{fracDEM}
FEM_j=\frac{DEM_j \Delta T}{\sum_j DEM_j \Delta T},$$ so the FEM in a given temperature bin gives the fraction of emission at that temperature compared to the total emission integrated over all temperatures. FEM maps are shown in figure \[aiadem1\]. These maps, for regions on the disk, are a powerful visualisation of the different general temperature dependencies of large-scale coronal features:
- At T$=0.5$MK, the FEM maps are dominated strongly by coronal holes and filament channels. This is an effective way of identifying these regions.
- At T$=1.5$MK, broad regions of the quiet corona and coronal holes have high FEM. Quiet regions surrounding active regions are particularly strong. Note that active regions have generally very low FEM at this temperature.
- At T$=4.1$MK, all regions except active regions have low FEM. Note in the original EM maps, that active regions have high EM at all temperatures compared to other regions due to their high mass. The FEM maps, through normalization by the total EM, removes this effect and shows that, despite the multithermality of active regions, their emission is dominated by high temperatures.
- At T$=5.6$MK, only the hot cores of the large active regions have high FEM. The quiet coronal regions have close to zero FEM at this temperature.
The DEMs in off-limb regions are hard to interpret and are subject to the bias towards high temperatures with increasing height, given the large height scale for hot structures, as explained by, e.g. . Solar rotational tomography offers a solution to this line-of-sight problem. A framework for tomography combined with a DEM analysis is given by , where the intensity from each channel, observed from several different viewpoints, is reconstructed in a 3D volume of emission, and a local DEM computed at each voxel.
Summary
=======
A new DEM method is presented which is reasonably fast, simple in concept, and simple to implement. It performs well on tests involving model DEMs and synthetic data based on the AIA/SDO instrument. In particular, the correlation between the model input DEMs and SITES inversions is excellent for a broad range of coronal temperatures. SITES performs less well on very narrow DEM peaks, and performs very poorly for temperatures below $\sim$0.5MK. This weakness is likely due to the limitations of the AIA/SDO instrumental temperature response curves rather than the SITES inversion itself, since other inversion methods show the same failing.
Applied to a set of AIA/SDO observations of the full-disk corona, SITES gives sensible values of emission as a function of temperature. Fractional emission measure is introduced as a simple yet powerful method to visualise DEM results within images, enabling straightforward comparison of different temperature regimes between regions.
The computational speed of the method compares well with most methods, but cannot compete with the sparse matrix approach of . However, the main advantages of SITES is its simplicity of concept and application, and its non-subjectiveness. Equations \[dem1\] and \[iexp\] form the core of the iterative procedure, and are simple to implement. The results of any DEM inversion method are subject to choices of fitting parameters. In the case of SITES, there is only one parameters which effects the result - the width of the smoothing kernel. Thus the method is relatively non-subjective.
The incentive for developing the method is to analyse large datasets, thus enabling large-scale studies of coronal changes over long time-scales using AIA/SDO. The method has therefore not been tested on flare-like temperatures. Reliable studies of such high temperatures need measurements by other instruments, possibly in combination with AIA/SDO. Given a set of temperature response functions and error estimates, the method presented here should work reliably - this will be investigated in the near future.
Future work by the authors (paper in preparation) involves a gridding method that may be used with any DEM inversion method to increase computational efficiency by one or two orders of magnitude. This will enable rapid processing of large datasets for AIA/SDO and other current or future instruments. The software for the DEM fitting method of this paper, plus the FEM visualisation method, written in IDL, is available by email request to the authors.
James Pickering is supported by an STFC studentship. Part of Huw Morgan’s work on this project is supported by an STFC consolidated grant to Aberystwyth University. CHIANTI is a collaborative project involving George Mason University, the University of Michigan (USA), University of Cambridge (UK) and NASA Goddard Space Flight Center (USA). The AIA/SDO data is courtesy of NASA/SDO and the AIA science team.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A conjecture on the relation between the cubic Hodge integrals and the topological vertex in topological string theory is resolved. A central role is played by the notion of generalized shift symmetries in a fermionic realization of the two-dimensional quantum torus algebra. These algebraic relations of operators in the fermionic Fock space are used to convert generating functions of the cubic Hodge integrals and the topological vertex to each other. As a byproduct, the generating function of the cubic Hodge integrals at special values of the parameters $\overrightarrow{w}$ therein is shown to be a tau function of the generalized KdV (aka Gelfand-Dickey) hierarchies.'
author:
- |
[ Toshio Nakatsu[^1] ]{}\
*Institute for Fundamental Sciences, Setsunan University*\
*17-8 Ikeda Nakamachi, Neyagawa, Osaka 572-8508, Japan*
- |
[ Kanehisa Takasaki[^2]]{}\
*Department of Mathematics, Kindai University*\
*3-4-1 Kowakae, Higashi-Osaka, Osaka 577-8502, Japan*
date:
-
- 'December, 2018'
title: |
**Three-partition Hodge integrals\
and\
the topological vertex\
**
---
Introduction
============
Let $\mathcal{M}_{g,n}$ denote the moduli space of connected complex algebraic curves of genus $g$ with $n$ marked points and $\overline{\mathcal{M}}_{g,n}$ the Deligne-Mumford compactification. Marked points on an algebraic curve $C$ of genus $g$ are referred to as $z_1,\dots,z_n$. The Hodge integrals are integrals on $\overline{\mathcal{M}}_{g,n}$ of the form $$\int_{\overline{\mathcal{M}}_{g, n}}
\psi_1^{i_1}\cdots\psi_n^{i_n}\,\lambda_1^{j_1}\cdots\lambda_g^{j_g},$$ where $\psi_i =c_1(\mathbb{L}_i)$ is the first Chern class of the bundle $\mathbb{L}_i$ of the cotangent space of $C$ at the $i$-th marked point $z_i$, and $\lambda_j =c_j(\mathbb{E})$ is the $j$-th Chern class of the Hodge bundle $\mathbb{E}$ over $\overline{\mathcal{M}}_{g,n}$. These Hodge integrals are building blocks of the localization computation of Gromov-Witten invariants [@Graber_Pandharipande_1999]. The localization formula expresses those invariants as a sum of weighted graphs [@Kontsevich_1995]. Integrals of the form $$\begin{aligned}
\int_{\overline{\mathcal{M}}_{g,n}}
\frac{\prod_{i=1}^s \bigwedge^\vee_g(u_i)}
{\prod_{j=1}^{n}\bigl(1-z_j \psi_j\bigr)},\end{aligned}$$ where $\mbox{$\bigwedge^\vee_g$}(u)$ is the characteristic polynomial $$\begin{aligned}
\mbox{$\bigwedge^\vee_g$}(u)
= \sum_{r=0}^g(-1)^r\lambda_ru^{g-r}
= u^g-\lambda_1u^{g-1}+\,\cdots\,+(-1)^g\lambda_g, \end{aligned}$$ arise as the vertex weights in the localization computation of Gromov-Witten invariants of an $s$-dimensional target space.
Li, Liu, Liu and Zhou [@Li_Liu_Liu_Zhou_2009] reformulated the topological vertex in topological string theory [@AKMV] on the basis of formal relative Gromov-Witten theory of non-compact Calabi-Yau threefolds. The topological vertex is a diagrammatic method to compute all-genus A-model topological string amplitudes on smooth non-compact toric Calabi-Yau threefolds. It is somewhat confusing that the vertex weight itself is called the topological vertex. All building blocks of the vertex weight are genuinely combinatorial objects such as partitions, the Littlewood-Richardson numbers and the Schur functions.
The geometric approach of Li et al. stems from their proof of the Mariño-Vafa conjecture and its generalization [@Liu_Liu_Zhou_2004; @Liu_Liu_Zhou_2003; @Liu_Liu_Zhou_2007]. [^3] The Mariño-Vafa conjecture [@Marino_Vafa_2001] is a variation of the so called ELSV formula [@ELSV_2001]. The ELSV formula expresses the Hurwitz numbers of ramified coverings of $\mathbb{P}^1$ in terms of linear ($s=1$) Hodge integrals. The Mariño-Vafa conjecture claims a similar relation between certain open Gromov-Witten invariants and cubic ($s=3$) Hodge integrals. Actually, the open Gromov-Witten invariants considered therein are related to a special case of the topological vertex.
The approach to the full (i.e., three-partition) topological vertex, just like the proof of the Mariño-Vafa conjecture, starts from a generating function of cubic Hodge integrals. The cubic Hodge integrals are labelled by a triple $\overrightarrow{\mu} = (\mu^{(1)},\mu^{(2)},\mu^{(3)})$ of partitions and depend on a triple $\overrightarrow{w} = (w_1,w_2,w_3)$ of parameters. Li et al. employ various skillful ideas to derive a combinatorial expression of the topological vertex from the generating function. This expression, however, takes a form apparently different from the topological vertex of Aganagic et al. Showing several pieces of evidence, Li et al. conjecture that the two expressions of the topological vertex are equivalent.
In this paper, we prove the equivalence of the two different expressions of the topological vertex. Our strategy is to construct a generating function of the topological vertex of Aganagic et al. (modified to depend on the parameters $\overrightarrow{w}$ of the cubic Hodge integrals), and to show that it coincides with (an exponentiated version of) the generating function of Li et al. To this end, we make use of tools developed in a series of our work on the melting crystal models [@Nakatsu_Takasaki_CMP_2009; @Nakatsu_Takasaki_ASPS] and topological string theory [@Takasaki_Nakatsu_JPA_2016; @Takasaki_2014; @Takasaki_Nakatsu_SIGMA_2017]. A central role is played by the notion of generalized shift symmetries [@Nakatsu_Takasaki_2019] in a fermionic realization of the two-dimensional quantum torus algebra. We start from a fermionic representation of the generating function of the topological vertex. The generalized shift symmetries enable us to convert this fermionic representation to the generating function of cubic Hodge integrals of Li et al. This implies the equivalence of the two expressions of the topological vertex.
As a byproduct, we find a new link between the topological vertex and integrable hierarchies. Just after the first proposal [@AKMV], particular generating functions of the topological vertex were pointed out to be tau functions of integrable hierarchies such as the KP and 2D Toda hierarchies [@ADKMV; @Zhou_2010]. We show that the generating function of the cubic Hodge integrals at special values of the parameters $\overrightarrow{w}$ is a tau function of the generalized KdV (aka Gelfand-Dickey) hierarchies. Such a link with the generalized KdV hierarchies seems to be unknown in the past literature.
Let us explain these backgrounds and results in more detail.
Partitions, representations of symmetric group and the Schur functions
----------------------------------------------------------------------
Our notations for partitions, representations of symmetric groups and the Schur functions are mostly borrowed from Macdonald’s book [@Macdonald_book]. Let us recall these notions and basic facts.
A partition is a sequence of non-negative integers $$\mu =(\mu_i)_{i=1}^\infty =(\mu_1,\mu_2,\,\dots\,)$$ satisfying $\mu_i\geq\mu_{i +1}$ for all $i\geq 1$. The number of the non-zero $\mu_i$ is the length of $\mu$, denoted by $l(\mu)$. The sum of the non-zero $\mu_i$ is the weight of $\mu$, denoted by $|\mu|=\mu_1+\mu_2+\cdots$. Partitions have another expression which indicates the number of times each non-negative integers occurs in a partition $$\mu =(1^{m_1}2^{m_2}\dots\,),$$ where $m_i=m_i(\mu)$ denotes the number of $i$ occurs in $\mu$. The automorphism group of $\mu$, denoted by $\mbox{Aut}(\mu)$, consists of possible permutations among the non-zero $\mu_i$’s that leave $\mu$. The number of elements in $\mbox{Aut}(\mu)$ is $$|\mbox{Aut}(\mu)| =\prod_{i=1}^\infty m_i(\mu)!.$$ Note that partitions are identified with the Young diagrams. The size of the Young diagram $\mu$ is $|\mu|$, which is the total number of boxes of the diagram, and $l(\mu)$ is the height of the diagram. The conjugate (or transpose) of $\mu$ is the partition $\ltrans{\,\mu}$ whose diagram is the transpose of the diagram $\mu$.
If $|\mu|=d$, we say that $\mu$ is a partition of $d$. Each partition of $d$ corresponds to a conjugacy class of the $d$-th symmetric group $S_d$. Let $C(\mu)$ denote the conjugacy class determined by a partition $\mu$ of $d$. It is determined by the cycle type $\mu$ of a representative $\sigma \in S_d$ of $C(\mu)$ as $$\sigma =(1,\dots,\mu_1)(\mu_1+1,\dots,\mu_1+\mu_2)
\cdots (\mu_1+\dots+\mu_{l-1},\dots,d),$$ where $l=l(\mu)$ and $(j_1,\dots, j_m)$ means the cyclic permutation sending $j_1\to j_2\to\cdots\to j_m\to j_1$. The number of elements in $C(\mu)$ is $$|C(\mu)| =\frac{d!}{z_\mu}, \quad z_\mu
=\prod_{i=1}^\infty m_i(\mu)!\,i^{m_i(\mu)}.$$ Each partition $\mu$ of $d$ determines the irreducible representation $(\rho_\mu,V_\mu)$ of $S_d$. $\chi_\mu$ denotes the character $\mbox{Tr}_{V_\mu}\rho_\mu$. The value of $\chi_\mu$ on the conjugacy class $C(\nu)\subset S_d$, denoted by $\chi_\mu (\nu)=\chi_\mu\left(C(\nu)\right)$, can be computed by the Frobenius formula $$\begin{aligned}
s_\mu (\bm{x}) =\sum_{|\nu|=d}\frac{\chi_{\mu}(\nu)}{z_\nu}p_\nu, \quad
\bm{x} =(x_i)_{i=1}^\infty =(x_1,x_2,\dots\,),
\label{Frobenius}\end{aligned}$$ where $p_\nu$’s are the monomials $$p_\nu =p_{\nu_1}p_{\nu_2}\dots$$ of the power sums $$p_k =\sum_{i=1}^\infty x_i^k,\quad k=1,2,\dots\,,$$ and $s_\mu(\bm{x})$ denotes the Schur function. (\[Frobenius\]) has the inversion formula $$\begin{aligned}
p_\mu =\sum_{|\nu|=d}\chi_\nu(\mu)s_\nu(\bm{x}).
\label{inverse_Frobenius}\end{aligned}$$
Let us recall the Jacobi-Trudi formula $$\begin{aligned}
s_\mu(\bm{x})
=\mbox{det}\left(h_{\mu_i-i+j}(\bm{x})\right)_{i, j=1}^n,
\label{JT_formula_Schur}\end{aligned}$$ where the length of $\mu=(\mu_i)_{i=1}^\infty$ is not larger than $n$ and $h_m(\bm{x})$’s are the complete symmetric functions defined by the generating function $$\sum_{m=0}^\infty h_m(\bm{x})z^m =\prod_{i=1}^\infty (1-x_iz)^{-1}.$$ Likewise, the skew Schur functions have a determinant formula similar to (\[JT\_formula\_Schur\]). Let $\nu =(\nu_i)_{i=1}^\infty$ be a partition such that $\mu\supset\nu$, [*i.e.*]{}, $\mu_i \geq \nu_i$ for all $i$. The determinant formula reads $$\begin{aligned}
s_{\mu /\nu}(\bm{x})
=\mbox{det}\left(h_{\mu_i-\nu_j-i+j }(\bm{x})\right)_{i, j=1}^n,
\label{JT_formula_skewSchur}\end{aligned}$$ where $\mu /\nu$ represents the skew diagram obtained by removing the Young diagram $\nu$ from the larger one $\mu$.
The Littlewood-Richardson numbers $c_{\mu\nu}^{\eta}$ are non-negative integers which are determined by the relation $$\begin{aligned}
s_\mu(\bm{x})s_\nu(\bm{x})
=\sum_{\eta\in\mathcal{P}}c_{\mu\nu}^{\eta}s_\eta(\bm{x}),
\label{def_LR}\end{aligned}$$ where the sum with respect to $\eta$ ranges over the set $\mathcal{P}$ of all partitions, and is a finite sum because $c_{\mu\nu}^{\eta} =0$ unless $|\eta| =|\mu|+|\nu|$, $\eta\supset\mu$ and $\eta\supset\nu$. The skew Schur functions can be expressed in a linear combination of the Schur functions weighted by the Littlewood-Richardson numbers as $$\begin{aligned}
s_{\mu/\nu}(\bm{x})
=\sum_{\eta\in\mathcal{P}}c_{\nu\eta}^{\mu}s_\eta(\bm{x}).
\label{skewSchur_LR}\end{aligned}$$
Three-partition Hodge integrals
-------------------------------
The three-partition Hodge integrals are Hodge integrals that depend on a triple of partitions in a specific manner. For a triple of partitions $\overrightarrow{\mu} =(\mu^{(1)},\mu^{(2)},\mu^{(3)})$ $\in$ $\mathcal{P}^3=\mathcal{P}\times\mathcal{P}\times\mathcal{P}$, we use the notations $$\begin{aligned}
l(\overrightarrow{\mu}) =\sum_{a=1}^3l(\mu^{(a)}),\quad
\mbox{Aut}(\overrightarrow{\mu}) =\prod_{a=1}^3\mbox{Aut}(\mu^{(a)}).\end{aligned}$$ $l(\overrightarrow{\mu})$ marked points on an algebraic curve of genus $g$ are referred to as $z^{(1)}_1,\dots,z^{(1)}_{l(\mu^{(1)})}$, $z^{(2)}_1,\dots,z^{(2)}_{l(\mu^{(2)})}$ and $z^{(3)}_1,\dots,z^{(3)}_{l(\mu^{(3)})}$. Cotangent lines of curves at the marked point $z^{(a)}_i$ are glued together and form the complex line bundle $\mathbb{L}^{(a)}_i$ over $\overline{\mathcal{M}}_{g,l(\overrightarrow{\mu})}$.
Let $\overrightarrow{w} =(w_1,w_2,w_3)$ be a triple of variables which satisfy the condition $w_1+w_2+w_3=0$. The indices of those variables are understood to be cyclic as $w_{i+3}\equiv w_i$. For $\overrightarrow{\mu}\in\mathcal{P}_+^3
=\mathcal{P}^3-\{(\emptyset,\emptyset,\emptyset)\}$, the three-partition Hodge integrals are defined by $$\begin{aligned}
G_{g,\overrightarrow{\mu}}(\overrightarrow{w})
& = \frac{(\sqrt{-1})^{l(\overrightarrow{\mu})}}
{|\mbox{Aut}(\overrightarrow{\mu})|}
\prod_{a=1}^3 \prod_{i=1}^{l(\mu^{(a)})}
\biggl\{
w_{a+1} \prod_{j=1}^{\mu^{(a)}_i}
\Bigl(1+\frac{\mu_i^{(a)}w_{a+1}}{jw_a}\Bigr)
\biggr\}
\\[1.5mm]
& \quad\,\times\,
\int_{\overline{\mathcal{M}}_{g,l(\overrightarrow{\mu})}}
\prod_{a=1}^3
\frac{\bigwedge^\vee_g(w_a)w_a^{l(\mu^{(a)})-1}}
{\prod_{i=1}^{l(\mu^{(a)})}\bigl(w_a-\mu_i^{(a)}\psi_i^{(a)}\bigr)},
\end{aligned}
\label{def_3-partition Hodge integral}$$ where $\psi^{(a)}_i =c_1(\mathbb{L}^{(a)}_i)$ are the $\psi$-classes associated with the marked points $z_i^{(a)}$.
We introduce generating functions of these three-partition Hodge integrals. Let $p^{(1)} =(p^{(1)}_k)_{k=1}^\infty$, $p^{(2)} =(p^{(2)}_k)_{k=1}^\infty$ and $p^{(3)} =(p^{(3)}_k)_{k=1}^\infty$ be formal variables, denoted collectively by $\overrightarrow{p} =(p^{(1)},p^{(2)},p^{(3)})$. We put $p^{(a)}_\mu =\prod_{i=1}^{l(\mu)} p^{(a)}_{\mu_i}$ for a non-zero partition $\mu=(\mu_i)_{i=1}^\infty$, and $p^{(a)}_\emptyset =1$ for the zero partition $\emptyset$. The generating functions are defined by $$\begin{aligned}
G_{\overrightarrow{\mu}}\left(\lambda;\overrightarrow{w}\right)
& = \sum_{g=0}^\infty\lambda^{2g-2+l(\overrightarrow{\mu})}
G_{g,\overrightarrow{\mu}}\left(\overrightarrow{w}\right),
\label{connected_G_genus} \\[1mm]
G\left(\lambda;\overrightarrow{p};\overrightarrow{w}\right)
& = \sum_{\overrightarrow{\mu}\in\mathcal{P}^3_+}
G_{\overrightarrow{\mu}}\left(\lambda;\overrightarrow{w}\right)
\prod_{a=1}^3p^{(a)}_{\mu^{(a)}},
\label{connected_G} \end{aligned}$$ where $\lambda$ amounts to the string coupling constant.
By degree counting of the RHS of (\[def\_3-partition Hodge integral\]), three-partition Hodge integrals turn out to be homogeneous of degree 0 with respect to $\overrightarrow{w}$: $$\begin{aligned}
G_{g,\overrightarrow{\mu}}\left(t\overrightarrow{w}\right)
= G_{g,\overrightarrow{\mu}}\left(\overrightarrow{w}\right), \quad
t\overrightarrow{w}=(tw_1,tw_2,tw_3),\quad
t\neq 0.\end{aligned}$$ This implies that $$\begin{aligned}
G_{\overrightarrow{\mu}}\left(\lambda;t\overrightarrow{w}\right)
= G_{\overrightarrow{\mu}}\left(\lambda;\overrightarrow{w}\right), \quad
G\left(\lambda;\overrightarrow{p};t\overrightarrow{w}\right)
= G\left(\lambda;\overrightarrow{p};\overrightarrow{w}\right).\end{aligned}$$ Therefore, $\overrightarrow{w}$ can be chosen as $$\begin{aligned}
\overrightarrow{w} = (1,\,\tau,\,-1-\tau)
\label{tau}\end{aligned}$$ where $\tau$ is a new variable. For $\overrightarrow{w}$ of the form (\[tau\]), the three-partition Hodge integrals are expressed in a shortened form as $G_{g,\overrightarrow{\mu}}(\tau) = G_{g,\overrightarrow{\mu}}(1,\tau,-1-\tau)$. The same abbreviation is also used for the generating functions (\[connected\_G\_genus\]) and (\[connected\_G\]) as $$\begin{aligned}
G_{\overrightarrow{\mu}}\left(\lambda;\tau\right)
= G_{\overrightarrow{\mu}}\left(\lambda;1,\tau ,-1-\tau\right),\quad
G\left(\lambda;\overrightarrow{p};\tau\right)
= G\left(\lambda;\overrightarrow{p};1,\tau ,-1-\tau\right). \end{aligned}$$
The generating function (\[connected\_G\]) can be expanded in an infinite series of the Schur functions by the inversion formula (\[inverse\_Frobenius\]). The disconnected version of (\[connected\_G\]) turns out to take the specific form [@Li_Liu_Liu_Zhou_2009] $$\begin{aligned}
\exp\left( G(\lambda;\overrightarrow{p};\overrightarrow{w}) \right)
= \sum_{\overrightarrow{\mu}\in\mathcal{P}^3}
\widetilde{\mathcal C}_{\overrightarrow{\mu}} (\lambda)
e^{\frac{\sqrt{-1}}{2}\lambda\sum_{a=1}^3\kappa(\mu^{(a)})w_{a+1}/w_a}
\prod_{a=1}^3 s^{(a)}_{\mu^{(a)}},
\label{expansion_disconnected_G}\end{aligned}$$ where $$\begin{aligned}
\kappa(\mu^{(a)})
=\sum_{i=1}^\infty\mu^{(a)}_i(\mu^{(a)}_i-2i+1)
$$ denotes the second Casimir invariant of $\mu^{(a)}$ and $$\begin{aligned}
s^{(a)}_{\mu^{(a)}}
=\sum_{\nu\in\mathcal{P}}\frac{\chi_{\mu^{(a)}}(\nu)}{z_\nu}p^{(a)}_\nu. \end{aligned}$$ $s^{(a)}_{\mu^{(a)}}$ is the Schur function $s_{\mu^{(a)}}$ obtained by substituting $p_\nu=p^{(a)}_\nu$.
(\[expansion\_disconnected\_G\]) shows that $\exp G(\lambda;\overrightarrow{p};\overrightarrow{w})$ depends on $\overrightarrow{w}$ in a particular form. It is a consequence of the [*invariance theorem*]{} [@Li_Liu_Liu_Zhou_2009] for a generating function of formal relative Gromov-Witten invariants of $\mathbb{C}^3$. These invariants count stable maps from possibly disconnected curves to $\mathbb{C}^3$. In the localization computation of these invariants, the fixed points are labelled by a triple of partitions. The contribution from each fixed point takes the form of three-partition Hodge integrals multiplied by the double Hurwitz numbers of $\mathbb{P}^1$. Thus, the generating function of the formal relative Gromov-Witten invariants is a sum of these products over the sets of partitions. Though these terms depend on $\overrightarrow{w}$, the invariance theorem implies that the generating function itself does not depend on $\overrightarrow{w}$ in total. Accordingly, $\overrightarrow{w}$-dependence of each term should cancel out after the summation. With the aid of a combinatorial expression [@Liu_Liu_Zhou_2007] of the double Hurwitz numbers, the cancellation eventually yields the expansion (\[expansion\_disconnected\_G\]).
Three-partition Hodge integrals at special values of $\tau$ reduce to two-partition Hodge integrals since the $\psi$-classes can be renumbered in terms of two partitions at such values of $\tau$. For instance, at $\tau=1$, $$\begin{aligned}
\begin{aligned}
G_{g,(\mu^{(1)},\,\mu^{(2)},\,\mu^{(3)})} (1)
& = (-1)^{|\mu^{(1)}|-l(\mu^{(1)})}
\frac{z_{\mu^{(1)}\cup\mu^{(2)}}}{z_{\mu^{(1)}}z_{\mu^{(2)}}}
G_{g,(\emptyset,\,\mu^{(1)}\cup\mu^{(2)},\,\mu^{(3)})} (1)
\\[0.2mm]
& \qquad
+\delta_{g, 0}\sum_{m=1}^\infty \frac{(-1)^{m+1}}{m}
\delta_{(\mu^{(1)},\,\mu^{(2)},\,\mu^{(3)}),\,(\emptyset,(m),(2m))},
\end{aligned}
\label{reduction}\end{aligned}$$ where $\mu^{(1)}\cup\mu^{(2)}$ denotes the partition obtained by the union of the parts of $\mu^{(1)}$ and $\mu^{(2)}$ which are rearranged in descending order. The second term in the RHS is an anomalous contribution from the unstable cases. In terms of the generating functions (\[connected\_G\]), the reduction formula (\[reduction\]) takes the form $$\begin{aligned}
\begin{aligned}
& \exp\left(
G\left(\lambda;(p^{(1)},p^{(2)},p^{(3)});1\right)
\right) \\[0.5mm]
& \qquad
=\exp\left( G\left(\lambda;(0,p^+,p^{(3)});1\right) \right)
\exp\left(
\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m} p_m^{(1)}p_{2m}^{(3)}
\right),
\end{aligned}
\label{reduction_formula_G} \end{aligned}$$ where $p^+ =(p^+_k)_{k=1}^\infty$, $p^+_k =(-1)^{k+1}p^{(1)}_k+p^{(2)}_k$.
An explicit expression of the coefficients $\widetilde{\mathcal C}_{\overrightarrow{\mu}}(\lambda)$ can be read out from (\[reduction\_formula\_G\]) by plugging (\[expansion\_disconnected\_G\]) into (\[reduction\_formula\_G\]): $$\begin{aligned}
\begin{aligned}
& e^{-\sqrt{-1}\lambda (-\kappa(\mu^{(1)})/2+\kappa(\mu^{(2)})+\kappa(\mu^{(3)})/4)}\,
\widetilde{\mathcal C}_{(\mu^{(1)},\,\mu^{(2)},\,\mu^{(3)})}\left(\lambda\right)
\\[2mm]
& \quad
= \sum_{\nu^1,\,\nu^3,\,\nu^+,\,\eta^1,\,\eta^3\,\in\mathcal{P}}
c_{\,\ltrans{\,\eta}^1\,\nu^1}^{\mu^{(1)}}\,
c_{\,\ltrans{\,\nu}^1\,\mu^{(2)}}^{\nu^+}\,
c_{\,\eta^3 \,\nu^3}^{\mu^{(3)}}\,
e^{-\sqrt{-1}\lambda (\kappa(\nu^+)+\kappa(\nu^3)/4)}\,
\widetilde{\mathcal C}_{(\emptyset,\,\nu^+,\,\nu^3)}\left(\lambda\right)
\\
& \qquad\qquad\qquad\qquad \times\,
\sum_{\xi\,\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi) \chi_{\eta^3}(2\xi)}{z_\xi},
\end{aligned}
\label{reduction_formula_C}\end{aligned}$$ where $2\xi =(2\xi_i)_{i=1}^\infty$ for $\xi=(\xi_i)_{i=1}^\infty$, and $c_{\,\ltrans{\,\eta}^1\,\nu^1}^{\mu^{(1)}}$, $c_{\,\ltrans{\,\nu}^1\,\mu^{(2)}}^{\nu^+}$ and $c_{\,\eta^3\,\nu^3}^{\mu^{(3)}}$ are the Littlewood-Richardson numbers (\[def\_LR\]). Furthermore, the [*cut-and-join equations*]{} for the two-partition Hodge integrals [@Liu_Liu_Zhou_2007] imply that $$\begin{aligned}
\widetilde{\mathcal C}_{(\emptyset,\,\nu^+,\,\nu^3)}(\lambda)
= q^{-\kappa(\nu^3)/2}
s_{\nu^+}(q^{-\rho})
s_{\ltrans{\,\nu}^3}(q^{-\nu^+-\rho}),
\label{two-partition_C}\end{aligned}$$ where $q=e^{-\sqrt{-1}\lambda}$, and $s_{\nu^+}(q^{-\rho})$ and $s_{\ltrans{\,\nu}^3}(q^{-\nu^+-\rho})$ are the special values of the infinite-variate Schur functions $s_{\nu^+}(\bm{x})$ and $s_{\ltrans{\,\nu}^3}(\bm{x})$ at $q^{-\rho}=\bigl(q^{i-1/2}\bigr)_{i=1}^\infty$ and $q^{-\nu^+-\rho}=\bigl(q^{-\nu^+_i+i-1/2}\bigr)_{i=1}^\infty$. As a consequence of (\[reduction\_formula\_C\]) and (\[two-partition\_C\]), $\widetilde{\mathcal C}_{\overrightarrow{\mu}}(\lambda)$ can be expressed in a closed form [@Li_Liu_Liu_Zhou_2009]: $$\begin{aligned}
\begin{aligned}
& \widetilde{\mathcal C}_{\overrightarrow{\mu}} (\lambda)
= q^{\kappa(\mu^{(1)})/2-\kappa(\mu^{(2)})-\kappa(\mu^{(3)})/4}
\\[1.5mm]
& \qquad \times\!\!
\sum_{\nu^1,\nu^3,\nu^+,\eta^1,\eta^3\,\in\mathcal{P}}\!\!
c_{\,\ltrans{\,\eta}^1\,\nu^1}^{~\mu^{(1)}}\,
c_{\,\ltrans{\,\nu}^1\,\mu^{(2)}}^{~\nu^+}\,
c_{\,\eta^3\,\ltrans{\,\nu}^3}^{~\mu^{(3)}}\,
q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho})
\\[0.2mm]
& \qquad \qquad \qquad \qquad \times
\sum_{\xi\,\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi)\chi_{\eta^3}(2\xi)}{z_\xi}.
\end{aligned}
\label{LLLZ_formula}\end{aligned}$$
Topological vertex and three-partition Hodge integrals
------------------------------------------------------
The toric data of a non-compact toric Calabi-Yau threefold are encoded in the associated fan of rational cones of dimension $\leq 3$ on $\mathbb{R}^3$. A plane section of this fan yields a triangulated polyhedron. Its dual graph is trivalent, and referred to as the “web” or “toric” diagram. Each vertex of this trivalent graph is given a vertex weight called the topological vertex. According to the proposal of Aganagic et al. [@AKMV], the topological string amplitudes on the Calabi-Yau threefold can be obtained by gluing these vertex weights along the edges of the graph.
Let $q$ be a parameter in the range $0 <|q|<1$. The vertex weight at each vertex is labelled by a triple of partitions $\overrightarrow{\mu}
=(\mu^{(1)},\mu^{(2)},\mu^{(3)})\in\mathcal{P}^3$, and defined as[^4] $$\begin{aligned}
\mathcal{C}_{\overrightarrow{\mu}}(q)
= q^{\kappa(\mu^{(1)})/2}s_{\ltrans{\,\mu}^{(2)}}(q^{-\rho})
\sum_{\eta\,\in\mathcal{P}}
s_{\mu^{(1)}/\eta}(q ^{-\ltrans{\,\mu}^{(2)}-\rho})
s_{\ltrans{\,\mu}^{(3)}/\eta}(q ^{-\mu^{(2)}-\rho}).
\label{topological_vertex_def}\end{aligned}$$ $\mu^{(1)},\mu^{(2)}$ and $\mu^{(3)}$ are assigned to the three legs of the trivalent vertex numbered in a counterclockwise direction. See Figure \[figure\_top\_vertex\].
![A diagrammatic representation of the topological vertex (\[topological\_vertex\_def\]).[]{data-label="figure_top_vertex"}](topological_vertex.eps)
$s_{\mu^{(1)}/\eta}(q^{-\ltrans{\,\mu}^{(2)}-\rho})$ and $s_{\ltrans{\,\mu}^{(3)}/\eta}(q ^{-\mu^{(2)}-\rho})$ are the special values of the infinite-variate skew Schur functions $s_{\mu^{(1)}/\eta} (\bm{x})$ and $s_{\ltrans{\,\mu}^{(3)}/\eta}(\bm{x})$ at $q ^{-\ltrans{\,\mu}^{(2)}-\rho} =(q^{-\ltrans{\,\mu}^{(2)}_i+i-1/2})_{i=1}^\infty$ and $q ^{-\mu^{(2)}-\rho} =(q^{-\mu^{(2)}_i +i-1/2})_{i=1}^\infty$.
We introduce a generating function of the topological vertex. Let $\bm{x}=\left(x_i\right)_{i=1}^\infty$, $\bm{y}=\left(y_i\right)_{i=1}^\infty$ and $\bm{z}=\left(z_i\right)_{i=1}^\infty$ be three sets of infinitely many variables. Define the generating function as $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};\overrightarrow{w}\right)
= \sum_{\overrightarrow{\mu}\in \mathcal{P}^3}
\mathcal{C}_{\overrightarrow{\mu}}\left(q\right)
q^{-\frac{1}{2}\sum_{a=1}^3\kappa(\mu^{(a)})(1+w_{a+1}/w_a)}
s_{\mu^{(1)}}(\bm{x}) s_{\mu^{(2)}}(\bm{y}) s_{\mu^{(3)}}(\bm{z}).
\label{generating_function_topological_vertex} \end{aligned}$$ In the case where $\overrightarrow{w}$ takes the form (\[tau\]), we use the abbreviated notation $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};\tau\right)
= \mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1,\tau,-1-\tau\right). \end{aligned}$$
Our goal is to show the equivalence between the aforementioned generating functions of the topological vertex and three-partition Hodge integrals. More precisely, we prove the following:
\[theorem\_conjecture\] Let $q=e^{-\sqrt{-1}\lambda}$. The generating function (\[generating\_function\_topological\_vertex\]) coincides with the exponential of the generating function (\[connected\_G\]) as $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};\overrightarrow{w}\right)
= \exp\left(G\left(\lambda;\overrightarrow{p};\overrightarrow{w}\right)\right),
\label{eq1_theorem_conjecture}\end{aligned}$$ where the variables $\bm{x}$, $\bm{y}$ and $\bm{z}$ are related to $\overrightarrow{p}=(p^{(1)},p^{(2)},p^{(3)})$ as $$\begin{aligned}
p^{(1)}_k =p_k(\bm{x}),\quad p^{(2)}_k=p_k(\bm{y}), \quad
p^{(3)}_k =p_k(\bm{z}), \quad
k=1,2,\dots ~.
\label{identify_p_x}\end{aligned}$$ In other words, the coefficients $\widetilde{\mathcal{C}}_{\overrightarrow{\mu}}(\lambda)$ of (\[expansion\_disconnected\_G\]) are connected with the topological vertex as $$\begin{aligned}
\widetilde{\mathcal{C}}_{\overrightarrow{\mu}}(\lambda)
= q^{-\frac{1}{2}\sum_{a=1}^3\kappa(\mu^{(a)})}
\mathcal{C}_{\overrightarrow{\mu}}(q), \quad
\forall\,\overrightarrow{\mu}=(\mu^{(1)},\mu^{(2)},\mu^{(3)})\in\mathcal{P}^3.
\label{tildeC=C}\end{aligned}$$
An immediate corollary of this result is cyclic symmetry of the topological vertex, which means that it is invariant under a cyclic permutation of three partitions assigned to the three legs, i.e., $$\begin{aligned}
\mathcal{C}_{(\mu^{(1)},\mu^{(2)},\mu^{(3)})}(q)
=\mathcal{C}_{(\mu^{(2)},\mu^{(3)},\mu^{(1)})}(q), \quad
\forall\,\overrightarrow{\mu} =(\mu^{(1)},\mu^{(2)},\mu^{(3)})\in\mathcal{P}^3.
\label{cyclicity_topological_vertex}\end{aligned}$$ Actually, three-partition Hodge integral is easily found to be invariant under simultaneous cyclic permutations of $\overrightarrow{\mu} =(\mu^{(1)},\mu^{(2)},\mu^{(3)})$ and $\overrightarrow{w}=(w_1,w_2,w_3)$ as $$\begin{aligned}
G_{g,\,\mu^{(1)},\mu^{(2)},\mu^{(3)}}(w_1,w_2,w_3)
= G_{g,\,\mu^{(2)},\mu^{(3)},\mu^{(1)}}(w_2,w_3,w_1). \end{aligned}$$ This fact implies the cyclic symmetry of $\mathcal{C}_{\overrightarrow{\mu}}(q)$ by (\[tildeC=C\]).
Integrable structure at positive integer values of $\tau$
---------------------------------------------------------
The reduction formula (\[reduction\_formula\_G\]) of three-partition Hodge integrals can be extended to the case where $\tau$ takes any positive integer value $N = 1,2,\ldots$. Moreover, this leads to an unexpected relation with the generalized KdV hierarchies.
The generalized KdV hierarchies are reduced systems of the KP hierarchy (see Dickey’s book [@Dickey_book] for details of these integrable hierarchies). The KP hierarchy in the Lax formalism is the system $$\begin{aligned}
\frac{\partial L}{\partial t_k} = [B_k,L], \quad k = 1,2,\ldots,\end{aligned}$$ of Lax equations for the pseudo-differential operator $$\begin{aligned}
L = \partial_x +\sum_{n=1}^\infty u_{n+1}\partial_x^{-n}, \quad
\partial_x = \frac{\partial}{\partial x}, \end{aligned}$$ where the coefficients $u_n$ are functions of the time variables ${\boldsymbol{t}}= (t_1,t_2,\ldots$), $t_1$ is identified with $x$, and $B_k$ is the differential operator part $$\begin{aligned}
B_k = (L^k)_{+} = \partial_x^k + b_{k2}\partial_x^{k-2} + \cdots + b_{kk}\end{aligned}$$ of the $k$-th power of $L$. If the $N+1$-st power ${\mathcal{L}}= L^{N+1}$ of $L$ is a differential operator, i.e., $$\begin{aligned}
{\mathcal{L}}= B_{N+1} = \partial_x^{N+1} + b_2\partial_x^{N-1} + \cdots + b_{N+1}, \end{aligned}$$ $L$ does not depend on $t_{m(N+1)}$, $m = 1,2,\dots$: $$\frac{\partial L}{\partial t_{m(N+1)}} = 0, \quad
m = 1,2,\ldots.
$$ This condition conversely characterizes the condition $L^{N+1} = B_{N+1}$. The remaining Lax equations can be reduced to the Lax equations $$\begin{aligned}
\frac{\partial{\mathcal{L}}}{\partial t_k} = [B_k,{\mathcal{L}}], \quad k = 1,2,\ldots, \end{aligned}$$ for ${\mathcal{L}}$. These Lax equations comprises the $N$-th generalized KdV hierarchy. The KdV hierarchy amounts to the case where $N = 1$.
This reduction procedure can be reformulated in terms of the tau function ${\mathcal{T}}= {\mathcal{T}}({\boldsymbol{t}})$. The tau function yields the dressing operator $$\begin{aligned}
W = 1 + \sum_{n=1}^\infty w_n\partial_x^{-n} \end{aligned}$$ via the generating function $$w({\boldsymbol{t}},z)
= 1 + \sum_{n=1}^\infty w_nz^{-n}
= \frac{{\mathcal{T}}(t_1-z^{-1},\ldots,t_k-z^{-k}/k,\ldots)}
{{\mathcal{T}}(t_1,\ldots,t_k,\ldots)}.
$$ The dressing operator, in turn, defines the Lax operator as $$\begin{aligned}
L = W\partial_xW^{-1}.\end{aligned}$$ Consequently, if the tau function satisfies the equations $$\frac{\partial^2\log{\mathcal{T}}}{\partial t_k\partial t_{m(N+1)}} = 0, \quad
k,m = 1,2,\ldots,
\label{red-cond2}$$ the coefficients of $W$ and $L$ do not depend on $t_{m(N+1)}$, $m = 1,2,\ldots$. This is a sufficient condition for the KP hierarchy to reduce to the $N$-th generalized KdV hierarchy. Note that this reduction condition is weaker than the commonly used condition $$\frac{\partial{\mathcal{T}}}{\partial t_{m(N+1)}} = 0, \quad
m = 1,2,\ldots.
$$ It is the weaker condition (\[red-cond2\]) that is relevant to the generating function (\[expansion\_disconnected\_G\]) of three-partition Hodge integrals.
Let ${\mathcal{T}}(\lambda,N,p^{(1)},p^{(2)},{\boldsymbol{t}})$ denote the function of ${\boldsymbol{t}}= (t_k)_{k=1}^\infty$ obtained from the generating function (\[expansion\_disconnected\_G\]) by substituting $t_k = p^{(3)}_k/k$, $k = 1,2,\ldots$: $${\mathcal{T}}(\lambda,N,p^{(1)},p^{(2)},{\boldsymbol{t}})
= \exp\left(G(\lambda;\,\vec{p};\,N)\right).
\label{def_calT}$$ $\lambda,N,p^{(1)}$ and $p^{(2)}$ are treated as parameters.
\[theorem\_NKdV\] Let $N$ be a positive integer. ${\mathcal{T}}(\lambda,N,p^{(1)},p^{(2)},{\boldsymbol{t}})$ is a tau function of the KP hierarchy that satisfies the condition (\[red-cond2\]). The associated Lax operator $L$ satisfies the reduction condition $L^{N+1} = B_{N+1}$ for the $N$-th generalized KdV hierarchy.\
### {#section .unnumbered}
We start Section 2 with a brief review on a two-dimensional charged free fermion system. Various operators on the fermionic Fock space, including a realization of the two-dimensional quantum torus algebra, are introduced. These operators are used as fundamental tools in the discussions through the article. In Proposition \[fermionic\_formula\_W(q;x,y,z;tau)\], we give a fermionic representation of the generating function of the topological vertex.
In Section 3, we explain the generalized shift symmetries. These symmetries act on a set of basis elements $V_m^{(k)}$ of the quantum torus algebra so as to shift the indices $k, m$ in a certain way. The shift symmeries are formulated by special values of the vertex operators $\Gamma_\pm(x)$ and $\Gamma'_\pm(x)$ and the exponential operator $q^{K/2}=\exp(\frac{K}{2}\log q)$, as $$\begin{aligned}
V^{(k)}_m\mathbb{L}_\emptyset = (-1)^k\mathbb{L}_\emptyset V^{(k)}_{m-k}, \quad
V^{(-k)}_m\mathbb{L}'_\emptyset = \mathbb{L}'_\emptyset V^{(-k)}_{m-k}, \quad
q^{K/2}V_m^{(k)}q^{-K/2} =V_m^{(k-m)}, \end{aligned}$$ where $\mathbb{L}_\emptyset =\Gamma_+ (q^{-\rho})\Gamma_-(q^{-\rho})$ and $\mathbb{L}'_\emptyset =\Gamma'_+ (q^{-\rho})\Gamma'_-(q^{-\rho})$ are the products of multi-variate vertex operators $\Gamma_\pm(\bm{x}) =\prod_{i=1}^\infty \Gamma_\pm(x_i)$ and $\Gamma'_\pm(\bm{x})=\prod_{i=1}^\infty \Gamma'_\pm(x_i)$ specialized to $q^{-\rho}$. These shift symmetries can be generalized by replacing $\mathbb{L}_\emptyset$ and $\mathbb{L}'_\emptyset$ with $$\begin{aligned}
\mathbb{L}_\alpha =s_\alpha(q^{-\rho})
\Gamma_-(q^{-\alpha-\rho})
\Gamma_+(q^{-\ltrans{\,\alpha}-\rho}), \quad
\mathbb{L}'_\alpha =s_\alpha(q^{-\rho})
\Gamma'_-(q^{-\alpha-\rho})
\Gamma'_+(q^{-\ltrans{\,\alpha}-\rho}) \end{aligned}$$ and modifying $q^{K/2}$ in a certain manner. The generalized shift symmetries are summarized in Propositions \[ext\_shift\_symmetry\_quote\_1\] and \[ext\_shift\_symmetry\_quote\_2\].
The fermionic representation of the generating function in Proposition \[fermionic\_formula\_W(q;x,y,z;tau)\] involves a weighted sum of operators of the form $$\begin{aligned}
\sum_{\alpha \in \mathcal{P}}\mathbb{G}_\alpha(\tau)s_\alpha, \end{aligned}$$ where $\displaystyle
\mathbb{G}_\alpha(\tau)
= q^{-\kappa(\alpha)/2} q^{-\tau K/2} \mathbb{L}_\alpha q^{\tau K/2(1+\tau)}$ is a particular linear combination of the generalized shift symmetries. The generalized shift symmetries imply that, at certain values of $\tau$, the foregoing sum is factorized into a triple product of operators. The factorization formula at $\tau=1$ is given in Theorem \[factorization\_formulas\_tau=1\].
In Section 4, we prove Theorem \[theorem\_conjecture\] using the results obtained in the preceding section. The aforementioned fermionic representation of the generating function of the topological vertex can be converted into a new representation by Theorem \[factorization\_formulas\_tau=1\]. The resulting representation is presented in Proposition \[fermionic\_formula\_W(q;x,y,z;1)\]. With the aid of the new representation of the generating function, we can reconsider its Schur function expansion and eventually find out the relation (\[tildeC=C\]). Theorem \[theorem\_conjecture\] is thus proved.
In Section 5, we prove Theorem \[theorem\_NKdV\]. The foregoing weighted sum of operators can be factorized in the cases where $\tau =1/N$, $N=2,3,\dots$, as well. The factorized expression is presented in Theorem \[factorization\_formulas\_1/N\]. This formula yields in Proposition \[fermionic\_formula\_W(q;x,y,z;N)\] a new representation of the generating functions of the topological vertex at positive integral values of $\tau$. Theorem \[theorem\_conjecture\] and Proposition \[fermionic\_formula\_W(q;x,y,z;N)\] lead to a reduction formula of the generating functions of three-partition Hodge integrals in Proposition \[reduction\_formula\_G\_tau=N\]. This formula gives a generalization of (\[reduction\]) at all positive integral values of $\tau$. We then consider integrable hierarchies underlying the generating functions of two-partition Hodge integrals. By combining these considerations, we eventually obtain Theorem \[theorem\_NKdV\].
### **** {#section-1 .unnumbered}
This work is supported by JSPS KAKENHI Grant Numbers JP15K04912 and JP18K03350.
Fermionic representations of generating functions
=================================================
Fermionic Fock space and operators
----------------------------------
Let $\psi_n,\psi^*_n,n\in\mathbb{Z}$, denote the Fourier modes of two-dimensional charged free fermion fields $$\psi (z) =\sum_{n\in\mathbb{Z}}\psi_nz^{-n-1},\quad
\psi^*(z) =\sum_{n\in\mathbb{Z}}\psi^*_nz^{-n}.$$ They satisfy the anti-commutation relations $$\psi_m\psi^*_n+\psi^*_n\psi_m =\delta_{m+n,0},\quad
\psi_m\psi_n+\psi_n\psi_m =0,\quad \psi^*_m\psi^*_n+\psi^*_n\psi^*_m =0.$$
The associated Fock space and its dual space are decomposed into the charge-$p$ sector for $p\in\mathbb{Z}$. Only the charge-$0$ sector is relevant to the subsequent computations. An orthonormal basis of the charge-$0$ sector is given by the ground states $$\begin{aligned}
& \langle0| =\langle-\infty|\dots\psi^*_{-i+1}\dots\psi^*_{-1}\psi^*_0,
\\[2mm]
& |0\rangle =\psi_0\psi_1\dots\psi_{i-1}\dots|-\infty\rangle \end{aligned}$$ and the excited states $$\begin{aligned}
& \langle\lambda|
=\langle-\infty| \dots
\psi^*_{\lambda_i -i+1}\dots\psi^*_{\lambda_2-1}\psi^*_{\lambda_1},
\\[2mm]
& |\lambda\rangle
=\psi_{-\lambda_1}\psi_{-\lambda_2+1}\dots\psi_{-\lambda_i+i-1}
\dots|-\infty\rangle \end{aligned}$$ labelled by partitions. The normal ordered product is prescribed by $$\begin{aligned}
:\psi_m\psi^*_n:\,\,=\,\psi_m\psi^*_n-\langle 0|\psi_m\psi^*_n|0\rangle.\end{aligned}$$
Fundamental tools in our computations are the following operators on the Fock space which preserve the charge.
1. The zero modes $$L_0 = \sum_{n\in\mathbb{Z}}n:\psi_{-n}\psi^*_n:,\quad
W_0 = \sum_{n\in\mathbb{Z}}n^2:\psi_{-n}\psi^*_n:$$ of the Virasoro and $\mbox{W}_3$ algebras and the Fourier modes $$J_m = \sum_{n\in\mathbb{Z}}:\psi_{m-n}\psi^*_{n}:,\quad
m\in\mathbb{Z}$$ of the $U(1)$ current $J(z) = \sum_{m\in\mathbb{Z}}J_mz^{-m-1}=:\psi (z)\psi^* (z):$.
2. The fermionic realization $$K = W_0-L_0+J_0/4
= \sum_{n\in\mathbb{Z}}(n-1/2)^2:\psi_{-n}\psi^*_n:$$ of the so called “cut-and-join operator” [@Goulden_Jackson_1997].
3. The basis elements $$V^{(k)}_m
=q^{-k(m+1)/2}\sum_{n\in\mathbb{Z}}q^{kn}:\psi_{m-n}\psi^*_n:
- \,\frac{q^{k/2}}{1-q^k}\,\delta_{m,0}, \quad
k\in\mathbb{Z}_{\neq 0},~ m\in\mathbb{Z}$$ and $$V^{(0)}_m =J_m, \quad m\in\mathbb{Z}$$ of a fermionic realization[^5] of the quantum torus algebra, which satisfy the commutation relations $$\left[\,V^{(k)}_m,\,V^{(l)}_n\,\right]
= \left(q^{-\frac{kn-lm}{2}}-q^{\frac{kn-lm}{2}}\right)
V^{(k+l)}_{m+n}+m\,\delta_{k+l,0}\,\delta_{m+n,0}, \quad
k,l,m,n\in\mathbb{Z}.$$
4. The vertex operators $$\Gamma_\pm(z)
= \exp\left(\sum_{k=1}^\infty\frac{z^k}{k}J_{\pm k}\right), \quad
\Gamma'_\pm(z)
= \exp\left(-\sum_{k=1}^\infty\frac{(-z)^k}{k}J_{\pm k}\right)$$ and the multi-variable extensions $$\Gamma_\pm(\bm{x})
= \prod_{i=1}^\infty\Gamma_{\pm}(x_i), \quad
\Gamma'_\pm(\bm{x})
= \prod_{i=1}^\infty\Gamma'_{\pm}(x_i).$$
The matrix elements of those operators are well-known. $J_0=V_0^{(0)}$, $L_0$ and $W_0$ are diagonal to the basis $\left\{|\lambda\rangle\right\}_{\lambda\in\mathcal{P}}$ in the charge-$0$ sector as $$\langle\lambda|J_0|\mu\rangle
= 0,\quad
\langle\lambda|L_0|\mu\rangle
= \delta_{\lambda\mu}|\lambda|,\quad
\langle\lambda|W_0|\mu\rangle
= \delta_{\lambda\mu}(\kappa(\lambda)+|\lambda|).$$ Consequently, we have $$\begin{aligned}
\langle\lambda|K|\mu\rangle
= \delta_{\lambda\mu}\kappa(\lambda).
\label{matrix_element_K}\end{aligned}$$ Other zero modes of the quantum torus algebra are likewise diagonal to the basis $\left\{|\lambda\rangle\right\}_{\lambda\in\mathcal{P}}$. Their matrix elements are of the form $$\langle\lambda|V^{(k)}_0|\mu\rangle
= \delta_{\lambda\mu}\varphi_k(\mu),
\quad k\in\mathbb{Z}_{\neq 0},$$ where $$\begin{aligned}
\varphi_k(\mu)
=\left\{\begin{array}{rl}
\displaystyle -\sum_{i=1}^\infty q^{k\left(-\ltrans{\,\mu}_i+i-1/2\right)}
& \mbox{for $k\geq 1$}, \\[2mm]
\displaystyle \sum_{i=1}^\infty q^{-k\left(-\mu_i+i-1/2\right)}
& \mbox{for $k\leq -1$}.
\end{array}\right.\end{aligned}$$
It is also well known that the matrix elements of $\Gamma_\pm(\bm{x})$ and $\Gamma'_\pm(\bm{x})$ are skew Schur functions $$\begin{aligned}
& \langle\lambda|\Gamma_-(\bm{x})|\mu\rangle
= \langle\mu|\Gamma_+(\bm{x})|\lambda\rangle =s_{\lambda /\mu}(\bm{x}),
\label{matrix_element_Gamma}
\\[2.5mm]
& \langle\lambda|\Gamma'_-(\bm{x})|\mu\rangle
= \langle\mu|\Gamma'_+(\bm{x})|\lambda\rangle
= s_{\,\ltrans{\,\lambda}/\ltrans{\,\mu}}(\bm{x}).
\label{matrix_element_Gamma'}\end{aligned}$$
Fermionic representation of generating function (\[generating\_function\_topological\_vertex\])
-----------------------------------------------------------------------------------------------
Let us provide a fermionic representation of the generating function (\[generating\_function\_topological\_vertex\]). We first introduce a set of operator-valued functions labeled by partitions as $$\begin{aligned}
\mathbb{G}_\alpha (\tau)
= q^{-\frac{\kappa(\alpha)}{2\tau}}s_\alpha (q^{-\rho})\,
q^{-\frac{\tau}{2}K}\Gamma_-\left(q^{-\alpha -\rho}\right)
\Gamma_+\bigl(q^{-\ltrans{\,\alpha}-\rho}\bigr)\,
q^{\frac{\tau}{2(1+\tau)}K},\quad
\alpha\in\mathcal{P},~ \tau\neq 0,-1,
\label{G_alpha(tau)} \end{aligned}$$ where $q^{-\tau K/2}$ and $q^{\tau K/2(1+\tau)}$ are the exponential operators $\exp\bigl(-\frac{\tau K}{2}\log q\bigr)$, $\exp\bigl(\frac{\tau K}{2(1+\tau)}\log q\bigr)$, and $\Gamma_-\left(q^{-\alpha -\rho}\right)$ and $\Gamma_+\bigl(q^{-\ltrans{\,\alpha}-\rho}\bigr)$ denote respectively the multi-variate operators $\Gamma_-(\bm{x})$ and $\Gamma_+(\bm{x})$ specialized at $q^{-\alpha -\rho}$ and $q^{-\ltrans{\,\alpha}-\rho}$. The matrix element of (\[G\_alpha(tau)\]) yields the topological vertex (\[topological\_vertex\_def\]) as follows:
\[prop\_fermion\_expression\_topological\_vertex\] $$\begin{aligned}
\mathcal{C}_{\overrightarrow{\mu}}(q)
= q^{ \frac{1+\tau}{2}\kappa (\mu^{(1)})
-\frac{1}{2\tau}\kappa(\mu^{(2)})
+\frac{\tau}{2(1+\tau)}\kappa(\mu^{(3)})}
\langle\mu^{(1)}|
\mathbb{G}_{\,\ltrans{\,\mu}^{(2)}}(\tau)
|\ltrans{\,\mu}^{(3)}\rangle,
\label{fermion_expression_topological_vertex}\end{aligned}$$ where $\overrightarrow{\mu}=(\mu^{(1)},\mu^{(2)},\mu^{(3)})$.
#### *Proof.*
We compute the matrix element in the RHS of (\[fermion\_expression\_topological\_vertex\]) by interposing the identity $\sum_{\eta\in\mathcal{P}}|\eta\rangle\langle\eta|=1$ in the charge-$0$ sector as follows: $$\begin{aligned}
& \langle\mu^{(1)}|
\mathbb{G}_{\,\ltrans{\,\mu}^{(2)}}(\tau)
|\ltrans{\,\mu}^{(3)}\rangle
\nonumber \\
& = q^{-\frac{\kappa(\ltrans{\,\mu}^{(2)})}{2\tau}}
s_{\,\ltrans{\,\mu}^{(2)}}(q^{-\rho})
\langle\mu^{(1)}|
q^{-\frac{\tau}{2}K}\,
\Gamma_-(q^{-\ltrans{\,\mu}^{(2)}-\rho})
\left(\sum_{\eta\in\mathcal{P}}|\eta\rangle\langle\eta|\right)
\Gamma_+(q^{-\mu^{(2)}-\rho})
q^{\frac{\tau}{2(1+\tau)}K}
|\ltrans{\,\mu}^{(3)}\rangle
\nonumber \\[1mm]
& =q^{\frac{\kappa(\mu^{(2)})}{2\tau}}s_{\,\ltrans{\,\mu}^{(2)}}(q^{-\rho})
\sum_{\eta\in\mathcal{P}}
\langle\mu^{(1)}|
q^{-\frac{\tau}{2}K}\,\Gamma_-(q^{-\ltrans{\,\mu}^{(2)}-\rho})
|\eta\rangle \langle\eta|
\Gamma_+(q^{-\mu^{(2)}-\rho})q^{\frac{\tau}{2(1+\tau)}K}
|\ltrans{\,\mu}^{(3)}\rangle.
\label{fermion_expression_topological_vertex_proof}\end{aligned}$$ By (\[matrix\_element\_K\]) and (\[matrix\_element\_Gamma\]), the matrix elements in the RHS can be expressed as $$\begin{aligned}
\langle\mu^{(1)}|
q^{-\frac{\tau}{2}K}\,\Gamma_-(q^{-\ltrans{\,\mu}^{(2)}-\rho})
|\eta\rangle
& = q^{-\frac{\tau}{2}\kappa(\mu^{(1)})}
s_{\mu^{(1)}/\eta}(q^{-\ltrans{\,\mu}^{(2)}-\rho}),
\\[2.5mm]
\langle\eta|
\Gamma_+(q^{-\mu^{(2)}-\rho})
q^{\frac{\tau}{2(1+\tau)}K}
|\ltrans{\,\mu}^{(3)}\rangle
& = q^{-\frac{\tau}{2(1+\tau)}\kappa(\mu^{(3)})}
s_{\,\ltrans{\,\mu}^{(3)}/\eta}(q^{-\mu^{(2)}-\rho}). \end{aligned}$$ By plugging those expressions into (\[fermion\_expression\_topological\_vertex\_proof\]), we find $$\begin{aligned}
& \langle\mu^{(1)}|
\mathbb{G}_{\,\ltrans{\,\mu}^{(2)}}(\tau)
|\ltrans{\,\mu}^{(3)}\rangle
\\[1mm]
& \quad
= q^{-\frac{\tau}{2}\kappa(\mu^{(1)})+\frac{1}{2\tau}\kappa(\mu^{(2)})
-\frac{\tau}{2(1+\tau)}\kappa(\mu^{(3)})}
s_{\ltrans{\,\mu}^{(2)}}(q^{-\rho})
\sum_{\eta \in \mathcal{P}}
s_{\mu^{(1)}/\eta}(q^{-\ltrans{\,\mu}^{(2)}-\rho})
s_{\,\ltrans{\,\mu}^{(3)}/\eta}(q^{-\mu^{(2)}-\rho})
\\
& \quad
= q^{-\frac{1+\tau}{2}\kappa(\mu^{(1)})+\frac{1}{2\tau}\kappa(\mu^{(2)})
-\frac{\tau}{2(1+\tau)}\kappa(\mu^{(3)})}
\mathcal{C}_{\overrightarrow{\mu}}(q). \end{aligned}$$ Thus we obtain (\[fermion\_expression\_topological\_vertex\]).\
The generating function (\[generating\_function\_topological\_vertex\]) has a fermionic expression of the following form due to (\[fermion\_expression\_topological\_vertex\]):
\[fermionic\_formula\_W(q;x,y,z;tau)\] $$\begin{aligned}
\mathcal{W}(q;\bm{x},\bm{y},\bm{z};\tau)
= \langle 0|
\Gamma_+(\bm{x})
\biggl(
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha (\tau)
s_{\,\ltrans{\,\alpha}}(\bm{y})
\biggr)
\Gamma'_-(\bm{z})|0\rangle.
\label{W(q;x,y,z;tau)_fermion_expression}\end{aligned}$$
#### *Proof.*
Using (\[fermion\_expression\_topological\_vertex\]), three sums in the RHS of (\[generating\_function\_topological\_vertex\]) for $\overrightarrow{w}=(1,\tau,-1-\tau)$ become $$\begin{aligned}
\mathcal{W}(q;\bm{x},\bm{y},\bm{z};\tau)
& = \sum_{\overrightarrow{\mu}\in\mathcal{P}^3}
\langle\mu^{(1)}|
\mathbb{G}_{\,\ltrans{\,\mu}^{(2)}}(\tau)
|\ltrans{\,\mu}^{(3)}\rangle
s_{\mu^{(1)}}(\bm{x})s_{\mu^{(2)}}(\bm{y})s_{\mu^{(3)}}(\bm{z})
\nonumber \\[1mm]
& = \Bigl(
\sum_{\mu^{(1)}\in\mathcal{P}}s_{\mu^{(1)}}(\bm{x})
\langle\mu^{(1)}|
\Bigr)
\sum_{\mu^{(2)}\in\mathcal{P}}
\mathbb{G}_{\,\ltrans{\,\mu}^{(2)}}(\tau)s_{\mu^{(2)}}(\bm{y})
\Bigl(
\sum_{\mu^{(3)}\in\mathcal{P}}
|\ltrans{\,\mu}^{(3)}\rangle s_{\mu^{(3)}}(\bm{z})
\Bigr)
\nonumber \\[1mm]
& = \langle 0|\Gamma_+(\bm{x})
\biggl(
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_{\alpha}(\tau)s_{\,\ltrans{\,\alpha}}(\bm{y})
\biggr)
\Gamma'_-(\bm{z})|0\rangle, \end{aligned}$$ where (\[matrix\_element\_Gamma\]) and (\[matrix\_element\_Gamma’\]) are used in the last line.
Generalized shift symmetry and factorization formula
====================================================
The shift symmetries [@Nakatsu_Takasaki_CMP_2009; @Nakatsu_Takasaki_ASPS] act on the basis elements of the quantum torus algebra. There are three different types of shift symmetries:
- For $k\geq 1$ and $m\in\mathbb{Z}$, $$\begin{aligned}
V^{(k)}_m\,\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
= (-1)^k\Gamma_-(q^{-\rho}) \Gamma_+(q^{-\rho})\,V^{(k)}_{m-k}.
\label{shift_symmetry_1}\end{aligned}$$
- For $k\geq 1$ and $m\in\mathbb{Z}$, $$\begin{aligned}
V^{(-k)}_{m}\,\Gamma'_-(q^{-\rho})\Gamma'_+(q^{-\rho})
=\Gamma'_-(q^{-\rho})\Gamma'_+(q^{-\rho})\,V^{(-k)}_{m-k}.
\label{shift_symmetry_2}\end{aligned}$$
- For $k,m\in\mathbb{Z}$, $$\begin{aligned}
q^{K/2}V^{(k)}_mq^{-K/2}=V^{(k-m)}_m.
\label{shift_symmetry_3}\end{aligned}$$
These symmetries together with their combinations are used to clarify the integrable structures of various melting crystal models [@Nakatsu_Takasaki_CMP_2009; @Nakatsu_Takasaki_ASPS] and open topological string amplitudes [@Takasaki_Nakatsu_JPA_2016], [@Takasaki_2014]. Likewise, quantum curves of open topological strings on the closed topological vertex and the strip geometries are derived [@Takasaki_Nakatsu_SIGMA_2017] using the symmetries. An extension of the symmetries is investigated in the companion paper [@Nakatsu_Takasaki_2019]. The generalized shift symmetries obtained therein provides a powerful tool to prove Theorem \[theorem\_conjecture\].
Generalized shift symmetries on quantum torus algebra
-----------------------------------------------------
The skew Young diagram $\alpha/\beta$ is called a ribbon iff there exist positive integers $r$ and $s$, $r>s$, such that partitions $\alpha=(\alpha_i)_{i=1}^\infty$ and $\beta=(\beta_i)_{i=1}^\infty$ correlate to each other through the condition $$\begin{aligned}
\bigl\{\alpha_i-i\bigr\}_{i=1}^N\cup\left\{\beta_r-r\right\}
=\bigl\{\beta_i-i\bigr\}_{i=1}^N\cup\left\{\alpha_s-s\right\}, \quad
\forall\,N>r, \end{aligned}$$ where $\bigl\{\alpha_i-i\bigr\}_{i=1}^N$ and $\bigl\{\beta_i-i\bigr\}_{i=1}^N$ are finite subsets of $\mathbb{Z}$ consisting of $N$ distinct integers. See Figure \[figure\_1\]. The semi-infinite subsets $\bigl\{\alpha_i-i\bigr\}_{i=1}^\infty$ and $\bigl\{\beta_i-i\bigr\}_{i=1}^\infty$ are referred to the Maya diagrams of $\alpha$ and $\beta$. The length of a ribbon $\alpha/\beta$ is given by $|\alpha|-|\beta|$ and the height, denoted by ${\sf ht}(\alpha/\beta)$, is given by $r-s$.
![ The Young diagrams $\alpha =(7,5,4,4,3,2)$ and $\beta=(7,5,3,2,1)$ are laid to overlap. The skew Young diagram $\alpha/\beta$ is indicated with hatched boxes and is a ribbon of length $7$ and height $3$.[]{data-label="figure_1"}](ribbon_ver.5.eps)
The set of partitions of which subtraction from $\alpha$ leaves a ribbon of length $k$ is denoted by $$\begin{aligned}
\mathcal{R}_{k,\alpha}^{(-)}
=\Bigl.\Bigl\{\beta\in\mathcal{P}\,\Bigr|\,
\mbox{$\alpha/\beta$ is a ribbon of length $k$.}\Bigr\}. \end{aligned}$$ Likewise, the set of partitions which leave a ribbon of length $k$ by subtraction of $\alpha$ is $$\begin{aligned}
\mathcal{R}_{k,\alpha}^{(+)}
=\Bigl.\Bigl\{\beta\in\mathcal{P}\,\Bigr|\,
\mbox{$\beta/\alpha$ is a ribbon of length $k$.}\Bigr\}. \end{aligned}$$ We describe these sets in a single symbol indexed by $\alpha\in\mathcal{P}$ and $k\in\mathbb{Z}$, $k\neq 0$, as $$\begin{aligned}
\mathcal{R}_{k,\alpha}
=\left\{\begin{array}{ll}
\displaystyle \mathcal{R}_{k,\alpha}^{(-)}
& \quad \mbox{for $k\geq 1$},
\\[2mm]
\displaystyle \mathcal{R}_{-k,\alpha}^{(+)}
& \quad \mbox{for $k\leq -1$}.
\end{array}\right. \end{aligned}$$ That means that $\mathcal{R}_{k,\alpha}$ consists of partitions which can be obtained from $\alpha$ by removing a ribbon of length $k$ for $k\geq 1$ or by adding a ribbon of length $-k$ for $k\leq -1$.
For $\alpha\in\mathcal{P}$ and $\beta\in\mathcal{R}_{k,\alpha}$, we introduce their relative sign by $$\begin{aligned}
{\sf sgn}(\alpha,\beta)
=\left\{\begin{array}{ll}
(-1)^{{\sf ht}(\alpha/\beta)}
& \quad \mbox{for $k\geq 1$},
\\[2mm]
(-1)^{{\sf ht}(\beta/\alpha)}
& \quad \mbox{for $k\leq -1$}.
\end{array}\right.\end{aligned}$$ Since $\beta\in\mathcal{R}_{k,\alpha}$ implies $\alpha\in\mathcal{R}_{-k,\beta}$, the relative sign is symmetric under the exchange of partitions $${\sf sgn}(\alpha,\beta) = {\sf sgn}(\beta,\alpha).$$
### Generalized shift symmetries
Let us explain the extension of the shift symmetries (i), (ii) and (iii) which is derived in [@Nakatsu_Takasaki_2019]. For $\alpha\in\mathcal{P}$, we introduce an operator of the form $$\begin{aligned}
\mathbb{L}_\alpha
= s_\alpha (q^{-\rho})
\Gamma_-\left(q^{-\alpha-\rho}\right)
\Gamma_+\bigl(q^{-\ltrans{\,\alpha}-\rho}\bigr).
\label{L_alpha} \end{aligned}$$ Note that (\[shift\_symmetry\_1\]) is written with the aid of $\mathbb{L}_\emptyset$ as $$V^{(k)}_m\,\mathbb{L}_\emptyset
= (-1)^k\mathbb{L}_\emptyset\,V^{(k)}_{m-k},\quad
k\geq 1,~m\in\mathbb{Z}.$$ We generalize the shift symmetry (i) by replacing $\mathbb{L}_\emptyset$ in the LHS with $\mathbb{L}_\alpha$. This yields an extension of the following form.
\[ext\_shift\_symmetry\_quote\_1\] [[@Nakatsu_Takasaki_2019]]{} For $k\in\mathbb{Z}$, $k\neq 0$ and $m\in\mathbb{Z}$, we have $$\begin{aligned}
V^{(k)}_m\,\mathbb{L}_\alpha
= (-1)^k\,\mathbb{L}_\alpha\,V^{(k)}_{m-k}
+ q^{k J_0}
\sum_{\beta\in\mathcal{R}_{k,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
q^{-m(\kappa(\alpha)-\kappa(\beta))/2k}
\,\mathbb{L}_\beta.
\label{ext_shift_symmetry_1}\end{aligned}$$ As regards the Heisenberg sub-algebra generated by $V^{(0)}_m$, we have $$\begin{aligned}
V^{(0)}_m\mathbb{L}_\alpha
= \mathbb{L}_\alpha V^{(0)}_m
+\varphi_{-m}(\alpha)\mathbb{L}_\alpha.
\label{ext_shift_symmetry_1_sub}\end{aligned}$$
(\[ext\_shift\_symmetry\_1\]) of $\alpha =\emptyset$ with $k\geq 1$ reproduces (\[shift\_symmetry\_1\]). Otherwise, there is no counterpart in (i). For instance, $\mathcal{R}_{-l,\emptyset}=\mathcal{R}^{(+)}_{l,\emptyset}$ $(l \geq 1)$ consists of hooks $(1^{l-j}j^1)$ of weight $l$. Thus, rewriting $k=-l$, (\[ext\_shift\_symmetry\_1\]) of $\alpha =\emptyset$ with $k\leq -1$ reads $$\begin{aligned}
V^{(-l)}_m\mathbb{L}_\emptyset
= (-1)^l\mathbb{L}_\emptyset V^{(-l)}_{m+l}
+ q^{-lJ_0}
\sum_{j=1}^{l}
(-1)^{l-j}q^{m(l+1-2j)/2}
\mathbb{L}_{(1^{l-j}j^1)},
\quad \mbox{$l\geq 1$.}\end{aligned}$$
The shift symmetry (ii) can be generalized in a manner similar to the case of (i) as described in Proposition \[ext\_shift\_symmetry\_quote\_1\]. With $\alpha\in\mathcal{P}$, we associate an operator of the form $$\begin{aligned}
\mathbb{L}'_\alpha = s_\alpha(q^{-\rho})
\Gamma'_-\left(q^{-\alpha-\rho}\right)
\Gamma'_+\bigl(q^{-\ltrans{\,\alpha}-\rho}\bigr).
\label{L'_alpha} \end{aligned}$$ Note that (\[shift\_symmetry\_2\]) is written with the aid of $\mathbb{L}'_\emptyset$ as $$V^{(-k)}_m\,\mathbb{L}_\emptyset
= (-1)^k\mathbb{L}_\emptyset\,V^{(-k)}_{m-k},
\quad k\geq 1,~m\in\mathbb{Z}.$$ The symmetry (ii) can be generalized by replacing $\mathbb{L}'_\emptyset$ in the LHS with $\mathbb{L}'_\alpha$. We then obtain an extension of the following form.
\[ext\_shift\_symmetry\_quote\_2\] [[@Nakatsu_Takasaki_2019]]{} For $k\in\mathbb{Z}$, $k\neq 0$ and $m\in\mathbb{Z}$, we have $$\begin{aligned}
V^{(-k)}_m\mathbb{L}'_\alpha
= \mathbb{L}'_\alpha V^{(-k)}_{m-k}
+(-1)^{m+1} q^{-kJ_0}
\sum_{\beta\in\mathcal{R}_{k,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
q^{-m\left(\kappa(\alpha)-\kappa(\beta)\right)/2k}
\,\mathbb{L}'_\beta.
\label{ext_shift_symmetry_2} \end{aligned}$$ As regards the generators of the Heisenberg sub-algebra, we have $$\begin{aligned}
V^{(0)}_m\mathbb{L}'_\alpha
=\mathbb{L}'_\alpha V^{(0)}_m
+(-1)^{m+1}
\varphi_{-m}(\alpha)\mathbb{L}'_\alpha.
\label{ext_shift_symmetry_2_sub}\end{aligned}$$
(\[ext\_shift\_symmetry\_2\]) of $\alpha =\emptyset$ with $k\geq 1$ reproduces (\[shift\_symmetry\_2\]). Otherwise, there is no counterpart in (ii).
The shift symmetry (iii) can be generalized as follows:
\[ext\_shift\_symmetry\_quote\_3\] For $k,m\in\mathbb{Z}$, we have $$\begin{aligned}
q^{\gamma K}V^{(k)}_m q^{-\gamma K}
= V^{\left(k-2m\gamma\right)}_m,
\label{ext_shift_symmetry_3}\end{aligned}$$ where $\gamma$ is chosen to be a rational number which satisfies $2m\gamma\in\mathbb{Z}$.
#### *Proof.*
The commutation relations $$\begin{aligned}
\left[K,\psi_n\right] =\bigl(n+\frac{1}{2}\bigr)^2\psi_n, \quad
\left[K,\psi^*_n\right] =-\bigl(n-\frac{1}{2}\bigr)^2\psi^*_n\end{aligned}$$ are integrated to give $$\begin{aligned}
q^{\gamma K}\psi_nq^{-\gamma K}
=q^{\gamma\left(n+1/2\right)^2}\psi_n, \quad
q^{\gamma K}\psi^*_nq^{-\gamma K}
=q^{-\gamma\left(n-1/2\right)^2}\psi_n.
\label{generalized_shift_symmetry_3_proof}\end{aligned}$$ Consider the case of $m\neq 0$. Using (\[generalized\_shift\_symmetry\_3\_proof\]), the LHS of (\[ext\_shift\_symmetry\_3\]) is computed as $$\begin{aligned}
q^{\gamma K}V^{(k)}_mq^{-\gamma K}
& =q^{-k(m+1)/2}\sum_nq^{kn}
q^{\gamma K}:\psi_{m-n}\psi^*_n:q^{-\gamma K} \\
& =q^{-k(m+1)/2}\sum_nq^{kn+\gamma\left(n-n+1/2\right)^2
-\gamma\left(n-1/2\right)^2}:\psi_{m-n}\psi^*_n: \\
& =q^{-\left(k-2m\gamma\right)\left(m+1\right)/2}
\sum_nq^{\left(k-2m\gamma\right)n}:\psi_{m-n}\psi^*_n: \\
& =V^{(k-2m\gamma)}_m.\end{aligned}$$ In the case of $m=0$, since $K$ commutes with $V^{(k)}_0$, we have $q^{\gamma K}V^{(k)}_0q^{-\gamma K}=V^{(k)}_0$. Thus, we obtain (\[ext\_shift\_symmetry\_3\]).
### Particular combinations of generalized shift symmetries
Combining (\[ext\_shift\_symmetry\_1\]) and (\[ext\_shift\_symmetry\_3\]), we obtain an operator-valued identity.
\[ext\_shift\_symmetry\_A\_tau\] For $m\in\mathbb{Z}$, $m\neq 0$, at values of $\tau$ satisfying $m\tau\in\mathbb{Z}$ and $m\tau\neq 0$, we have the operator-valued identity $$\begin{aligned}
J_{-m}\mathbb{G}_\alpha (\tau)
=\mathbb{G}_\alpha (\tau)(-1)^{m\tau}J_{-m(1+\tau)}
+q^{m\tau J_0}
\sum_{\beta\in\mathcal{R}_{m\tau,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
\mathbb{G}_\beta(\tau).
\label{ext_shift_symmetry_a_tau}\end{aligned}$$
#### *Proof.*
Using $\mathbb{L}_\alpha$, we express $\mathbb{G}_\alpha (\tau)$ in (\[G\_alpha(tau)\]) as $$\mathbb{G}_\alpha (\tau)
=q^{-\kappa(\alpha)/2\tau}q^{-\tau K/2}
\mathbb{L}_\alpha q^{\tau K/2(1+\tau)}.$$ The symmetry (\[ext\_shift\_symmetry\_3\]) with $\gamma =\tau/2$ reads $$q^{\tau K/2}J_{-m}q^{-\tau K/2} =V_{-m}^{(m\tau)},$$ which converts $J_{-m}\mathbb{G}_\alpha (\tau)$ to $$\begin{aligned}
J_{-m}\mathbb{G}_\alpha (\tau)
& =q^{-\kappa(\alpha)/2\tau}
J_{-m}q^{-\tau K/2}
\mathbb{L}_\alpha q^{\tau K/2(1+\tau)}
\nonumber \\[2mm]
& =q^{-\kappa(\alpha)/2\tau}
q^{-\tau K/2}V_{-m}^{(m\tau)}
\mathbb{L}_\alpha q^{\tau K/2(1+\tau)}.
\label{proof_symmetry_A_tau_1}\end{aligned}$$ Using (\[ext\_shift\_symmetry\_1\]), the RHS of (\[proof\_symmetry\_A\_tau\_1\]) is further converted to $$\begin{aligned}
& \mbox{RHS of (\ref{proof_symmetry_A_tau_1})}
\nonumber \\[1mm]
& =q^{-\kappa(\alpha)/2\tau}
q^{-\tau K/2}
\biggl\{
(-1)^{m\tau}\mathbb{L}_\alpha V_{-m(1+\tau)}^{(m\tau)}
+q^{m\tau J_0}
\sum_{\beta\in\mathcal{R}_{m\tau,\alpha}}
{\sf sgn}(\alpha,\beta)
q^{\frac{\kappa(\alpha)-\kappa(\beta)}{2\tau}}
\mathbb{L}_\beta
\biggr\}
q^{\tau K/2(1+\tau)}
\nonumber \\[2mm]
& =\mathbb{G}_\alpha (\tau)
(-1)^{m\tau}q^{-\tau K/2(1+\tau)}
V_{-m(1+\tau)}^{(m\tau)}
q^{\tau K/2(1+\tau)}
+q^{m\tau J_0}
\sum_{\beta\in\mathcal{R}_{m\tau,\alpha}}
{\sf sgn}(\alpha,\beta) \mathbb{G}_\beta (\tau).
\label{proof_symmetry_A_tau_2}\end{aligned}$$ Here, the symmetry (\[ext\_shift\_symmetry\_3\]) gives $$q^{-\tau K/2(1+\tau)}V_{-m(1+\tau)}^{(m\tau)}q^{\tau K/2(1+\tau)}
=V_{-m(1+\tau)}^{(0)}
=J_{-m(1+\tau)}.$$ Thus, the last line in (\[proof\_symmetry\_A\_tau\_2\]) becomes $$\mbox{RHS of (\ref{proof_symmetry_A_tau_2})}
=\mathbb{G}_\alpha (\tau)
(-1)^{m\tau}J_{-m(1+\tau)}
+q^{m\tau J_0}
\sum_{\beta \in \mathcal{R}_{m\tau,\alpha}}
{\sf sgn}(\alpha,\beta)\mathbb{G}_\beta (\tau).$$ We therefore obtain (\[ext\_shift\_symmetry\_a\_tau\]).\
Operator-valued identity analogous to (\[ext\_shift\_symmetry\_a\_tau\]) can be obtained if we focus on (\[ext\_shift\_symmetry\_2\]) instead of (\[ext\_shift\_symmetry\_1\]). With $\alpha\in\mathcal{P}$, let us introduce an operator-valued function of the form $$\begin{aligned}
\mathbb{G}'_\alpha(\tau)
= q^{-\kappa(\alpha)/2\tau}q^{\tau K/2}
\mathbb{L}'_\alpha
q^{-\tau K/2(1+\tau)},\quad
\tau\neq 0,-1.
\label{G'_alpha(tau)} \end{aligned}$$ We note that the topological vertex (\[topological\_vertex\_def\]) can be written in terms of the matrix element of (\[G’\_alpha(tau)\]) as $$\begin{aligned}
\mathcal{C}_{\overrightarrow{\mu}}(q)
=q^{\frac{1+\tau}{2}\kappa (\mu^{(1)})
-\frac{1}{2\tau}\kappa(\mu^{(2)})
+\frac{\tau}{2(1+\tau)}\kappa(\mu^{(3)})}
\langle\ltrans{\,\mu}^{(1)}|
\mathbb{G}'_{\,\ltrans{\,\mu}^{(2)}}(\tau)
|\mu^{(3)}\rangle. \end{aligned}$$
\[ext\_shift\_symmetry\_B\_tau\] For $m\in\mathbb{Z}$, $m\neq 0$, at values of $\tau$ satisfying $m\tau\in\mathbb{Z}$, $m\tau\neq 0$, we have the operator-valued identity $$\begin{aligned}
J_{-m}\mathbb{G}'_\alpha(\tau)
=\mathbb{G}'_\alpha (\tau)J_{-m(1+\tau)}
+(-1)^{m+1}q^{-m\tau J_0}
\sum_{\beta\in\mathcal{R}_{m\tau,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
\mathbb{G}'_\beta(\tau).
\label{ext_shift_symmetry_b_tau}\end{aligned}$$
#### *Proof.*
Using (\[ext\_shift\_symmetry\_2\]) instead of (\[ext\_shift\_symmetry\_1\]), the same computations as in the proof of Proposition \[ext\_shift\_symmetry\_A\_tau\] give rise to (\[ext\_shift\_symmetry\_b\_tau\]).\
Factorization formulas at $\tau =1$
-----------------------------------
We consider generating functions of the operator-valued functions (\[G\_alpha(tau)\]) and (\[G’\_alpha(tau)\]). These generating functions can be factorized at $\tau=1$. Let us introduce infinite-variate functions $$\begin{aligned}
s_\alpha[\bm{t}]
=\langle 0|
\exp\left(\sum_{k=1}^\infty t_kJ_k\right)
|\alpha\rangle, \quad
\bm{t}=(t_k)_{k=1}^\infty.\end{aligned}$$ It is well known that the functions $s_\alpha[\bm{t}]$ are converted to the Schur functions $s_\alpha(\bm{x})$ by the Miwa transformation $$\begin{aligned}
t_k = \frac{1}{k}\sum_{i=1}^\infty x_i^k,\quad k=1,2,\ldots\,, \end{aligned}$$ and the Schur functions $s_{\ltrans{\,\alpha}}(\bm{x})$ by another version of the transformation $$\begin{aligned}
t_k = -\frac{1}{k}\sum_{i=1}^\infty (-x_i)^k, \quad k=1,2,\dots\,.\end{aligned}$$
We note that the partial derivatives of $s_\alpha[\bm{t}]$ with respect to $\bm{t}$ become as follows.
\[prop\_partial\_t\_s\[t\]\] $$\begin{aligned}
\partial_{t_m}s_\alpha[\bm{t}]
=\sum_{\beta\in\mathcal{R}_{m,\alpha}^{(-)}}
{\sf sgn}(\alpha,\beta)
s_\beta[\bm{t}], \quad
m=1,2, \dots \,.
\label{partial_t_s[t]}\end{aligned}$$
#### *Proof.*
We rewrite the partial derivatives as $$\begin{aligned}
\partial_{t_m}s_\alpha[\bm{t}]
=\partial_{t_m}
\langle 0|\exp\left(\sum_{k=1}^\infty t_kJ_k\right)|\alpha\rangle
=\langle 0|\exp\left(\sum_{k=1}^\infty t_kJ_k\right)J_m|\alpha\rangle,
\label{proof_partial_t_s[t]}\end{aligned}$$ which can be computed by use of a version of Murnaghan-Nakayama’s rule: $$\begin{aligned}
J_k|\alpha\rangle
= \sum_{\beta\in\mathcal{R}_{k,\alpha}}
{\sf sgn}(\alpha,\beta)|\beta\rangle, \quad
k\in\mathbb{Z}_{\neq 0}. $$ Applying the rule into the RHS of (\[proof\_partial\_t\_s\[t\]\]), we see $$\begin{aligned}
\mbox{RHS of (\ref{proof_partial_t_s[t]})}
=\sum_{\beta\in\mathcal{R}_{m,\alpha}^{(-)}}
{\sf sgn}(\alpha,\beta)
\langle 0|
\exp\left(\sum_{k=1}^\infty t_kJ_k\right)
|\beta\rangle
=\sum_{\beta\in\mathcal{R}_{m,\alpha}^{(-)}}
{\sf sgn}(\alpha,\beta)s_\beta[\bm{t}]. \end{aligned}$$ Thus we obtain (\[partial\_t\_s\[t\]\]).
### Operator-valued generating functions
We define operator-valued generating functions by $$\begin{aligned}
\mathbb{G}[\bm{t};\tau]
& =\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha (\tau)s_\alpha[\bm{t}],
\label{G[t;tau]}
\\[1mm]
\mathbb{G}'[\bm{t};\tau]
& =\sum_{\alpha\in\mathcal{P}}
\mathbb{G}'_\alpha (\tau)s_\alpha[\bm{t}].
\label{G'[t;tau]}\end{aligned}$$
\[partial-differentials\_G\[t;tau\]\] Partial derivatives of (\[G\[t;tau\]\]) and (\[G’\[t;tau\]\]) with respect to $\bm{t}=(t_k)_{k=1}^\infty$ are expressed in the following forms. $$\begin{aligned}
\partial_{t_m}\mathbb{G}[\bm{t};\tau]
& =\sum_{\alpha\in\mathcal{P}}
\Biggl(
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn} (\alpha,\beta)
\mathbb{G}_\beta(\tau)
\Biggr)
s_\alpha[\bm{t}],
\label{partial_t_G[t;tau]}
\\[1mm]
\partial_{t_m}\mathbb{G}'[\bm{t};\tau]
& =\sum_{\alpha\in\mathcal{P}}
\Biggl(
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\alpha,\beta)
\mathbb{G}'_\beta(\tau)
\Biggr)
s_\alpha[\bm{t}],
\label{partial_t_G'[t;tau]}\end{aligned}$$ where $m=1,2,\dots$.
#### *Proof.*
By (\[partial\_t\_s\[t\]\]), the partial derivatives of $\mathbb{G}[\bm{t};\tau]$ read $$\begin{gathered}
\partial_{t_m}\mathbb{G}[\bm{t};\tau]
=\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha(\tau)\partial_{t_m}s_\alpha[\bm{t}]
\nonumber \\[1mm]
=\sum_{\alpha\in\mathcal{P}}
\sum_{\beta\in\mathcal{R}^{(-)}_{m,\alpha}}
{\sf sgn}(\alpha,\beta)
\mathbb{G}_\alpha(\tau)s_\beta[\bm{t}]
=\sum_{\alpha\in\mathcal{P}}
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\beta,\alpha)
\mathbb{G}_\beta(\tau)s_\alpha[\bm{t}],
\label{proof_partial_t_G[t]}\end{gathered}$$ where the roles of $\alpha$ and $\beta$ are exchanged in the last line. Similarly, the partial derivatives of $\mathbb{G}'[\bm{t};\tau]$ can be expressed as $$\begin{aligned}
\partial_{t_m}\mathbb{G}'[\bm{t};\tau]
=\sum_{\alpha\in\mathcal{P}}
\mathbb{G}'_\alpha (\tau)\partial_{t_m}s_\alpha[\bm{t}]
=\sum_{\alpha\in\mathcal{P}}
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\beta,\alpha)
\mathbb{G}'_\beta(\tau)s_\alpha[\bm{t}].
\label{proof_partial_t_G'[t]}\end{aligned}$$ Taking account of the symmetry ${\sf sgn}(\beta,\alpha)={\sf sgn}(\alpha,\beta)$, we find that (\[proof\_partial\_t\_G\[t\]\]) and (\[proof\_partial\_t\_G’\[t\]\]) yield (\[partial\_t\_G\[t;tau\]\]) and (\[partial\_t\_G’\[t;tau\]\]) respectively.
### Factorization formulas at $\tau =1$
The operator-valued generating functions $\mathbb{G}[\bm{t};\tau]$ and $\mathbb{G}'[\bm{t};\tau]$ at $\tau=1$ can be factorized into a triple product of operators.
[(factorization formulas at $\tau=1$)]{} \[factorization\_formulas\_tau=1\] $\mathbb{G}[\bm{t};\tau]$ and $\mathbb{G}'[\bm{t};\tau]$ at $\tau=1$ satisfy the identities $$\begin{aligned}
\mathbb{G}[\bm{t};1]
& = \exp\left(\sum_{k=1}^\infty t_kq^{kJ_0}J_k\right)
\mathbb{G}_\emptyset (1)
\exp\left(\sum_{k=1}^\infty (-1)^{k+1}t_kq^{kJ_0}J_{2k}\right),
\label{factorization_tau=1}
\\[2mm]
\mathbb{G}'[\bm{t};1]
& = \exp\left(\sum_{k=1}^\infty (-1)^{k+1}t_kq^{-kJ_0}J_k\right)
\mathbb{G}'_\emptyset (1)
\exp\left(\sum_{k=1}^\infty (-1)^kt_kq^{-kJ_0}J_{2k}\right),
\label{factorization'_tau=1}\end{aligned}$$ where $\mathbb{G}_\emptyset (1)$ and $\mathbb{G}'_\emptyset (1)$ are operators of $\mathbb{G}_\emptyset (\tau)$ and $\mathbb{G}'_\emptyset (\tau)$ specialized to $\tau=1$.
#### *Proof.*
Let us derive differential equations for $\mathbb{G}[\bm{t};1]$ and $\mathbb{G}'[\bm{t};1]$. We first consider the case of $\mathbb{G}\left[\bm{t};1\right]$. Partial derivatives of $\mathbb{G}\left[\bm{t};1\right]$ with respect to $\bm{t}$ are of the form (\[partial\_t\_G\[t;tau\]\]). Note here that (\[ext\_shift\_symmetry\_a\_tau\]) now reads $$\begin{aligned}
J_{-k}\mathbb{G}_\alpha(1)
=\mathbb{G}_\alpha(1)(-1)^kJ_{-2k}
+q^{kJ_0}\sum_{\beta\in\mathcal{R}_{k,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
\mathbb{G}_\beta(1),
\quad k\in\mathbb{Z}_{\neq 0}.
\label{ext_shift_symmetry_a1}\end{aligned}$$ We rewrite (\[ext\_shift\_symmetry\_a1\]) for the cases of $k\leq -1$. By letting $k=-m$, these become $$\begin{aligned}
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\alpha,\beta)\mathbb{G}_\beta(1)
=q^{mJ_0}J_m\mathbb{G}_\alpha(1)
+\mathbb{G}_\alpha(1)(-1)^{m+1}q^{m J_0}J_{2m},
\label{ext_shift_symmetry_a2}\end{aligned}$$ where $m=1,2,\dots$. By plugging (\[ext\_shift\_symmetry\_a2\]) into the double sum, the RHS of (\[partial\_t\_G\[t;tau\]\]) turns out to be $$\begin{aligned}
\mbox{RHS of (\ref{partial_t_G[t;tau]}) at $\tau\!=\!1$} \quad
& =\sum_{\alpha\in\mathcal{P}}
\Bigl(
q^{mJ_0}J_m\mathbb{G}_\alpha(1)
+\mathbb{G}_\alpha(1)(-1)^{m+1}q^{mJ_0}J_{2m}
\Bigr)
s_\alpha[\bm{t}]
\\[1mm]
& =q^{mJ_0}J_m
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha(1)s_\alpha\left[\bm{t}\right]
+\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha(1)s_\alpha[\bm{t}]
(-1)^{m+1}q^{mJ_0}J_{2m}
\\[1mm]
& =q^{mJ_0}J_m\mathbb{G}[\bm{t};1]
+\mathbb{G}[\bm{t};1](-1)^{m+1}q^{mJ_0}J_{2m}. \end{aligned}$$ We thus find that $\mathbb{G}\left[\bm{t};1\right]$ satisfies the first-order differential equations $$\begin{aligned}
\partial_{t_m}\mathbb{G}[\bm{t};1]
=q^{mJ_0}J_m\mathbb{G}[\bm{t};1]
+\mathbb{G}[\bm{t};1](-1)^{m+1}q^{mJ_0}J_{2m}
\label{PDE_G[t;1]}\end{aligned}$$ for $m=1,2,\dots$.
By the uniqueness of solutions of the initial value problem, (\[PDE\_G\[t;1\]\]) can be solved as $$\begin{aligned}
\mathbb{G}[\bm{t};1]
=\exp\left(\sum_{k=1}^\infty t_kq^{kJ_0}J_k\right)
\mathbb{G}[\bm{0};1]
\exp\left(\sum_{k=1}^\infty(-1)^{k+1}t_kq^{kJ_0}J_{2k}\right),
\label{GS_PDE_G[t;1]}\end{aligned}$$ where $\mathbb{G}[\bm{0};1]$ denotes an operator at the initial time. Since $\mathbb{G}[\bm{0};1]=\mathbb{G}_\emptyset(1)$, (\[GS\_PDE\_G\[t;1\]\]) yields (\[factorization\_tau=1\]).
(\[factorization’\_tau=1\]) is likewise obtained by using Proposition \[ext\_shift\_symmetry\_B\_tau\] in place of Proposition \[ext\_shift\_symmetry\_A\_tau\]. Partial derivatives of $\mathbb{G}'[\bm{t};1]$ with respect to $\bm{t}$ are of the form (\[partial\_t\_G’\[t;tau\]\]). Specialized to $\tau=1$, the double sum therein can be computed by Proposition \[ext\_shift\_symmetry\_B\_tau\]. (\[ext\_shift\_symmetry\_b\_tau\]) therein reads $$\begin{aligned}
J_{-k}\mathbb{G}'_\alpha(1)
=\mathbb{G}'_\alpha(1)J_{-2k}
+(-1)^{k+1}q^{-kJ_0}
\sum_{\beta\in\mathcal{R}_{k,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
\mathbb{G}'_\beta(1),
\quad k\in\mathbb{Z}_{\neq 0}.
\label{ext_shift_symmetry_b1}\end{aligned}$$ We rewrite (\[ext\_shift\_symmetry\_b1\]) for the cases of $k\leq -1$. By letting $k=-m$, these become $$\begin{aligned}
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\alpha,\beta)\mathbb{G}'_\beta(1)
=(-1)^{m+1}
\Bigl(
q^{-mJ_0}J_m\mathbb{G}'_\alpha(1)
-\mathbb{G}'_\alpha(1)q^{-m J_0}J_{2m}
\Bigr),
\label{ext_shift_symmetry_b2}\end{aligned}$$ where $m=1,2,\dots$. By plugging (\[ext\_shift\_symmetry\_b2\]) into the double sum, the RHS of (\[partial\_t\_G’\[t;tau\]\]) is converted to $$\begin{aligned}
\mbox{RHS of (\ref{partial_t_G'[t;tau]}) at $\tau\!=\!1$} \quad
& = (-1)^{m+1}
\sum_{\alpha\in\mathcal{P}}
\Bigl(
q^{-mJ_0}J_m\mathbb{G}'_\alpha(1)
-\mathbb{G}'_\alpha(1)q^{-mJ_0}J_{2m}
\Bigr)
\\[1mm]
& = (-1)^{m+1}q^{-mJ_0}J_m\mathbb{G}'[\bm{t};1]
+ \mathbb{G}'[\bm{t};1](-1)^mq^{-mJ_0}J_{2m}. \end{aligned}$$ We thus find that $\mathbb{G}'[\bm{t};1]$ satisfies the first-order differential equations $$\begin{aligned}
\partial_{t_m}\mathbb{G}'[\bm{t};1]
= (-1)^{m+1}q^{-mJ_0} J_m\mathbb{G}'[\bm{t};1]
+ \mathbb{G}'[\bm{t};1](-1)^mq^{-mJ_0}J_{2m}
\label{PDE_G'[t;1]}\end{aligned}$$ for $m=1,2,\dots$.
General solutions of the differential equations (\[PDE\_G’\[t;1\]\]) are of the form $$\begin{aligned}
\mathbb{G}'[\bm{t};1]
=\exp\left(\sum_{k=1}^\infty(-1)^{k+1}t_kq^{-kJ_0}J_k\right)
\mathbb{G}'[\bm{0};1]
\exp\left(\sum_{k=1}^\infty(-1)^kt_kq^{-kJ_0}J_{2k}\right),
\label{GS_PDE_G'[t;1]}\end{aligned}$$ where $\mathbb{G}'[\bm{0};1]$ denotes an operator at the initial time. In the current case, we have $\mathbb{G}'[\bm{0};1]=\mathbb{G}'_\emptyset(1)$. Therefore, (\[GS\_PDE\_G’\[t;1\]\]) yields (\[factorization’\_tau=1\]).\
Proof of Theorem \[theorem\_conjecture\]
========================================
Yet another representation at $\tau=1$
--------------------------------------
In the case of $\tau=1$, another expression of generating function of the topological vertex can be derived from Theorem \[factorization\_formulas\_tau=1\]. We conveniently start with an operator identity of the following form.
$$\begin{aligned}
\mathbb{G}_\emptyset (1)\Gamma'_-(\bm{z})
= \exp\left(
\sum_{m=1}^\infty \frac{(-1)^{m+1}p_{2m}(\bm{z})}{m}J_{-m}
\right)
\mathbb{G}_{\emptyset} (1)\Gamma_-(\bm{z}),
\label{lemma_rewriting_W(q;x,y,z;1)}\end{aligned}$$
#### *Proof.*
Let us rewrite $\Gamma'_-(\bm{z})$ as $$\begin{aligned}
\Gamma'_-(\bm{z})
& =\exp\left(
\sum_{m=1}^\infty\frac{p_{2m-1}(\bm{z})}{2m-1}J_{-2m+1}
- \sum_{m=1}^\infty\frac{p_{2m}(\bm{z})}{2m}J_{-2m}
\right)
\nonumber \\[1mm]
& =\exp\left(
- \sum_{m=1}^\infty\frac{p_{2m}(\bm{z})}{m}J_{-2m}
\right)
\Gamma_-(\bm{z}).
\label{proof_lemma_W(q;x,y,z;1)_1}\end{aligned}$$ (\[ext\_shift\_symmetry\_a1\]) implies the relation $$\begin{aligned}
J_{-m}\mathbb{G}_\emptyset (1)
= (-1)^m\mathbb{G}_\emptyset (1)J_{-2m}, \quad
m=1,2,\dots,\,.
\label{proof_lemma_W(q;x,y,z;1)_2}\end{aligned}$$ Combining (\[proof\_lemma\_W(q;x,y,z;1)\_1\]) with (\[proof\_lemma\_W(q;x,y,z;1)\_2\]), we see $$\begin{aligned}
\mathbb{G}_\emptyset(1)\Gamma'_-(\bm{z})
& = \mathbb{G}_\emptyset(1)
\exp\left(
- \sum_{m=1}^\infty\frac{p_{2m}(\bm{z})}{m}J_{-2m}
\right)
\Gamma_-(\bm{z})
\\[1.5mm]
& = \exp\left(
\sum_{m=1}^\infty\frac{(-1)^{m+1}p_{2m}(\bm{z})}{m}J_{-m}
\right)
\mathbb{G}_\emptyset(1)\Gamma_-(\bm{z}).\end{aligned}$$ Thus we obtain (\[lemma\_rewriting\_W(q;x,y,z;1)\]).
\[fermionic\_formula\_W(q;x,y,z;1)\] The generating function (\[generating\_function\_topological\_vertex\]) at $\tau =1$ has a fermionic representation of the form $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)
= \langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset (1)\Gamma_-(\bm{z})
|0\rangle
\exp\left(
\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p_m(\bm{x})p_{2m}(\bm{z})
\right).
\label{W(q;x,y,z;1)_fermion_expression}\end{aligned}$$
#### *Proof.*
We adopt the fermionic representation (\[W(q;x,y,z;tau)\_fermion\_expression\]) for a description of the generating function (\[generating\_function\_topological\_vertex\]). Letting $\tau=1$, we rewrite (\[W(q;x,y,z;tau)\_fermion\_expression\]) by using Theorem \[factorization\_formulas\_tau=1\]. Actually, substituting $t_k=(-1)^{k+1}p_k(\bm{y})/k$ in (\[factorization\_tau=1\]), we find $$\begin{aligned}
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha (1)s_{\ltrans{\,\alpha}}(\bm{y})
=\exp\left(
\sum_{k=1}^\infty
\frac{(-1)^{k+1}p_k(\bm{y})}{k}q^{kJ_0}J_k
\right)
\mathbb{G}_\emptyset (1)
\exp\left(
\sum_{k=1}^\infty
\frac{p_k(\bm{y})}{k}q^{kJ_0}J_{2k}
\right). $$ The sum of operators in (\[W(q;x,y,z;tau)\_fermion\_expression\]) can be thus converted into a triple product of operators as $$\begin{aligned}
& \mathcal{W}(q;\bm{x},\bm{y},\bm{z};1)
\nonumber \\
& =\langle 0|
\Gamma_+(\bm{x})
\exp\left(
\sum_{k=1}^\infty\frac{(-1)^{k+1}p_k(\bm{y})}{k}J_k
\right)
\mathbb{G}_\emptyset (1)
\exp\left(
\sum_{k=1}^\infty\frac{p_k(\bm{y})}{k}J_{2k}
\right)
\Gamma'_- (\bm{z})
|0\rangle
\nonumber \\
& =\langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})\mathbb{G}_\emptyset (1)
\exp\left(
\sum_{k=1}^\infty\frac{p_k(\bm{y})}{k}J_{2k}
\right)
\Gamma'_-(\bm{z})
|0\rangle.
\label{rewriting_W(q;x,y,z;1)_1}\end{aligned}$$
We further rewrite the RHS of (\[rewriting\_W(q;x,y,z;1)\_1\]). The commutation relations $$\begin{aligned}
\Gamma'_-(\bm{z})J_{2k}
=J_{2k}\Gamma'_-(\bm{z})
+\Gamma'_-(\bm{z})p_{2k}(\bm{z}), \end{aligned}$$ and the annihilation property $J_{2k}|0\rangle =0$ for $k\geq 1$ yield $$\begin{aligned}
\exp\left(
\sum_{k=1}^\infty\frac{p_k(\bm{y})}{k}J_{2k}
\right)
\Gamma'_-(\bm{z})|0\rangle
=\Gamma'_-(\bm{z})|0\rangle
\exp\left(
-\sum_{k=1}^\infty\frac{p_k(\bm{y})p_{2k}(\bm{z})}{k}
\right).
\label{rewriting_W(q;x,y,z;1)_2}\end{aligned}$$ By (\[rewriting\_W(q;x,y,z;1)\_2\]), the RHS of (\[rewriting\_W(q;x,y,z;1)\_1\]) can be converted to $$\begin{aligned}
\mathcal{W}(q;\bm{x},\bm{y},\bm{z};1)
=\langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset(1)\Gamma'_-(\bm{z})
|0\rangle
\exp\left(
-\sum_{k=1}^\infty\frac{p_k(\bm{y})p_{2k}(\bm{z})}{k}
\right).
\label{rewriting_W(q;x,y,z;1)_3}\end{aligned}$$ The bra vector in (\[rewriting\_W(q;x,y,z;1)\_3\]) can be expressed with the aid of (\[lemma\_rewriting\_W(q;x,y,z;1)\]) as follows: $$\begin{aligned}
\begin{aligned}
& \langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset(1)\Gamma'_-(\bm{z})
\\[1mm]
& =\langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\exp\left(
\sum_{m=1}^\infty\frac{(-1)^{m+1}p_{2m}(\bm{z})}{m}J_{-m}
\right)
\mathbb{G}_{\emptyset} (1)\Gamma_-(\bm{z})
\\[1mm]
& =\langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_{\emptyset} (1)\Gamma_-(\bm{z})
\exp\biggl\{
\sum_{m=1}^\infty
\frac{1}{m}\Bigl(
p_m(\bm{y})p_{2m}(\bm{z})
+(-1)^{m+1}p_m(\bm{x})p_{2m}(\bm{z})
\Bigr)
\biggr\},
\end{aligned}
\label{eq_rewriting_W(q;x,y,z;1)_3}\end{aligned}$$ where the negative modes $J_{-m}$ have been moved to the left by the commutation relations $$\begin{aligned}
\Gamma_+(\bm{x})J_{-m}
& =J_{-m}\Gamma_+(\bm{x})
+\Gamma_+(\bm{x})p_{m}(\bm{x}),
\\[2mm]
\Gamma'_+(\bm{y})J_{-m}
& =J_{-m}\Gamma'_+(\bm{y})
+\Gamma'_+(\bm{y})(-1)^{m+1}p_{m}(\bm{y}), \end{aligned}$$ together with $\langle 0|J_{-m}=0$ at the end. Designating the last part of (\[eq\_rewriting\_W(q;x,y,z;1)\_3\]) as the bra vector and its pairing with the ket vector $|0\rangle$ being substituted for the matrix element in (\[rewriting\_W(q;x,y,z;1)\_3\]), we eventually obtain (\[W(q;x,y,z;1)\_fermion\_expression\]).\
The matrix element in the RHS of (\[W(q;x,y,z;1)\_fermion\_expression\]) has an expansion in terms of the Schur functions and the skew Schur functions.
\[proposition\_Schur\_expansion\_fermion\_rep\] The matrix element in the fermionic expression (\[W(q;x,y,z;1)\_fermion\_expression\]) has a Schur function expansion of the form $$\begin{aligned}
\begin{aligned}
& \langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset (1)\Gamma_-(\bm{z})
|0\rangle
\\[1mm]
& =\sum_{\nu^1,\nu^3,\nu^+\in\mathcal{P}}
s_{\nu^1}(\bm{x})
s_{\nu^+/\ltrans{\,\nu}^1}(\bm{y})
s_{\,\ltrans{\,\nu}^3}(\bm{z})
q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho}).
\end{aligned}
\label{Schur_expansion_fermion_rep}\end{aligned}$$
#### *Proof.*
By inserting the identity $\sum_{\nu\in\mathcal{P}}|\nu\rangle\langle\nu|=1$ between all pairs of adjacent operators in $\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset (1)\Gamma_-(\bm{z})$, the LHS of (\[Schur\_expansion\_fermion\_rep\]) reads $$\begin{aligned}
& \langle 0|
\Gamma_+(\bm{x})\Gamma'_+(\bm{y})
\mathbb{G}_0\Gamma_-(\bm{z})
|0\rangle
\nonumber \\[1mm]
& =\langle 0|
\Gamma_+(\bm{x})
\left(
\sum_{\nu^1\in\mathcal{P}}
|\nu^1\rangle\langle\nu^1|
\right)
\Gamma'_+(\bm{y})
\left(
\sum_{\nu^+\in\mathcal{P}}
|\ltrans{\,\nu}^+\rangle\langle\ltrans{\,\nu}^+|
\right)
\mathbb{G}_\emptyset (1)
\left(
\sum_{\nu^3\in\mathcal{P}}
|\ltrans{\,\nu}^3\rangle\langle\ltrans{\,\nu}^3|
\right)
\Gamma_-(\bm{z})
|0\rangle
\nonumber \\[1mm]
& =\sum_{\nu^1,\nu^+,\nu^3\in\mathcal{P}}
s_{\nu^1}(\bm{x})
s_{\nu^+/\ltrans{\,\nu}^1}(\bm{y})
s_{\ltrans{\,\nu}^3}(\bm{z})
\langle\ltrans{\,\nu}^+|
\mathbb{G}_\emptyset(1)
|\ltrans{\,\nu}^3\rangle.
\label{Schur_expansion_fermion_rep_proof1}\end{aligned}$$ Note that (\[matrix\_element\_Gamma\]) and (\[matrix\_element\_Gamma’\]) have been used in the last line.
The matrix element of $\mathbb{G}_\emptyset(1) =
q^{-K/2}\Gamma_-(q^\rho)\Gamma_+(q^\rho)q^{K/4}$ can be expressed in terms of special values of the Schur function as $$\begin{aligned}
\langle\ltrans{\,\nu}^+|
\mathbb{G}_\emptyset(1)
|\ltrans{\,\nu}^3\rangle
=q^{\kappa(\nu^+)+\kappa(\nu^3 )/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho}).
\label{C_two_leggs}\end{aligned}$$ Pluging (\[C\_two\_leggs\]) into the RHS of (\[Schur\_expansion\_fermion\_rep\_proof1\]), we obtain (\[Schur\_expansion\_fermion\_rep\]).
Let us derive (\[C\_two\_leggs\]). It is convenient to use the formula [@Takasaki_Nakatsu_JPA_2016] $$\begin{aligned}
q^{K/2}\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
|\ltrans{\,\alpha}\rangle
= \Gamma'_- (q^{-\alpha-\rho})
|0\rangle s_{\,\ltrans{\,\alpha}}(q^{-\rho}), \quad
\forall\,\alpha\in\mathcal{P}.
\label{exchange_formula}\end{aligned}$$ By taking the transpose and using (\[exchange\_formula\]), we can rewrite the LHS of (\[C\_two\_leggs\]) as $$\begin{aligned}
\langle\ltrans{\,\nu}^+|\mathbb{G}_\emptyset(1)|\ltrans{\,\nu}^3\rangle
& =q^{\kappa(\nu^+)/2}
\langle\ltrans{\,\nu}^+|
\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})q^{K/2}
|\ltrans{\,\nu}^3\rangle q^{\kappa(\nu^3)/4}
\nonumber \\[2mm]
& =q^{\kappa(\nu^+)/2}
\left(
s_{\ltrans{\,\nu}^+}(q^{-\rho})
\langle 0|\Gamma'_+(q^{-\nu^+-\rho})
\right)
|\ltrans{\,\nu}^3\rangle
q^{\kappa(\nu^3)/4}
\nonumber \\[2mm]
& =q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho}). \end{aligned}$$ Note that the identity $s_{\ltrans{\,\nu}^+}(q^{-\rho})=q^{\kappa(\nu^+)/2}s_{\nu^+}(q^{-\rho})$ is used in the last line. Thus, we obtain (\[C\_two\_leggs\]).\
$\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})$ is a self-adjoint operator, that is, satisfies $$\begin{aligned}
\langle\ltrans{\,\beta}|
\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
|\ltrans{\,\alpha}\rangle
=\langle\ltrans{\,\alpha}|
\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
|\ltrans{\,\beta}\rangle,
\quad\forall\,\alpha,\beta\in\mathcal{P}.
\label{two-legged_identity_matrix}\end{aligned}$$ The matrix elements in (\[two-legged\_identity\_matrix\]) can be expressed in terms of special values of the Schur function by (\[exchange\_formula\]). Actually, the LHS of (\[two-legged\_identity\_matrix\]) becomes $$\begin{aligned}
& \langle\ltrans{\,\beta}|
\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
|\ltrans{\,\alpha}\rangle
=q^{\kappa(\beta)/2}
\langle\ltrans{\,\beta}|
q^{K/2}\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
|\ltrans{\,\alpha}\rangle
\\[2mm]
& \qquad \qquad
= q^{\kappa(\beta)/2}
\langle\ltrans{\,\beta}|
q^{K/2} \Gamma'_-(q^{-\alpha-\rho})
|0\rangle
s_{\,\ltrans{\,\alpha}}(q^{-\rho})
= q^{\frac{\kappa(\alpha)+\kappa(\beta)}{2}}
s_\alpha(q^{-\rho})s_\beta(q^{-\alpha-\rho}), \end{aligned}$$ whereas the RHS of (\[two-legged\_identity\_matrix\]) reads $$\begin{aligned}
\langle\ltrans{\,\alpha}|
\Gamma_-(q^{-\rho})\Gamma_+(q^{-\rho})
|\ltrans{\,\beta}\rangle
=q^{\frac{\kappa(\alpha)+\kappa(\beta)}{2}}
s_\alpha(q^{-\beta-\rho})s_\beta(q^{-\rho}). \end{aligned}$$ Thus, (\[two-legged\_identity\_matrix\]) implies the identity $$\begin{aligned}
s_\alpha(q^{-\rho})s_\beta(q^{-\alpha-\rho})
=s_\alpha(q^{-\beta-\rho}) s_\beta (q^{-\rho}),
\quad \forall\,\alpha,\beta\in\mathcal{P}.
\label{two-legged_identity_Schur}\end{aligned}$$ The identity (\[two-legged\_identity\_Schur\]) is exploited in [@Takasaki_Nakatsu_JPA_2016] to simplify open topological string amplitudes including the case of open topological string on closed vertex.\
From (\[fermion\_expression\_topological\_vertex\]), the matrix element of $\mathbb{G}_\emptyset(1)$ can be written in terms of the two-legged topological vetrex as $$\begin{aligned}
\langle\ltrans{\,\nu}^+|
\mathbb{G}_\emptyset(1)
|\ltrans{\,\nu}^3\rangle
=q^{\kappa(\nu^+)-\kappa(\nu^3)/4}
\mathcal{C}_{\left(\ltrans{\,\nu}^+,\emptyset,\,\nu^3\right)}(q). \end{aligned}$$ On the other hand, (\[topological\_vertex\_def\]) and (\[C\_two\_leggs\]) yield another expression $$\begin{aligned}
\langle\ltrans{\,\nu}^+|
\mathbb{G}_\emptyset(1)
|\ltrans{\,\nu}^3\rangle
= q^{\kappa(\nu^+)-\kappa(\nu^3)/4}
\mathcal{C}_{\left(\nu^3,\ltrans{\,\nu}^+,\emptyset\right)}(q). \end{aligned}$$ Furthermore, by (\[topological\_vertex\_def\]) and (\[two-legged\_identity\_Schur\]), we find $$\begin{aligned}
& \mathcal{C}_{\left(\nu^3,\ltrans{\,\nu}^+,\emptyset\right)}(q)
=q^{\kappa(\nu^3)/2} s_{\nu^3}(q^{-\nu^+-\rho}) s_{\nu^+} (q^{-\rho})
\nonumber \\[1.5mm]
& \qquad \qquad
=q^{\kappa(\nu^3)/2} s_{\nu^3}(q^{-\rho}) s_{\nu^+} (q^{-\nu^3-\rho})
=\mathcal{C}_{\left(\emptyset, \nu^3,\ltrans{\,\nu}^+\right)}(q).\end{aligned}$$ These identities imply the cyclic symmetry of the two-legged topological vertex: $$\begin{aligned}
\mathcal{C}_{\left(\ltrans{\,\nu}^+,\emptyset,\,\nu^3\right)}(q)
=\mathcal{C}_{\left(\nu^3,\ltrans{\,\nu}^+,\emptyset\right)}(q)
=\mathcal{C}_{\left(\emptyset,\nu^3,\ltrans{\,\nu}^+\right)}(q). \end{aligned}$$
Proof of Theorem \[theorem\_conjecture\]
----------------------------------------
We note that the exponentiation of quadratic form of the power sums in (\[W(q;x,y,z;1)\_fermion\_expression\]) has a Schur function expansion of the following form:
$$\begin{aligned}
\exp\left(
\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p_m(\bm{x})p_{2m}(\bm{z})
\right)
=\sum_{\eta^1,\eta^3\in\mathcal{P}}
s_{\,\ltrans{\,\eta}^1}(\bm{x})s_{\eta^3}(\bm{z})
\sum_{\xi\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi)\chi_{\eta^3}(2\xi)}{z_\xi}.
\label{eq3_proof_conjecture}\end{aligned}$$
#### *Proof.*
We rewrite the LHS of (\[eq3\_proof\_conjecture\]) into $$\begin{aligned}
& \exp\left(
\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p_m(\bm{x})p_{2m}(\bm{z})
\right)
\nonumber \\[1.5mm]
& \quad
=\prod_{k=1}^\infty
\exp\left(
\frac{(-1)^{k-1}}{k}p_k(\bm{x})p_{2k}(\bm{z})
\right)
=\prod_{k=1}^\infty
\sum_{m_k=0}^\infty
\frac{(-1)^{(k-1)m_k}}{m_k! k^{m_k}}
\left(p_k(\bm{x})p_{2k}(\bm{z})\right)^{m_k}
\nonumber \\[1.5mm]
& \quad
=\sum_{m_1,m_2,\cdots}
\frac{(-1)^{\sum_{k=1}^\infty (k-1)m_k}}{\prod_{k=1}^\infty m_k!k^{m_k}}
\prod_{k=1}^\infty\bigl(p_k(\bm{x})p_{2k}(\bm{z})\bigr)^{m_k}
=\sum_{\xi\in\mathcal{P}}
\frac{(-1)^{|\xi|-l(\xi)}}{z_\xi}p_\xi(\bm{x})p_{2\xi}(\bm{z}),
\label{eq4_proof_conjecture}\end{aligned}$$ where the summation over non-negative integers $m_k$’s is replaced with the summation over partitions $\xi$ by arranging $\xi=(1^{m_1}2^{m_2}\cdots)$. Plugging the expansions $$\begin{aligned}
(-1)^{|\xi|-l(\xi)}p_\xi(\bm{x})
=\sum_{\eta^1\in\mathcal{P}}
\chi_{\eta^1}(\xi)s_{\,\ltrans{\,\eta}^1}(\bm{x}), \quad
p_{2\xi}(\bm{z})
=\sum_{\eta^3\in\mathcal{P}}
\chi_{\eta^3}(2\xi)s_{\eta^3}(\bm{z})\end{aligned}$$ into the RHS of (\[eq4\_proof\_conjecture\]), we obtain (\[eq3\_proof\_conjecture\]).
We can now rewrite $\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)$ as follows.
\[prop:proof\_conjecture\] $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)
& =\sum_{\overrightarrow{\mu}\in\mathcal{P}^3}
\mathcal{\widetilde{C}}_{\overrightarrow{\mu}}(\lambda)
q^{-\kappa(\mu^{(1)})/2+\kappa(\mu^{(2)})+\kappa(\mu^{(3)})/4}
s_{\mu^{(1)}}(\bm{x})s_{\mu^{(2)}}(\bm{y})s_{\mu^{(3)}}(\bm{z}),
\label{eq1_proof_conjecture} \end{aligned}$$ where $\mathcal{\widetilde{C}}_{\overrightarrow{\mu}}(\lambda)$ is given by (\[LLLZ\_formula\]) with $q=e^{-\sqrt{-1}\lambda}$.\
Theorem \[theorem\_conjecture\] is an immediate consequence of Proposition \[prop:proof\_conjecture\]. By (\[generating\_function\_topological\_vertex\]), $\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)$ is a generating function of $\mathcal{C}_{\overrightarrow{\mu}}(q)$ of the form $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)
=\sum_{\overrightarrow{\mu}\in\mathcal{P}^3}
\mathcal{C}_{\overrightarrow{\mu}}(q)
q^{-\kappa(\mu^{(1)})+\kappa(\mu^{(2)})/2-\kappa(\mu^{(3)})/4}
s_{\mu^{(1)}}(\bm{x})s_{\mu^{(2)}}(\bm{y})s_{\mu^{(3)}}(\bm{z}).\end{aligned}$$ Equating the coefficients of the Schur function products in this expression with those of (\[eq1\_proof\_conjecture\]), we obtain (\[tildeC=C\]). This also implies that (\[eq1\_theorem\_conjecture\]) holds.
#### *Proof of Proposition \[prop:proof\_conjecture\].*
We derive an expansion of the fermionic representation (\[W(q;x,y,z;1)\_fermion\_expression\]) in the basis of the Schur functions. By (\[Schur\_expansion\_fermion\_rep\]) and (\[eq3\_proof\_conjecture\]), $\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)$ can be expressed as $$\begin{aligned}
& \mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)
\nonumber \\[1mm]
& = \sum_{\nu^1,\nu^3,\nu^+\in\mathcal{P}}
s_{\nu^1}(\bm{x})
s_{\nu^+/\ltrans{\,\nu}^1}(\bm{y})
s_{\,\ltrans{\,\nu}^3}(\bm{z})
q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho})
\nonumber \\
& \qquad\qquad \qquad \times
\sum_{\eta^1,\eta^3\in\mathcal{P}}
s_{\,\ltrans{\,\eta}^1}(\bm{x})
s_{\eta^3}(\bm{z})
\sum_{\xi\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi)\chi_{\eta^3}(2\xi)}{z_\xi}
\nonumber \\[1.5mm]
& = \sum_{\nu^1,\nu^3,\nu^+,\eta^1,\eta^3\in\mathcal{P}}
s_{\,\ltrans{\,\eta}^1}(\bm{x})
s_{\nu^1}(\bm{x})
s_{\nu^+/\ltrans{\,\nu}^1}(\bm{y})
s_{\eta^3}(\bm{z})
s_{\,\ltrans{\,\nu}^3}(\bm{z})
\nonumber \\
& \qquad\qquad\qquad \times
q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho})
\sum_{\xi\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi)\chi_{\eta^3}(2\xi)}{z_\xi}.
\label{eq5_proof_conjecture}\end{aligned}$$ The products of the Schur funcions and the skew Schur function can be expressed in terms of the Littlewood-Richardson numbers as $$\begin{aligned}
s_{\,\ltrans{\,\eta}^1}(\bm{x}) s_{\nu^1}(\bm{x})
& =\sum_{\mu^{(1)}\in\mathcal{P}}
c_{\,\ltrans{\,\eta}^1\nu^1}^{~\mu^{(1)}}
s_{\mu^{(1)}}(\bm{x}),
\\[0.5mm]
s_{\nu^+/\ltrans{\,\nu}^1}(\bm{y})
& =\sum_{\mu^{(2)}\in\mathcal{P}}
c_{\,\ltrans{\,\nu}^1\mu^{(2)}}^{~\nu^+}
s_{\mu^{(2)}}(\bm{y}),
\\[0.5mm]
s_{\eta^3}(\bm{z})s_{\,\ltrans{\,\nu}^3}(\bm{z})
& =\sum_{\mu^{(3)}\in\mathcal{P}}
c_{\eta^3\,\ltrans{\,\nu}^3}^{~\mu^{(3)}}
s_{\mu^{(3)}}(\bm{z}).\end{aligned}$$ Consequently, $$\begin{aligned}
& \mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};1\right)
\\[0.5mm]
& =\sum_{\nu^1,\nu^3,\nu^+,\eta^1,\eta^3\in\mathcal{P}}
\Biggl(
\sum_{\mu^{(1)}\in\mathcal{P}}
c_{\,\ltrans{\,\eta}^1\nu^1}^{~\mu^{(1)}}
s_{\mu^{(1)}}(\bm{x})
\Biggr)
\Biggl(
\sum_{\mu^{(2)}\in\mathcal{P}}
c_{\,\ltrans{\,\nu}^1\mu^{(2)}}^{~\nu^+}
s_{\mu^{(2)}}(\bm{y})
\Biggr)
\Biggl(
\sum_{\mu^{(3)}\in\mathcal{P}}
c_{\eta^3\,\ltrans{\,\nu}^3}^{~\mu^{(3)}}
s_{\mu^{(3)}}(\bm{z})
\Biggr) \\
& \qquad\qquad\qquad \times
q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho})
\sum_{\xi\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi)\chi_{\eta^3}(2\xi)}{z_\xi} \\
& =\sum_{(\mu^{(1)},\mu^{(2)},\mu^{(3)})\,\in\mathcal{P}^3}
s_{\mu^{(1)}}(\bm{x})
s_{\mu^{(2)}}(\bm{y})
s_{\mu^{(3)}}(\bm{z}) \\
& \times
\sum_{\nu^1,\nu^3,\nu^+,\eta^1,\eta^3\in\mathcal{P}}
c_{\,\ltrans{\,\eta}^1\nu^1}^{~\mu^{(1)}}
c_{\,\ltrans{\,\nu}^1\mu^{(2)}}^{~\nu^+}
c_{\eta^3\,\ltrans{\,\nu}^3}^{~\mu^{(3)}}
q^{\kappa(\nu^+)+\kappa(\nu^3)/4}
s_{\nu^+}(q^{-\rho})s_{\nu^3}(q^{-\nu^+-\rho})
\sum_{\xi\in\mathcal{P}}
\frac{\chi_{\eta^1}(\xi)\chi_{\eta^3}(2\xi)}{z_\xi}. \end{aligned}$$ Thus we obtain (\[eq1\_proof\_conjecture\]).\
Proof of Theorem \[theorem\_NKdV\]
==================================
Factorization formulas at $\tau=1/N$
------------------------------------
Just like $\mathbb{G}[\bm{t};1]$ and $\mathbb{G}'[\bm{t};1]$, the operator-valued generating functions (\[G\[t;tau\]\]) and (\[G’\[t;tau\]\]) specialized to $\tau=1/N$ ($N=1,2,\dots$) can be factorized to a triple product of operators.
[(factorization formulas at $\tau=1/N$)]{} \[factorization\_formulas\_1/N\] Let $N$ be a positive integer. The operator-valued generating functions $\mathbb{G}[\bm{t};\tau]$ and $\mathbb{G}'[\bm{t};\tau]$ satisfy the following identities at $\tau=1/N$: $$\begin{gathered}
\mathbb{G}\bigl[\bm{t};1/N\bigr]
=\exp\left(
\sum_{k=1}^\infty t_kq^{kJ_0}J_{kN}
\right)
\mathbb{G}_\emptyset \bigl(1/N\bigr)
\exp\left(
\sum_{k=1}^\infty
(-1)^{k+1}t_kq^{kJ_0}J_{k(N+1)}
\right),
\label{factorization_1/N}
\\[3mm]
\mathbb{G}'\bigl[\bm{t};1/N\bigr]
=\exp\left(
\sum_{k=1}^\infty
(-1)^{kN+1}t_kq^{-kJ_0}J_{kN}
\right)
\mathbb{G}'_\emptyset \bigl(1/N\bigr)
\exp\left(
\sum_{k=1}^\infty
(-1)^{kN}t_kq^{-kJ_0}J_{k(N+1)}
\right).
\label{factorization'_1/N}\end{gathered}$$
#### *Proof.*
The proof is mostly parallel to the proof of Theorem \[factorization\_formulas\_tau=1\]. We derive differential equations that the operator-valued generating functions satisfy at $\tau=1/N$. We describe the case of $\mathbb{G}\left[\bm{t};1/N\right]$ in detail. Partial derivatives of $\mathbb{G}\left[\bm{t};1/N\right]$ with respect to $\bm{t}$ can be expressed in the form of (\[partial\_t\_G\[t;tau\]\]). The double sum in (\[partial\_t\_G\[t;tau\]\]) can be computed by Proposition \[ext\_shift\_symmetry\_A\_tau\]. By specializing $\tau$ to $\tau=1/N$, (\[ext\_shift\_symmetry\_a\_tau\]) reads $$\begin{aligned}
J_{-kN}\mathbb{G}_\alpha(1/N)
= \mathbb{G}_\alpha(1/N)(-1)^kJ_{-k(N+1)}
+q^{kJ_0}
\sum_{\beta\in\mathcal{R}_{k,\alpha}}
\mbox{\sf sgn}(\alpha,\beta)
\mathbb{G}_\beta(1/N)
\label{ext_shift_symmetry_c1}\end{aligned}$$ for $k\in\mathbb{Z}, k \neq 0$. By letting $k=-m$, we have $$\begin{aligned}
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\alpha,\beta)
\mathbb{G}_\beta(1/N)
=q^{mJ_0}J_{mN}
\mathbb{G}_\alpha(1/N)
+\mathbb{G}_\alpha(1/N)(-1)^{m+1}q^{m J_0}J_{m(N+1)}
\label{ext_shift_symmetry_c2}\end{aligned}$$ for $m=1,2,\dots$. The RHS of (\[partial\_t\_G\[t;tau\]\]) thereby becomes $$\begin{aligned}
& \mbox{RHS of (\ref{partial_t_G[t;tau]}) at $\displaystyle \tau\!=\!1/N$}
\\[1mm]
& \qquad
=\sum_{\alpha\in\mathcal{P}}
\Bigl(
q^{mJ_0}J_{mN}\mathbb{G}_\alpha (1/N)
+\mathbb{G}_\alpha(1/N)(-1)^{m+1}q^{mJ_0}J_{m(N+1)}
\Bigr)
s_\alpha[\bm{t}]
\\[1mm]
& \qquad
=q^{mJ_0}J_{mN}
\mathbb{G}[\bm{t};1/N]
+\mathbb{G}[\bm{t};1/N](-1)^{m+1}q^{mJ_0}J_{m(N+1)}. \end{aligned}$$ We thus find that $\mathbb{G}\left[\bm{t};1/N\right]$ satisfies the first-order differential equations $$\begin{aligned}
\partial_{t_m}\mathbb{G}[\bm{t};1/N]
=q^{mJ_0}J_{mN}
\mathbb{G}[\bm{t};1/N]
+\mathbb{G}[\bm{t};1/N](-1)^{m+1}q^{mJ_0}J_{m(N+1)}
\label{PDE_G[t;1/N]}\end{aligned}$$ for $m=1,2,\dots$. General solutions of the differential equations (\[PDE\_G\[t;1/N\]\]) are given by $$\begin{aligned}
\mathbb{G}[\bm{t};1/N]
= \exp\left(
\sum_{k=1}^\infty t_kq^{kJ_0}J_{kN}
\right)
\mathbb{G}[\bm{0};1/N]
\exp\left(
\sum_{k=1}^\infty (-1)^{k+1}t_kq^{kJ_0}J_{k(N+1)}
\right).
\label{GS_PDE_G[t;1/N]}\end{aligned}$$ Since $\mathbb{G}[\bm{0};1/N]=\mathbb{G}_\emptyset(1/N)$, (\[GS\_PDE\_G\[t;1/N\]\]) yields (\[factorization\_1/N\]).
(\[factorization’\_1/N\]) can be likewise derived from Proposition \[ext\_shift\_symmetry\_B\_tau\]. Partial derivatives of $\mathbb{G}'[\bm{t};1]$ with respect to $\bm{t}$ take the form of (\[partial\_t\_G’\[t;tau\]\]). The double sum therein can be computed using Proposition \[ext\_shift\_symmetry\_B\_tau\]. By letting $\tau=1/N$, (\[ext\_shift\_symmetry\_b\_tau\]) yields $$\begin{aligned}
\sum_{\beta\in\mathcal{R}^{(+)}_{m,\alpha}}
{\sf sgn}(\alpha,\beta)\mathbb{G}'_\beta(1/N)
= (-1)^{mN}\Bigl(
-q^{-mJ_0}J_{mN}\mathbb{G}'_\alpha(1/N)
+\mathbb{G}'_\alpha(1/N)q^{-mJ_0}J_{m(N+1)}
\Bigr)
\label{ext_shift_symmetry_d2}\end{aligned}$$ for $m=1,2,\dots$. By plugging (\[ext\_shift\_symmetry\_d2\]) into the RHS of (\[partial\_t\_G’\[t;tau\]\]), we eventually find out that $\mathbb{G}'[\bm{t};1/N]$ satisfies the first-order differential equations $$\begin{aligned}
& \partial_{t_m}\mathbb{G}'[\bm{t};1/N]
\nonumber \\[1mm]
& \quad =(-1)^{mN+1}
q^{-mJ_0}J_{mN}
\mathbb{G}'[\bm{t};1/N]
+\mathbb{G}'[\bm{t};1/N]
(-1)^{mN}q^{-mJ_0}J_{m(N+1)}
\label{PDE_G'[t;1/N]}\end{aligned}$$ for $m=1,2,\dots$. General solutions of (\[PDE\_G’\[t;1/N\]\]) are given by $$\begin{aligned}
& \mathbb{G}'[\bm{t};1/N]
\nonumber \\
& \quad =\exp\left(
\sum_{k=1}^\infty(-1)^{kN+1}t_kq^{-kJ_0}J_{kN}
\right)
\mathbb{G}'[\bm{0};1/N]
\exp\left(
\sum_{k=1}^\infty(-1)^{kN}t_kq^{-kJ_0}J_{k(N+1)}
\right).
\label{GS_PDE_G'[t;1/N]}\end{aligned}$$ Since $\mathbb{G}'[\bm{0};1/N]=\mathbb{G}'_\emptyset(1/N)$, (\[GS\_PDE\_G’\[t;1/N\]\]) yields the formula (\[factorization’\_1/N\]).
Representation of generating function at $\tau=N$
-------------------------------------------------
The foregoing factorization formula gives a new fermionic representation of the generating function of the topological vertex.
\[fermionic\_formula\_W(q;x,y,z;N)\] Let $N$ be a positive integer. The generating function (\[generating\_function\_topological\_vertex\]) at $\tau=N$ has a fermionic representation of the form $$\begin{aligned}
\mathcal{W}\left(q;\bm{x},\bm{y},\bm{z};N\right)
& =\langle 0|
\exp\left(
\sum_{k=1}^\infty \frac{p_k(\bm{x})}{k}J_{kN}
\right)
\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset \bigl(1/N\bigr)\Gamma_-(\bm{z})
|0\rangle
\nonumber \\
& \quad \times
\exp\left(
\sum_{m=1}^\infty
\frac{(-1)^{m+1}}{m}
p_m(\bm{x})p_{m(N+1)}(\bm{z})\right).
\label{W(q;x,y,z;N)_fermion_expression}\end{aligned}$$
#### *Proof.*
The cyclic symmetry (\[cyclicity\_topological\_vertex\]) can be embodied in the language of the generating functions (\[generating\_function\_topological\_vertex\]) as $$\begin{aligned}
\mathcal{W}\bigl(q;\bm{x},\bm{y},\bm{z};\tau\bigr)
=\mathcal{W}\Bigl(q;\bm{z},\bm{x},\bm{y};\frac{-1}{1+\tau}\Bigr). \end{aligned}$$ In view of (\[W(q;x,y,z;tau)\_fermion\_expression\]), this relation leads to the expression $$\begin{aligned}
\mathcal{W}(q;\bm{x},\bm{y},\bm{z};\tau)
=\langle 0|\Gamma_+(\bm{z})
\left(
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha\Bigl(\frac{-1}{1+\tau}\Bigr)
s_{\,\ltrans{\,\alpha}}(\bm{x})
\right)
\Gamma'_-(\bm{y})|0\rangle.
\label{W(q;x,y,z;N)_fermion_representation_1}\end{aligned}$$ We can replace the VEV $\langle 0|\,\dots\,|0\rangle$ in the RHS of (\[W(q;x,y,z;N)\_fermion\_representation\_1\]) with the transpose. Actually, since the operator-valued functions $\mathbb{G}_\alpha (-1/1+\tau)$ satisfy $$\begin{aligned}
\langle\beta|
\mathbb{G}_\alpha\Bigl(\frac{-1}{1+\tau}\Bigr)
|\gamma\rangle
=\langle\gamma|
\mathbb{G}_{\,\ltrans{\,\alpha}}\bigl(1/\tau\bigr)
|\beta\rangle,
\quad \forall \,\beta,\gamma\in\mathcal{P},
$$ we can rewrite (\[W(q;x,y,z;N)\_fermion\_representation\_1\]) as $$\begin{aligned}
\mathcal{W}(q;\bm{x},\bm{y},\bm{z};\tau)
=\langle 0|\Gamma'_+(\bm{y})
\left(
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha\bigl(1/\tau\bigr)
s_{\alpha}(\bm{x})
\right)
\Gamma_-(\bm{z})|0\rangle.
$$ In particular, by specialization to $\tau=N$, we find $$\begin{aligned}
\mathcal{W}(q;\bm{x},\bm{y},\bm{z};N)
=\langle 0|\Gamma'_+(\bm{y})
\left(
\sum_{\alpha\in\mathcal{P}}
\mathbb{G}_\alpha \bigl(1/N\bigr)
s_{\alpha}(\bm{x})
\right)
\Gamma_-(\bm{z})|0\rangle.
\label{W(q;x,y,z;N)_fermion_representation_2}\end{aligned}$$
By substituting $t_k=p_k(\bm{x})/k$, (\[factorization\_1/N\]) converts the sum of operators in (\[W(q;x,y,z;N)\_fermion\_representation\_2\]) into a triple product of operators as $$\begin{aligned}
& \mathcal{W}(q;\bm{x},\bm{y},\bm{z};N)
\\[1mm]
& =\langle 0|\Gamma'_+(\bm{y})
\exp\left(
\sum_{k=1}^\infty\frac{p_k(\bm{x})}{k}J_{kN
}\right)
\mathbb{G}_\emptyset\bigl(1/N\bigr)
\exp\left(
\sum_{k=1}^\infty
\frac{(-1)^{k+1}p_k(\bm{x})}{k}J_{k(N+1)}
\right)
\Gamma_- (\bm{z})|0\rangle
\\[1mm]
& =\langle 0|
\exp\left(
\sum_{k=1}^\infty \frac{p_k(\bm{x})}{k}J_{kN}
\right)
\Gamma'_+(\bm{y})
\mathbb{G}_\emptyset \bigl(1/N\bigr)
\Gamma_-(\bm{z})
|0\rangle
\exp\left(
\sum_{m=1}^\infty
\frac{(-1)^{m+1}}{m}
p_m(\bm{x})p_{m(N+1)}(\bm{z})
\right). \end{aligned}$$ We thus obtain (\[W(q;x,y,z;N)\_fermion\_expression\]).
Proof of Theorem \[theorem\_NKdV\]
----------------------------------
We first derive a reduction formula of the generating functions of three-partition Hodge integrals at positive integral values of $\tau$. This formula gives a generalization of (\[reduction\]) at all positive integral values of $\tau$. We then consider integrable hierarchies that capture a certain aspect of integrable structure underlying the generating functions of two-partition Hodge integrals. By combining these considerations, we eventually obtain Theorem \[theorem\_NKdV\].
Let $N$ be a positive integer.
\[reduction\_formula\_G\_tau=N\] The generating function (\[connected\_G\]) satisfies the relation $$\begin{aligned}
G\bigl(\lambda;(p^{(1)},p^{(2)},p^{(3)});N\bigr)
=G\bigl(\lambda;(0,p^+,p^{(3)});N\bigr)
+ \sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p^{(1)}_mp^{(3)}_{m(N+1)},
\label{reduction_G_tau=N}\end{aligned}$$ where $p^+=(p^+_k)_{k=1}^\infty$ is a linear combination of $p^{(1)}$ and $p^{(2)}$ given by $$\begin{aligned}
p^+_k
& =\left\{\begin{array}{cl}
(-1)^{k+1}Np^{(1)}_{k/N}+p^{(2)}_k, & \mbox{$k\equiv 0$ mod $N$},
\\[2mm]
p^{(2)}_k, &\mbox{$k\not\equiv 0$ mod $N$}.
\end{array}\right.
\label{p+_tau=N}\end{aligned}$$
#### *Proof.*
By Theorem \[theorem\_conjecture\] and Proposition \[fermionic\_formula\_W(q;x,y,z;N)\], we can rewrite the exponentiated generating function as $$\begin{aligned}
& \exp\left(G(\lambda;\overrightarrow{p};N)\right)
\nonumber \\[2mm]
& \quad = \langle 0|\exp\left(\sum_{k=1}^\infty \frac{p^{(1)}_k}{k}J_{kN}
+ \sum_{k=1}^\infty\frac{(-1)^{k+1}p^{(2)}_k}{k}J_{k}\right)
\mathbb{G}_\emptyset \bigl(1/N\bigr)
\exp\left(\sum_{k=1}^\infty\frac{p^{(3)}_k}{k}J_{-k}\right)|0\rangle
\nonumber \\[0.5mm]
& \qquad\qquad \times
\exp\left(\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p^{(1)}_mp^{(3)}_{m(N+1)}\right)
\nonumber \\[2mm]
& \quad = \langle 0|\exp\left(\sum_{k=1}^\infty\frac{(-1)^{k+1}p^+_k}{k}J_{k}\right)
\mathbb{G}_\emptyset\bigl(1/N\bigr)
\exp\left(\sum_{k=1}^\infty \frac{p^{(3)}_k}{k}J_{-k}\right)|0\rangle
\nonumber \\[0.5mm]
& \qquad\qquad \times
\exp\left(\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p^{(1)}_mp^{(3)}_{m(N+1)}\right) .
\label{exp(G)_tau=N}\end{aligned}$$ By letting $\overrightarrow{p}=(0,p^{(2)},p^{(3)})$, (\[exp(G)\_tau=N\]) takes the simplified form $$\begin{aligned}
& \exp\left(G(\lambda;(0,p^{(2)},p^{(3)});N)\right) \\[1.0mm]
&\qquad = \langle 0|\exp\left(\sum_{k=1}^\infty\frac{(-1)^{k+1}p^{(2)}_k}{k}J_{k}\right)
\mathbb{G}_\emptyset\bigl(1/N\bigr)
\exp\left(\sum_{k=1}^\infty \frac{p^{(3)}_k}{k}J_{-k}\right)|0\rangle. \end{aligned}$$ We can replace the VEV $\langle 0|\,\dots\,|0\rangle$ in the last part of (\[exp(G)\_tau=N\]) with this expression. This yields the relation $$\begin{aligned}
& \exp\left(G(\lambda;(p^{(1)},p^{(2)},p^{(3)});N)\right) \\
& \qquad = \exp\left(G(\lambda;(0,p^+,p^{(3)});N)\right)
\exp\left(\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p^{(1)}_mp^{(3)}_{m(N+1)}\right),\end{aligned}$$ hence (\[reduction\_G\_tau=N\]).\
\[integrable\_structure\_two\_partition\_tau=N\] ${\mathcal{T}}(\lambda,N,0,p^{(2)},{\boldsymbol{t}})$ is a tau function of the KP hierarchy with respect to the time variables ${\boldsymbol{t}}$, and satisfies the equations $$\begin{aligned}
\frac{\partial^2\log{\mathcal{T}}(\lambda,N,0,p^{(2)},{\boldsymbol{t}})}{\partial t_k\partial t_{m(N+1)}}
= 0,\quad k,m = 1,2,\ldots.
\label{531}\end{aligned}$$
#### *Proof.*
We start from the fermionic expression $${\mathcal{T}}(\lambda,N,0,p^{(2)},{\boldsymbol{t}})
= \langle 0|\mathbf{g}(\lambda,N;p^{(2)})
\exp\left(\sum_{k=1}^\infty t_kJ_{-k}\right)|0\rangle,
\label{531proof-1}$$ where $\mathbf{g}(\lambda,N;p^{(2)})$ is an element of $GL(\infty)$ of the form $$\mathbf{g}(\lambda,N;p^{(2)})
= \exp\left(\sum_{k=1}^\infty\frac{(-1)^{k+1}p^{(2)}_k}{k}J_k\right)
\mathbb{G}_\emptyset(1/N).$$ (\[ext\_shift\_symmetry\_c1\]) contains the special subset $$\mathbb{G}_\emptyset(1/N)J_{-m(N+1)}
= (-1)^mJ_{-mN}\mathbb{G}_\emptyset(1/N), \quad m = 1,2,\ldots,$$ which yields the algebraic relation $$\begin{gathered}
\mathbf{g}(\lambda,N;p^{(2)})\exp\left(\sum_{m=1}^\infty t_{m(N+1)}J_{-m(N+1)}\right)
\notag\\[0.5mm]
= \exp\left(\sum_{m=1}^\infty(-1)^mt_{m(N+1)}J_{-mN}\right)
\mathbf{g}(\lambda,N;p^{(2)})
\exp\left(- \sum_{m=1}^\infty(-1)^{m(N+1)}p^{(2)}_{mN}t_{m(N+1)}\right) \end{gathered}$$ satisfied by $\mathbf{g}(\lambda,N;p^{(2)})$. We can use this algebraic relation to rewrite (\[531proof-1\]) as $$\begin{aligned}
{\mathcal{T}}(\lambda,N,0,p^{(2)},{\boldsymbol{t}})
& = \langle 0|\mathbf{g}(\lambda,N;p^{(2)})
\exp\left(\sum_{r=1}^N\sum_{m=1}^\infty t_{m(N+1)-r}J_{-m(N+1)+r}
\right)|0\rangle
\notag\\
& \quad\mbox{}\times
\exp\left(- \sum_{m=1}^\infty(-1)^{m(N+1)}p^{(2)}_{mN}t_{m(N+1)}\right). \end{aligned}$$ This implies that ${\mathcal{T}}(\lambda,N,0,p^{(2)},{\boldsymbol{t}})$ satisfies (\[531\]).\
#### *Proof of Theorem \[theorem\_NKdV\].*
By (\[reduction\_G\_tau=N\]), ${\mathcal{T}}(\lambda,N,p^{(1)},p^{(2)},{\boldsymbol{t}})$ can be factorized as $$\begin{aligned}
{\mathcal{T}}(\lambda,N,p^{(1)},p^{(2)},{\boldsymbol{t}})
&= \exp\left(G(\lambda;\,0,p^{+},p^{(3)};\,N)\right)
\exp\left(\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p^{(1)}_mp^{(3)}_{m(N+1)}\right)\\
&= {\mathcal{T}}(\lambda,N,0,p^{+},{\boldsymbol{t}})
\exp\left(\sum_{m=1}^\infty\frac{(-1)^{m+1}}{m}p^{(1)}_mp^{(3)}_{m(N+1)}\right).
\end{aligned}$$ This relation and (\[531\]) imply that ${\mathcal{T}}(\lambda,N,p^{(1)},p^{(2)},{\boldsymbol{t}})$ satisfies (\[red-cond2\]).\
[99]{}
T. Graber and R. Pandharipande, Localization of virtual classes, Invent. Math. [**135**]{} (1999) 487–518. M. Kontsevich, Enumeration of rational curves via torus actions, Progr. Math. [**129**]{}, Birkhäuser, Boston, Boston, MA, 1995, pp. 335–368.
J. Li, C.-C. M. Liu, K. Liu and J. Zhou, A mathematical theory of the topological vertex, Geom. Topol. [**13**]{} (2009), 527–621, arXiv:math/0408426v3.
M. Aganagic, A. Klemm, M. Mari$\tilde{\mbox{n}}$o and C. Vafa, The topological vertex, Commun. Math. Phys. [**254**]{} (2005), 425–478, arXiv:hep-th/0305132.
C.-C. Liu, K. Liu and J. Zhou, On a proof of a conjecture of Mariño-Vafa on Hodge Integrals, Math. Res. Lett. [**11**]{} (2004), no. 2-3, 259–272, arXiv:math/0306257. C.-C. Liu, K. Liu and J. Zhou, A proof of a conjecture of Mariño–Vafa on Hodge Integrals, J. Differential Geom. [**65**]{} (2003), 289–340, arXiv:math/0306434. C.-C. M. Liu, K. Liu and J. Zhou, A formula of two-partition Hodge integrals, J. Amer. Math. Soc. [**20**]{} (2007), no. 1, 149–184, arXiv:math/0310272v3.
M. Mariño and C. Vafa, Framed knots at large N, Contemp. Math. [**310**]{} (Amer. Math. Soc., Providence, RI, 2002), pp. 185–204, arXiv:hep-th/0108064. T. Ekedahl, S. Lando, M. Shapiro and A. Vainshtein, Hurwitz numbers and intersections on moduli spaces of curves, Invent. Math. 1[**46**]{} (2001), 297–327, arXiv:math/0004096. T. Nakatsu and K. Takasaki, Melting crystal, quantum torus and Toda hierarchy, Comm. Math. Phys. [**285**]{} (2009), 445–156, arXiv:0710.5339 \[hep-th\].
T. Nakatsu and K. Takasaki, Integrable structure of melting crystal model with external potentials, Adv. Stud. Pure Math. vol. [**59**]{}, 201–223 (Mathematical Society of Japan, Tokyo, 2010), arXiv:0807.4970 \[math-ph\].
K. Takasaki and T. Nakatsu, Open string amplitudes of closed topological vertex, J. Phys. A: Math. Theor. [**49**]{} (2016), 025403, arXiv:1507.07053 \[math-ph\].
K. Takasaki, Generalized Ablowitz-Ladik hierarchy in topological string theory, J. Phys. A: Math. Theor. [**47**]{} (2014), 165201, arXiv:1312.7184 \[math-ph\].
K. Takasaki and T. Nakatsu, $q$-difference Kac-Schwarz operators in topological string theory, SIGMA [**13**]{} (2017), 009, arXiv:1609.00882 \[math-ph\].
T. Nakatsu and K. Takasaki, to appear.
M. Aganagic, R. Dijkgraaf, A. Klemn, M. Mari$\tilde{\mbox{n}}$o and C. Vafa, Topological strings and integrable hierarchies, Commun. Math. Phys. [**261**]{} (2006), 451–516, arXiv:hep-th/0312085.
J. Zhou, Hodge integrals and integrable hierarchies, Lett. Math. Phys. [**93**]{} (2010), 55–71, arXiv:math/0310408. I. G Macdonald, Symmetric functions and Hall polynomials, Oxford University Press 1995. A. Iqbal and C. Kz$\mbox{\c{c}}$az and C. Vafa, The refined topological vertex, JHEP [**0910**]{} (2009), 069, arXiv:hep-th/0701156.
M. Taki, Refined topological vertex and instanton counting, JHEP [**0803**]{} (2008), 048, arXiv:0710.1776 \[hep-th\].
L. A. Dickey, Soliton Equations and Hamiltonian System, World Scientific, Singapore, 2003. I. P. Goulden and D. M. Jackson, Transitive factorizations into transpositions and holomorphic mappings on the sphere, Proc. Amer. Math. Soc. [**125**]{} (1997), 51-60.
[^1]: E-mail: `[email protected]`
[^2]: E-mail: `[email protected]`
[^3]: The Mariño-Vafa conjecture was also resolved by Okounkov and Pandharipande using a quite different method.
[^4]: We follow a definition used in the recent literature [@Iqbal-Kzcaz-Vafa; @Taki]. This definition differs from the earlier one [@AKMV; @ADKMV] in that $q$ is replaced by $q^{-1}$ and an overall factor of the form $q^{\sum_{a=1}^3\kappa(\mu^{(a)})/2}$ is multiplied.
[^5]: The normalization of $V^{(k)}_m$ for $k\neq 0$ differs from [@Nakatsu_Takasaki_CMP_2009; @Nakatsu_Takasaki_ASPS] in that an overall factor $q^{-k/2}$ is multiplied and that a constant term $q^{k/2}(1-q^k)^{-1}\delta_{m,0}$ is subtracted.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Kondo effect is the many-body screening of a local spin by a cloud of electrons at very low temperature. It has been proposed as an explanation of the zero-bias anomaly in quantum point contacts where interactions drive a spontaneous charge localization. However, the Kondo origin of this anomaly remains under debate, and additional experimental evidence is necessary. Here we report on the first phase-sensitive measurement of the zero-bias anomaly in quantum point contacts using a scanning gate microscope to create an electronic interferometer. We observe an abrupt shift of the interference fringes by half a period in the bias range of the zero-bias anomaly, a behavior which cannot be reproduced by single-particle models. We instead relate it to the phase shift experienced by electrons scattering off a Kondo system. Our experiment therefore provides new evidence of this many-body effect in quantum point contacts.'
author:
- |
B. Brun,$^{1,2}$ F. Martins,$^3$ S. Faniel,$^3$ B. Hackens,$^3$ A. Cavanna,$^4$ C. Ulysse,$^4$ A. Ouerghi,$^4$ U. Gennser,$^4$\
D. Mailly,$^4$ P. Simon,$^5$ S. Huant,$^{1,2}$ V. Bayot,$^{1,3}$ M. Sanquer,$^{1,6}$ and H. Sellier$^{1,2}$
title: 'Electron phase shift at the zero-bias anomaly of quantum point contacts'
---
Quantum point contacts [@wharam-88-jpc; @vanwees-88-prl] (QPCs) are small constrictions in high-mobility two-dimensional electron gases (2DEGs) controlled by a metallic split gate at the surface of a semiconductor heterostructure. Despite their apparent simplicity, they reveal complex many-body phenomena which defy our understanding. When these quasi-one-dimensional ballistic channels are sufficiently open, electrons are perfectly transmitted via each available transverse mode [@buttiker-90-prb], and the conductance is quantized in units of the conductance quantum $2e^2/h$. Below the first conductance plateau however, this single-particle picture fails due to the increasing importance of many-body effects. An additional shoulder shows up in the linear conductance curve around $0.7\times 2e^2/h$, called the 0.7 anomaly [@thomas-96-prl], and a narrow peak of enhanced conductance appears around zero bias in the non-linear conductance curves at low enough temperature, called the zero-bias anomaly [@cronenwett-02-prl] (ZBA). The peak behavior versus temperature and magnetic field was shown to share strong similarities with the Kondo effect in quantum dots [@goldhabergordon-98-nat; @cronenwett-98-sci] (QDs), i.e. the many-body screening of a local spin by conduction electrons below a characteristic temperature [@kondo-64-ptp; @glazman-88-jetp; @ng-88-prl]. However, deviations of the ZBA from the established Kondo effect have been reported [@cronenwett-02-prl; @sfigakis-08-prl; @sarkozy-09-prb; @ren-10-prb], and the occurrence of this effect in QPCs remains a debated issue [@micolich-11-jpcm; @bauer-13-nat; @heyder-15-prb].
Because of enhanced electron interactions at low density, a spontaneous charge localization is predicted in QPCs below the first plateau [@rejec-06-nat; @guclu-09-prb], showing similarities with the one-dimensional Wigner crystallization [@wigner-34-pr; @matveev-04-prl]. This phenomenon is supported by two recent experiments where localized states with even and odd numbers of charges have been observed [@iqbal-13-nat; @brun-14-ncom]. The development of a Kondo effect is therefore expected at very low temperature, but its specific properties for a self-consistently localized state have not been calculated yet, due to the complexity of the problem. In this unsettled situation, the ZBA remains the subject of intensive investigations, and any new information pointing to a Kondo origin is important.
![(a) QPC conductance versus gate voltage at 30 mK (different cool down than other figures). Inset: image of the metallic split gate. (b) Differential conductance versus source-drain bias at 25 mK and different gate voltages. (c) Temperature dependence of the 0.7 anomaly from 50 to 900 mK. (d) Temperature dependence of the ZBA from 25 to 870 mK.[]{data-label="figure1"}](figure1.pdf){width="7.5cm"}
Here we use a scanning gate microscope [@topinka-01-nat] (SGM) to create a Fabry-Pérot (FP) cavity between the QPC and the tip [@jura-09-prb; @freyn-08-prl], and measure the phase of the electron wave function scattered by the QPC in the ZBA regime. Phase-sensitive experiments indeed provide unique information on quantum phenomena and, in the case of QPCs, will help us to clarify the microscopic origin of the ZBA. Recently, a phase measurement on a QPC has been reported [@kobayashi-13-arxiv], but no significant deviation from the single-particle prediction has been found [@lackenby-14-prb]. In the past, the transmission phase of QDs in the Kondo regime was measured by embedding them in Aharonov-Bohm (AB) rings [@ji-00-sci; @zaffalon-08-prl; @takada-14-prl]. Here we measure instead the reflection phase of the system and observe a phase shift by $\pi$ of the interference fringes in the bias voltage range of the ZBA. This shift occurs via two phase jumps, and disappears with gate voltage and temperature in the same way as the ZBA. Calculations of the reflection phase for a single-particle resonant level give a smooth shift across the resonance [@buks-96-prl], in strong contrast with the two phase jumps observed in our experiment, thereby indicating a many-body origin. The observed behavior shows characteristic signatures of the Kondo effect, where the transmission phase at the Fermi energy is locked at $\pi/2$ in the Kondo valleys [@glazman-88-jetp], and where a “sharp Kondo double phase lapse” is predicted as a function of source-drain bias [@gerland-00-prl]. We therefore attribute the observed phase shift to the Kondo effect, thus providing new evidence of this effect as the origin of the zero-bias anomaly in QPCs.
![(a) Principle of scanning gate interferometry. (b) Potential landscape created by the split gate and the tip. (c) SGM image of the conductance at 25 mK when the QPC is tuned to the first conductance plateau. The QPC center is located at the coordinates ($-500$ nm, 650 nm).[]{data-label="figure2"}](figure2.pdf){width="8.5cm"}
) versus source-drain bias at $-0.67$ V gate voltage in (b). The conductance is differentiated with respect to tip position. Top panel: conductance curve for the tip position indicated by the arrow. Right panel: 3D plot showing the position of the phase shift at the bottom of the zero-bias peak. (b) Interference fringes along the same line versus gate voltage (at zero bias). The pinch-off voltage is shifted by 40 mV in the presence of the polarized tip with respect to Fig. \[figure1\](c). Top panel: conductance curve for the tip position indicated by the arrow. Right panel: 3D plot showing the position of the phase shift at the border of the plateau.[]{data-label="figure3"}](figure3.pdf){width="8.5cm"}
. Third panel: conductance curve at the tip position of the arrow, below the plateau (red curve) and on the plateau (blue curve). Bottom panel: the phase of the fringes exhibits a shift in the bias range of the ZBA (red curve) and evolves linearly on the plateau (blue curve). (b) Interference fringes along line 2 at a gate voltage below the plateau at 760 mK (first panel) and 25 mK (second panel). Third panel: conductance curve at the tip position of the arrow, at 25, 240, 440, and 760 mK from top to bottom. Bottom panel: phase of the fringes at the same temperatures.[]{data-label="figure4"}](figure4.pdf){width="8.5cm"}
The QPC is defined in a GaAs/AlGaAs heterostructure by the 270 nm long and 300 nm wide gap of a Ti/Au split gate (inset of Fig. \[figure1\](a)). The 2DEG located 105 nm below the surface has a $2.5 \times 10^{11}$ cm$^{-2}$ electron density and a $1.0 \times 10^6$ cm$^2$V$^{-1}$s$^{-1}$ electron mobility. The device is fixed to the mixing chamber of a dilution fridge in front of a cryogenic scanning probe microscope [@hackens-10-ncom; @martins-13-sr] and cooled down to a base temperature of 25 mK at zero gate voltage. The four-probe differential conductance is measured by a lock-in technique using a 10 $\mu$V excitation. A series resistance of 1600 $\Omega$ is subtracted from all data in order to have the first conductance plateau at $2e^2/h$.
At the base temperature (25 mK), the linear conductance shows quantized plateaus and smooth transitions versus gate voltage (Fig. \[figure1\](a)), while at higher temperatures, the conductance exhibits the well-known 0.7 anomaly [@thomas-96-prl] below the first plateau (Fig. \[figure1\](c)). The non-linear differential conductance versus source-drain bias shows a narrow peak at zero bias (Fig. \[figure1\](b)), the so-called ZBA [@cronenwett-02-prl], which vanishes rapidly at higher temperatures (Fig. \[figure1\](d)). The temperature dependence of the peak height can be rescaled on the universal scaling law of Kondo QDs [@goldhabergordon-98-prl; @vanderwiel-00-sci; @nygard-00-nat] with a single scaling parameter $T_K$, called Kondo temperature, for all gate voltages (Fig. S1 in the Supplemental Material [@supplemental]).
We now investigate the scattering phase of the QPC in the ZBA regime at very low temperature (25 mK) using a SGM-based interferometry experiment (Fig. \[figure2\](a)). The SGM tip is scanned above the 2DEG at finite distance from the QPC, with a tip voltage of $-6$ V and a tip-to-surface height of 30 nm, chosen such as to locally deplete the 2DEG (Fig. \[figure2\](b)). Electrons propagating out of the QPC are scattered by this tip-induced perturbation and partially reflected towards the QPC. Interference fringes show up in the SGM images (Fig. \[figure2\](c)) due to the coherent superposition of waves reflected by the QPC and the tip, forming together a FP cavity. To probe the scattering phase at the ZBA, the tip is scanned along individual lines where regular fringe patterns are observed (red lines). In the ZBA region below the first plateau, a shift of the interference fringes appears around zero source-drain bias, with abrupt jumps on each side of the ZBA (Fig. \[figure3\](a)). When the fringes are recorded while sweeping the gate voltage (Fig. \[figure3\](b)), a similar shift is observed when the conductance drops below the first plateau, i.e. when the QPC enters the ZBA region. This phase shift reveals the non-trivial scattering phase of the ZBA and constitutes a new experimental signature of this many-body effect.
The phase of the interference fringes in various situations is extracted in Fig. \[figure4\] from a Fourier transform performed along the scan axis. When the QPC is tuned to the first plateau (Fig. \[figure4\](a), top panel), the fringes evolve monotonically with source-drain bias due to a change in wavelength for electrons injected at higher energy [@gorini-14-prb], and the extracted phase is linear (blue curve, bottom panel). Below the first plateau (second panel), the fringes exhibit a sharp phase jump at negative bias and a smooth one at positive bias, also visible on the extracted phase (red curve, bottom panel). These phase jumps occur when the conductance increases above the background to build the zero-bias peak (red curve, third panel). Figure S2 in the Supplemental Material [@supplemental] presents additional data.
In order to measure the zero-bias phase shift, it is necessary to have a reference phase at the same gate voltage for a situation without the ZBA. This can be obtained by recording the interference pattern at different temperatures and fixed gate voltage (Fig. \[figure4\](b)). At a temperature where the ZBA has disappeared (top panel), the phase evolves linearly (blue curve, bottom panel), whereas at the lowest temperature where the ZBA is at maximum (second panel), the phase shows two jumps with a shift of about $\pi$ (red curve, bottom panel). At intermediate temperatures, the phase jumps remain at the same bias voltages, but the shift disappears progressively, in a non-uniform way, explaining larger fluctuations in the extracted phase (Fig. S3 [@supplemental]).
A better accuracy on the phase determination can be obtained by choosing a longer scanning line with more interference fringes, but the difficulty is to find such a long line where the ZBA remains relatively constant along the entire scan. Indeed, as reported in Ref. [@brun-14-ncom], the ZBA splits up into finite bias peaks due to a periodic change of the localized state occupancy with tip distance, and this limits the available scan lengths. However, when the tip is scanned along the red line 3, the interference fringes are regularly spaced (Fig. \[figure5\](a), top panel) and the ZBA is only slightly disturbed. The phase (bottom panel) shows an abrupt jump at negative bias and a smoother change at positive bias, with a zero-bias shift close to $\pi$. The phase shift is also observed versus gate voltage along this scanning line (Fig. S4 [@supplemental]).
In our experiment, the sensitivity of the interference pattern to the ZBA, which is an intrinsic QPC property, demonstrates that the QPC is part of the interferometric cavity. The QPC represents one of the cavity mirrors, as also realized in Refs. [@jura-09-prb; @kozikov-13-njp], but in contrast to experiments where interference was attributed to impurities in the 2DEG [@topinka-01-nat; @jura-07-natp]. This situation is consistent with the fact that the interference fringes are observed within the thermal length $L_T=\hbar v_F/k_B T$ which is 1.5 $\mu$m at 1 K and much more below (Fig. S5 [@supplemental]). In addition, the zero-bias phase shift is observed for all scanning lines that have been investigated, showing that it really corresponds to the scattering phase of the QPC, and does not result from specific scatterers in the 2DEG region between the QPC and the tip. It has also been observed in a second device (Fig. S6 [@supplemental]).
![(a) Interference fringes when the tip is scanned along line 3. Bottom panel: average conductance curve (blue) and phase of the fringes (red) showing a shift by $\pi$ within the ZBA. (b) Model of the SGM-based interferometry experiment where the QPC is represented by an asymmetric QD. Bottom panel: conductance in the symmetric case (blue) and phase of the interference fringes versus energy, for different asymmetries of the tunneling rates $\Gamma_L$ and $\Gamma_R$ (red to yellow). Central panel: tip-induced interference fringes for the asymmetry of the red curve.[]{data-label="figure5"}](figure5.pdf){width="8cm"}
For quantitative analysis, it is important to note that our SGM experiment realizes a FP cavity [@rossler-15-prl] and therefore probes the reflection phase of the QPC. This situation differs from previous experiments on QDs using AB rings [@schuster-97-nat] which probe the transmission phase of the embedded device. In the case of a single-particle resonant level in a QD, the transmission phase presents a smooth shift by $\pi$ across the resonance [@schuster-97-nat], while the reflection phase of an asymmetric QD presents a shift by 0 or $2\pi$ depending on which side the highest barrier is located [@buks-96-prl]. The reflection phase measured in our SGM experiment is therefore between zero and twice the transmission phase of the QPC and should be interpreted carefully.
The spontaneous charge localization in QPCs results from the formation of self-consistent barriers along the channel [@rejec-06-nat; @guclu-09-prb]. The QPC can thus be modeled by a small QD with two asymmetric barriers on top of the main potential barrier controlled by the gate [@ren-10-prb]. The phase of tip-induced interference fringes has been calculated for non-interacting electrons using this simple model (Fig. \[figure5\](b) and Fig. S7 [@supplemental]). For all barrier asymmetries, the calculated phase exhibits a single smooth shift across the resonance (Fig. \[figure5\](b), bottom panel), in contrast with the experimental behavior showing phase jumps on both sides of the resonance (Fig. \[figure5\](a)). This difference indicates that the observed phase shift does not result from scattering on a localized state. The spontaneously localized states are indeed expected at larger energy and to survive up to much higher temperatures [@yoon-07-prl]. Here, we are dealing with a low-energy phenomenon, that we attribute to the screening of the localized states by the Kondo effect at very low temperature [@kondo-64-ptp; @glazman-88-jetp; @ng-88-prl]. This screening produces a narrow resonance in the density of states at the Fermi level and gives rise to a conductance peak at zero bias [@hershfield-91-prl].
Below the Kondo temperature, the transmission phase of a symmetric QD equals $\pi/2$ in the gate voltage range of a Kondo valley [@ji-00-sci; @zaffalon-08-prl; @takada-14-prl], and the conductance reaches $2e^2/h$ [@vanderwiel-00-sci; @kretinin-11-prb]. The phase shift observed in our experiment at zero bias may correspond to this Kondo scattering phase, but in the reflection coefficient, which can be twice the value of the transmission coefficient [@buks-96-prl]. This situation arises if the smallest barrier is located on the cavity side, which is likely to occur since the main gate-controlled barrier induces this asymmetry on the self-consistent confinement potential [@ren-10-prb]. A phase shift by $\pi$ is therefore expected at zero bias, which is close to the value found experimentally.
At finite bias voltage, the Kondo phase shift has been calculated in Ref. [@gerland-00-prl] for a QD in equilibrium (Fig. S8 [@supplemental]). It exhibits three switches from 0 to $\pi$ corresponding to the transmission through the single-particle level (first and second electrons) and through the Kondo resonance (always centered at zero bias). A “sharp Kondo double phase lapse” has been predicted around the Kondo peak at low enough temperature [@gerland-00-prl], and the double phase jump seen in our experiment around the ZBA may correspond to such an effect. Phase lapses by $\pi$ are usually observed *versus gate voltage* between the successive charge states of QDs in the Coulomb blockade regime, and explained by the coupling of the different orbitals to the leads [@oreg-97-prb; @karrasch-07-prl; @hecht-09-prb]. But to our knowledge, phase lapses *versus source-drain bias* have not been reported before. In addition, decoherence of the Kondo correlations at finite bias voltage [@hershfield-91-prl; @meir-93-prl] is also an effect that should be considered, but no theoretical prediction of the Kondo phase shift out of equilibrium exists at the moment. We expect our experiment to stimulate theoretical works in this direction.
To conclude, we performed the first phase-sensitive measurements on the QPC conductance anomalies using scanning gate interferometry. Whenever the ZBA is present, a phase shift of the interference fringes is observed around zero bias, and we interpret it as the Kondo phase shift experienced by electrons at the Fermi level. In addition, the two phase jumps around the conductance peak may correspond to the predicted phase lapses around the Kondo resonance. These results reinforce our understanding of the ZBA in terms of a Kondo effect on spontaneously localized states.
We thank J.-L. Pichard, D. Weinmann, X. Waintal, H. Baranger, M. Lavagna, S. Florens, and T. Meunier for discussions. This work was supported by the French Agence Nationale de la Recherche (“ITEM-exp” project), by FRFC Grant No. 2.4503.12, and by FRS-FNRS Grants No. 1.5.044.07.F and No. J.0067.13. F.M. and B.H. acknowledge support from the Belgian FRS-FNRS, S.F. received support from the FSR at UCL, and V.B. acknowledges the award of a “chair of excellence” by the Nanosciences foundation in Grenoble.
[99]{} D. A. Wharam, T. J. Thornton, R. Newbury, M. Pepper, H. Ahmed, J. E. F. Frost, D. G. Hasko, D. C. Peacock, D. A. Ritchie, and G. A. C. Jones, J. Phys. C **21**, L209 (1988).
B. J. van Wees, H. van Houten, C. W. J. Beenakker, J. G. Williamson, L. P. Kouwenhoven, D. van der Marel, and C. T. Foxon, Phys. Rev. Lett. **60**, 848 (1988).
M. Büttiker, Phys. Rev. B **41**, 7906 (1990).
K. J. Thomas, J. T. Nicholls, M. Y. Simmons, M. Pepper, D. R. Mace, and D. A. Ritchie, Phys. Rev. Lett. **77**, 135 (1996).
S. M. Cronenwett, H. J. Lynch, D. Goldhaber-Gordon, L. P. Kouwenhoven, C. M. Marcus, K. Hirose, N. S. Wingreen, and V. Umansky, Phys. Rev. Lett. **88**, 226805 (2002).
D. Goldhaber-Gordon, H. Shtrikman, D. Mahalu, D. Abusch-Magder, U. Meirav, and M. A. Kastner, Nature (London) **391**, 156 (1998).
S. M. Cronenwett, T. H. Oosterkamp, and L. P. Kouwenhoven, Science **281**, 540 (1998).
J. Kondo, Prog. Theor. Phys. **32**, 37 (1964).
L. I. Glazman and M. E. Raikh, JETP Lett. **47**, 452 (1988).
T. K. Ng and P. A. Lee, Phys. Rev. Lett. **61**, 1768 (1988).
F. Sfigakis, C. J. B. Ford, M. Pepper, M. Kataoka, D. A. Ritchie, and M. Y. Simmons, Phys. Rev. Lett. **100**, 026807 (2008).
S. Sarkozy, F. Sfigakis, K. Das Gupta, I. Farrer, D. A. Ritchie, G. A. C. Jones, and M. Pepper, Phys. Rev. B **79**, 161307(R) (2009).
Y. Ren, W. W. Yu, S. M. Frolov, J. A. Folk, and W. Wegscheider, Phys. Rev. B **82**, 045313 (2010).
A. P. Micolich, J. Phys. Condens. Matter **23**, 443201 (2011).
F. Bauer, J. Heyder, E. Schubert, D. Borowsky, D. Taubert, B. Bruognolo, D. Schuh, W. Wegscheider, J. von Delft, and S. Ludwig, Nature (London) **501**, 73 (2013).
J. Heyder, F. Bauer, E. Schubert, D. Borowsky, D. Schuh, W. Wegscheider, J. von Delft, and S. Ludwig, Phys. Rev. B **92**, 195401 (2015).
T. Rejec and Y. Meir, Nature (London) **442**, 900 (2006).
A. D. Guçlu, C. J. Umrigar, H. Jiang, and H. U. Baranger, Phys. Rev. B **80**, 201302(R) (2009).
E. Wigner, Phys. Rev. **46**, 1002 (1934).
K. A. Matveev, Phys. Rev. Lett. **92**, 106801 (2004).
M. J. Iqbal, R. Levy, E. J. Koop, J. B. Dekker, J. P. de Jong, J. H. M. van der Velde, D. Reuter, A. D. Wieck, R. Aguado, Y. Meir, and C. H. van der Wal, Nature (London) **501**, 79 (2013).
B. Brun, F. Martins, S. Faniel, B. Hackens, G. Bachelier, A. Cavanna, C. Ulysse, A. Ouerghi, U. Gennser, D. Mailly, S. Huant, V. Bayot, M. Sanquer, and H. Sellier, Nat. Commun. **5**, 4290 (2014).
M. A. Topinka, B. J. LeRoy, R. M. Westervelt, S. E. J. Shaw, R. Fleischmann, E. J. Heller, K. D. Maranowski, and A. C. Gossard, Nature (London) **410**, 183 (2001).
M. P. Jura, M. A. Topinka, M. Grobis, L. N. Pfeiffer, K. W. West, and D. Goldhaber-Gordon, Phys. Rev. B **80**, 041303(R) (2009).
A. Freyn, I. Kleftogiannis, and J.-L. Pichard, Phys. Rev. Lett. **100**, 226802 (2008).
T. Kobayashi, S. Tsuruta, S. Sasaki, H. Tamura, and T. Akazaki, arXiv:1306.6689.
B. G. C. Lackenby and O. P. Sushkov, Phys. Rev. B **90**, 155434 (2014).
Y. Ji, M. Heiblum, D. Sprinzak, D. Mahalu, and H. Shtrikman, Science **290**, 779 (2000).
M. Zaffalon, A. Bid, M. Heiblum, D. Mahalu, and V. Umansky, Phys. Rev. Lett. **100**, 226601 (2008).
S. Takada, C. Bäuerle, M. Yamamoto, K. Watanabe, S. Hermelin, T. Meunier, A. Alex, A. Weichselbaum, J. von Delft, A. Ludwig, A. D. Wieck, and S. Tarucha, Phys. Rev. Lett. **113**, 126601 (2014).
E. Buks, R. Schuster, M. Heiblum, D. Mahalu, V. Umansky, and H. Shtrikman, Phys. Rev. Lett. **77**, 4664 (1996).
U. Gerland, J. von Delft, T. A. Costi, and Y. Oreg, Phys. Rev. Lett. **84**, 3710 (2000).
B. Hackens, F. Martins, S. Faniel, C. A. Dutu, H. Sellier, S. Huant, M. Pala, L. Desplanque, X. Wallart, and V. Bayot, Nat. Commun. **1**, 39 (2010).
F. Martins, S. Faniel, B. Rosenow, H. Sellier, S. Huant, M. G. Pala, L. Desplanque, X. Wallart, V. Bayot, and B. Hackens, Sci. Rep. **3**, 1416 (2013).
D. Goldhaber-Gordon, J. Göres, M. A. Kastner, H. Shtrikman, D. Mahalu, and U. Meirav, Phys. Rev. Lett. **81**, 5225 (1998).
W. G. van der Wiel, S. De Franceschi, T. Fujisawa, J. M. Elzerman, S. Tarucha, and L. P. Kouwenhoven, Science **289**, 2105 (2000).
J. Nyg[å]{}rd, D. H. Cobden, and P. E. Lindelof, Nature (London) **408**, 342 (2000).
See Supplemental Material (below) for additional data and analysis, which also includes Refs. [@costi-94-jpcm; @pletyukhov-12-prl; @kretinin-12-prb; @klochan-13-prb; @costi-09-prl].
T. A. Costi, A. C. Hewson, and V. Zlatic, J. Phys. Condens. Matter **6**, 2519 (1994).
M. Pletyukhov and H. Schoeller, Phys. Rev. Lett. **108**, 260601 (2012).
A. V. Kretinin, H. Shtrikman, and D. Mahalu, Phys. Rev. B **85**, 201301(R) (2012).
O. Klochan, A. P. Micolich, A. R. Hamilton, D. Reuter, A. D. Wieck, F. Reininghaus, M. Pletyukhov, and H. Schoeller, Phys. Rev. B **87**, 201104(R) (2013).
T. A. Costi, L. Bergqvist, A. Weichselbaum, J. von Delft, T. Micklitz, A. Rosch, P. Mavropoulos, P. H. Dederichs, F. Mallet, L. Saminadayar, and C. Bäuerle, Phys. Rev. Lett. **102**, 056802 (2009).
C. Gorini, D. Weinmann, and R. A. Jalabert, Phys. Rev. B **89**, 115414 (2014).
A. A. Kozikov, C. Rössler, T. Ihn, K. Ensslin, C. Reichl, and W. Wegscheider, New J. Phys. **15**, 013056 (2013).
M. P. Jura, M. A. Topinka, L. Urban, A. Yazdani, H. Shtrikman, L. N. Pfeiffer, K. W. West, and D. Goldhaber-Gordon, Nat. Phys. **3**, 841 (2007).
C. Rössler, D. Oehri, O. Zilberberg, G. Blatter, M. Karalic, J. Pijnenburg, A. Hofmann, T. Ihn, K. Ensslin, C. Reichl, and W. Wegscheider, Phys. Rev. Lett. **115**, 166603 (2015).
R. Schuster, E. Buks, M. Heiblum, D. Mahalu, V. Umansky, and H. Shtrikman, Nature (London) **385**, 417 (1997).
Y. Yoon, L. Mourokh, T. Morimoto, N. Aoki, Y. Ochiai, J. L. Reno, and J. P. Bird, Phys. Rev. Lett. **99**, 136805 (2007).
S. Hershfield, J. H. Davies, and J. W. Wilkins, Phys. Rev. Lett. **67**, 3720 (1991).
A. V. Kretinin, H. Shtrikman, D. Goldhaber-Gordon, M. Hanl, A. Weichselbaum, J. von Delft, T. Costi, and D. Mahalu, Phys. Rev. B **84**, 245316 (2011).
Y. Oreg and Y. Gefen, Phys. Rev. B **55**, 13726 (1997).
C. Karrasch, T. Hecht, A. Weichselbaum, Y. Oreg, J. von Delft, and V. Meden, Phys. Rev. Lett. **98**, 186802 (2007).
T. Hecht, A. Weichselbaum, Y. Oreg, and J. von Delft, Phys. Rev. B **80**, 115330 (2009).
Y. Meir, N. S. Wingreen, and P. A. Lee, Phys. Rev. Lett. **70**, 2601 (1993).
**Supplemental Material for\
Electron phase shift at the zero-bias anomaly of quantum point contacts**
B. Brun,$^{1,2}$ F. Martins,$^3$ S. Faniel,$^3$ B. Hackens,$^3$ A. Cavanna,$^4$ C. Ulysse,$^4$ A. Ouerghi,$^4$ U. Gennser,$^4$\
D. Mailly,$^4$ P. Simon,$^5$ S. Huant,$^{1,2}$ V. Bayot,$^{1,3}$ M. Sanquer,$^{1,6}$ and H. Sellier$^{1,2}$
*$^1$Université Grenoble Alpes, F-38000 Grenoble, France\
$^2$CNRS, Institut NEEL, F-38042 Grenoble, France\
$^3$IMCN/NAPS, Université catholique de Louvain, B-1348 Louvain-la-Neuve, Belgium\
$^4$CNRS, Laboratoire de Photonique et de Nanostructures, UPR20, F-91460 Marcoussis, France\
$^5$Laboratoire de Physique des Solides, Université Paris-Sud, F-91405 Orsay, France\
$^6$CEA, INAC-SPSMS, F-38054 Grenoble, France*
Scaling analysis of the zero-bias conductance versus temperature
================================================================
The temperature dependence of the linear conductance at zero bias is shown in Fig. \[figure-S1\](a). The 0.7 anomaly observed at high temperature is related to the suppression of the conductance peak observed versus source-drain bias as shown in Fig. \[figure-S1\](b). The zero-bias conductance $G$ at finite temperature, normalized to its value $G_{\rm max}$ at the lowest temperature, is plotted in Fig. \[figure-S1\](c) as a function of the temperature $T$, rescaled by a parameter $T_{\rm K}$ called Kondo temperature. This parameter is chosen such that the data points follow the universal scaling law [@costi-94-jpcm] of Kondo quantum dots (QDs) which can be approximated by the phenomenological formula [@goldhabergordon-98-prl; @vanderwiel-00-sci; @nygard-00-nat]: $$\begin{aligned}
G(T/T_{\rm K})=G_{\rm max}(1+(2^{1/s}-1)(T/T_{\rm K})^2)^{-s}\end{aligned}$$ where $s$ is a fixed parameter close to 0.2 for $S=1/2$ Kondo QDs. The single scaling parameter $T_{\rm K}$ is shown in Fig. \[figure-S1\](d) and depends on the gate voltage. Note that the conductance $G_{\rm max}$ at zero temperature is less than $2e^2/h$ and might be expressed as [@ng-88-prl]: $$\begin{aligned}
G_{\rm max}=(2e^2/h)4\Gamma_{\rm L}\Gamma_{\rm R}/(\Gamma_{\rm L}+\Gamma_{\rm R})^2\end{aligned}$$ using asymmetric tunneling rates $\Gamma_{\rm L,R}$ for the main gate-controlled QPC barrier and the weak self-consistent barrier resulting from Coulomb interactions [@ren-10-prb].
In Kondo QDs [@vanderwiel-00-sci; @nygard-00-nat] and in QPCs [@ren-10-prb; @cronenwett-02-prl; @sarkozy-09-prb], the full width at half-maximum (FWHM) of the zero-bias peak is often assumed to equal $2k_{\rm B}T_{\rm K}/e$, but recent investigations [@pletyukhov-12-prl; @kretinin-12-prb; @klochan-13-prb] have shown that the full width at 2/3 of the maximum (FW2/3M) gives a better estimate of $T_{\rm K}$. In our QPC, the peak width (almost temperature independent) is plotted in Fig. \[figure-S1\](d) as a function of gate voltage. The FW2/3M values (open symbols and dotted line) show a slightly better agreement with the $T_{\rm K}$ values from the scaling analysis than the FWHM values (open symbols and dashed line).
In the absence of a generally accepted theory of the Kondo effect in QPCs, we have also investigated the possibility of a $S=1$ Kondo effect, for which the parameter $s$ is close to 0.16 [@costi-09-prl]. However, the quality of the fit is not better, and the new set of Kondo temperatures is found to be 1.4 times larger, in worse agreement with the values of the peak width, in particular with the FW2/3M values. Since the $S=1/2$ Kondo model describes better the behavior of the ZBA, we have here an indication of the Kondo mechanism responsible for the ZBA in QPCs.
![(a) Conductance at zero bias versus gate voltage from 50 mK (blue) to 900 mK (red). (b) Differential conductance versus source-drain bias from 25 mK (blue) to 870 mK (red), at the four gate voltages indicated by the colored symbols in (a). Temperatures are 25, 90, 145, 230, 380, 650, 870 mK. (c) The conductance at zero-bias from (b) is normalized to the lowest temperature value and plotted versus a rescaled temperature $T/T_K$. The symbol color corresponds to the data taken from (b) at the gate voltage indicated in (a). The solid line indicates the universal Kondo behavior. (d) The Kondo temperatures $T_K$ used in the scaling analysis in (c) are plotted versus gate voltage as colored symbols. The graph also shows (open symbols) the characteristic temperatures $T_K^*=e\Delta V/2k_B$ given by the full width $\Delta V$ of the zero-bias peaks shown in (b) at 25 mK. Top symbols (connected by a dashed line) are obtained using the full width at 1/2 of the peak maximum (FWHM). Bottom symbols (connected by a dotted line) are obtained using the full width at 2/3 of the peak maximum (FW2/3M).[]{data-label="figure-S1"}](figure-S1.pdf){width="18cm"}
Evolution of the phase shift with gate voltage (line 1)
=======================================================
![(a) Same SGM image as in Fig. 2(c). The QPC is on the first conductance plateau. (b) Same data as in Fig. 3(b). The interference fringes at zero bias along the red line in (a) are plotted versus gate voltage. Top panel: conductance curve at 200 nm tip distance. (c) Interference fringes along the red line in (a) versus source-drain bias, for gate voltages every 5 mV from -0.675 to -0.650 V (from left to right). Central panels: raw conductance plot. Bottom panels: derivative of the conductance with respect to tip distance. Top panels: conductance curves at 0, 200, and 400 nm (from bottom to top). A phase shift with two phase jumps are visible around the ZBA when the conductance is below $0.8\times 2e^2/h$. Above this value, the ZBA splits up in two finite bias peaks, as reported in Ref. [@brun-14-ncom], and the interference fringes shows an additional phase jump at zero-bias.[]{data-label="figure-S2"}](figure-S2.pdf){width="17cm"}
Evolution of the phase shift with temperature (line 2)
======================================================
![Interference fringes versus source-drain bias for different temperatures. This figure shows the data at 240 and 440 mK mentioned in Fig. 4(b), together with the data at 25 and 760 mK already shown in Fig. 4(b). The gate voltage is -0.68 V and the tip is scanned along line 2 in Fig. 2(c). Top panels: conductance curves at 0, 100, 200, and 300 nm (from bottom to top). The phase shift disappears with increasing temperature in the same way as the ZBA.[]{data-label="figure-S3"}](figure-S3.pdf){width="15cm"}
Phase shift versus bias voltage and gate voltage (line 3)
=========================================================
![Interference fringes when the tip is scanned along line 3 in Fig. 2(c). (a) Interference fringes versus source-drain bias at -0.69 V gate voltage (same data as in Fig. 5(a)). Top panel: conductance curves every 100 nm from 0 to 900 nm (from bottom to top). (b) Interference fringes versus gate voltage at zero bias. Top panel: conductance curves every 100 nm from 0 to 900 nm (from right to left). The fringes show a phase shift at the border of the conductance plateau, similar to that shown in Fig. 3(b) which was measured along line 1.[]{data-label="figure-S4"}](figure-S4.pdf){width="11cm"}
Temperature dependence of the fringes visibility
================================================
![(a) Interference fringes versus source-drain bias (below the first conductance plateau) when the tip is scanned along a line close to the end of line 3 in Fig. 2(c). Top panel: conductance curve at 0 nm. (b) Interference fringes at zero-bias along the same scanning line, measured as a function of temperature. Top panel: evolution of the fringes visibility, showing an exponential decay on a characteristic temperature of about 800 mK. Above this temperature, the thermal length $L_T=\hbar v_F/k_B T$ becomes shorter than the distance to the QPC (about 1 $\mu$m) and the interference fringes are smeared out by thermal averaging [@topinka-01-nat]. The conductance is differentiated with respect to tip position in (a) and (b).[]{data-label="figure-S5"}](figure-S5.pdf){width="11cm"}
Phase shift at the ZBA measured in another QPC device
=====================================================
![Results obtained on a similar QPC device fabricated on the same 2DEG (this sample was studied in Ref. [@brun-14-ncom]). All data are recorded at the base temperature of 20 mK. (a) Conductance versus gate voltage. (b) SGM map at $-0.95$ V gate voltage, $-6$ V tip voltage, and 40 nm tip height. (c,d) Source-drain bias spectroscopy of the interference fringes along the line indicated in (b) at two gate voltages: $-1$ V for (c) and $-0.95$ V for (d). The interference fringes exhibit a phase shift in the bias range of the zero-bias peak in (c) while the fringes are linear on the conductance plateau in (d).[]{data-label="figure-S6"}](figure-S6.pdf){width="18cm"}
Model of the scanning gate interferometry experiment
====================================================
![(a) Model of the SGM-based interferometry experiment where the QPC is represented by an asymmetric QD with a single level $\epsilon_0$. (b) Conductance of the QD (top) and phase of the tip-induced interference (bottom) as a function of energy $\epsilon$ for different asymmetries of the QD tunneling rates $\Gamma_L$ and $\Gamma_R$. (c) Pattern of tip-induced interference fringes for the blue (top) and red (bottom) curves in (b). The conductance is differentiated with respect to $D_{\rm tip}$ to show only the fringes.[]{data-label="figure-S7"}](figure-S7.pdf){width="9cm"}
We consider a one-dimensional Fabry-Pérot (FP) cavity formed between an asymmetric quantum dot (QD), that represents the QPC, and a local potential perturbation, corresponding to the SGM tip (Fig. \[figure-S7\](a)). At zero temperature, the conductance is given by the transmission of the FP cavity: $$\begin{aligned}
T_{\rm FP} = \left|\frac{t_{\rm dot}t_{\rm tip}}{1-r_{\rm dot}r_{\rm tip}\exp(2ikD_{\rm tip})}\right|^2\end{aligned}$$ where the cavity length $D_{\rm tip}$ is changed by moving the tip, $k$ is the electron wave-vector, $t_{\rm dot},r_{\rm dot}$ and $t_{\rm tip},r_{\rm tip}$ are the complex transmission and reflexion amplitudes of the dot and the tip, respectively. The scattering amplitudes of the dot can be calculated exactly in the non-interacting case, using a tight-binding model made of a single site at energy $\epsilon_0$ connected via two tunnel barriers (hoping terms $V_{\rm L}$ and $V_{\rm R}$) to two semi-infinite leads (hoping term $t$ and lattice parameter $a$): $$\begin{aligned}
t_{\rm dot} = \frac{ 2 i \sin(ka) V_{\rm L}V_{\rm R}/t }{ \epsilon(k) - \epsilon_0 + A\cos(ka) + i A\sin(ka) } \\
r_{\rm dot} = - \: \frac{ \epsilon(k) - \epsilon_0 + A\cos(ka) + i B\sin(ka) }{ \epsilon(k) - \epsilon_0 + A\cos(ka) + i A\sin(ka) }\end{aligned}$$ where $\epsilon(k)=-2t\cos(ka)$, $A=(V_{\rm L}^2+V_{\rm R}^2)/t$, and $B=(V_{\rm L}^2-V_{\rm R}^2)/t$. Note that the reflection coefficient is for waves arriving from the cavity, i.e. on the right side of the dot. The transmission $T_{\rm FP}$ of the interferometer is calculated for a weak tip back-scattering ($t_{\rm tip}=0.99$) and various asymmetries of the dot barriers $(V_{\rm L}/t,V_{\rm R}/t)=(0.132,0.05)$, (0.109,0.09), (0.1,0.1), (0.093,0.107), (0.0515,0.132). The interference fringes are plotted in Fig. \[figure-S7\](c) as a function of tip distance and electron energy, in two situations corresponding to opposite barrier asymmetries. The simulations are plotted as a function of $-\epsilon$ to be easily compared with the experiments where the conductance is plotted as a function of source-drain bias ($eV_{\rm bias}=-\epsilon$). The global slope of the fringes simply results from the change of the electron wavelength with energy, and the shift around $\epsilon=\epsilon_0$ results from the resonant level in the QD. The phase of the fringes is extracted from the Fourier transform of $T_{\rm FP}$ with respect to $D_{\rm tip}$ and plotted in Fig. \[figure-S7\](b), bottom panel, for the five tunneling rate ratios $\Gamma_{\rm R}/\Gamma_{\rm L}=V_{\rm R}^2/V_{\rm L}^2$ indicated above. Because of the weak tip-induced reflection ($t_{\rm tip}=0.99$), the phase extracted from the interference fringes is close to the reflection phase of the dot: $$\begin{aligned}
\phi_{\rm dot} = \arctan\left(\frac{B\sin(ka)}{\epsilon(k)-\epsilon_0+A\cos(ka)}\right) - \arctan\left(\frac{A\sin(ka)}{\epsilon(k)-\epsilon_0+A\cos(ka)}\right) + \pi\end{aligned}$$ which was also discussed in Ref. [@buks-96-prl]. In contrast to the transmission $|t_{\rm dot}|^2$ of the dot which follows the well-known Breit-Wigner formula for any barrier asymmetry (Fig. \[figure-S7\](b), top panel), the phase shift that occurs across the resonance strongly depends on the relative values of the tunneling rates (Fig. \[figure-S7\](b), bottom panel). The shift is $2\pi$ when the right barrier is the most transparent because the dot belongs to the cavity (Fig. \[figure-S7\](c), bottom panel), it is only $\pi$ for the symmetric case, and it reduces to zero when the right barrier is the least transparent because the dot is outside the cavity (Fig. \[figure-S7\](c), top panel). As a consequence, the reflection phase measured by scanning gate interferometry is between zero and twice the transmission phase and should be interpreted carefully. This simple model also shows that only smooth phase shifts are expected for non-interacting electrons in contrast to the abrupt phase jumps observed in our experiment. To discuss the theoretical predictions for QDs in the Kondo regime, the solution of the Anderson model with finite Coulomb interaction $U$ in the dot, calculated in Ref. [@gerland-00-prl], is reported schematically in Fig. \[figure-S8\] and discussed below.
Expected transmission phase for a Kondo quantum dot
===================================================
When a QD is in a Coulomb blocked region with an odd number of electrons, the Kondo effect gives rise to an enhanced transmission amplitude that reaches $2e^2/h$ at low enough temperature and for symmetric couplings to the leads (Fig. \[figure-S8\](a), top panel, red line). The transmission phase of the electrons at low bias is locked at $\pi/2$ in this Kondo valley (Fig. \[figure-S8\](a), bottom panel, red line), according to the Friedel sum rule which relates the phase to the occupation probability per spin (this quantity equals $1/2$ in the Kondo regime due to screening of the unpaired spin by the surrounding conduction electrons). At finite bias voltage, the problem is more complex due to decoherence of the Kondo correlations between the two reservoirs. Assuming that the QD remains in equilibrium, the spectral properties of the transmission coefficient have been discussed in Ref. [@gerland-00-prl]. In the Kondo regime, the transmission amplitude shows a sharp resonance at the Fermi level, whose width is given by the Kondo temperature $T_K$, in addition to the two peaks of width $\Gamma$, separated by the Coulomb charging energy $U$, and corresponding to the spin-degenerate single-particle energy levels (Fig. \[figure-S8\](b), top panel). The transmission phase exhibits a smooth shift from 0 to $\pi$ when the energy is swept across the single-particle energy levels and across the Kondo resonance at zero bias (Fig. \[figure-S8\](b), bottom panel). Between each of these peaks, the phase shows a sharp lapse by $-\pi$, resulting in the “Kondo double phase lapse” predicted in Ref. [@gerland-00-prl] at low enough temperature. In absence of Kondo correlations, the phase shows only one lapse by $-\pi$ between the Coulomb blockade peaks corresponding to the spin-degenerate single-particle energy levels (Fig. \[figure-S8\](a), blue lines).
![Red lines show the amplitude and phase of the transmission coefficient through a QD in the Kondo regime at zero temperature according to Ref. [@gerland-00-prl] versus (a) gate voltage and (b) bias voltage at fixed gate voltage $U/2$. Blue lines in (a) show the Coulomb blockade regime without Kondo correlations.[]{data-label="figure-S8"}](figure-S8.pdf){width="10cm"}
[2]{} T. A. Costi, A. C. Hewson, and V. Zlatic, J. Phys.: Condens. Matter **6**, 2519 (1994).
D. Goldhaber-Gordon, J. Göres, M. A. Kastner, H. Shtrikman, D. Mahalu, and U. Meirav, Phys. Rev. Lett. **81**, 5225 (1998).
W. G. van der Wiel, S. De Franceschi, T. Fujisawa, J. M. Elzerman, S. Tarucha, and L. P. Kouwenhoven, Science **289**, 2105 (2000).
J. Nyg[å]{}rd, D. H. Cobden, and P. E. Lindelof, Nature **408**, 342 (2000).
T. K. Ng and P. A. Lee, Phys. Rev. Lett. **61**, 1768 (1988).
Y. Ren, W. W. Yu, S. M. Frolov, J. A. Folk, and W. Wegscheider, Phys. Rev. B **82**, 045313 (2010).
S. M. Cronenwett, H. J. Lynch, D. Goldhaber-Gordon, L. P. Kouwenhoven, C. M. Marcus, K. Hirose, N. S. Wingreen, and V. Umansky, Phys. Rev. Lett. **88**, 226805 (2002).
S. Sarkozy, F. Sfigakis, K. Das Gupta, I. Farrer, D. A. Ritchie, G. A. C. Jones, and M. Pepper, Phys. Rev. B **79**, 161307(R) (2009).
M. Pletyukhov and H. Schoeller, Phys. Rev. Lett. **108**, 260601 (2012).
A. V. Kretinin, H. Shtrikman, and D. Mahalu, Phys. Rev. B **85**, 201301(R) (2012).
O. Klochan, A. P. Micolich, A. R. Hamilton, D. Reuter, A. D. Wieck, F. Reininghaus, M. Pletyukhov, and H. Schoeller, Phys. Rev. B **87**, 201104(R) (2013).
T. A. Costi, L. Bergqvist, A. Weichselbaum, J. von Delft, T. Micklitz, A. Rosch, P. Mavropoulos, P. H. Dederichs, F. Mallet, L. Saminadayar, and C. Bäuerle, Phys. Rev. Lett. **102**, 056802 (2009).
B. Brun, F. Martins, S. Faniel, B. Hackens, G. Bachelier, A. Cavanna, C. Ulysse, A. Ouerghi, U. Gennser, D. Mailly, S. Huant, V. Bayot, M. Sanquer, and H. Sellier, Nat. Commun. **5**, 4290 (2014).
M. A. Topinka, B. J. LeRoy, R. M. Westervelt, S. E. J. Shaw, R. Fleischmann, E. J. Heller, K. D. Maranowski, and A. C. Gossard, Nature **410**, 183 (2001).
E. Buks, R. Schuster, M. Heiblum, D. Mahalu, V. Umansky, and H. Shtrikman, Phys. Rev. Lett. **77**, 4664 (1996).
U. Gerland, J. von Delft, T. A. Costi, and Y. Oreg, Phys. Rev. Lett. **84**, 3710 (2000).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Radu Ioan Boţ [^1]'
- 'Christopher Hendrich [^2]'
title: 'Convergence analysis for a primal-dual monotone + skew splitting algorithm with applications to total variation minimization'
---
[**Abstract.**]{} In this paper we investigate the convergence behavior of a primal-dual splitting method for solving monotone inclusions involving mixtures of composite, Lipschitzian and parallel sum type operators proposed by Combettes and Pesquet in [@ComPes12]. Firstly, in the particular case of convex minimization problems, we derive convergence rates for the sequence of objective function values by making use of conjugate duality techniques. Secondly, we propose for the general monotone inclusion problem two new schemes which accelerate the sequences of primal and/or dual iterates, provided strong monotonicity assumptions for some of the involved operators are fulfilled. Finally, we apply the theoretical achievements in the context of different types of image restoration problems solved via total variation regularization.
[**Keywords.**]{} splitting method, Fenchel duality, convergence statements, image processing
[**AMS subject classification.**]{} 90C25, 90C46, 47A52
Introduction and preliminaries {#sectionIntro}
==============================
The last few years have shown a rising interest in solving structured nondifferentiable convex optimization problems within the framework of the theory of conjugate functions. Applications in fields like signal and image processing, location theory and supervised machine learning motivate these efforts.
In this article we investigate and improve the convergence behavior of the primal-dual monotone + skew splitting method for solving monotone inclusions which was proposed by Combettes and Pesquet in [@ComPes12], itself being an extension of the algorithmic scheme from [@BriCom11] obtained by allowing also Lipschitzian monotone operators and parallel sums in the problem formulation. In the mentioned works, by means of a product space approach, the problem is reduced to the one of finding the zeros of the sum of a Lipschitzian monotone operator with a maximally monotone operator. The latter is solved by using an error-tolerant version of Tseng’s algorithm which has forward-backward-forward characteristics and allows to access the monotone Lipschitzian operators via explicit forward steps, while set-valued maximally monotone operators are processed via their resolvents. A notable advantage of this method is given by both its highly parallelizable character, most of its steps could be executed independently, and by the fact that allows to process maximal monotone operators and linear bounded operators separately, whenever they occur in the form of precompositions in the problem formulation.
Before coming to the description of the problem formulation and of the algorithm from [@ComPes12], we introduce some preliminary notions and results which are needed throughout the paper.
We are considering the real Hilbert spaces ${\ensuremath{\mathcal{H}}}$ and ${\ensuremath{\mathcal{G}}}_i$, $i=1,\ldots,m$, endowed with the *inner product* $\left\langle \cdot ,\cdot \right\rangle$ and associated *norm* $\left\| \cdot \right\| = \sqrt{\left\langle \cdot, \cdot \right\rangle}$, for which we use the same notation, respectively, as there is no risk of confusion. The symbols $\rightharpoonup$ and $\rightarrow$ denote weak and strong convergence, respectively. By ${\ensuremath{\mathbb{R}}}_{++}$ we denote the set of strictly positive real numbers, while the *indicator function* of a set $C \subseteq {\ensuremath{\mathcal{H}}}$ is $\delta_C : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}:= {\ensuremath{\mathbb{R}}}\cup \left\{ \pm \infty \right\}$, defined by $\delta_C(x) = 0$ for $x \in C$ and $\delta_C(x) = +\infty$, otherwise. For a function $f: {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$ we denote by $\operatorname*{dom}f := \left\{ x \in {\ensuremath{\mathcal{H}}}: f(x) < +\infty \right\}$ its *effective domain* and call $f$ *proper* if $\operatorname*{dom}f \neq \varnothing$ and $f(x)>-\infty$ for all $x \in {\ensuremath{\mathcal{H}}}$. Let be $$\Gamma({\ensuremath{\mathcal{H}}}) := \{f: {\ensuremath{\mathcal{H}}}\rightarrow \overline {\ensuremath{\mathbb{R}}}: f \ \mbox{is proper, convex and lower semicontinuous}\}.$$ The *conjugate function* of $f$ is $f^*:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, $f^*(p)=\sup{\left\{ \left\langle p,x \right\rangle -f(x) : x\in{\ensuremath{\mathcal{H}}}\right\}}$ for all $p \in {\ensuremath{\mathcal{H}}}$ and, if $f \in \Gamma({\ensuremath{\mathcal{H}}})$, then $f^* \in \Gamma({\ensuremath{\mathcal{H}}})$, as well. The *(convex) subdifferential* of $f: {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$ at $x \in {\ensuremath{\mathcal{H}}}$ is the set $\partial f(x) = \{p \in {\ensuremath{\mathcal{H}}}: f(y) - f(x) \geq \left\langle p,y-x \right\rangle \ \forall y \in {\ensuremath{\mathcal{H}}}\}$, if $f(x) \in {\ensuremath{\mathbb{R}}}$, and is taken to be the empty set, otherwise. For a linear continuous operator $L_i: {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{G}}}_i$, the operator $L_i^*: {\ensuremath{\mathcal{G}}}_i \rightarrow {\ensuremath{\mathcal{H}}}$, defined via ${\left\langle}L_ix,y {\right\rangle}= {\left\langle}x,L_i^*y {\right\rangle}$ for all $x \in {\ensuremath{\mathcal{H}}}$ and all $y \in {\ensuremath{\mathcal{G}}}_i$, denotes its *adjoint operator*, for $i\in \{1,\ldots,m\}$.
Having two functions $f,\,g : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, their *infimal convolution* is defined by $f {\ensuremath{\mbox{\small$\,\square\,$}}}g : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, $(f {\ensuremath{\mbox{\small$\,\square\,$}}}g) (x) = \inf_{y \in {\ensuremath{\mathcal{H}}}}\left\{ f(y) + g(x-y) \right\}$ for all $x \in {\ensuremath{\mathcal{H}}}$, being a convex function when $f$ and $g$ are convex.
Let $M:{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ be a set-valued operator. We denote by $\operatorname*{gra}M = \{ (x,u) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\mathcal{H}}}: u \in Mx\}$ its *graph* and by $\operatorname*{ran}M =\{u \in {\ensuremath{\mathcal{H}}}: \exists x \in {\ensuremath{\mathcal{H}}},\ u\in Mx\}$ its *range*. The *inverse operator* of $M$ is defined as $M^{-1} :{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}, M^{-1}(u) = \{x \in H: u \in Mx\}$. The operator $M$ is called *monotone* if ${\left\langle}x-y,u-v {\right\rangle}\geq 0$ for all $(x,u),\,(y,v) \in \operatorname*{gra}M$ and it is called *maximally monotone* if there exists no monotone operator $M':{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ such that $\operatorname*{gra}M'$ properly contains $\operatorname*{gra}M$ . The operator $M$ is called $\rho$-strongly monotone, for $\rho \in {\ensuremath{\mathbb{R}}}_{++}$, if $M-\rho {\ensuremath{\text{Id}}}$ is monotone, i.e. ${\left\langle}x-y,u-v {\right\rangle}\geq \rho \| x-y \|^2$ for all $(x,u),\,(y,v) \in \operatorname*{gra}M$, where ${\ensuremath{\text{Id}}}$ denotes the identity on ${\ensuremath{\mathcal{H}}}$. The operator $M : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{H}}}$ is called $\nu$-Lipschitzian for $\nu \in {\ensuremath{\mathbb{R}}}_{++}$ if it is single-valued and it fulfills $\|Mx-My\| \leq \nu \|x-y\|$ for all $x,y \in {\ensuremath{\mathcal{H}}}$.
The resolvent of a set-valued operator $M:{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ is $J_M:{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}, J_M = \left( {\ensuremath{\text{Id}}}+ M \right)^{-1}$. When $M$ is maximally monotone, the resolvent is a single-valued, $1$-Lipschitzian and maximal monotone operator. Moreover, when $f \in \Gamma({\ensuremath{\mathcal{H}}})$ and $\gamma \in {\ensuremath{\mathbb{R}}}_{++}$, $\partial (\gamma f)$ is maximally monotone (cf. [@Zalinescu02 Theorem 3.2.8]) and it holds $J_{\gamma \partial f} = \left({\ensuremath{\text{Id}}}+ \gamma \partial f \right)^{-1} = {\ensuremath{\text{Prox}}}_{\gamma f}$. Here, ${\ensuremath{\text{Prox}}}_{\gamma f}(x)$ denotes the *proximal point* of parameter $\gamma$ of $f$ at $x \in {\ensuremath{\mathcal{H}}}$ and it represents the unique optimal solution of the optimization problem $$\label{prox-def}
\inf_{y\in {\ensuremath{\mathcal{H}}}}\left \{f(y)+\frac{1}{2\gamma}\|y-x\|^2\right\}.$$ For a nonempty, convex and closed set $C \subseteq {\ensuremath{\mathcal{H}}}$ and $\gamma \in {\ensuremath{\mathbb{R}}}_{++}$ we have ${\ensuremath{\text{Prox}}}_{\gamma \delta_C} = \mathcal{P}_C$, where $\mathcal{P}_C : {\ensuremath{\mathcal{H}}}\rightarrow C$, $\mathcal{P}_C(x) = \operatorname*{arg\,min}_{z\in C}\left\| x-z \right\|$, denotes the *projection operator* on $C$.
Finally, the *parallel sum* of two set-valued operators $M_1, M_2: {\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ is defined as $$M_1 {\ensuremath{\mbox{\small$\,\square\,$}}}M_2 : {\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}, M_1 {\ensuremath{\mbox{\small$\,\square\,$}}}M_2 = \left(M_1^{-1} + M_2^{-1}\right)^{-1}.$$ We can formulate now the monotone inclusion problem which we investigate in this paper (see [@ComPes12]).
\[opt-problem-inclusion-full\] Consider the real Hilbert spaces ${\ensuremath{\mathcal{H}}}$ and ${\ensuremath{\mathcal{G}}}_i, i=1,...,m,$ $A:{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ a maximally monotone operator and $C:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{H}}}$ a monotone and $\mu$-Lipschitzian operator for some $\mu \in {\ensuremath{\mathbb{R}}}_{++}$. Furthermore, let $z \in {\ensuremath{\mathcal{H}}}$ and for every $i\in\{1,\ldots,m\}$, let $r_i \in {\ensuremath{\mathcal{G}}}_i$, let $B_i : {\ensuremath{\mathcal{G}}}_i \rightarrow 2^{{\ensuremath{\mathcal{G}}}_i}$ be maximally monotone operators, let $D_i:{\ensuremath{\mathcal{G}}}_i \rightarrow 2^{{\ensuremath{\mathcal{G}}}_i}$ be monotone operators such that $D_{i}^{-1}$ is $\nu_i$-Lipschitzian for some $\nu_i \in {\ensuremath{\mathbb{R}}}_{++}$, and let $L_i : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{G}}}_i$ be a nonzero linear continuous operator. The problem is to solve the primal inclusion $$\begin{aligned}
\label{opt-problem-primal-inclusion-full}
\text{find }{\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}\text{ such that } z \in A{\ensuremath{\overline{x}}}+ \sum_{i=1}^m L_i^* \left( (B_i{\ensuremath{\mbox{\small$\,\square\,$}}}D_i)(L_i {\ensuremath{\overline{x}}}-r_i) \right) + C{\ensuremath{\overline{x}}},\end{aligned}$$ together with the dual inclusion $$\begin{aligned}
\label{opt-problem-dual-inclusion-full}
\text{find }{\ensuremath{\overline{v}}}_1 \in {\ensuremath{\mathcal{G}}}_1,\ldots,{\ensuremath{\overline{v}}}_m \in {\ensuremath{\mathcal{G}}}_m \text{ such that }(\exists x\in{\ensuremath{\mathcal{H}}})\left\{
\begin{array}{l}
z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in Ax + Cx \\
\!\!{\ensuremath{\overline{v}}}_i \!\in \!(B_i {\ensuremath{\mbox{\small$\,\square\,$}}}D_i)(L_ix-r_i), \!i=1,\ldots,m.
\end{array}
\right.\end{aligned}$$
Throughout this paper we denote by ${\ensuremath{\boldsymbol{\mathcal{G}}}}:={\ensuremath{\mathcal{G}}}_1 \times... \times {\ensuremath{\mathcal{G}}}_m$ the Hilbert space equipped with the inner product $$\langle (p_1,\ldots,p_m), (q_1,\ldots,q_m) \rangle = \sum_{i=1}^m \langle p_i,q_i \rangle \ \forall (p_1,\ldots,p_m) \ \forall (q_1,\ldots,q_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$$ and the associated norm $\|(p_1,\ldots,p_m)\| = \sqrt{\sum_{i=1}^m \| p_i \|^2}$ for all $(p_1,\ldots,p_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$. We introduce also the nonzero linear continuous operator ${\ensuremath{\boldsymbol{L}}}:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\boldsymbol{\mathcal{G}}}}$, ${\ensuremath{\boldsymbol{L}}}x = (L_1 x,\ldots,L_m x)$, its adjoint being ${\ensuremath{\boldsymbol{L}}}^* : {\ensuremath{\boldsymbol{\mathcal{G}}}}\rightarrow {\ensuremath{\mathcal{H}}}$, ${\ensuremath{\boldsymbol{L}}}^* {\ensuremath{\boldsymbol{v}}}= \sum_{i=1}^m L_i^* v_i$.
We say that $({\ensuremath{\overline{x}}}, {\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\], if $$\label{operator-proof-conditions-full}
z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in A{\ensuremath{\overline{x}}}+ C{\ensuremath{\overline{x}}}\ \mbox{and} \ {\ensuremath{\overline{v}}}_i \in (B_i {\ensuremath{\mbox{\small$\,\square\,$}}}D_i)(L_i{\ensuremath{\overline{x}}}-r_i), \,i=1,\ldots,m.$$ If $({\ensuremath{\overline{x}}}, {\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\], then ${\ensuremath{\overline{x}}}$ is a solution to and $({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$ is a solution to . Notice also that $$\begin{aligned}
{\ensuremath{\overline{x}}}\text{ solves }\eqref{opt-problem-primal-inclusion-full} & \Leftrightarrow z - \sum_{i=1}^m L_i^* (B_i{\ensuremath{\mbox{\small$\,\square\,$}}}D_i)(L_i {\ensuremath{\overline{x}}}-r_i) \in A{\ensuremath{\overline{x}}}+ C{\ensuremath{\overline{x}}}\Leftrightarrow\\
\exists\, {\ensuremath{\overline{v}}}_1\in{\ensuremath{\mathcal{G}}}_1,\ldots,{\ensuremath{\overline{v}}}_m\in{\ensuremath{\mathcal{G}}}_m \ & \mbox{such that} \ \left\{
\begin{array}{l} z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in A{\ensuremath{\overline{x}}}+ C{\ensuremath{\overline{x}}}, \\ {\ensuremath{\overline{v}}}_i \in (B_i{\ensuremath{\mbox{\small$\,\square\,$}}}D_i)(L_i {\ensuremath{\overline{x}}}-r_i) , \ i=1,\ldots,m. \end{array}\right.\end{aligned}$$ Thus, if ${\ensuremath{\overline{x}}}$ is a solution to , then there exists $({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ such that $({\ensuremath{\overline{x}}}, {\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\] and if $({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$ is a solution to , then there exists ${\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}$ such that $({\ensuremath{\overline{x}}}, {\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\].
The next result provides the error-free variant of the primal-dual algorithm in [@ComPes12] and the corresponding convergence statements, as given in [@ComPes12 Theorem 3.1].
\[theorem-inclusion-prelim\] For Problem \[opt-problem-inclusion-full\] suppose that $$z \in \operatorname*{ran}\left( A + \sum_{i=1}^m L_i^*\left( B_i {\ensuremath{\mbox{\small$\,\square\,$}}}D_i \right)(L_i\cdot -r_i) + C \right).$$ Let $x_0 \in {\ensuremath{\mathcal{H}}}$ and $(v_{1,0}, \ldots, v_{m,0}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$, set $$\beta = \max\{\mu,\nu_1\ldots,\nu_m\} + \sqrt{\sum_{i=1}^m \|L_i\|^2},$$ choose $\varepsilon \in (0,\frac{1}{\beta+1})$ and $(\gamma_n)_{n\geq 0}$ a sequence in $\left[\varepsilon, \frac{1-\varepsilon}{\beta}\right]$ and set $$\begin{aligned}
\label{A0}
\left(\forall n\geq 0\right) \ \left\lfloor \begin{array}{l}
p_{1,n} = J_{\gamma_n A}\left(x_n - \gamma_n \left( Cx_n +\sum_{i=1}^mL_i^* v_{i,n} -z\right) \right) \\
\text{For }i=1,\ldots,m \\
\ \left\lfloor \begin{array}{l}
p_{2,i,n} = J_{\gamma_n B_i^{-1}}\left(v_{i,n} +\gamma_n (L_i x_n -D_i^{-1}v_{i,n} -r_i)\right) \\
v_{i,n+1} = \gamma_n L_i( p_{1,n} - x_n) + \gamma_n(D_i^{-1}v_{i,n} - D_i^{-1}p_{2,i,n}) + p_{2,i,n} \\
\end{array} \right.\\
x_{n+1} = \gamma_n \sum_{i=1}^m L_i^*(v_{i,n}-p_{2,i,n}) + \gamma_n(C x_n - C p_{1,n}) +p_{1,n} .
\end{array}
\right.
\end{aligned}$$ Then the following statements are true:
1. $\sum_{n\in{\ensuremath{\mathbb{N}}}}\|x_n-p_{1,n}\|^2 < +\infty$ and $\forall i \in \{1,\ldots,m\}$ $\sum_{n\in{\ensuremath{\mathbb{N}}}}\|v_{i,n}-p_{2,i,n}\|^2 < +\infty$.
2. There exists a primal-dual solution $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$ to Problem \[opt-problem-inclusion-full\] such that the following hold:
1. $x_n \rightharpoonup {\ensuremath{\overline{x}}}$ and $p_{1,n}\rightharpoonup {\ensuremath{\overline{x}}}$.
2. $(\forall i \in \{1,\ldots,m\})\ v_{i,n} \rightharpoonup {\ensuremath{\overline{v}}}_i$ and $p_{2,i,n}\rightharpoonup {\ensuremath{\overline{v}}}_i$.
In this paper we consider first Problem \[opt-problem-inclusion-full\] in its particular formulation as a primal-dual pair of convex minimization problems, approach which relies on the fact that the subdifferential of a proper, convex and lower semicontinuous function is maximally monotone, and show that the convergence rate of the sequence of objective function values on the iterates generated by is of ${\ensuremath{\mathcal{O}}}(\frac{1}{n})$, where $n \in {\ensuremath{\mathbb{N}}}$ is the number of passed iterations. Further, in Section \[sectionOperator\], we provide for the general monotone inclusion problem, as given in Problem \[opt-problem-inclusion-full\], two new acceleration schemes which generate under strong monotonicity assumptions sequences of primal and/or dual iterates converge with improved convergence properties. The feasibility of the proposed methods is explicitly shown in Section \[sectionApp\] by means of numerical experiments in the context of solving image denoising, image deblurring and image inpainting problems via total variation regularization.
One of the iterative schemes to which we compare our algorithms is a primal-dual splitting method for solving highly structured monotone inclusions, as well, and it was provided by Vũ in [@Vu11]. Here, instead of monotone Lipschitzian operators, cocoercive operators were used and, consequently, instead of Tseng’s splitting, the forward-backward splitting method has been used. The primal-dual method due to Chambolle and Pock described in [@ChaPoc11 Algorithm 1] is a particular instance of Vũ’s algorithm.
Convex minimization problems {#sectionFval}
============================
The aim of this section is to provide a rate of convergence for the sequence of the values of the objective function at the iterates generated by the algorithm when solving a convex minimization problem and its conjugate dual. The primal-dual pair under investigation is described in the following.
\[opt-problem:general\] Consider the real Hilbert spaces ${\ensuremath{\mathcal{H}}}$ and ${\ensuremath{\mathcal{G}}}_i, i=1,...,m,$ $f\in \Gamma({\ensuremath{\mathcal{H}}})$ and $h : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathbb{R}}}$ a convex and differentiable function with $\mu$-Lipschitzian gradient for some $\mu \in {\ensuremath{\mathbb{R}}}_{++}$. Furthermore, let $z \in {\ensuremath{\mathcal{H}}}$ and for every $i\in\{1,\ldots,m\}$, let $r_i \in {\ensuremath{\mathcal{G}}}_i$, $g_i, l_i \in \Gamma({\ensuremath{\mathcal{G}}}_i)$ such that $l_i$ is $\nu_i^{-1}$-strongly convex for some $\nu_i \in {\ensuremath{\mathbb{R}}}_{++}$, and let $L_i : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{G}}}_i$ be a nonzero linear continuous operator. We consider the convex minimization problem $$\begin{aligned}
\label{opt-problem:general-primal}
(P) \quad \inf_{x \in {\ensuremath{\mathcal{H}}}}{\left\{f(x)+\sum_{i=1}^m (g_i {\ensuremath{\mbox{\small$\,\square\,$}}}l_i)(L_ix-r_i) +h(x) -{\left\langle}x,z{\right\rangle}\right\}}\end{aligned}$$ and its dual problem $$\begin{aligned}
\label{opt-problem:general-dual}
(D) \quad \sup_{(v_i,\ldots,v_m) \in {\ensuremath{\mathcal{G}}}_1\times\ldots\times{\ensuremath{\mathcal{G}}}_m}{\left\{-\left( f^*{\ensuremath{\mbox{\small$\,\square\,$}}}h^*\right)\left( z - \sum_{i=1}^m L_i^*v_i\right) - \sum_{i=1}^m \left( g_i^*(v_i) + l_i^*(v_i) + {\left\langle}v_i,r_i {\right\rangle}\right) \right\} }.\end{aligned}$$
In order to investigate the primal-dual pair - in the context of Problem \[opt-problem:general\], one has to take $$A = \partial f, \ C=\nabla h, \text{ and, for }i=1,\ldots,m, \ B_i=\partial g_i \text{ and } D_i = \partial l_i.$$ Then $A$ and $B_i, i=1,...,m$ are maximal monotone, $C$ is monotone, by [@BauschkeCombettes11 Proposition 17.10], and $D_i^{-1} = \nabla l_i^*$ is monotone and $\nu_i$-Lipschitz continuous for $i=1,\ldots,m$, according to [@BauschkeCombettes11 Proposition 17.10, Theorem 18.15 and Corollary 16.24]. One can easily see that (see, for instance, [@ComPes12 Theorem 4.2]) whenever $({\ensuremath{\overline{x}}}, {\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\], with the above choice of the involved operators, ${\ensuremath{\overline{x}}}$ is an optimal solution to $(P)$, $({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$ is an optimal solution to $(D)$ and for $(P)$-$(D)$ strong duality holds, thus the optimal objective values of the two problems coincide.
The primal-dual pair in Problem \[opt-problem:general\] captures various different types of optimization problems. One such particular instance is formulated as follows and we refer for more examples to [@ComPes12].
In Problem \[opt-problem:general\] take $z=0$, let $l_i:{\ensuremath{\mathcal{G}}}_i \rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, $l_i=\delta_{\{0\}}$ and $r_i=0$ for $i=1,\ldots,m$, and set $h:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathbb{R}}}$, $h(x)=0$ for all $x\in {\ensuremath{\mathcal{H}}}$. Then reduces to $$(P) \quad \inf_{x \in {\ensuremath{\mathcal{H}}}}{\left\{f(x)+\sum_{i=1}^m g_i(L_ix) \right\}},$$ while the dual problem becomes $$(D) \quad \sup_{(v_i,\ldots,v_m) \in {\ensuremath{\mathcal{G}}}_1\times\ldots\times{\ensuremath{\mathcal{G}}}_m}{\left\{-f^*\left(- \sum_{i=1}^m L_i^*v_i\right) - \sum_{i=1}^m g_i^*(v_i) \right\}}.$$
In order to simplify the upcoming formulations and calculations we introduce the following more compact notations. With respect to Problem \[opt-problem:general\], let ${\ensuremath{F}}:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}},\ {\ensuremath{F}}(x) = f(x) + h(x) -{\left\langle}x,z{\right\rangle}$. Then $\operatorname*{dom}{\ensuremath{F}}= \operatorname*{dom}f$ and its conjugate ${\ensuremath{F}}^* :{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$ is given by ${\ensuremath{F}}^*(p) = (f + h)^*(z+p) = (f^* {\ensuremath{\mbox{\small$\,\square\,$}}}h^*)(z+p)$, since $\operatorname*{dom}h ={\ensuremath{\mathcal{H}}}$. Further, we set $$\begin{aligned}
{\ensuremath{\boldsymbol{v}}}= (v_1,\ldots,v_m), \quad {\ensuremath{\boldsymbol{\overline{v}}}}= ({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m), \quad {\ensuremath{\boldsymbol{p}}}_{2,n} = (p_{2,1,n},\ldots,p_{2,m,n}), \quad {\ensuremath{\boldsymbol{r}}}= (r_1,\ldots,r_m).\end{aligned}$$ We define the function ${\ensuremath{G}}:{\ensuremath{\boldsymbol{\mathcal{G}}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, ${\ensuremath{G}}({\ensuremath{\boldsymbol{y}}}) = \sum_{i=1}^m (g_i{\ensuremath{\mbox{\small$\,\square\,$}}}l_i)(y_i)$ and observe that its conjugate ${\ensuremath{G}}^* :{\ensuremath{\boldsymbol{\mathcal{G}}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$ is given by ${\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) = \sum_{i=1}^m (g_i{\ensuremath{\mbox{\small$\,\square\,$}}}l_i)^*(v_i)=\sum_{i=1}^m(g_i^* + l_i^*)(v_i)$. Notice that, as $l_i^*, i=1,\ldots,m$, has full domain (cf. [@BauschkeCombettes11 Theorem 18.15]), we get $$\begin{aligned}
\operatorname*{dom}{\ensuremath{G}}^* &= (\operatorname*{dom}g_1^* \cap \operatorname*{dom}l_1^*) \times \ldots \times (\operatorname*{dom}g_m^* \cap \operatorname*{dom}l_m^*) = \operatorname*{dom}g_1^* \times \ldots \times \operatorname*{dom}g_m^*,\end{aligned}$$
The primal and the dual optimization problems given in Problem \[opt-problem:general\] can be equivalently represented as $$\begin{aligned}
(P) \quad \inf_{x \in {\ensuremath{\mathcal{H}}}}{\left\{{\ensuremath{F}}(x)+{\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}x -{\ensuremath{\boldsymbol{r}}})\right\}},\end{aligned}$$ and, respectively, $$\begin{aligned}
(D) \quad \sup_{{\ensuremath{\boldsymbol{v}}}\in {\ensuremath{\boldsymbol{\mathcal{G}}}}}{\left\{-{\ensuremath{F}}^*(-{\ensuremath{\boldsymbol{L}}}^* {\ensuremath{\boldsymbol{v}}})-{\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) - {\left\langle}{\ensuremath{\boldsymbol{v}}},{\ensuremath{\boldsymbol{r}}}{\right\rangle}\right\}}.\end{aligned}$$ Then ${\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}$ solves $(P)$, ${\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ solves $(D)$ and for $(P)$-$(D)$ strong duality holds if and only if (cf. [@Bot10; @BotGradWanka09]) $$\begin{aligned}
\label{opt-problem-optimality-condition}
-{\ensuremath{\boldsymbol{L}}}^*{\ensuremath{\boldsymbol{\overline{v}}}}&\in \partial {\ensuremath{F}}({\ensuremath{\overline{x}}}) \ \mbox{and} \ {\ensuremath{\boldsymbol{L}}}{\ensuremath{\overline{x}}}- {\ensuremath{\boldsymbol{r}}}\in \partial {\ensuremath{G}}^*({\ensuremath{\boldsymbol{\overline{v}}}}).\end{aligned}$$ Let us mention also that for ${\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}$ and ${\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ fulfilling it holds $$\left[{\left\langle}{\ensuremath{\boldsymbol{L}}}x-{\ensuremath{\boldsymbol{r}}}, {\ensuremath{\boldsymbol{\overline{v}}}}{\right\rangle}+ {\ensuremath{F}}(x) -{\ensuremath{G}}^*({\ensuremath{\boldsymbol{\overline{v}}}}) \right]- \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}{\ensuremath{\overline{x}}}-{\ensuremath{\boldsymbol{r}}}, {\ensuremath{\boldsymbol{v}}}{\right\rangle}+ {\ensuremath{F}}({\ensuremath{\overline{x}}}) -{\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) \right] \geq 0 \ \forall x\in {\ensuremath{\mathcal{H}}}\ \forall {\ensuremath{\boldsymbol{v}}}\in {\ensuremath{\boldsymbol{\mathcal{G}}}}.$$
For given sets $B_1\subseteq{\ensuremath{\mathcal{H}}}$ and $B_2 \subseteq {\ensuremath{\boldsymbol{\mathcal{G}}}}$ we introduce the so-called *primal-dual gap function* $$\begin{aligned}
\label{primal-dual-gap}
\mathcal{G}_{B_1 \times B_2}(x,{\ensuremath{\boldsymbol{v}}}) &= \sup_{\tilde{{\ensuremath{\boldsymbol{v}}}}\in B_2}\left\{ {\left\langle}{\ensuremath{\boldsymbol{L}}}x-{\ensuremath{\boldsymbol{r}}}, \tilde{{\ensuremath{\boldsymbol{v}}}} {\right\rangle}+ {\ensuremath{F}}(x) -{\ensuremath{G}}^*(\tilde{{\ensuremath{\boldsymbol{v}}}}) \right\} \notag \\
& \quad- \inf_{\tilde{x} \in B_1}{\left\{ {\left\langle}{\ensuremath{\boldsymbol{L}}}\tilde{x}-{\ensuremath{\boldsymbol{r}}}, {\ensuremath{\boldsymbol{v}}}{\right\rangle}+ {\ensuremath{F}}(\tilde{x}) -{\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) \right\}}.\end{aligned}$$ We consider the following algorithm for solving $(P)$-$(D)$, which differs from the one given in Theorem \[theorem-inclusion-prelim\] by the fact that we are asking the sequence $(\gamma_n)_{n\geq 0} \subseteq {\ensuremath{\mathbb{R}}}_{++}$ to be nondecreasing.
\[alg1\] Let $x_0 \in {\ensuremath{\mathcal{H}}}$ and $(v_{1,0}, \ldots, v_{m,0}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$, set $$\beta = \max\{ \mu, \nu_1,\ldots, \nu_m \} + \sqrt{\sum_{i=1}^n\| L_i \|^2},$$ choose $\varepsilon \in \left(0, \frac{1}{\beta +1} \right)$ and $(\gamma_n)_{n\geq 0}$ a nondecreasing sequence in $\left[ \varepsilon, \frac{1-\varepsilon}{\beta}\right]$ and set $$\begin{aligned}
\label{A1}
\left(\forall n\geq 0\right) \ \left\lfloor \begin{array}{l}
p_{1,n} = {\ensuremath{\text{Prox}}}_{\gamma_n f}\left(x_n - \gamma_n \left( \nabla h(x_n) +\sum_{i=1}^mL_i^*v_{i,n} -z\right) \right) \\
\text{For }i=1,\ldots,m \\
\ \left\lfloor \begin{array}{l}
p_{2,i,n} = {\ensuremath{\text{Prox}}}_{\gamma_n g_i^*}\left(v_{i,n} +\gamma_n (L_i x_n - \nabla l_i^*(v_{i,n}) -r_i)\right) \\
v_{i,n+1} = \gamma_n L_i( p_{1,n} - x_n) +\gamma_n(\nabla l_i^*(v_{i,n})-\nabla l_i^*(p_{2,i,n})) + p_{2,i,n}
\end{array} \right.\\
x_{n+1} = \gamma_n \sum_{i=1}^mL_i^*(v_{i,n}-p_{2,i,n}) + \gamma_n(\nabla h(x_n) - \nabla h(p_{1,n})) +p_{1,n}.
\end{array}
\right.
\end{aligned}$$
\[theorem-fvalues\] For Problem \[opt-problem:general\] suppose that $$z \in \operatorname*{ran}\left( \partial f + \sum_{i=1}^m L_i^*\left( \partial g_i {\ensuremath{\mbox{\small$\,\square\,$}}}\partial l_i \right)(L_i\cdot -r_i) + \nabla h \right).$$ Then there exists an optimal solution ${\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}$ to $(P)$ and an optimal solution $({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ to $(D)$, such that the following holds for the sequences generated by Algorithm \[alg1\]:
1. $z-\sum_{i=1}^mL_i^*{\ensuremath{\overline{v}}}_i \in \partial f({\ensuremath{\overline{x}}}) + \nabla h({\ensuremath{\overline{x}}})$ and $L_i {\ensuremath{\overline{x}}}- r_i \in \partial g_i^*({\ensuremath{\overline{v}}}_i) + \nabla l_i^*({\ensuremath{\overline{v}}}_i)$ $\forall i\in\{1,\ldots,m\}$.
2. $x_n \rightharpoonup {\ensuremath{\overline{x}}}$, $p_{1,n} \rightharpoonup {\ensuremath{\overline{x}}}$ and $v_{i,n} \rightharpoonup {\ensuremath{\overline{v}}}_i$, $p_{2,i,n} \rightharpoonup {\ensuremath{\overline{v}}}_i$ $\forall i\in\{1,\ldots,m\}$.
3. For $n\geq 0$ it holds $$\frac{\| x_n - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} + \sum_{i=1}^m\frac{\| v_{i,n} - {\ensuremath{\overline{v}}}_i \|^2}{2\gamma_n} \leq \frac{\| x_0 - {\ensuremath{\overline{x}}}\|^2}{2\gamma_0} + \sum_{i=1}^m\frac{\| v_{i,0} - {\ensuremath{\overline{v}}}_i \|^2}{2\gamma_0}.$$
4. If $B_1\subseteq{\ensuremath{\mathcal{H}}}$ and $B_2\subseteq {\ensuremath{\boldsymbol{\mathcal{G}}}}$ are bounded, then for $x^N:=\frac{1}{N}\sum_{n=0}^{N-1} p_{1,n}$ and $v_i^N:=\frac{1}{N}\sum_{n=0}^{N-1} p_{2,i,n}$, $i=1,\ldots,m$, the primal-dual gap has the upper bound $$\begin{aligned}
\label{fval-global-gap} \mathcal{G}_{B_1 \times B_2}(x^N,v_1^N,\ldots,v_m^N) \leq \frac{C(B_1,B_2)}{N}, \end{aligned}$$ where $$C(B_1,B_2) = \sup_{(x,v_1,\ldots,v_m) \in B_1 \times B_2}\left\{ \frac{\| x_0 - x \|^2}{2\gamma_0} + \sum_{i=1}^m\frac{\| v_{i,0} - v_i \|^2}{2\gamma_0} \right\}.$$
5. The sequence $(x^N,v_1^N,\ldots,v_m^N)$ converges weakly to $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$.
Theorem 4.2 in [@ComPes12] guarantees the existence of an optimal solution ${\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}$ to and of an optimal solution $({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ to such that strong duality holds, $x_n \rightharpoonup {\ensuremath{\overline{x}}}$, $p_{1,n} \rightharpoonup {\ensuremath{\overline{x}}}$, as well as $v_{i,n} \rightharpoonup {\ensuremath{\overline{v}}}_i$ and $p_{2,i,n} \rightharpoonup {\ensuremath{\overline{v}}}_i$ for $i=1,\ldots,m$, when $n$ converges to $+\infty$. Hence *(a)* and *(b)* are true. Thus, the solutions ${\ensuremath{\overline{x}}}$ and ${\ensuremath{\boldsymbol{\overline{v}}}}=({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m)$ fulfill .
Regarding the sequences $(p_{1,n})_{n \geq 0}$ and $(p_{2,i,n})_{n \geq 0}$, $i=1,\ldots,m$, generated in Algorithm \[alg1\] we have for every $n \geq 0$ $$\begin{aligned}
p_{1,n} = \left( {\ensuremath{\text{Id}}}+ \gamma_n \partial f\right)^{-1}&\left(x_n - \gamma_n \left( \nabla h(x_n) +L^*v_n -z\right) \right) \\
&\Leftrightarrow \frac{x_n-p_{1,n}}{\gamma_n} -\nabla h(x_n) - L^*v_n +z \in \partial f(p_{1,n})\end{aligned}$$ and, for $i=1,...,m$, $$\begin{aligned}
p_{2,i,n} = \left({\ensuremath{\text{Id}}}+ \gamma_n \partial g_i^*\right)^{-1}&\left(v_{i,n} +\gamma_n (L_i x_n - \nabla l_i^*(v_{i,n}) -r_i)\right) \\
&\Leftrightarrow \frac{v_{i,n}-p_{2,i,n}}{\gamma_n} + L_i x_n - \nabla l_i^*(v_{i,n}) -r_i \in \partial g_i^*(p_{2,i,n}).\end{aligned}$$ In other words, it holds for every $n \geq 0$ $$\begin{aligned}
\label{sub-ineq-f}
f(x) &\geq f(p_{1,n}) + {\left\langle}\frac{x_n-p_{1,n}}{\gamma_n} -\nabla h(x_n) - L^*v_n +z, x-p_{1,n} {\right\rangle}\, \forall x\in {\ensuremath{\mathcal{H}}}\end{aligned}$$ and, for $i=1,\ldots,m$, $$\begin{aligned}
\label{sub-ineq-g_i^*}
g_i^*(v_i) &\geq g_i^*(p_{2,i,n}) + {\left\langle}\frac{v_{i,n}-p_{2,i,n}}{\gamma_n} + L_i x_n - \nabla l_i^*(v_{i,n}) -r_i, v_i - p_{2,i,n} {\right\rangle}\, \forall v_i \in {\ensuremath{\mathcal{G}}}_i.\end{aligned}$$ In addition to that, using that $h$ and $l_i^*, i=1,...,m$, are convex and differentiable, it holds for every $n \geq 0$ $$\begin{aligned}
\label{sub-ineq-h}
h(x) &\geq h(p_{1,n}) + {\left\langle}\nabla h(p_{1,n}), x-p_{1,n} {\right\rangle}\, \forall x\in {\ensuremath{\mathcal{H}}}\end{aligned}$$ and, for $i=1,\ldots,m$, $$\begin{aligned}
\label{sub-ineq-l_i^*}
l_i^*(v_i) &\geq l_i^*(p_{2,i,n}) + {\left\langle}\nabla l_i^*(p_{2,i,n}), v_i - p_{2,i,n} {\right\rangle}\, \forall v_i \in {\ensuremath{\mathcal{G}}}_i.\end{aligned}$$ Consider arbitrary $x \in {\ensuremath{\mathcal{H}}}$ and ${\ensuremath{\boldsymbol{v}}}=({\ensuremath{\boldsymbol{v}}}_1,\ldots,{\ensuremath{\boldsymbol{v}}}_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$. Since $$\begin{aligned}
{\left\langle}\frac{x_n-p_{1,n}}{\gamma_n}, x-p_{1,n} {\right\rangle}&= \frac{\| x_n-p_{1,n}\|^2}{2\gamma_n} + \frac{\| x-p_{1,n}\|^2}{2\gamma_n} -\frac{\| x_n-x\|^2}{2\gamma_n} \\
{\left\langle}\frac{v_{i,n}-p_{2,i,n}}{\gamma_n} , v_i - p_{2,i,n} {\right\rangle}&= \frac{\| v_{i,n}-p_{2,i,n}\|^2}{2\gamma_n} + \frac{\| v_i-p_{2,i,n}\|^2}{2\gamma_n} -\frac{\| v_{i,n}-v_i\|^2}{2\gamma_n}, i=1,...,m,\end{aligned}$$ we obtain for every $n \geq 0$, by using the more compact notation of the elements in ${\ensuremath{\boldsymbol{\mathcal{G}}}}$ and by summing up the inequalities –, $$\begin{aligned}
\frac{\| x_n-x\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n}- {\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_n} \geq
\frac{\| x_n-p_{1,n}\|^2}{2\gamma_n} + \frac{\| x-p_{1,n}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\gamma_n}+ \frac{\| {\ensuremath{\boldsymbol{v}}}-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\gamma_n} \\
+ \sum_{i=1}^m {\left\langle}L_i x_n +\nabla l_i^*(p_{2,i,n}) - \nabla l_i^*(v_{i,n}) -r_i, v_i - p_{2,i,n} {\right\rangle}- \sum_{i=1}^m (g_i^*+l_i^*)(v_i) +(f+h)(p_{1,n}) \\
+ {\left\langle}\nabla h(p_{1,n})-\nabla h(x_n) - L^*v_n +z, x-p_{1,n} {\right\rangle}- \left[ \sum_{i=1}^m-(g_i^* +l_i^*)(p_{2,i,n}) + (f+h)(x) \right].\end{aligned}$$ Further, using again the update rules in Algorithm \[alg1\] and the equations $$\begin{aligned}
{\left\langle}\frac{p_{1,n}-x_{n+1}}{\gamma_n}, x-p_{1,n} {\right\rangle}&= \frac{\| x_{n+1}-x\|^2}{2\gamma_n} - \frac{\| x_{n+1} -p_{1,n}\|^2}{2\gamma_n} -\frac{\| x-p_{1,n}\|^2}{2\gamma_n}\end{aligned}$$ and, for $i=1,...,m$, $$\begin{aligned}
{\left\langle}\frac{p_{2,i,n}-v_{i,n+1}}{\gamma_n} , v_i - p_{2,i,n} {\right\rangle}&= \frac{\| v_{i,n+1}-v_i\|^2}{2\gamma_n} - \frac{\| v_{i,n+1}-p_{2,i,n}\|^2}{2\gamma_n} -\frac{\| v_i-p_{2,i,n}\|^2}{2\gamma_n},\end{aligned}$$ we obtain for every $n \geq 0$ $$\begin{aligned}
\label{ineq-fval1}
\frac{\| x_n-x\|^2}{2\gamma_n} &+ \frac{\| {\ensuremath{\boldsymbol{v}}}_{n}- {\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_n} \geq \frac{\|x_{n+1}-x \|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_n}+\frac{\| x_n-p_{1,n}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\gamma_n} \notag \\
& -\frac{\| x_{n+1}-p_{1,n} \|^2}{2\gamma_n} - \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} -{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\gamma_n} + \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}p_{1,n} -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{v}}}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) + {\ensuremath{F}}(p_{1,n})\right] \notag \\
&- \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}x -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{p}}}_{2,n}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{p}}}_{2,n}) + {\ensuremath{F}}(x)\right].\end{aligned}$$ Further, we equip the Hilbert space ${\ensuremath{\boldsymbol{\mathcal{H}}}}={\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$ with the inner product $$\label{inprodfH}
\langle (y,{\ensuremath{\boldsymbol{p}}}),(z,{\ensuremath{\boldsymbol{q}}})\rangle = \langle y,z\rangle + \langle {\ensuremath{\boldsymbol{p}}},{\ensuremath{\boldsymbol{q}}}\rangle \ \forall (y,{\ensuremath{\boldsymbol{p}}}), (z,{\ensuremath{\boldsymbol{q}}}) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$$ and the associated norm $\|(y,{\ensuremath{\boldsymbol{p}}})\| = \sqrt{\|y\|^2 + \|{\ensuremath{\boldsymbol{p}}}\|^2}$ for every $(y,{\ensuremath{\boldsymbol{p}}}) \in {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$. For every $n \geq 0$ it holds $$\begin{aligned}
\frac{\| x_{n+1}-p_{1,n} \|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} -{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\gamma_n}
= \frac{\| (x_{n+1},{\ensuremath{\boldsymbol{v}}}_{n+1})-(p_{1,n},{\ensuremath{\boldsymbol{p}}}_{2,n}) \|^2}{2\gamma_n}\end{aligned}$$ and, consequently, by making use of the Lipschitz continuity of $\nabla h$ and $\nabla l_i^*$, $i=1,\ldots,m$, it shows that [$$\begin{aligned}
&\| (x_{n+1},{\ensuremath{\boldsymbol{v}}}_{n+1})-(p_{1,n},{\ensuremath{\boldsymbol{p}}}_{2,n}) \| \notag\\
&= \gamma_n \| ( {\ensuremath{\boldsymbol{L}}}^*({\ensuremath{\boldsymbol{v}}}_n-{\ensuremath{\boldsymbol{p}}}_{2,n}) , L_1( p_{1,n} - x_n), \ldots, L_m(p_{1,n}-x_n) ) \notag\\
& \qquad \quad + (\nabla h(x_n) - \nabla h(p_{1,n}), \nabla l_1^*(v_{1,n})-\nabla l_1^*(p_{2,1,n}), \ldots, \nabla l_m^*(v_{m,n})-\nabla l_1^*(p_{2,m,n}))\| \notag\\
&\leq \gamma_n \| ( {\ensuremath{\boldsymbol{L}}}^*({\ensuremath{\boldsymbol{v}}}_n-{\ensuremath{\boldsymbol{p}}}_{2,n}) , L_1( p_{1,n} - x_n), \ldots, L_m(p_{1,n}-x_n) ) \| \notag\\
& \quad + \gamma_n \|(\nabla h(x_n) - \nabla h(p_{1,n}), \nabla l_1^*(v_{1,n})-\nabla l_1^*(p_{2,1,n}), \ldots, \nabla l_m^*(v_{m,n})-\nabla l_1^*(p_{2,m,n}))\| \notag\\
&= \gamma_n \sqrt{ \left\| \sum_{i=1}^m L_i^*(v_{i,n}-p_{2,i,n}) \right\|^2 + \sum_{i=1}^m \left\| L_i( p_{1,n} - x_n) \right\|^2} \notag\\
& \quad + \gamma_n \sqrt{ \| \nabla h(x_n) - \nabla h(p_{1,n}) \|^2 + \sum_{i=1}^m \| \nabla l_i^*(v_{i,n})-\nabla l_i^*(p_{2,i,n}) \|^2} \notag\\
&\leq \gamma_n \sqrt{ \left( \sum_{i=1}^m \| L_i\|^2 \right) \sum_{i=1}^m \left\| v_{i,n}-p_{2,i,n} \right\|^2 + \left( \sum_{i=1}^m \| L_i \|^2 \right) \left\| p_{1,n} - x_n \right\|^2} \notag\\
& \quad + \gamma_n \sqrt{ \mu^2 \| x_n - p_{1,n} \|^2 + \sum_{i=1}^m \nu_i^2 \| v_{i,n}-p_{2,i,n} \|^2} \notag\\
\label{ineq-lipschitz-continuity}
&\leq \gamma_n \left( \sqrt{\sum_{i=1}^m \| L_i \|^2} + \max\{ \mu, \nu_1,\ldots, \nu_m \} \right) \| (x_n,{\ensuremath{\boldsymbol{v}}}_n) - (p_{1,n},{\ensuremath{\boldsymbol{p}}}_{2,n}) \|.\end{aligned}$$]{}
Hence, by taking into consideration the way in which $(\gamma_n)_{n\geq 0}$ is chosen, we have for every $n \geq 0$ $$\begin{aligned}
&\frac{1}{2\gamma_n} \left[ \| x_n-p_{1,n}\|^2 + \| {\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2 -\| x_{n+1}-p_{1,n} \|^2 - \| {\ensuremath{\boldsymbol{v}}}_{n+1} -{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2 \right] \\
&\geq \frac{1}{2\gamma_n} \left(1-\gamma_n^2 \left( \sqrt{\sum_{i=1}^m \| L_i \|^2} + \max\{ \mu, \nu_1,\ldots, \nu_m \} \right)^2\right) \| (x,{\ensuremath{\boldsymbol{v}}}_n) - (p_{1,n},{\ensuremath{\boldsymbol{p}}}_{2,n}) \|^2 \geq 0.\end{aligned}$$ and, consequently, reduces to $$\begin{aligned}
\frac{\| x_n-x\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n}- {\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_n} &\geq \frac{\gamma_{n+1}}{\gamma_n}\frac{\|x_{n+1}-x \|^2}{2\gamma_{n+1}} + \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}p_{1,n} -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{v}}}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) + {\ensuremath{F}}(p_{1,n})\right] \\ &\quad + \frac{\gamma_{n+1}}{\gamma_n}\frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_{n+1}} - \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}x -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{p}}}_{2,n}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{p}}}_{2,n}) + {\ensuremath{F}}(x)\right].\end{aligned}$$ Let $N \geq 1$ be an arbitrary natural number. Summing the above inequality up from $n=0$ to $N-1$ and using the fact that $(\gamma_n)_{n\geq 0}$ is nondecreasing, it follows that $$\begin{aligned}
\label{ineq-sum-fval}
\frac{\| x_0-x\|^2}{2\gamma_0} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{0}- {\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_0}
&\geq \frac{\|x_{N}-x \|^2}{2\gamma_{N}} + \sum_{n=0}^{N-1}\left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}p_{1,n} -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{v}}}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) + {\ensuremath{F}}(p_{1,n})\right] \notag\\
&\quad + \frac{\| {\ensuremath{\boldsymbol{v}}}_{N}-{\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_{N}}- \sum_{n=0}^{N-1}\left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}x -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{p}}}_{2,n}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{p}}}_{2,n}) + {\ensuremath{F}}(x)\right].\end{aligned}$$
Replacing $x={\ensuremath{\overline{x}}}$ and ${\ensuremath{\boldsymbol{v}}}= {\ensuremath{\boldsymbol{\overline{v}}}}$ in the above estimate, since they fulfill , we obtain $$\sum_{n=0}^{N-1}\left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}p_{1,n} -{\ensuremath{\boldsymbol{r}}},\boldsymbol{{\ensuremath{\overline{v}}}} {\right\rangle}- {\ensuremath{G}}^*(\boldsymbol{{\ensuremath{\overline{v}}}}) + {\ensuremath{F}}(p_{1,n})\right] - \sum_{n=0}^{N-1}\left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}{\ensuremath{\overline{x}}}-{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{p}}}_{2,n}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{p}}}_{2,n}) + {\ensuremath{F}}({\ensuremath{\overline{x}}})\right] \geq 0.$$ Consequently, $$\frac{\| x_0-{\ensuremath{\overline{x}}}\|^2}{2\gamma_0} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{0}- \boldsymbol{{\ensuremath{\overline{v}}}}\|^2}{2\gamma_0} \geq \frac{\|x_{N}-{\ensuremath{\overline{x}}}\|^2}{2\gamma_{N}}+\frac{\| {\ensuremath{\boldsymbol{v}}}_{N}-\boldsymbol{{\ensuremath{\overline{v}}}}\|^2}{2\gamma_{N}}$$ and statement *(c)* follows. On the other hand, dividing by $N$, using the convexity of ${\ensuremath{F}}$ and ${\ensuremath{G}}^*$, and denoting $x^N:=\frac{1}{N}\sum_{n=0}^{N-1} p_{1,n}$ and $v_i^N:=\frac{1}{N}\sum_{n=0}^{N-1} p_{2,i,n}$, $i=1,\ldots,m$, we obtain $$\begin{aligned}
\frac{1}{N}\left(\frac{\| x_0-x\|^2}{2\gamma_0} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{0}- {\ensuremath{\boldsymbol{v}}}\|^2}{2\gamma_0} \right)
&\geq \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}x^N -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{v}}}{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}) + {\ensuremath{F}}(x^N)\right] \\ &\quad - \left[ {\left\langle}{\ensuremath{\boldsymbol{L}}}x -{\ensuremath{\boldsymbol{r}}},{\ensuremath{\boldsymbol{v}}}^N{\right\rangle}- {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}^N) + {\ensuremath{F}}(x)\right], \notag\end{aligned}$$ which shows when passing to the supremum over $x\in B_1$ and ${\ensuremath{\boldsymbol{v}}}\in B_2$. In this way statement *(d)* is verified. The weak convergence of $(x^N,{\ensuremath{\boldsymbol{v}}}^N)$ to $({\ensuremath{\overline{x}}},\boldsymbol{{\ensuremath{\overline{v}}}})$ when $N$ converges to $+\infty$ is an easy consequence of the Stolz–Ces[à]{}ro Theorem, fact which shows *(e)*.
\[remark1\] In the situation when the functions $g_i$ are Lipschitz continuous on ${\ensuremath{\mathcal{G}}}_i, i=1,...,m,$ inequality provides for the sequence of the values of the objective of $(P)$ taken at $(x^N)_{N \geq 1}$ a convergence rate of ${\ensuremath{\mathcal{O}}}(\frac{1}{N})$, namely, it holds $$\begin{aligned}
\label{ineq-primal-convergence}
{\ensuremath{F}}(x^N) + {\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}x^N -{\ensuremath{\boldsymbol{r}}}) - {\ensuremath{F}}({\ensuremath{\overline{x}}}) -{\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}{\ensuremath{\overline{x}}}-{\ensuremath{\boldsymbol{r}}}) \leq \frac{C(B_1,B_2)}{N} \ \forall N \geq 1.\end{aligned}$$ Indeed, due to statement *(b)* of the previous theorem, the sequence $(p_{1,n})_{n\geq 0} \subseteq {\ensuremath{\mathcal{H}}}$ is bounded and one can take $B_1 \subset {\ensuremath{\mathcal{H}}}$ being a bounded, convex and closed set containing this sequence. Obviously, $\bar x \in B_1$. On the other hand, we take $B_2=\operatorname*{dom}g_1^* \times \ldots \times \operatorname*{dom}g_m^*$, which is in this situation a bounded set. Then it holds, using the Fenchel-Moreau Theorem and the Young-Fenchel inequality, that $$\begin{aligned}
\mathcal{G}_{B_1 \times B_2}(x^N,{\ensuremath{\boldsymbol{v}}}^N) &= {\ensuremath{F}}(x^N) + {\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}x^N -{\ensuremath{\boldsymbol{r}}}) + {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}^N) - \inf_{\tilde{x} \in B_1}{\left\{ {\left\langle}{\ensuremath{\boldsymbol{L}}}\tilde{x}-{\ensuremath{\boldsymbol{r}}}, {\ensuremath{\boldsymbol{v}}}^N {\right\rangle}+ {\ensuremath{F}}(\tilde{x}) \right\}} \\
&\geq {\ensuremath{F}}(x^N) + {\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}x^N -{\ensuremath{\boldsymbol{r}}}) + {\ensuremath{G}}^*({\ensuremath{\boldsymbol{v}}}^N) -{\left\langle}{\ensuremath{\boldsymbol{L}}}{\ensuremath{\overline{x}}}-{\ensuremath{\boldsymbol{r}}}, {\ensuremath{\boldsymbol{v}}}^N {\right\rangle}- {\ensuremath{F}}({\ensuremath{\overline{x}}}) \\
&\geq {\ensuremath{F}}(x^N) + {\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}x^N -{\ensuremath{\boldsymbol{r}}}) - {\ensuremath{F}}({\ensuremath{\overline{x}}}) -{\ensuremath{G}}({\ensuremath{\boldsymbol{L}}}{\ensuremath{\overline{x}}}-{\ensuremath{\boldsymbol{r}}}).\end{aligned}$$ Hence, follows by statement *(d)* in Theorem \[theorem-fvalues\].
In a similar way, one can show that, whenever $f$ is Lipschitz continuous, provides for the sequence of the values of the objective of $(D)$ taken at $(v^N)_{N \geq 1}$ a convergence rate of ${\ensuremath{\mathcal{O}}}(\frac{1}{N})$.
\[remark11\] If ${\ensuremath{\mathcal{G}}}_i$, $i=1,\ldots,m$, are finite-dimensional real Hilbert spaces, then is true, even under the weaker assumption that the convex functions $g_i, i=1,...,m$, have full domain, without necessarily being Lipschitz continuous. The set $B_1 \subset {\ensuremath{\mathcal{H}}}$ can be chosen as in Remark \[remark1\], but this time we take $B_2 = \bigtimes _{i=1}^m \bigcup_{n\geq 0}\partial g_i\left( L_i p_{1,n}\right) \subset {\ensuremath{\boldsymbol{\mathcal{G}}}}$, by noticing also that the functions $g_i, i=1,...,m$, are everywhere subdifferentiable.
The set $B_2$ is bounded, as for every $i=1,\ldots,m$ the set $\bigcup_{n\geq 0}\partial g_i\left( L_i p_{1,n}\right)$ is bounded. Let be $i\in \{1,...,m\}$ fixed. Indeed, as $p_{1,n} \rightharpoonup {\ensuremath{\overline{x}}}$, it follows that $L_i p_{1,n} \rightarrow L_i\bar x$ for $i=1,...,m$. Using the fact that the subdifferential of $g_i$ is a locally bounded operator at $L_i \bar x$, the boundedness of $\bigcup_{n\geq 0}\partial g_i\left( L_i p_{1,n}\right)$ follows automatically.
For this choice of the sets $B_1$ and $B_2$, by using the same arguments as in the previous remark, it follows that is true.
Zeros of sums of monotone operators {#sectionOperator}
===================================
In this section we turn our attention to the primal-dual monotone inclusion problems formulated in Problem \[opt-problem-inclusion-full\] with the aim to provide accelerations of the iterative method described in Theorem \[theorem-inclusion-prelim\] under the additional strong monotonicity assumptions.
The case when A+C is strongly monotone
--------------------------------------
We focus first on the case when $A+C$ is $\rho$-strongly monotone for some $\rho \in {\ensuremath{\mathbb{R}}}_{++}$ and investigate the impact of this assumption on the convergence rate of the sequence of primal iterates. The condition $A+C$ is $\rho$-strongly monotone is fulfilled when either $A:{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ or $C:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{H}}}$ is $\rho$-strongly monotone. In case that $A$ is $\rho_1$-monotone and $C$ is $\rho_2$-monotone, the sum $A+C$ is $\rho$-monotone with $\rho=\rho_1+\rho_2$.
\[rm1\] The situation when $B_i^{-1} + D_i^{-1}$ is $\tau_i$-strongly monotone with $\tau_i \in {\ensuremath{\mathbb{R}}}_{++}$ for $i=1,\ldots,m$, which improves the convergence rate of the sequence of dual iterates, can be handled with appropriate modifications.
Due to technical reasons we assume in the following that the operators $D_i^{-1}$ in Problem \[opt-problem-inclusion-full\] are zero for $i=1,\ldots,m$, thus, $D_i(0) = {\ensuremath{\mathcal{G}}}_i$ and $D_i(x) = \emptyset$ for $x \neq 0$, for $i=1,...,m$. In Remark \[rm2\] we show how the results given in this particular context can be employed when treating the primal-dual pair of monotone inclusions -. Consequently, the problem we deal with in this subsection is as follows.
\[opt-problem-inclusion\] Consider the real Hilbert spaces ${\ensuremath{\mathcal{H}}}$ and ${\ensuremath{\mathcal{G}}}_i, i=1,...,m,$ $A:{\ensuremath{\mathcal{H}}}\rightarrow 2^{{\ensuremath{\mathcal{H}}}}$ a maximally monotone operator and $C:{\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{H}}}$ a monotone and $\mu$-Lipschitzian operator for some $\mu \in {\ensuremath{\mathbb{R}}}_{++}$. Furthermore, let $z \in {\ensuremath{\mathcal{H}}}$ and for every $i\in\{1,\ldots,m\}$, let $r_i \in {\ensuremath{\mathcal{G}}}_i$, let $B_i : {\ensuremath{\mathcal{G}}}_i \rightarrow 2^{{\ensuremath{\mathcal{G}}}_i}$ be maximally monotone operators and let $L_i : {\ensuremath{\mathcal{H}}}\rightarrow {\ensuremath{\mathcal{G}}}_i$ be a nonzero linear continuous operator. The problem is to solve the primal inclusion $$\begin{aligned}
\label{opt-problem-primal-inclusion}
\text{find }{\ensuremath{\overline{x}}}\in {\ensuremath{\mathcal{H}}}\text{ such that } z \in A{\ensuremath{\overline{x}}}+ \sum_{i=1}^m L_i^* B_i(L_i {\ensuremath{\overline{x}}}-r_i) + C{\ensuremath{\overline{x}}},\end{aligned}$$ together with the dual inclusion $$\begin{aligned}
\label{opt-problem-dual-inclusion}
\text{find }{\ensuremath{\overline{v}}}_1 \in {\ensuremath{\mathcal{G}}}_1,\ldots,{\ensuremath{\overline{v}}}_m \in {\ensuremath{\mathcal{G}}}_m \text{ such that }(\exists x\in{\ensuremath{\mathcal{H}}})\left\{
\begin{array}{l}
z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in Ax + Cx \\
{\ensuremath{\overline{v}}}_i \in B_i(L_ix-r_i), \,i=1,\ldots,m.
\end{array}
\right.\end{aligned}$$
The subsequent algorithm represents an accelerated version of the one given in Theorem \[theorem-inclusion-prelim\] and relies on the fruitful idea of using a second sequence of variable step length parameters $(\sigma_n)_{n \geq 0} \subseteq {\ensuremath{\mathbb{R}}}_{++}$, which, together with the sequence of parameters $(\gamma_n)_{n \geq 0} \subseteq {\ensuremath{\mathbb{R}}}_{++}$, play an important role in the convergence analysis.
\[alg2\] Let $x_0 \in {\ensuremath{\mathcal{H}}}$, $(v_{1,0}, \ldots, v_{m,0}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$, $$\gamma_0 \in \left(0,\min\left\{1, \frac{\sqrt{1+4\rho}}{2(1+2\rho)\mu}\right\}\right) \text{ and set }
\sigma_0 = \frac{1}{2\gamma_0(1+2\rho)\sum_{i=1}^m \|L_i \|^2}.$$ Consider the following updates: $$\begin{aligned}
\label{A2}
\left(\forall n\geq 0\right) \ \left\lfloor \begin{array}{l}
p_{1,n} = J_{\gamma_n A}\left(x_n - \gamma_n \left( Cx_n +\sum_{i=1}^mL_i^* v_{i,n} -z\right) \right) \\
\text{For }i=1,\ldots,m \\
\ \left\lfloor \begin{array}{l}
p_{2,i,n} = J_{\sigma_n B_i^{-1}}\left(v_{i,n} +\sigma_n (L_i x_n -r_i)\right) \\
v_{i,n+1} = \sigma_n L_i( p_{1,n} - x_n) + p_{2,i,n} \\
\end{array} \right.\\
x_{n+1} = \gamma_n \sum_{i=1}^m L_i^*(v_{i,n}-p_{2,i,n}) + \gamma_n(C x_n - C p_{1,n}) +p_{1,n} \\
\theta_n=1/\sqrt{1+2\rho\gamma_n(1-\gamma_n)}, \ \gamma_{n+1} = \theta_n\gamma_n, \ \sigma_{n+1} = \sigma_n/\theta_n.
\end{array}
\right.
\end{aligned}$$
\[theorem-inclusion\] In Problem \[opt-problem-inclusion\] suppose that $A+C$ is $\rho$-strongly monotone with $\rho \in {\ensuremath{\mathbb{R}}}_{++}$ and let $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times{\ensuremath{\boldsymbol{\mathcal{G}}}}$ be a primal-dual solution to Problem \[opt-problem-inclusion\]. Then for every $n\geq 0$ it holds $$\begin{aligned}
\label{ineq-th2}
\| x_{n} - {\ensuremath{\overline{x}}}\|^2 + \gamma_n \sum_{i=1}^m \frac{\| v_{i,n} - {\ensuremath{\overline{v}}}_i \|^2}{\sigma_{n}} \leq \gamma_n^2 \left( \frac{\| x_0 - {\ensuremath{\overline{x}}}\|^2}{\gamma_0^2} + \sum_{i=1}^m\frac{\| v_{i,0} - {\ensuremath{\overline{v}}}_i \|^2}{\gamma_0\sigma_0}\right),\end{aligned}$$ where $\gamma_n,\,\sigma_n \in {\ensuremath{\mathbb{R}}}_{++}$, $x_n \in {\ensuremath{\mathcal{H}}}$ and $(v_{1,n},\ldots,v_{m,n}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ are the iterates generated by Algorithm \[alg2\].
Taking into account the definitions of the resolvents occurring in Algorithm \[alg2\] we obtain $$\begin{aligned}
\text{ and }\ \left.
\begin{aligned}
\frac{x_n-p_{1,n}}{\gamma_n} -Cx_n - \sum_{i=1}^m L_i^*v_{i,n} +z &\in A p_{1,n} \\
\frac{v_{i,n}-p_{2,i,n}}{\sigma_{n}} + L_i x_n -r_i &\in B_i^{-1} p_{2,i,n}, i=1,\ldots,m,
\end{aligned}
\right.\end{aligned}$$ which, in the light of the updating rules in , furnishes for every $n \geq 0$ $$\begin{aligned}
\label{inclusion-alg2-prelim}
\text{ and } \ \left.
\begin{aligned}
\frac{x_n-x_{n+1}}{\gamma_n} - \sum_{i=1}^m L_i^*p_{2,i,n} +z &\in (A+C) p_{1,n} \\
\frac{v_{i,n}-v_{i,n+1}}{\sigma_{n}} + L_i p_{1,n} -r_i &\in B_i^{-1} p_{2,i,n}, i=1,\ldots,m.
\end{aligned}
\right.\end{aligned}$$ The primal-dual solution $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times{\ensuremath{\boldsymbol{\mathcal{G}}}}$ to Problem \[opt-problem-inclusion\] fulfills (see , where $D_i^{-1}$ are taken to be zero for $i=1,...,m$) $$z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in A{\ensuremath{\overline{x}}}+ C{\ensuremath{\overline{x}}}\ \mbox{and} \ {\ensuremath{\overline{v}}}_i \in B_i(L_i{\ensuremath{\overline{x}}}-r_i), \,i=1,\ldots,m.$$ Since the sum $A+C$ is $\rho$-strongly monotone, we have for every $n \geq 0$ $$\begin{aligned}
\label{operator-monotone1}
{\left\langle}p_{1,n} - {\ensuremath{\overline{x}}}, \frac{x_n-x_{n+1}}{\gamma_n} - \sum_{i=1}^m L_i^*p_{2,i,n} +z -\left( z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \right) {\right\rangle}\geq \rho \| p_{1,n} - {\ensuremath{\overline{x}}}\|^2\end{aligned}$$ while, due to the monotonicity of $B_i^{-1}:{\ensuremath{\mathcal{G}}}_i \rightarrow 2^{{\ensuremath{\mathcal{G}}}_i}$, we obtain for every $n \geq 0$ $$\begin{aligned}
\label{operator-monotone3}
{\left\langle}p_{2,i,n} - {\ensuremath{\overline{v}}}_i , \frac{v_{i,n}-v_{i,n+1}}{\sigma_{n}} + L_i p_{1,n} -r_i - \left( L_i {\ensuremath{\overline{x}}}-r_i \right){\right\rangle}\geq 0, \ i=1,\ldots,m.\end{aligned}$$ Further, we set $${\ensuremath{\boldsymbol{\overline{v}}}}= ({\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m), \quad {\ensuremath{\boldsymbol{v}}}_n = (v_{1,n},\ldots,v_{m,n}), \quad {\ensuremath{\boldsymbol{p}}}_{2,n}=(p_{2,1,n},\ldots,p_{2,m,n}).$$ Summing up the inequalities and , it follows that $$\begin{aligned}
\label{ineq-sum-monotone}
&{\left\langle}p_{1,n} - {\ensuremath{\overline{x}}}, \frac{x_n-x_{n+1}}{\gamma_n} {\right\rangle}+ {\left\langle}{\ensuremath{\boldsymbol{p}}}_{2,n} - {\ensuremath{\boldsymbol{\overline{v}}}}, \frac{{\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{v}}}_{n+1}}{\sigma_{n}} {\right\rangle}+ {\left\langle}p_{1,n} - {\ensuremath{\overline{x}}}, {\ensuremath{\boldsymbol{L}}}^*({\ensuremath{\boldsymbol{\overline{v}}}}-{\ensuremath{\boldsymbol{p}}}_{2,n}){\right\rangle}\notag\\ &+ {\left\langle}{\ensuremath{\boldsymbol{p}}}_{2,n} - {\ensuremath{\boldsymbol{\overline{v}}}}, {\ensuremath{\boldsymbol{L}}}(p_{1,n}-{\ensuremath{\overline{x}}}) {\right\rangle}\geq \rho \| p_{1,n} - {\ensuremath{\overline{x}}}\|^2.\end{aligned}$$ and, from here, $$\begin{aligned}
\label{ineq-monotone}
{\left\langle}p_{1,n} - {\ensuremath{\overline{x}}}, \frac{x_n-x_{n+1}}{\gamma_n} {\right\rangle}+ {\left\langle}{\ensuremath{\boldsymbol{p}}}_{2,n} - {\ensuremath{\boldsymbol{\overline{v}}}}, \frac{{\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{v}}}_{n+1}}{\sigma_{n}} {\right\rangle}\geq \rho \| p_{1,n} - {\ensuremath{\overline{x}}}\|^2 \ \forall n \geq 0.\end{aligned}$$ In the light of the equations $$\begin{aligned}
{\left\langle}p_{1,n} - {\ensuremath{\overline{x}}}, \frac{x_n - x_{n+1}}{\gamma_n} {\right\rangle}&= {\left\langle}p_{1,n} - x_{n+1}, \frac{x_n - x_{n+1}}{\gamma_n} {\right\rangle}+ {\left\langle}x_{n+1} - {\ensuremath{\overline{x}}}, \frac{x_n - x_{n+1}}{\gamma_n} {\right\rangle}\\
&= \frac{\| x_{n+1} - p_{1,n} \|^2}{2\gamma_n} - \frac{\|x_n- p_{1,n}\|^2}{2\gamma_n} + \frac{\| x_n - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} - \frac{\| x_{n+1} - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n},\end{aligned}$$ and $$\begin{aligned}
{\left\langle}{\ensuremath{\boldsymbol{p}}}_{2,n} - {\ensuremath{\boldsymbol{\overline{v}}}}, \frac{{\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{v}}}_{n+1}}{\sigma_n} {\right\rangle}&= {\left\langle}{\ensuremath{\boldsymbol{p}}}_{2,n} - {\ensuremath{\boldsymbol{v}}}_{n+1}, \frac{{\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{v}}}_{n+1}}{\sigma_n} {\right\rangle}+ {\left\langle}{\ensuremath{\boldsymbol{v}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{v}}}}, \frac{{\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{v}}}_{n+1}}{\sigma_n} {\right\rangle}\\
&= \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\sigma_n} - \frac{\|{\ensuremath{\boldsymbol{v}}}_n-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\sigma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n} - \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n}\end{aligned}$$ inequality reads for every $n\geq 0$ $$\begin{aligned}
\frac{\| x_n - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n} &\geq \rho \| p_{1,n} - {\ensuremath{\overline{x}}}\|^2 + \frac{\| x_{n+1} - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n} + \frac{\|x_n- p_{1,n}\|^2}{2\gamma_n} \notag\\
\label{ineq-m1}
&\quad + \frac{\|{\ensuremath{\boldsymbol{v}}}_n-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\sigma_n} - \frac{\| x_{n+1} - p_{1,n} \|^2}{2\gamma_n} - \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\sigma_n}.\end{aligned}$$ Using that $2ab \leq \alpha a^2 + \frac{b^2}{\alpha}$ for all $a,b \in {\ensuremath{\mathbb{R}}}$, $\alpha \in {\ensuremath{\mathbb{R}}}_{++}$, we obtain for $\alpha:=\gamma_n$, $$\begin{aligned}
\rho \| p_{1,n} - {\ensuremath{\overline{x}}}\|^2 &\geq \rho \left( \| x_{n+1} - {\ensuremath{\overline{x}}}\|^2 - 2 \| x_{n+1} - {\ensuremath{\overline{x}}}\| \| x_{n+1} - p_{1,n} \| + \|x_{n+1} - p_{1,n} \|^2\right) \\
&\geq \frac{2\rho\gamma_n(1-\gamma_n)}{2\gamma_n} \| x_{n+1} - {\ensuremath{\overline{x}}}\|^2 - \frac{2\rho(1-\gamma_n)}{2\gamma_n} \|x_{n+1} - p_{1,n} \|^2,\end{aligned}$$ which, in combination with , yields for every $n\geq 0$ $$\begin{aligned}
&\frac{\| x_n - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n} \geq \frac{(1+2\rho\gamma_n(1-\gamma_n))\| x_{n+1} - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n} \notag\\
\label{ineq-m2}
& + \frac{\|x_n- p_{1,n}\|^2}{2\gamma_n} + \frac{\|{\ensuremath{\boldsymbol{v}}}_n-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\sigma_n} - \frac{(1+2\rho(1-\gamma_n))\| x_{n+1} - p_{1,n} \|^2}{2\gamma_n} - \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\sigma_n}.\end{aligned}$$ Investigating the last two terms in the right-hand side of the above estimate it shows for every $n \geq 0$ that $$\begin{aligned}
&- \frac{(1+2\rho(1-\gamma_n))\| x_{n+1} - p_{1,n} \|^2}{2\gamma_n} \\
&\geq-\frac{(1+2\rho)\gamma_n}{2}\left\| \sum_{i=1}^m L_i^*(v_{i,n}-p_{2,i,n}) + (C x_n - C p_{1,n}) \right\|^2 \\
&\geq -\frac{2(1+2\rho)\gamma_n}{2}\left( \left(\sum_{i=1}^m \|L_i \|^2\right) \|{\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2 + \mu^2 \|x_n - p_{1,n}\|^2 \right),\end{aligned}$$ and $$\begin{aligned}
- \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\sigma_n}
=-\frac{\sigma_n}{2} \left( \sum_{i=1}^m \| L_i( p_{1,n} - x_n) \|^2 \right)
\geq -\frac{\sigma_n}{2} \left(\sum_{i=1}^m \|L_i\|^2\right) \| p_{1,n} - x_n\|^2 .\end{aligned}$$ Hence, for every $n \geq 0$ it holds $$\begin{aligned}
&\frac{\|x_n- p_{1,n}\|^2}{2\gamma_n} + \frac{\|{\ensuremath{\boldsymbol{v}}}_n-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2}{2\sigma_n} - \frac{(1+2\rho(1-\gamma_n))\| x_{n+1} - p_{1,n} \|^2}{2\gamma_n} - \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1}-{\ensuremath{\boldsymbol{p}}}_{2,n} \|^2}{2\sigma_n} \\
&\geq \frac{\left(1 - \gamma_n\sigma_n \sum_{i=1}^m \|L_i\|^2 - 2(1+2\rho)\gamma_n^2 \mu^2 \right)}{2\gamma_n} \| p_{1,n} - x_n\|^2 \\
&\quad + \frac{\left(1-2\gamma_n\sigma_n(1+2\rho)\sum_{i=1}^m \|L_i \|^2\right) }{2\sigma_n} \|{\ensuremath{\boldsymbol{v}}}_{n}-{\ensuremath{\boldsymbol{p}}}_{2,n}\|^2\\
&\geq 0.\end{aligned}$$ The nonnegativity of the expression in the above relation follows because of the sequence $(\gamma_n)_{n\geq 0}$ is nonincreasing, $\gamma_n \sigma_n = \gamma_0 \sigma_0$ for every $n\geq 0$ and $$\begin{aligned}
\gamma_0 \in \left(0,\min\left\{1, \frac{\sqrt{1+4\rho}}{2(1+2\rho)\mu}\right\}\right) \ \mbox{and} \ \sigma_0 = \frac{1}{2\gamma_0(1+2\rho)\sum_{i=1}^m \|L_i \|^2}.\end{aligned}$$ Consequently, inequality becomes $$\begin{aligned}
\label{ineq-m3}
\frac{\| x_n - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n} + \frac{\| {\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_n} \geq \frac{(1+2\rho\gamma_n(1-\gamma_n))\| x_{n+1} - {\ensuremath{\overline{x}}}\|^2}{2\gamma_{n}} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_{n}} \ \forall n \geq 0.\end{aligned}$$ Notice that we have $\gamma_{n+1} < \gamma_n$, $\sigma_{n+1}>\sigma_n$ and $\gamma_{n+1}\sigma_{n+1}=\gamma_n\sigma_n$ for every $n \geq 0$. Dividing by $\gamma_n$ and making use of $$\begin{aligned}
\theta_n = \frac{1}{\sqrt{1+2\rho\gamma_n(1-\gamma_n)}}, \quad \gamma_{n+1} = \theta_n \gamma_n, \quad \sigma_{n+1}=\frac{\sigma_n}{\theta_n},\end{aligned}$$ we obtain $$\begin{aligned}
\frac{\| x_n - {\ensuremath{\overline{x}}}\|^2}{2\gamma_n^2} + \frac{\| {\ensuremath{\boldsymbol{v}}}_n - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\gamma_n\sigma_n} \geq \frac{\| x_{n+1} - {\ensuremath{\overline{x}}}\|^2}{2\gamma_{n+1}^2} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\gamma_{n+1}\sigma_{n+1}} \ \forall n \geq 0.\end{aligned}$$ Let be $N \geq 1$. Summing this inequalities from $n=0$ to $N-1$, we finally get $$\begin{aligned}
\label{ineq-m4}
\frac{\| x_0 - {\ensuremath{\overline{x}}}\|^2}{2\gamma_0^2} + \frac{\| {\ensuremath{\boldsymbol{v}}}_0 - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\gamma_0\sigma_0} \geq \frac{\| x_{N} - {\ensuremath{\overline{x}}}\|^2}{2\gamma_{N}^2} + \frac{\| {\ensuremath{\boldsymbol{v}}}_{N} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\gamma_{N}\sigma_{N}}.\end{aligned}$$ In conclusion, $$\begin{aligned}
\label{ineq-m5}
\frac{\| x_{n} - {\ensuremath{\overline{x}}}\|^2}{2} + \gamma_n\frac{\| {\ensuremath{\boldsymbol{v}}}_{n} - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\sigma_{n}} \leq \gamma_n^2 \left( \frac{\| x_0 - {\ensuremath{\overline{x}}}\|^2}{2\gamma_0^2} + \frac{\| {\ensuremath{\boldsymbol{v}}}_0 - {\ensuremath{\boldsymbol{\overline{v}}}}\|^2}{2\gamma_0\sigma_0}\right) \ \forall n \geq 0,\end{aligned}$$ which completes the proof.
Next we show that $\rho \gamma_n$ converges like $\frac{1}{n}$ as $n \rightarrow +\infty$.
Let $\gamma_0 \in (0,1)$ and consider the sequence $(\gamma_n)_{n\geq 0} \subseteq {\ensuremath{\mathbb{R}}}_{++}$, where $$\begin{aligned}
\label{prop-eq1}
\gamma_{n+1} = \frac{\gamma_n}{\sqrt{1+2\rho\gamma_n(1-\gamma_n)}} \ \forall n \geq 0.
\end{aligned}$$ Then $\lim_{n \rightarrow +\infty} n\rho\gamma_n=1$.
Since the sequence $(\gamma_n)_{n\geq 0} \subseteq \left(0,1\right)$ is bounded and decreasing, it converges towards some $l \in \left[0,1\right)$ as $n \rightarrow +\infty$. We let $n \rightarrow + \infty$ in and obtain $$\begin{aligned}
l^2 (1+2\rho l (1-l)) = l^2 \Leftrightarrow 2\rho l^3 (1-l) = 0,\end{aligned}$$ which shows that $l=0$, i.e. $\gamma_n \rightarrow 0 \ (n \rightarrow +\infty)$. On the other hand, implies that $\frac{\gamma_n}{\gamma_{n+1}} \rightarrow 1 (n \rightarrow +\infty)$. As $(\frac{1}{\gamma_n})_{n\geq 0}$ is a strictly increasing and unbounded sequence, by applying the Stolz–Ces[à]{}ro Theorem it shows that $$\begin{aligned}
\lim_{n \rightarrow + \infty} n\gamma_n
&=\lim_{n \rightarrow + \infty} \frac{n}{\frac{1}{\gamma_n}}
= \lim_{n \rightarrow + \infty} \frac{ n+1- n}{\frac{1}{\gamma_{n+1}} - \frac{1}{\gamma_n}}
= \lim_{n \rightarrow + \infty} \frac{ \gamma_n \gamma_{n+1}}{\gamma_n - \gamma_{n+1}}\\
&= \lim_{n \rightarrow + \infty} \frac{ \gamma_n \gamma_{n+1} (\gamma_n + \gamma_{n+1})}{\gamma_n^2 - \gamma_{n+1}^2}
\overset{\mathclap{\eqref{prop-eq1}}}{=} \lim_{n \rightarrow + \infty} \frac{ \gamma_n \gamma_{n+1} (\gamma_n + \gamma_{n+1})}{2\rho \gamma_{n+1}^2 \gamma_n (1-\gamma_n)}\\
&= \lim_{n \rightarrow + \infty} \frac{ \gamma_n + \gamma_{n+1}}{2\rho \gamma_{n+1}(1-\gamma_n)}
= \lim_{n \rightarrow + \infty} \frac{ \frac{\gamma_n}{\gamma_{n+1}} + 1}{2\rho (1-\gamma_n)}
= \frac{2}{2\rho} = \frac{1}{\rho},\end{aligned}$$ which completes the proof.
Hence, we have shown the following result.
\[theorem-inclusion2\] In Problem \[opt-problem-inclusion\] suppose that $A+C$ is $\rho$-strongly monotone and let $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times{\ensuremath{\boldsymbol{\mathcal{G}}}}$ be a primal-dual solution to Problem \[opt-problem-inclusion\]. Then, for any $\varepsilon > 0$, there exists some $n_0\in{\ensuremath{\mathbb{N}}}$ (depending on $\varepsilon$ and $\rho\gamma_0$) such that for any $n\geq n_0$ $$\begin{aligned}
\label{ineq-th2-1}
\| x_{n} - {\ensuremath{\overline{x}}}\|^2 \leq \frac{1+\varepsilon}{n^2} \left( \frac{\| x_0 - {\ensuremath{\overline{x}}}\|^2}{\rho^2\gamma_0^2} + \sum_{i=1}^m\frac{\| v_{i,0} - {\ensuremath{\overline{v}}}_i \|^2}{\rho^2\gamma_0\sigma_0}\right),\end{aligned}$$ where $\gamma_n,\,\sigma_n \in {\ensuremath{\mathbb{R}}}_{++}$, $x_n \in {\ensuremath{\mathcal{H}}}$ and $(v_{1,n},\ldots,v_{m,n}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ are the iterates generated by Algorithm \[alg2\].
\[rm2\] In Algorithm \[alg2\] and Theorem \[theorem-inclusion2\] we assumed that $D_i^{-1} = 0$ for $i=1,\ldots,m$, however, similar statements can be also provided for Problem \[opt-problem-inclusion-full\] under the additional assumption that the operators $D_i : {\ensuremath{\mathcal{G}}}_i \rightarrow 2^{{\ensuremath{\mathcal{G}}}_i}$ are such that $D_i^{-1}$ is $\nu_i^{-1}$-cocoercive with $\nu_i \in {\ensuremath{\mathbb{R}}}_{++}$ for $i=1,\ldots,m$. This assumption is in general stronger than assuming that $D_i$ is monotone and $D_i^{-1}$ is $\nu_i$-Lipschitzian for $i=1,...,m$. However, it guarantees that $D_i$ is $\nu_i^{-1}$-strongly monotone and maximally monotone for $i=1,...,m$ (see [@BauschkeCombettes11 Example 20.28, Proposition 20.22 and Example 22.6]). We introduce the Hilbert space ${\ensuremath{\boldsymbol{\mathcal{H}}}}= {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$, the element ${\ensuremath{\boldsymbol{z}}}=(z,0,\ldots,0) \in {\ensuremath{\boldsymbol{\mathcal{H}}}}$ and the maximally monotone operator ${\ensuremath{\boldsymbol{A}}}: {\ensuremath{\boldsymbol{\mathcal{H}}}}\rightarrow 2^{{\ensuremath{\boldsymbol{\mathcal{H}}}}}$, ${\ensuremath{\boldsymbol{A}}}(x,y_1,\ldots,y_m) = (Ax,D_1y_1,\ldots,D_m y_m)$ and the monotone and Lipschitzian operator ${\ensuremath{\boldsymbol{C}}}: {\ensuremath{\boldsymbol{\mathcal{H}}}}\rightarrow {\ensuremath{\boldsymbol{\mathcal{H}}}}$, ${\ensuremath{\boldsymbol{C}}}(x,y_1,\ldots,y_m) = (Cx,0,\ldots,0)$. Notice also that ${\ensuremath{\boldsymbol{A}}}+ {\ensuremath{\boldsymbol{C}}}$ is strongly monotone. Furthermore, we introduce the element ${\ensuremath{\boldsymbol{r}}}=(r_1,\ldots,r_m)\in{\ensuremath{\boldsymbol{\mathcal{G}}}}$, the maximally monotone operator ${\ensuremath{\boldsymbol{B}}}: {\ensuremath{\boldsymbol{\mathcal{G}}}}\rightarrow 2^{{\ensuremath{\boldsymbol{\mathcal{G}}}}}$, ${\ensuremath{\boldsymbol{B}}}(y_1,\ldots,y_m) = (B_1y_1,\ldots,B_my_m)$, and the linear continuous operator ${\ensuremath{\boldsymbol{L}}}: {\ensuremath{\boldsymbol{\mathcal{H}}}}\rightarrow {\ensuremath{\boldsymbol{\mathcal{G}}}}$, ${\ensuremath{\boldsymbol{L}}}(x,y_1\ldots,y_m) = (L_1 x-y_1,\ldots,L_m x - y_m),$ having as adjoint ${\ensuremath{\boldsymbol{L}}}^* : {\ensuremath{\boldsymbol{\mathcal{G}}}}\rightarrow {\ensuremath{\boldsymbol{\mathcal{H}}}}$, ${\ensuremath{\boldsymbol{L}}}^* (q_1,\ldots,q_m) = (\sum_{i=1}^m L_i^* q_i, -q_1,\ldots,-q_m)$. We consider the primal problem $$\begin{aligned}
\label{primal-inc-pspace}
\text{find } {\ensuremath{\boldsymbol{\overline{x}}}}= ({\ensuremath{\overline{x}}},{\ensuremath{\overline{p}}}_1\ldots{\ensuremath{\overline{p}}}_m) \in {\ensuremath{\boldsymbol{\mathcal{H}}}}\text{ such that } {\ensuremath{\boldsymbol{z}}}\in {\ensuremath{\boldsymbol{A}}}{\ensuremath{\boldsymbol{\overline{x}}}}+ {\ensuremath{\boldsymbol{L}}}^*{\ensuremath{\boldsymbol{B}}}\left({\ensuremath{\boldsymbol{L}}}{\ensuremath{\boldsymbol{\overline{x}}}}- {\ensuremath{\boldsymbol{r}}}\right) + {\ensuremath{\boldsymbol{C}}}{\ensuremath{\boldsymbol{\overline{x}}}},
\end{aligned}$$ together with the dual inclusion problem $$\begin{aligned}
\label{dual-inc-pspace}
\text{find } {\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{\mathcal{G}}}}\text{ such that }(\exists {\ensuremath{\boldsymbol{x}}}\in{\ensuremath{\boldsymbol{\mathcal{H}}}})\left\{
\begin{array}{l}
{\ensuremath{\boldsymbol{z}}}- {\ensuremath{\boldsymbol{L}}}^* {\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{A}}}{\ensuremath{\boldsymbol{x}}}+ {\ensuremath{\boldsymbol{C}}}{\ensuremath{\boldsymbol{x}}}\\
{\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{B}}}({\ensuremath{\boldsymbol{L}}}{\ensuremath{\boldsymbol{x}}}-{\ensuremath{\boldsymbol{r}}})
\end{array}
\right..\end{aligned}$$ We notice that Algorithm \[alg2\] can be employed for solving this primal-dual pair of monotone inclusion problems and, by separately involving the resolvents of $A, B_i$ and $D_i, i=1,...,m$, as for $\gamma \in {\ensuremath{\mathbb{R}}}_{++}$ $$\begin{aligned}
J_{\gamma {\ensuremath{\boldsymbol{A}}}}(x,y_1,\ldots,y_m) &= (J_{\gamma A}x, J_{\gamma D_1} y_1, \ldots, J_{\gamma D_m} y_m) \ \forall (x,y_1,\ldots,y_m) \in {\ensuremath{\boldsymbol{\mathcal{H}}}}\\
J_{\gamma {\ensuremath{\boldsymbol{B}}}}(q_1,\ldots,q_m) &= (J_{\gamma B_1} q_1, \ldots, J_{\gamma B_m} q_m) \ \forall (q_1,\ldots,q_m) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}.\end{aligned}$$ Having $({\ensuremath{\boldsymbol{\overline{x}}}},{\ensuremath{\boldsymbol{\overline{v}}}}) \in {\ensuremath{\boldsymbol{\mathcal{H}}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$ a primal-dual solution to the primal-dual pair of monotone inclusion problems -, Algorithm \[alg2\] generates a sequence of primal iterates fulfilling in ${\ensuremath{\boldsymbol{\mathcal{H}}}}$. Moreover, $({\ensuremath{\boldsymbol{\overline{x}}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ is a a primal-dual solution to - if and only if $${\ensuremath{\boldsymbol{z}}}- {\ensuremath{\boldsymbol{L}}}^*{\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{A}}}{\ensuremath{\boldsymbol{\overline{x}}}}+ {\ensuremath{\boldsymbol{C}}}{\ensuremath{\boldsymbol{\overline{x}}}}\ \mbox{and} \ {\ensuremath{\boldsymbol{\overline{v}}}}\in {\ensuremath{\boldsymbol{B}}}\left({\ensuremath{\boldsymbol{L}}}{\ensuremath{\boldsymbol{\overline{x}}}}- {\ensuremath{\boldsymbol{r}}}\right)$$ $$\Leftrightarrow z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in A {\ensuremath{\overline{x}}}+ C {\ensuremath{\overline{x}}}\ \mbox{and} \ {\ensuremath{\overline{v}}}_i \in D_i {\ensuremath{\overline{p}}}_i, {\ensuremath{\overline{v}}}_i \in B_i \left(L_i {\ensuremath{\overline{x}}}- {\ensuremath{\overline{p}}}_i - r_i \right), i=1,\ldots,m$$ $$\Leftrightarrow z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in A {\ensuremath{\overline{x}}}+ C {\ensuremath{\overline{x}}}\ \mbox{and} \ {\ensuremath{\overline{v}}}_i \in D_i {\ensuremath{\overline{p}}}_i, L_i {\ensuremath{\overline{x}}}- r_i \in B_i^{-1}{\ensuremath{\overline{v}}}_i + {\ensuremath{\overline{p}}}_i, i=1,\ldots,m.$$ Thus, if $({\ensuremath{\boldsymbol{\overline{x}}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ is a primal-dual solution to -, then $({\ensuremath{\overline{x}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\]. Viceversa, if $({\ensuremath{\overline{x}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ is a primal-dual solution to Problem \[opt-problem-inclusion-full\], then, choosing ${\ensuremath{\overline{p}}}_i \in D_i^{-1}{\ensuremath{\overline{v}}}_i, i=1,...,m$, and ${\ensuremath{\boldsymbol{\overline{x}}}}= ({\ensuremath{\overline{x}}},{\ensuremath{\overline{p}}}_1\ldots{\ensuremath{\overline{p}}}_m)$, it yields that $({\ensuremath{\boldsymbol{\overline{x}}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ is a primal-dual solution to -. In conclusion, the first component of every primal iterate in ${\ensuremath{\boldsymbol{\mathcal{H}}}}$ generated by Algorithm \[alg2\] for finding the primal-dual solution $({\ensuremath{\boldsymbol{\overline{x}}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ to - will furnish a sequence of iterates verifying in ${\ensuremath{\mathcal{H}}}$ for the primal-dual solution $({\ensuremath{\overline{x}}},{\ensuremath{\boldsymbol{\overline{v}}}})$ to Problem \[opt-problem-inclusion-full\].
The case when A+C and Bi(inverse)+Di(inverse), i=1,...,m are strongly monotone
------------------------------------------------------------------------------
Within this subsection we consider the case when $A+C$ is $\rho$-strongly monotone with $\rho \in {\ensuremath{\mathbb{R}}}_{++}$ and $B_i^{-1}+D_i^{-1}$ is $\tau_i$-strongly monotone with $\tau_i \in {\ensuremath{\mathbb{R}}}_{++}$ for $i=1,\ldots,m,$ and provide an accelerated version of the algorithm in Theorem \[theorem-inclusion-prelim\] which generates sequences of primal and dual iterates that converge to the primal-dual solution to Problem \[opt-problem-inclusion-full\] with an improved rate of convergence.
\[alg3\] Let $x_0 \in {\ensuremath{\mathcal{H}}}$, $(v_{1,0}, \ldots, v_{m,0}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$, and $\gamma \in (0,1)$ such that $$\gamma \leq \frac{1}{\sqrt{1+2\min\left\{\rho,\tau_1,\ldots,\tau_m\right\}}\left(\sqrt{\sum_{i=1}^m \| L_i\|^2} + \max \left\{ \mu, \nu_1,\ldots,\nu_m \right\}\right)}.$$ Consider the following updates: $$\begin{aligned}
\label{A3}
\left(\forall n\geq 0\right) \ \left\lfloor \begin{array}{l}
p_{1,n} = J_{\gamma A}\left(x_n - \gamma \left( C x_n +\sum_{i=1}^mL_i^*v_{i,n} -z\right) \right) \\
\text{For }i=1,\ldots,m \\
\ \left\lfloor \begin{array}{l}
p_{2,i,n} = J_{\gamma B_i^{-1}}\left(v_{i,n} +\gamma (L_i x_n - D_i^{-1} v_{i,n} -r_i)\right) \\
v_{i,n+1} = \gamma L_i( p_{1,n} - x_n) +\gamma (D_i^{-1} v_{i,n} -D_i^{-1} p_{2,i,n}) + p_{2,i,n}
\end{array} \right.\\
x_{n+1} = \gamma \sum_{i=1}^mL_i^*(v_{i,n}-p_{2,i,n}) + \gamma (C x_n - C p_{1,n}) +p_{1,n}.
\end{array}
\right.
\end{aligned}$$
\[theorem-inclusion-strongly2\] In Problem \[opt-problem-inclusion-full\] suppose that $A+C$ is $\rho$-strongly monotone with $\rho \in {\ensuremath{\mathbb{R}}}_{++}$, $B_i^{-1}+D_i^{-1}$ is $\tau_i$-strongly monotone with $\tau_i \in {\ensuremath{\mathbb{R}}}_{++}$ for $i=1,\ldots,m,$ and let $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times{\ensuremath{\boldsymbol{\mathcal{G}}}}$ be a primal-dual solution to Problem \[opt-problem-inclusion-full\]. Then for every $n\geq 0$ it holds $$\begin{aligned}
\| x_{n} - {\ensuremath{\overline{x}}}\|^2 + \sum_{i=1}^m \| v_{i,n} - {\ensuremath{\overline{v}}}_i \|^2 \leq \left(\frac{1}{1+2\rho_{\min}\gamma(1-\gamma)}\right)^n \left( \| x_{0} - {\ensuremath{\overline{x}}}\|^2 + \sum_{i=1}^m \| v_{i,0} - {\ensuremath{\overline{v}}}_i \|^2\right),\end{aligned}$$ where $\rho_{\min}=\min\left\{\rho,\tau_1,\ldots,\tau_m\right\}$ and $x_n \in {\ensuremath{\mathcal{H}}}$ and $(v_{1,n},\ldots,v_{m,n}) \in {\ensuremath{\boldsymbol{\mathcal{G}}}}$ are the iterates generated by Algorithm \[alg3\].
Taking into account the definitions of the resolvents occurring in Algorithm \[alg3\] we obtain for every $n \geq 0$ $$\begin{aligned}
\frac{x_n-x_{n+1}}{\gamma} - \sum_{i=1}^m L_i^*p_{2,i,n} +z &\in (A+C)p_{1,n}\end{aligned}$$ and $$\begin{aligned}
\frac{v_{i,n}-v_{i,n+1}}{\gamma} + L_i p_{1,n} - r_i &\in (B_i^{-1} + D_i^{-1}) p_{2,i,n}, \ i=1,\ldots,m.\end{aligned}$$ The primal-dual solution $({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m) \in {\ensuremath{\mathcal{H}}}\times{\ensuremath{\boldsymbol{\mathcal{G}}}}$ to Problem \[opt-problem-inclusion-full\] fulfills (see ) $$z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \in A{\ensuremath{\overline{x}}}+ C{\ensuremath{\overline{x}}}\ \mbox{and} \ {\ensuremath{\overline{v}}}_i \in (B_i {\ensuremath{\mbox{\small$\,\square\,$}}}D_i)(L_i{\ensuremath{\overline{x}}}-r_i), \,i=1,\ldots,m.$$ By the strong monotonicity of $A+C$ and $B_i^{-1}+D_i^{-1}$, $i=1,\ldots,m$, we obtain for every $n \geq 0$ $$\begin{aligned}
\label{ineq-sm1}
{\left\langle}p_{1,n} - {\ensuremath{\overline{x}}}, \frac{x_n-x_{n+1}}{\gamma} - \sum_{i=1}^m L_i^*p_{2,i,n} +z - \left( z - \sum_{i=1}^m L_i^*{\ensuremath{\overline{v}}}_i \right) {\right\rangle}&\geq \rho \| p_{1,n} - {\ensuremath{\overline{x}}}\|^2\end{aligned}$$ and, respectively, $$\begin{aligned}
\label{ineq-sm2}
{\left\langle}p_{2,i,n} - {\ensuremath{\overline{v}}}_i, \frac{v_{i,n}-v_{i,n+1}}{\gamma} + L_i p_{1,n} - r_i - \left( L_i {\ensuremath{\overline{x}}}-r_i \right){\right\rangle}&\geq \tau_i \| p_{2,i,n} -{\ensuremath{\overline{v}}}_i \|^2, i=1,...,m.
\end{aligned}$$ Consider the Hilbert space $ {\ensuremath{\boldsymbol{\mathcal{H}}}}= {\ensuremath{\mathcal{H}}}\times {\ensuremath{\boldsymbol{\mathcal{G}}}}$, equipped with the inner product defined in and associated norm, and set $${\ensuremath{\boldsymbol{\overline{x}}}}= ({\ensuremath{\overline{x}}},{\ensuremath{\overline{v}}}_1,\ldots,{\ensuremath{\overline{v}}}_m), \quad {\ensuremath{\boldsymbol{x}}}_n=(x_n,v_{1,n},\ldots,v_{m,n}), \quad {\ensuremath{\boldsymbol{p}}}_n=(p_{1,n},p_{2,1,n},\ldots,p_{2,m,n}).$$ Summing up the inequalities and and using $$\begin{aligned}
{\left\langle}{\ensuremath{\boldsymbol{p}}}_n - {\ensuremath{\boldsymbol{\overline{x}}}}, \frac{{\ensuremath{\boldsymbol{x}}}_n - {\ensuremath{\boldsymbol{x}}}_{n+1}}{\gamma} {\right\rangle}= \frac{\| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{p}}}_{n} \|^2}{2\gamma} - \frac{\|{\ensuremath{\boldsymbol{x}}}_n- {\ensuremath{\boldsymbol{p}}}_{n}\|^2}{2\gamma} + \frac{\| {\ensuremath{\boldsymbol{x}}}_n - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2}{2\gamma} - \frac{\| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2}{2\gamma},\end{aligned}$$ we obtain for every $n \geq 0$ $$\begin{aligned}
\label{ineq-sm3}
\frac{\| {\ensuremath{\boldsymbol{x}}}_n - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2}{2\gamma} \geq \rho_{\min} \| {\ensuremath{\boldsymbol{p}}}_{n} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 + \frac{\| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2}{2\gamma} + \frac{\|{\ensuremath{\boldsymbol{x}}}_n- {\ensuremath{\boldsymbol{p}}}_{n}\|^2}{2\gamma} - \frac{\| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{p}}}_{n} \|^2}{2\gamma}.\end{aligned}$$ Further, using the estimate $2ab\leq \gamma a^2 + \frac{b^2}{\gamma}$ for all $a,b \in {\ensuremath{\mathbb{R}}}$, we obtain $$\begin{aligned}
\rho_{\min} \| {\ensuremath{\boldsymbol{p}}}_{n} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 &\geq \frac{2\rho_{\min}\gamma(1-\gamma)}{2\gamma} \| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 - \frac{2\rho_{\min}(1-\gamma)}{2\gamma} \| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{p}}}_{n}\|^2 \\
&\geq \frac{2\rho_{\min}\gamma(1-\gamma)}{2\gamma} \| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 - \frac{2\rho_{\min}}{2\gamma} \| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{p}}}_{n}\|^2 \ \forall n \geq 0.\end{aligned}$$ Hence, reduces to $$\begin{aligned}
\frac{\| {\ensuremath{\boldsymbol{x}}}_n - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2}{2\gamma} &\geq \frac{ (1+2\rho_{\min}\gamma(1-\gamma) )\| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2}{2\gamma} \\
&\quad + \frac{\|{\ensuremath{\boldsymbol{x}}}_n- {\ensuremath{\boldsymbol{p}}}_{n}\|^2}{2\gamma} - \frac{(1+2\rho_{\min}) \| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{p}}}_{n} \|^2}{2\gamma} \ \forall n \geq 0.\end{aligned}$$ Using the same arguments as in , it is easy to check that for every $n \geq 0$ $$\begin{aligned}
&\frac{\|{\ensuremath{\boldsymbol{x}}}_n- {\ensuremath{\boldsymbol{p}}}_{n}\|^2}{2\gamma} - \frac{(1+2\rho_{\min}) \| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{p}}}_{n} \|^2}{2\gamma} \\
&\geq \left(1-(1+2\rho_{\min})\gamma^2\left(\sqrt{\sum_{i=1}^m \| L_i\|^2} + \max \left\{ \mu, \nu_1,\ldots,\nu_m \right\}\right)^2\right) \frac{\|{\ensuremath{\boldsymbol{x}}}_n- {\ensuremath{\boldsymbol{p}}}_{n}\|^2}{2\gamma}\\
&\geq 0,\end{aligned}$$ whereby the nonnegativity of this term is ensured by the assumption that $$\gamma \leq \frac{1}{\sqrt{1+2\rho_{\min}}\left(\sqrt{\sum_{i=1}^m \| L_i\|^2} + \max \left\{ \mu, \nu_1,\ldots,\nu_m \right\}\right)}.$$ Therefore, we obtain $$\begin{aligned}
\| {\ensuremath{\boldsymbol{x}}}_n - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 \geq (1+2\rho_{\min}\gamma(1-\gamma) )\| {\ensuremath{\boldsymbol{x}}}_{n+1} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 \ \forall n \geq 0,\end{aligned}$$ which leads to $$\begin{aligned}
\| {\ensuremath{\boldsymbol{x}}}_{n} - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 \leq \left(\frac{1}{1+2\rho_{\min}\gamma(1-\gamma)}\right)^n \| {\ensuremath{\boldsymbol{x}}}_0 - {\ensuremath{\boldsymbol{\overline{x}}}}\|^2 \ \forall n \geq 0.\end{aligned}$$
Numerical experiments in imaging {#sectionApp}
================================
In this section we test the feasibility of Algorithm \[alg1\] and of its accelerated version Algorithm \[alg2\] in the context of different problem formulations occurring in imaging and compare their performances to the ones of two other popular primal-dual algorithms introduced in [@ChaPoc11]. For all applications discussed in this section the images have been normalized, in order to make their pixels range in the closed interval from $0$ (pure black) to $1$ (pure white).
TV-based image denoising {#subsectionDenoise}
------------------------
Our first numerical experiment aims the solving of an image denoising problem via total variation regularization. More precisely, we deal with the convex optimization problem $$\begin{aligned}
\label{ex-denoise}
\inf_{x\in{\ensuremath{\mathbb{R}}}^n} \left\{ \lambda \, TV(x) + \frac{1}{2} \|x-b\|^2 \right\},\end{aligned}$$ where $\lambda\in{\ensuremath{\mathbb{R}}}_{++}$ is the regularization parameter, $TV:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$ is a discrete total variation functional and $b\in{\ensuremath{\mathbb{R}}}^n$ is the observed noisy image.
In this context, $x \in {\ensuremath{\mathbb{R}}}^n$ represents the vectorized image $X\in{\ensuremath{\mathbb{R}}}^{M\times N}$, where $n = M\cdot N$ and $x_{i,j}$ denotes the normalized value of the pixel located in the $i$-th row and the $j$-th column, for $i=1,\ldots,M$ and $j=1,\ldots,N$. Two popular choices for the discrete total variation functional are the *isotropic total variation* $TV_{{\ensuremath{\text{iso}}}}:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$, $$\begin{aligned}
TV_{{\ensuremath{\text{iso}}}}(x) &= \sum_{i=1}^{M-1}\sum_{j=1}^{N-1}\sqrt{ (x_{i+1,j}-x_{i,j})^2 + (x_{i,j+1}-x_{i,j})^2 } \\
&\quad + \sum_{i=1}^{M-1} \left| x_{i+1,N}-x_{i,N} \right| + \sum_{j=1}^{N-1} \left| x_{M,j+1}-x_{M,j} \right| ,\end{aligned}$$ and the *anisotropic total variation* $TV_{{\ensuremath{\text{aniso}}}}:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$, $$\begin{aligned}
TV_{{\ensuremath{\text{aniso}}}}(x) &= \sum_{i=1}^{M-1}\sum_{j=1}^{N-1} \left|x_{i+1,j}-x_{i,j}\right| + \left|x_{i,j+1}-x_{i,j}\right| \\
&\quad + \sum_{i=1}^{M-1} \left| x_{i+1,N}-x_{i,N} \right| + \sum_{j=1}^{N-1} \left| x_{M,j+1}-x_{M,j} \right| ,\end{aligned}$$ where in both cases reflexive (Neumann) boundary conditions are assumed.
We denote ${\ensuremath{\mathcal{Y}}}={\ensuremath{\mathbb{R}}}^n \times {\ensuremath{\mathbb{R}}}^n$ and define the linear operator $L:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathcal{Y}}}$, $x_{i,j} \mapsto (L_1x_{i,j}, L_2x_{i,j})$, where $$\begin{aligned}
L_1x_{i,j} = \left\{ \begin{array}{ll} x_{i+1,j}-x_{i,j}, & \text{if }i<M\\ 0, &\text{if }i=M\end{array}\right. \ \mbox{and} \
L_2x_{i,j} = \left\{ \begin{array}{ll} x_{i,j+1}-x_{i,j}, & \text{if }j<N\\ 0, &\text{if }j=N\end{array}\right. .\end{aligned}$$ The operator $L$ represents a discretization of the gradient using reflexive (Neumann) boundary conditions and standard finite differences. One can easily check that $\| L \|^2 \leq 8$ and that its adjoint $L^* : {\ensuremath{\mathcal{Y}}}\rightarrow {\ensuremath{\mathbb{R}}}^n$ is as easy to implement as the operator itself (cf. [@Cha04]).
Within this example we will focus on the anisotropic total variation function which is nothing else than the composition of the $l_1$-norm in ${\ensuremath{\mathcal{Y}}}$ with the linear operator $L$. Due to the full splitting characteristics of the iterative methods presented in this paper, we need only to compute the proximal point of the conjugate of the $l_1$-norm, the latter being the indicator function of the dual unit ball. Thus, the calculation of the proximal point will result in the computation of a projection, which has an easy implementation. The more challenging isotropic total variation functional is employed in the forthcoming subsection in the context of an image deblurring problem.
Thus, problem reads equivalently $$\inf_{x\in{\ensuremath{\mathbb{R}}}^n} \left\{ h(x) + g(Lx) \right\},$$ where $h: {\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$, $h(x)=\frac{1}{2} \|x-b\|^2$, is $1$-strongly monotone and differentiable with $1$-Lipschitzian gradient and $g:{\ensuremath{\mathcal{Y}}}\rightarrow {\ensuremath{\mathbb{R}}}$ is defined as $g(y_1,y_2)=\lambda \|(y_1,y_2)\|_1$. Then its conjugate $g^*:{\ensuremath{\mathcal{Y}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$ is nothing else than $$g^*(p_1,p_2) = \left(\lambda\|\cdot\|_1\right)^*(p_1,p_2)= \lambda \left\|\left(\frac{p_1}{\lambda}, \frac{p_2}{\lambda}\right)\right\|_1^* = \delta_{S}(p_1,p_2),$$ where $S=\left[-\lambda, \lambda\right]^n \times \left[-\lambda, \lambda\right]^n$. Taking $x_0 \in {\ensuremath{\mathcal{H}}}$, $v_0 \in {\ensuremath{\mathcal{Y}}}$, $$\gamma_0 \in \left(0,\min\left\{1, \frac{\sqrt{1+4\rho}}{2(1+2\rho)\mu}\right\}\right) \text{ and }
\sigma_0 = \frac{1}{2\gamma_0(1+2\rho)\sum_{i=1}^m \|L_i \|^2},$$ Algorithm \[alg2\] looks for this particular problem like $$\begin{aligned}
\left(\forall n\geq 0\right) \ \left\lfloor \begin{array}{l}
p_{1,n} = x_n - \gamma_n \left( x_n -b +L^* v_{n} \right) \\
p_{2,n} = {\ensuremath{\mathcal{P}}}_{S}\left(v_{n} +\sigma_n L x_n \right) \\
v_{n+1} = \sigma_n L( p_{1,n} - x_n) + p_{2,n} \\
x_{n+1} = \gamma_n L^*(v_{n}-p_{2,n}) + \gamma_n(x_n - p_{1,n}) +p_{1,n} \\
\theta_n=1/\sqrt{1+2\rho\gamma_n(1-\gamma_n)}, \ \gamma_{n+1} = \theta_n\gamma_n, \ \sigma_{n+1} = \sigma_n/\theta_n.
\end{array}
\right.
\end{aligned}$$
------ ------------------------- --------------------------- -- -- ------------------------- ---------------------------
$\varepsilon=10^{-4}$ $\varepsilon=10^{-6}$ $\varepsilon=10^{-4}$ $\varepsilon=10^{-6}$
ALG1 $350\ (7.03 \text{ s})$ $2989\ (59.82 \text{ s})$ $184\ (3.69 \text{ s})$ $1454\ (29.07 \text{ s})$
ALG2 $101\ (2.28 \text{ s})$ $442\ (9.91 \text{ s})$ $72\ (1.62 \text{ s})$ $298\ (6.68 \text{ s})$
PD1 $342\ (3.59 \text{ s})$ $3133\ (32.68 \text{ s})$ $180\ (1.91 \text{ s})$ $1427\ (14.87 \text{ s})$
PD2 $96\ (1.02 \text{ s})$ $442\ (4.67 \text{ s})$ $69\ (0.76 \text{ s})$ $319\ (3.39 \text{ s})$
------ ------------------------- --------------------------- -- -- ------------------------- ---------------------------
: Performance evaluation for the images in Figure \[fig:lichtenstein\]. The entries represent to the number of iterations and the CPU times in seconds, respectively, needed in order to attain a root mean squared error for the iterates below the tolerance $\varepsilon$.[]{data-label="table:performance"}
However, we solved the regularized image denoising problem with Algorithm \[alg1\], the primal-dual iterative scheme from [@ChaPoc11] (see, also, [@Vu11]) and the accelerated version of the latter presented in [@ChaPoc11 Theorem 2], as well, and refer the reader to Table \[table:performance\] for a comparison of the obtained results:
- ALG1: Algorithm \[alg1\] with $\gamma = \frac{1-\tilde\varepsilon}{\sqrt{8}}$, small $\tilde\varepsilon>0$ and by taking the last iterate instead of the averaged sequence.
- ALG2: Algorithm \[alg2\] with $\rho=0.3$, $\mu=1$ and $\gamma_0=\frac{\sqrt{1+4\rho}}{2(1+2\rho)\mu}$.
- PD1: Algorithm 1 in [@ChaPoc11] with $\tau=\frac{1}{\sqrt{8}}$, $\tau \sigma 8 = 1$ and by taking the last iterate instead of the averaged sequence.
- PD2: Algorithm 2 in [@ChaPoc11] with $\rho=0.3$, $\tau_0=\frac{1}{\sqrt{8}}$, $\tau_0\sigma_0 8=1$.
From the point of view of the number of iterations, one can notice similarities between both the primal-dual algorithms ALG1 and PD1 and the accelerated versions ALG2 and PD2. From this point of view they behave almost equal. When comparing the CPU times, it shows that the methods in this paper need almost twice amount of time. This is since ALG1 and ALG2 lead back to a forward-backward-forward splitting, whereas PD1 and PD2 rely on a forward-backward splitting scheme, meaning that ALG1 and ALG2 process the double amount of forward steps than PD1 and PD2. In this example the evaluation of forward steps (i.e. which constitute in matix-vector multiplications involving the linear operators and their adjoints) is, compared with the calculation of projections when computing the resolvents, the most costly step.
\[\]
TV-based image deblurring {#subsectionDeblur}
-------------------------
The second numerical experiment that we consider concerns the solving of an extremely ill-conditioned linear inverse problem which arises in image deblurring and denoising. For a given matrix $A \in \mathbb{R}^{n \times n}$ describing a blur operator and a given vector $b \in {\ensuremath{\mathbb{R}}}^n$ representing the blurred and noisy image, the task is to estimate the unknown original image ${\ensuremath{\overline{x}}}\in{\ensuremath{\mathbb{R}}}^n$ fulfilling $$A{\ensuremath{\overline{x}}}=b.$$ To this end we basically solve the following regularized convex nondifferentiable problem $$\label{probimageproc}
\inf_{x \in {\ensuremath{\mathbb{R}}}^n}{\left\{ \left\| Ax-b \right\|_1 +\lambda_1 TV_{{\ensuremath{\text{iso}}}}(x) + \lambda_2 \left\| x \right\|_1 + \delta_{\left[0,1\right]^n}(x) \right\}},$$ where $\lambda_1,\,\lambda_2 \in {\ensuremath{\mathbb{R}}}_{++}$ are regularization parameters and $TV_{{\ensuremath{\text{iso}}}}:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$ is the discrete isotropic total variation function. Notice that none of the functions occurring in is differentiable, while the regularization is done by a combination of two regularization functionals with different properties.
The blurring operator is constructed by making use of the Matlab routines [imfilter]{} and [fspecial]{} as follows:
``` {.numberLines numbers="left" numberstyle="\tiny" frame="tlrb" showstringspaces="false"}
H=fspecial('gaussian',9,4); % gaussian blur of size 9 times 9
% and standard deviation 4
B=imfilter(X,H,'conv','symmetric'); % B=observed blurred image
% X=original image
```
The function [fspecial]{} returns a rotationally symmetric Gaussian lowpass filter of size $9 \times 9$ with standard deviation $4$, the entries of $H$ being nonnegative and their sum adding up to $1$. The function [imfilter]{} convolves the filter $H$ with the image $X$ and furnishes the blurred image $B$. The boundary option “symmetric” corresponds to reflexive boundary conditions. Thanks to the rotationally symmetric filter $H$, the linear operator $A$ defined via the routine [imfilter]{} is symmetric, too. By making use of the real spectral decomposition of $A$, it shows that $\left\| A \right\|^2=1$.
For $(y,z),\,(p,q) \in {\ensuremath{\mathcal{Y}}}$, we introduce the inner product $${\left\langle}(y,z),(p,q) {\right\rangle}= \sum_{i=1}^M\sum_{j=1}^N y_{i,j}p_{i,j} + z_{i,j}q_{i,j}$$ and define $\| (y,z)\|_{\times} = \sum_{i=1}^M\sum_{j=1}^N \sqrt{y_{i,j}^2 + z_{i,j}^2}$. One can check that $\|\cdot\|_{\times}$ is a norm on ${\ensuremath{\mathcal{Y}}}$ and that for every $x\in{\ensuremath{\mathbb{R}}}^n$ it holds $TV_{{\ensuremath{\text{iso}}}}(x) = \| L x \|_{\times}$, where $L$ is the linear operator defined in the previous section. The conjugate function $(\|\cdot\|_{\times})^*:{\ensuremath{\mathcal{Y}}}\rightarrow {\ensuremath{\overline{\mathbb{R}}}}$ of $\|\cdot\|_{\times}$ is for every $(p,q) \in {\ensuremath{\mathcal{Y}}}$ given by (see, for instance, [@BotGradWanka09]) $$(\|\cdot\|_{\times})^* (p,q) = \left\{ \begin{array}{ll}0, & \text{if }\|(p,q)\|_{\times *} \leq 1 \\ +\infty, & \text{otherwise} \end{array}\right.,$$ where $$\|(p,q)\|_{\times *} = \sup_{\|(y,z)\|_{\times} \leq 1} {\left\langle}(p,q),(y,z) {\right\rangle}= \max_{\substack{1 \leq i \leq M \\ 1 \leq j \leq N}}\sqrt{p_{i,j}^2 + q_{i,j}^2}.$$ Therefore, the optimization problem can be written in the form of $$\inf_{x \in {\ensuremath{\mathbb{R}}}^n}{\left\{f(x) + g_1(Ax) + g_2(Lx)\right\}},$$ where $f: {\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, $f(x)=\lambda_2 \|x\|_1 + \delta_{\left[0,1\right]^n}(x)$, $g_1:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$, $g_1(y)=\left\| y-b \right\|_1$ and $g_2:{\ensuremath{\mathcal{Y}}}\rightarrow {\ensuremath{\mathbb{R}}}$, $g_2(y,z)=\lambda_1 \left\| (y,z) \right\|_{\times}$. For every $p \in {\ensuremath{\mathbb{R}}}^n$ it holds $g_1^*(p)=\delta_{\left[ -1, 1\right]^n}(p)+p^Tb$ (see, for instance, [@Bot10]), while for any $(p,q)\in{\ensuremath{\mathcal{Y}}}$ we have $g_2^*(p,q)=\delta_{S}(p,q)$, with $S=\{(p,q) \in {\ensuremath{\mathcal{Y}}}: \|(p,q)\|_{\times *} \leq \lambda_1\}$.
We solved this problem by Algorithm \[alg1\] and to this end we made use of the following formulae for the proximal points involved in the formulation of this iterative scheme: $$\begin{aligned}
{\ensuremath{\text{Prox}}}_{\gamma f}(x)&=\operatorname*{arg\,min}_{z\in\left[0,1\right]^n}\left\{\gamma\lambda_2\|z\|_1 + \frac{1}{2}\|z-x\|^2\right\} = {\ensuremath{\mathcal{P}}}_{\left[0,1\right]^n} \left(x-\gamma \lambda_2 \mathbbm{1}^n\right) \ \forall x\in{\ensuremath{\mathbb{R}}}^n,\\
{\ensuremath{\text{Prox}}}_{\gamma g_1^*}(p) &= \operatorname*{arg\,min}_{z\in\left[-1,1\right]^n}\left\{ \gamma {\left\langle}z,b{\right\rangle}+ \frac{1}{2}\| z-p \|^2 \right\} = {\ensuremath{\mathcal{P}}}_{\left[-1,1\right]^n} \left(p-\gamma b\right) \ \forall p\in{\ensuremath{\mathbb{R}}}^n,\end{aligned}$$ and $$\begin{aligned}
{\ensuremath{\text{Prox}}}_{\gamma g_2^*}(p,q)&=\operatorname*{arg\,min}_{(y,z)\in S} \frac{1}{2}\|(y,z)-(p,q)\|^2 = {\ensuremath{\mathcal{P}}}_{S} \left(p,q\right) \ \forall (p,q)\in{\ensuremath{\mathcal{Y}}},\end{aligned}$$ where $\gamma \in {\ensuremath{\mathbb{R}}}_{++}$, $\mathbbm{1}^n$ is the vector in ${\ensuremath{\mathbb{R}}}^n$ with all entries equal to $1$ and the projection operator ${\ensuremath{\mathcal{P}}}_S:{\ensuremath{\mathcal{Y}}}\rightarrow S$ is defined as $$(p_{i,j},q_{i,j}) \mapsto \left(\frac{p_{i,j}}{\max\left\{1,\frac{\sqrt{p_{i,j}^2+q_{i,j}^2}}{\lambda_1}\right\}}, \frac{q_{i,j}}{\max\left\{1,\frac{\sqrt{p_{i,j}^2+q_{i,j}^2}}{\lambda_1}\right\}}\right).$$
Taking $x_0 \in {\ensuremath{\mathcal{H}}}$, $(v_{1,0},v_{2,0}) \in {\ensuremath{\mathbb{R}}}^n \times {\ensuremath{\mathcal{Y}}}$, $ \beta = \sqrt{1 + 8}=3$, $\varepsilon \in \left(0, \frac{1}{\beta +1} \right)$ and $(\gamma_n)_{n\geq 0}$ a nondecreasing sequence in $\left[ \varepsilon, \frac{1-\varepsilon}{\beta}\right]$, Algorithm \[alg1\] looks for this particular problem like $$\begin{aligned}
\left(\forall n\geq 0\right) \ \left\lfloor \begin{array}{l}
p_{1,n} = {\ensuremath{\mathcal{P}}}_{\left[0,1\right]^n} \left(x_n - \gamma_n \left( A^*v_{1,n} + L^* v_{2,n} + \lambda_2 \mathbbm{1}^n \right) \right) \\
p_{2,1,n} = {\ensuremath{\mathcal{P}}}_{\left[-1,1\right]^n} \left(v_{1,n} + \gamma_n (A x_n - b)\right) \\
p_{2,2,n} = {\ensuremath{\mathcal{P}}}_S \left(v_{2,n} +\gamma_n L x_n \right) \\
v_{1,n+1} = \gamma_n A( p_{1,n} - x_n) + p_{2,1,n} \\
v_{2,n+1} = \gamma_n L( p_{1,n} - x_n) + p_{2,2,n} \\
x_{n+1} = \gamma_n (A^*(v_{1,n}-p_{2,1,n}) + L^*(v_{2,n}-p_{2,2,n})) +p_{1,n}.
\end{array}
\right.
\end{aligned}$$ Figure \[fig:cameraman\] shows the original cameraman test image, which is part of the image processing toolbox in Matlab, the image obtained after multiplying it with the blur operator and adding after that normally distributed white Gaussian noise with standard deviation $10^{-3}$ and the image reconstructed by Algorithm \[alg1\] when taking as regularization parameters $\lambda_1=3\text{e-}3$ and $\lambda_2=2\text{e-}5$.
TV-based image inpainting {#subsectionInpaint}
-------------------------
In the last section of the paper we show how image inpainting problems, which aim for recovering lost information, can be efficiently solved via the primal-dual methods investigated in this work. To this end, we consider the following $TV$-$l_1$ model $$\begin{aligned}
\label{ex-inpaint}
\inf_{x\in{\ensuremath{\mathbb{R}}}^n}{\left\{ \lambda TV_{{\ensuremath{\text{iso}}}}(x) + \|Kx-b\|_1 + \delta_{\left[0,1\right]^n}(x)\right\}},\end{aligned}$$ where $\lambda \in {\ensuremath{\mathbb{R}}}_{++}$ is the regularization parameter and $TV_{{\ensuremath{\text{iso}}}}:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$ is the isotropic total variation functional and $K \in {\ensuremath{\mathbb{R}}}^{n \times n}$ is the diagonal matrix, where for $i=1,...,n$, $K_{i,i} = 0$, if the pixel $i$ in the noisy image $b\in {\ensuremath{\mathbb{R}}}^n$ is lost (in our case pure black) and $K_{i,i} = 1$, otherwise. The induced linear operator $K : {\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}^n$ fulfills $\|K\|=1$, while, in the light of the considerations made in the previous two subsections, we have that $TV_{{\ensuremath{\text{iso}}}}(x)=\|Lx\|_{\times}$ for all $x \in {\ensuremath{\mathbb{R}}}^n$.
Thus, problem can be formulated as $$\inf_{x\in{\ensuremath{\mathbb{R}}}^n}{\left\{ f(x) + g_1(Lx) + g_2(Kx)\right\}},$$ where $f:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\overline{\mathbb{R}}}}$, $f(x)=\delta_{\left[0,1\right]^n}$, $g_1:{\ensuremath{\mathcal{Y}}}\rightarrow {\ensuremath{\mathbb{R}}}$, $g_1(y_1,y_2)=\|(y_1,y_2)\|_{\times}$ and $g_2:{\ensuremath{\mathbb{R}}}^n \rightarrow {\ensuremath{\mathbb{R}}}$, $g_2(y)=\|y-b\|_1$. We solve it by Algorithm \[alg1\], the formulae for the proximal points involved in this iterative scheme been already given in Subsection \[subsectionDeblur\]. Figure \[fig:fruits\] shows the original fruit image, the image obtained from it after setting to pure black $80$% randomly chosen pixels and the image reconstructed by Algorithm \[alg1\] when taking as regularization parameter $\lambda=0.05$.
[10]{}
H.H. Bauschke and P.L. Combettes. . CMS Books in Mathematics, Springer, 2011.
R.I. Boţ. . Lecture Notes in Economics and Mathematical Systems, Vol. 637, Springer-Verlag Berlin Heidelberg, 2010.
R.I. Bo[ţ]{}, S.M. Grad and G. Wanka. . Springer-Verlag Berlin Heidelberg, 2009.
L.M. Briceño-Arias and P.L. Combettes. A monotone + skew splitting model for composite monotone inclusions in duality. 21(4):1230–1250, 2011.
A. Chambolle. An algorithm for total variation minimization and applications. 20(1–2):89–97, 2004.
A. Chambolle and T. Pock. A first-order primal-dual algorithm for convex problems with applications to imaging. 40(1):120–145, 2011.
P.L. Combettes and J.-C. Pesquet. Primal-dual splitting algorithm for solving inclusions with mixtures of composite, Lipschitzian, and parallel-sum type monotone operators. 20(2):307–330, 2012.
B.C. Vũ. A splitting algorithm for dual monotone inclusions involving cocoercive operators. , 2011. <http://dx.doi.org/10.1007/s10444-011-9254-8>
C. Z[ă]{}linescu. . World Scientific, 2002.
[^1]: Faculty of Mathematics, Chemnitz University of Technology, D-09107 Chemnitz, Germany, e-mail: [email protected]. Research partially supported by DFG (German Research Foundation), project BO 2516/4-1.
[^2]: Faculty of Mathematics, Chemnitz University of Technology, D-09107 Chemnitz, Germany, e-mail: [email protected]. Research supported by a Graduate Fellowship of the Free State Saxony, Germany.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The solar system’s terrestrial planets are thought to have accreted over millions of years out of a sea of smaller embryos and planetesimals. Because it is impossible to know the surface density profile for solids and size frequency distribution in the primordial solar nebula, distinguishing between the various proposed evolutionary schemes has been historically difficult. Nearly all previous simulations of terrestrial planet formation assume that Moon to Mars massed embryos formed throughout the inner solar system during the primordial gas-disk phase. However, validating this assumption through models of embryo accretion is computationally challenging because of the large number of bodies required. Here, we reevaluate this problem with GPU-accelerated, direct N-body simulations of embryo growth starting from $r\sim$100 km planetesimals. We find that embryos emerging from the primordial gas phase at a given radial distance already have masses similar to the largest objects at the same semi-major axis in the modern solar system. Thus, Earth and Venus attain $\sim$50$\%$ of their modern mass, Mars-massed embryos form in the Mars region, and Ceres-massed objects are prevalent throughout asteroid belt. Consistent with other recent work, our new initial conditions for terrestrial accretion models produce markedly improved solar system analogs when evolved through the giant impact phase of planet formation. However, we still conclude that an additional dynamical mechanism such as giant planet migration is required to prevent Earth-massed Mars analogs from growing. **[Keywords:]{} Inner planets, Solar system planets, Planetary system formation, Solar system formation, Planetary science, Planetesimals, Protoplanetary disks**'
author:
- 'Matthew S. Clement, Nathan A. Kaib, & John E. Chambers'
bibliography:
- 'embryo.bib'
title: 'Embryo formation with GPU acceleration: reevaluating the initial conditions for terrestrial accretion'
---
Introduction
============
Since advances in computing power led to the widespread availability of open source symplectic integrators [@wisdomholman; @duncan98; @chambers99], numerous theoretical studies have been dedicated to understanding the origins of the solar system’s terrestrial architecture [for recent reviews on the topic see: @izidoro18_book_review; @ray18_rev]. Broadly speaking, a successful model for the inner solar system must reconcile the differences in mass distribution and orbital excitation between the modern system and the presumed primordial solar nebula [e.g.: @mmsn]. Of particular interest in the literature are the order of magnitude differences between the masses of Mercury and Mars and the neighboring Earth and Venus [@chambers01; @ray09a; @lykawka19], the four orders of magnitude mass disparity between the asteroid belt and planetary regimes [@obrien07; @izidoro15; @clement18_ab], and the plethora of high orbital eccentricities and inclinations in the asteroid belt [@petit01; @morby10; @deienno16]. While there are many compelling models and solutions to these issues, non-uniform disk conditions [@hansen09; @iz14], the influence of the giant planets [@walsh09; @ray09a; @lykawaka13; @bromley17] and different modes of accretion [@levison15; @morby15; @draz16] feature prominently in most proposed evolutionary scenarios. However, many models are based on an assumption that Moon to Mars massed planet-forming embryos were abundant throughout the terrestrial disk [@koko_ida02] that is in conflict with modern, high-resolution simulations of embryo formation [@carter15; @walsh19]. This is partially a consequence of the fact that, due to limits in computing power, the various phases of planet growth (planetesimal formation, embryo formation and the giant impact phase) are often treated separately. We briefly summarize each phase of accretion below [not discussed here, but still relevant for the solar system are gas accretion and planet migration; see reviews in: @morb12; @ray18_rev]:
Planetesimal Formation
----------------------
Large infrared excesses in observed proto-planetary nebulae [@brice01] imply that the majority of the solid mass in young planet forming disks is concentrated in sub-micron sized dust grains. Population studies of gaseous disk ages [@haisch01] indicate that they do not persist for longer than a few Myr. Because ages of iron meteorites indicate that their primitive parent bodies accreted rapidly [$\lesssim$5 Myr, apparently pre-dating the appearance of chondrules: @kleine05], the transition from dust to 10-100 kilometer-scale planetesimals must have occurred rapidly, while gas was still present in the system. Additionally, surveys of proto-$stellar$ disks indicate that they possess significantly higher dust masses than proto-$planetary$ disks [@tobin20]. This result seems to imply that the conversion of dust to larger solid bodies occurs rapidly, and in conjunction with the earliest epochs of star formation.
Dust grains can grow via various processes including coagulation [@xiang19], aggregation and compaction [@wetherill80_rev; @dominik07]. However, explaining growth beyond meter-scales [the so-called meter-barrier: @weidenschilling77a; @birnstiel12] is difficult because millimeter sized bodies experience significant aero-dynamic drag, causing them to orbit at sub-Keplerian velocities and migrate inward [@whipple72]. One intriguing solution to this issue might be direct formation via gravitational collapse. If dust particles are sufficiently concentrated relative to the gas, they can clump together and form $D\sim$100km planetesimals rapidly via gravitational instability [@johansen15]. While the “streaming instability” scenario offers a compelling resolution to the meter-barrier problem, the specific properties (radial location, final masses and formation time) of the resulting primordial generation of planetesimals are dependent on unconstrained disk parameters [@draz16; @carrera17; @abod19].
Embryo Formation
----------------
Planetesimals continue to grow larger throughout the gas disk phase by direct accretion of other planetesimals [@wetherill93; @koko_ida_98] and inward-drifting $\sim$meter-scale “pebbles” [the radial flux of which is still debated: @johansen10; @chambers16; @ida16]. This growth phase is highly efficient as long as the relative velocities in the region are low. The largest local planetesimal gravitationally focuses [@chambers06] the incoming flux of planetesimals and pebbles, causing runaway growth to ensue [@koko_ida_96]. Eventually, the oligarch planetesimals grow large enough to excite the orbits of nearby planetesimals and inhibit accretion [@koko_ida00]. Growth can continue beyond this stage if the pebble flux is great enough [@lambrechts14b], as pebbles are sufficiently small for gas disk interactions to damp their orbits. Eventually, pebble accretion shuts off when an embryo grows large enough [“pebble isolation mass:” @morb_nesvory_12] to induce a pressure gradient in the disk that prohibits inward pebble drift. Recent high-resolution simulations in @wallace19 uncovered an additional growth mode that occurs in the vicinity of massive oligarchs as small planetesimals are stacked inside of nearby first order mean motion resonances (MMR); thereby accelerating their growth towards intermediate masses and inducing a corresponding bump in the planetesimal size frequency distribution (SFD).
Giant Impact Phase
------------------
In the final phase of terrestrial growth, the embryo and planetesimal populations emerging from the gas disk collisionally assemble over $\sim$100 Myr timescales [@wetherill78; @chambers01; @ray09a]. By and large, numerical models of the “giant impact” phase report timescales for the epoch’s completion consistent with the geologically inferred timing of the Moon-forming impact [$\sim$ 50-150 Myr; @kleine09]. Moreover, it is within this ultimate stage of accretion that many authors have proposed solutions to the aforementioned mass and orbital excitation disparities in the inner solar system. Accounting for hit-and-run collisions [@chambers13; @clement18_frag] and dynamical friction induced by small bodies [@obrien06; @ray06] can result in analog terrestrial planets with more realistic, dynamically cold orbits. Furthermore, the giant planet instability [the “Nice Model” of @Tsi05; @levison08; @nesvorny12] is typically invoked to explain the asteroid belt’s excited state [@roig15; @deienno16; @deienno18] and (at least some of) its primordial depletion [@morby10; @clement18_ab]. Reconciling the Earth-Mars mass disparity, however, has led to the development of a multitude of different models. It should also be mentioned here that collisional fragmentation [@chambers13] plays a role in the giant impact phase of terrestrial growth, the degree to which is a topic of continued debate [for recent works espousing differing viewpoints, we direct the reader to: @clement18_frag; @deienno19; @kobayashi19]. Specifically, a large, fragmenting collision [@asphaug14] is a potentially viable explanation for Mercury’s large, iron-rich core [@jackson18].
Small Mars Problem
------------------
Solutions to the small Mars problem [@wetherill91] generally fall in to one of two categories. In the first class of models [e.g.: @hansen09; @ray17sci], the outer terrestrial disk is already depleted during the primordial gas-disk phase, and the terrestrial planets form out of a narrow annulus of material. In one such scenario [the so-called “Grand Tack” model of @walsh11; @jacobson14; @walsh16], the inner disk is truncated when Jupiter migrates into, and subsequently back out of the terrestrial region. Thus, Mars forms rapidly as a “stranded embryo” [@Dauphas11; @iz14]. In contrast, the second class of models invoke a dynamical mechanism to starve the region of material during the planet formation process [@ray09a]. Typically, the influence of Jupiter and Saturn perturbs objects in the region, and inhibits Mars’ formation. Resonance sweeping or crossing [@lykawaka13; @bromley17], primordially excited giant planet orbits [@ray09a; @lykawka19] and the Nice Model instability [@clement18] have all been shown to substantially restrict Mars’ growth. In particular, the “Early Instability” scenario argues that a Nice Model timed $\sim$1-5 Myr [@clement18_frag] after gas disk dispersal explains Mars’ rapid inferred geologic formation time [@Dauphas11; @kruijer17_mars]. However, each model should be considered in the appropriate context given the fact that Mercury’s low mass and orbit are still very low probability outcomes of numerical simulations [@ls14; @clement19_merc].
This Work
---------
With few exceptions, the aforementioned N-body studies of the giant impact phase all place large embryos throughout the terrestrial disk and modern asteroid belt. However, such initial conditions are at odds with semi-analytic predictions of oligarchic growth [@koko_ida_98; @koko_ida00], as well as recent high-resolution studies of embryo formation within gaseous disks [@carter15; @wallace19; @walsh19]. In particular, it appears unlikely that the primordial asteroid belt region attained such an advanced evolutionary state during the gas disk phase. In this paper we follow the complete growth of the terrestrial system starting from $r\sim$100 km planetesimals accreting in a decaying gas disk [@morishma10]. It should be noted that, as a tangentially related alternative to our proposed scenario, self-consistent pebble accretion (not considered in our work) simulations [e.g.: @morby15] form embryos directly throughout the terrestrial disk.
Our current study is perhaps most similar to the recent work of @carter15 and @walsh19, and we compare our results with both authors’ findings throughout this manuscript. While @carter15 used a parallelized N-body code [$PKDGRAV$: @pkdgrav1; @pkdgrav2] with inflated planet radii and @walsh19 employed a Lagrangian integrator and tracer particles [the $LIPAD$ code: @lipad], we opt for a direct, GPU accelerated N-body scheme [$GENGA$: @genga] that fully resolves close encounters. Notably, we investigate the effects of Jupiter and Saturn’s presence, and use self-interacting planetesimals for much of the runaway growth phase. While we leave the full-resolution evolution of our generated distributions of embryos and planetesimals through the giant impact phase to a future paper, we perform an additional suite of simplified simulations of the final stage of accretion for a first order approximation of the final system architectures.
Methods
=======
To achieve sufficiently high particle resolution throughout the terrestrial disk, we begin by numerically integrating different radial annuli separately. As oligarchic growth ensues in each annulus, we begin to merge our simulations; combining and interpolating between individual annuli until the entire terrestrial disk is assembled in a single simulation at the $t=$1 Myr. For each of the simulations that include gas drag, we use the GPU parallelized hybrid symplectic integrator $GENGA$ [@genga]. $GENGA$ is based on the $Mercury$ code of @chambers99, runs on all $Nvidia$ GPUs, and is available to the public in an open source format. In all of our simulations we employ a time-step equal to $\sim$5$\%$ that of the shortest orbital period, remove objects that pass within 0.1 au of the Sun [common practice in N-body studies of planet formation, see @chambers01], and consider particles ejected from the system at 100 au. We also incorporate a simple, analytic gas disk model [@morishma10] that mimics the effects of aerodynamic drag [@adachi76], $e/i$ damping induced by tidal interactions between the gas disk and proto-planets [@tanaka02], and the global nebular gravitational [@nagasawa00] force by applying an additional acceleration to each body after the Keplerian drift kicks. In this model, gas dissipates exponentially in time and uniformly in space as:
$$\Sigma_{gas}(r,t) = \Sigma_{gas}(1\medspace au,0)\bigg(\frac{r}{1\medspace au}\bigg)^{-\alpha}exp\bigg(-\frac{t}{\tau_{gas}}\bigg)$$
While previous authors [@morishma10; @walsh19] have investigated different gas densities and decay times, because we are constricted by the availability of GPUs, we limit our study to the nominal minimum mass solar nebula [MMSN: @hayashi81]. Thus, for all of our integrations, we use a decay time of $\tau_{gas}=$ 3 Myr, an initial surface density of $\Sigma_{gas}(1\medspace au,0)=$ 2,000 $g\thinspace cm^{-2}$, and set $\alpha=$1 [based on the nominal $\alpha$ of @morishma10] .
Runaway growth in the inner disk: 0-100 Kyr {#sect:meth_ann}
-------------------------------------------
We begin by following the evolution of five, 0.1$M_{\oplus}$ annuli located at 0.5, 1.0, 1.5, 2.0 and 3.0 au (see table \[table:ann\]) for 100 Kyr. Each annulus is composed of 5000, fully self-gravitating, equal-massed planetesimals on nearly circular, co-planar orbits [we draw initial eccentricities and inclinations from Rayleigh distributions with $\sigma_{e}=$0.002 and $\sigma_{i}=$0.2$^{\circ}$ as described in @ida90; @koko_ida_98]. Thus, assuming a nominal planetesimal density of 3.0 $g\thinspace cm^{-3}$, each object has D$\approx$200 km. Our annuli are derived from a terrestrial disk (0.5-4.0 au) that contains 5 $M_{\oplus}$ of solid material with a surface density profile that falls off radially as $r^{-3/2}$ [consistent with studies of terrestrial planet formation: @wetherill96; @chambers01; @ray09a; @walsh19]. To account for edge effects, we employ a boundary condition similar to @koko_ida00. If a particle’s semi-major axis passes beyond 0.005 au of either annulus edge, it is removed from the simulation and a new object is added at the opposite edge with inclination and eccentricity drawn from the annulus’ in situ distributions. The largest error introduced by this choice of boundary condition is an artificial damping of orbits near the boundary edge because the objects with higher eccentricities are being removed. However, the gas-driven migration timescale for $r=$ 100 km planetesimals [for example: @weidenschilling77a] is greater than the nebular lifetime for our gas disk model [@morishma10]. Moreover, radial scattering due to close encounters is significantly subdued in this phase of our simulations as a result of eccentricities being highly damped. Therefore, in practice, this exchange of particles rarely occurs because orbits are highly damped.
Annulus $a_{in}$ (au) $a_{center}$ (au) $a_{out}$ (au) $N_{f}/N_{i}$
--------- --------------- ------------------- ---------------- ---------------
1 0.48 0.50 0.52 0.06
2 0.9745 1.0 1.0255 0.27
3 1.4685 1.5 1.5315 0.51
4 1.9635 2.0 2.0365 0.68
5 2.9505 3.0 3.0495 0.88
: Summary of annulus edges and centers for simulations of oligarchic growth from 0-100 Kyr. Each annulus contains 5,000 fully self-gravitating objects and has a total mass of 0.1 $M_{\oplus}$. Note that the different annulus widths are a result of the $\Sigma \propto r^{-3/2}$ surface density profile.[]{data-label="table:ann"}
Runaway growth in the outer disk: 100-1000 Kyr {#sect:meth_interpolate}
----------------------------------------------
{width=".9\textwidth"}
Since runaway growth ensues rapidly [@koko_ida_98] at small radial distances, the total number of particles in our $r=$ 0.5, 1.0 and 1.5 au annuli is small enough at $t=$ 100 Kyr ($N_{f}/N_{i}=$ 0.06, 0.27 and 0.51, respectively) to begin interpolating between annuli. We generate 7 wider annuli (table \[table:new\] and figure \[fig:interp\]) based on the mass and orbital distributions within our inner 3 annuli after 100 Kyr of evolution. Combined, these new simulations span the entire radial range of 0.48-1.65 au. To accomplish this interpolation, we first separate each annulus into three classes of particles: oligarch embryos (the most massive object in the annulus), proto-embryos (all other objects larger than 25 times the initial planetesimal mass), and planetesimals. These class divisions are purely used for our interpolation methodology, and do not affect how the particles are treated in the actual integration. In the intra-annulus regions, we assume that the percentage of total mass concentrated in embryos (both oligarchs and proto-embryos) and planetesimals is equal to the arithmetic mean of the corresponding percentages in the neighboring annuli. We then place new oligarch embryos in the region with linearly decreasing masses, and semi major axes that maintain a surface density profile proportional to $r^{-3/2}$. Inclinations and eccentricities for the new oligarch embryos are chosen at random from the original embryo distributions. Finally, we add new proto-embryos and planetesimals by randomly drawing masses, eccentricities and inclinations from the respective distributions, and semi-major axes that maintain $\Sigma_{disk} \propto r^{-3/2}$. Therefore, as radial distance increases in the intra-annulus regions, total embryo mass is concentrated in a greater number of smaller embryos. All three classes of objects (planetesimals, proto-embryos and embryos) interact with one another gravitationally for this phase of our study.
To verify the effectiveness of our interpolation method, we generate a new, artificial, annulus 2 utilizing the outputs of annuli 1 and 3 at $t=$ 100 Kyr. When we compare our new artificial annulus with the actual state of annulus 2 at 100 Kyr, we find that the two systems are remarkably similar. Annulus 2 contains 1,378 total particles at this stage of evolution, and our interpolation method creates a system of 1,354 particles. Additionally, our scheme overestimates the mass of the oligarch in annulus 2 by just 3.9$\%$. The ratio of total embryo (oligarchs and proto-embryos) to planetesimal mass in the actual annulus 2 is 0.54, as compared to 0.60 in our artificial system. The only manner in which the two systems are significantly dissimilar is in the mass distribution of the proto-embryos. Because we are interpolating over such a wide radial range, our artificial system contains around twice as many proto-embryos that are, on average, half as massive as those in the real annulus 2. As this issue is lessened when the intra-annulus distance is reduced, we argue that our interpolation prescription is an adequate means for accelerating our calculations while minimizing additional error terms.
Annulus $a_{in}$ (au) $a_{out}$ (au) $N_{part}$
--------- --------------- ---------------- ------------
1a 0.48 0.76 4984
1b 0.74 1.02 5106
2a 0.98 1.16 5622
2b 1.14 1.30 5123
2c 1.28 1.42 3852
2d 1.40 1.54 3758
3a 1.52 1.65 5355
: Summary of annulus edges and total particle numbers for simulations of embryo growth from 0.1-1.0 Myr.[]{data-label="table:new"}
Through a process of trial and error, we determine the largest radial bins $GENGA$ can efficiently integrate with fully-interacting particles. These seven new, slightly overlapping annuli (table \[table:new\] and figure \[fig:interp\]) are integrated for 900 Kyr utilizing the same gas disk and boundary conditions described in section \[sect:meth\_ann\]. Conversely, our two original outermost annuli (2.0 and 3.0 au) are integrated up to $t=$1 Myr as is.
Gas dispersal and the influence of Jupiter and Saturn: 1-3 Myr
--------------------------------------------------------------
![State of the terrestrial disk at 1 Myr; the point where we combine all annuli and add giant planets. At this stage, the system contains 43,608 total disk particles. The state of these systems at $t=$ 3 Myr is depicted in figure \[fig:ic\_comp\] The black line represents the total disk mass profile, while the red triangles plot each individual embryo with $M>$0.01 $M_{\oplus}$.[]{data-label="fig:1myr"}](1myr_disk){width=".5\textwidth"}
At $t=$1 Myr, we combine all annuli in one large simulation containing 43,608 disk particles using the same interpolation method described in section \[sect:meth\_interpolate\] (however, all objects with $a>$2.0 au and $M>$25$M_{init}$ are treated as proto-embryos). Figure \[fig:1myr\] plots the radial mass profile of our fully constructed terrestrial formation disk. Because we do not include the 3.0$<r<$4.0 au region of the disk in any of our $t=$ 0-1 Myr integrations given the lengthy timescales for collisions to occur at large radial distances, we approximate this section of the disk in these simulations with planetesimals five times as massive as our initial planetesimals. Since this region is quickly excited and eroded by perturbations from the giant planets following gas disk dispersal [e.g.: @ray06], we populate the region with unrealistically large asteroids only for the purposes of modeling the exterior mass’ effect on the interior disk regions. To further accelerate our calculation, we also treat all planetesimals ($M<25M_{init}$) as semi-interacting (interact gravitationally with the embryos, but not with one another) for this phase of our study [e.g.: @ray09a]. However, all particles still feel the effects of the decaying gas disk.
At this stage of analysis, we begin to consider the effects of the growing gas giants with 3 separate models:
- No giant planets (NOJS)
- Jupiter and Saturn each with $M=$ 8.0 $M_{\oplus}$ (8JS).
- Jupiter and Saturn begin with $M=$ 1.0 $M_{\oplus}$ and grow logarithmically to 95$\%$ their modern masses at $t=$ 3 Myr (GROW).
In each case, the giant planets are placed on near-circular orbits in mutual 3:2 [$a_{Jup}=$5.6 au, consistent with the planets’ presumed per-instability orbits, see @nesvorny11; @deienno17] MMR [e.g.: @lee02; @clement18]. We then integrate each system up to $t=$ 3 Myr as described above. While none of these giant planet mass configurations are akin to that of the actual solar system, we include them for the purposes of testing, to first order, Jupiter and Saturn’s effect on this phase of terrestrial evolution.
The giant impact phase: 3-200 Myr {#sect:nbody}
---------------------------------
While we plan to continue the full resolution study of our complete terrestrial disk (figure \[fig:1myr\]) in a future paper, we present a suite of simplified CPU simulations of the giant impact phase here to briefly comment on the implications of our generated disks. These integrations make use of the $Mercury6$ hybrid integrator [@chambers99], employ a 6 day time-step, and remove bodies with $r>$ 100 au and $r<$ 0.1 au. Modern massed versions of Jupiter and Saturn are placed in a 3:2 MMR as described above. All embryos (here $M\geq$ 0.01 $M_{\oplus}$) from our $GENGA$ simulations are included as fully interacting bodies. The remaining disk mass is replaced by 1,000, equal-mass, semi-interacting planetesimals with orbits drawn randomly from the remaining $GENGA$ particles such that $\Sigma_{disk} \propto r^{-3/2}$ is maintained. We perform 24 (as these are run on a cluster with 24 CPU cores per node) separate simulations in this manner for our NOJS, 8JS and GROW disks. Additionally, we completely remove all gas disk interactions in these simplified integrations. While a step change in the masses of Jupiter and Saturn and gas abundance at $t=$3 Myr is obviously not realistic, we present these simulations here to provide a zeroth order approximation of the final system architectures.
It should be noted here that the remnant planetesimal population can significantly affect the system’s evolution within the giant impact phase [see @ray06; @ray07; @jacobson14 where the ratio of total embryo to planetesimal mass is varied]. Therefore, our study is inherently biased by our initial planetesimal masses. If the first generation of planetesimals indeed formed large and rapidly [e.g.: @morby09_ast; @johansen15; @dermott18], then our $D=$ 200 km initial bodies might be realistic. Therefore, we conclude our study with a discussion of how the remnant planetesimal SFD can affect growing embryos with an additional suite of 50 simplified simulations of terrestrial accretion. These simulations are performed with the $Mercury6$ hybrid integrator as described above utilizing embryo and planetesimal distributions akin to those supposed in classic N-body studies of terrestrial planet formation [@chambers01; @chamb_weth01; @ray09a]. Each simulation assumes a 5$M_{\oplus}$ disk with half its mass concentrated in 50 equal-mass embryos, and the other 50$\%$ distributed equally between either 1,000 or 2,000 planetesimals (25 integrations each). Semi-major axes are selected to achieve $\Sigma_{disk} \propto r^{-3/2}$, while eccentricities and inclinations are drawn from Rayleigh distributions ($\sigma_{e}=$0.002 and $\sigma_{i}=$0.2$^{\circ}$).
Results and Discussion
======================
We present the results of our GPU-accelerated simulations of embryo growth in the subsequent sections \[sect:runaway\]-\[sect:ics\]. The following sections, \[sect:tp\_form\]-\[sect:control\], discuss the outcomes of our additional, CPU-only, simulations of the giant impact phase of terrestrial planet formation.
Oligarchic Growth {#sect:runaway}
-----------------
![Comparison of varied initial conditions for our $r=$ 0.5 au annulus. In the first four runs (color coded black, blue, red and green), the initial planetesimal eccentricities and total number of annulus particles are varied. In the final simulation (“clone,” grey line), the annulus is tripled in size once the total particle number drops by a factor of three.[]{data-label="fig:emb_compare"}](emb_time2){width=".5\textwidth"}
We begin our analysis with a brief validation of our initial conditions and methods. We perform four additional integrations of our innermost ($r=$ 0.5 au) annulus (plotted in figure \[fig:emb\_compare\]). In two simulations, we increase the RMS eccentricity of our planetesimals ($\bar{e}_{pln}= $0.01 and 0.02; red and green lines in figure \[fig:emb\_compare\], respectively) and verify that our results are independent of the particular initial orbits of the planetesimals. Indeed, both simulations experience nearly identical runaway growth sequences. This is because planetesimal orbits are rapidly damped to nearly zero eccentricity at the beginning of our simulations when the gas disk is particularly dense. Next, we perform an integration where the annulus is represented by 2,500 equal-massed objects, rather than 5,000. While the embryo growth sequence and planetesimal SFD in this run (blue line in figure \[fig:emb\_compare\]) are obviously different, we find that the net result at the end of the runaway growth phase is largely the same as in our nominal run in terms of $\bar{e}_{pln}$ and final embryo mass. Finally, we scrutinize the effectiveness of our boundary condition by tripling the size of our nominal annulus once the total number of particles decreases by a factor of three. This is accomplished by generating two, identical exterior annuli where the semi-major axis of each particle is shifted by the annulus width (0.04 au) while holding the other orbital elements constant. The additional dynamical friction of the new surrounding planetesimals has the immediate effect of briefly damping planetesimal eccentricities (grey line in figure \[fig:emb\_compare\]), however the net result of the oligarchic growth scheme in terms of $M_{iso}$ and $T_{grow}$ is the same as in our nominal run after 100 Kyr.
![Growth of the largest object in each of our five initial annuli. Oligarchic growth ensues rapidly in the inner disk, while our outermost annulus experiences few accretion events in the first 100 Kyr of evolution.[]{data-label="fig:emb_time"}](emb_time){width=".5\textwidth"}
The mass evolution from $t=$ 0-100 Kyr of the largest object in each of our five initial annuli (1-5, table \[table:ann\]) is plotted in figure \[fig:emb\_time\]. Runaway growth is self-limiting in the sense that it is effectively only dependent on the available mass to be accreted (a function of the local planetesimal surface density and, not considered here, the inward pebble flux) and the RMS eccentricity [e.g.: @lissauer87] in the region as:
$$\frac{1}{M}\frac{dM}{dt}\propto \Sigma_{pln} M^{1/3}\bar{e}_{pln}^{-2}$$
The growth timescale while in the runaway regime is
$$T_{grow} = \frac{M}{dM/dt}.$$
Taking the “kinetic gas” approach [@wetherill80], and assuming that dM/dt goes as
$$\frac{dM}{dt} = \pi R^{2} \rho_{pln} \bar{v}_{pln} \bigg[1 + \bigg(\frac{v_{esc}}{\bar{v}_{pln}}\bigg)^{2}\bigg],$$
(where $v_{esc}$ is the escape velocity at the surface of a growing embryo, $\rho_{pln}$ is the volume density of planetesimals and $\bar{v}_{pln}$ is their velocity dispersion) it can be shown that:
$$T_{grow} \simeq 2 \times 10^4 \bar{e}_{pln}^{2}\bigg(\frac{\Sigma_{pln}}{\Sigma_{o}}\bigg)^{-1}\bigg(\frac{M}{10^{26}\medspace g}\bigg)^{1/3}\bigg(\frac{a}{1\medspace au}\bigg)^{2} yr.
\label{eqn:tgrow}$$
where $\Sigma_{o} =10\thinspace g\thinspace cm^{-2}$ at 1 au [see @k_and_i95 for a full derivation]. Note that, here, the RMS planetesimal eccentricity is scaled by the reduced Hill radius:
$$h_{r} = \bigg(\frac{M}{3M_{\odot}}\bigg)^{1/3}
\label{eqn:miso}$$
The runaway growth regime is only relevant when $\Sigma_{pln}$ is large, and ample planetesimals are available to feed the growing embryo. Therefore, this relationship (\[eqn:tgrow\]) is typically cited in reference to growth towards the “isolation mass” [e.g.: @koko_ida02; @kobayashi13]:
$$M_{iso} = 0.14 \chi^{3/2}\bigg(\frac{a}{1.5 au}\bigg)^{3/4}$$
where $\chi$ is the scaling of the classic @hayashi81 MMSN. For $\chi=$1, this relationship predicts Mars massed embryos accreting at $\sim$1.5 au, and larger embryos in the asteroid belt region. In practice, however, the isolation mass is likely never reached in the outer terrestrial disk as the timescales for giant planet growth and gas dispersal are significantly shorter than $T_{grow}$ for a$\gtrsim$1.5 au. Thus, other dynamical processes likely begin perturbing this region long before $M_{iso}$ is reached. In contrast to the outer disk, accretion in our innermost annuli is indicative of runaway growth (equation \[eqn:tgrow\] and figure \[fig:added\]. For our purposes, however, we are most interested in the time required to accrete embryos of different masses as we seek to infer the conditions of the terrestrial disk around the time of nebular gas dispersal. Figure \[fig:powerlaw\] depicts this relationship for our five annuli for two different growth masses. Assuming no evolution in $\bar{e}_{pln}^{2}$, from equation \[eqn:tgrow\], we would expect $T_{grow}$ in our $\Sigma \propto r^{-3/2}$ disk to scale as $a^{2}$ with increasing radial distance. In our simulations, however, we find that it scales closer to $\sim a^{3}$. Additionally, growth towards larger masses (bottom panel of figure \[fig:powerlaw\]) is further curtailed in the outermost annulus and better fit by an $\sim a^{3.5}$ radial dependency. These results are largely consistent with previous studies [@koko_ida02; @chambers06; @kobayashi13], and analytical derivations incorporating the radial dependencies of $\bar{e}_{pln}$ (mainly a result of gas dynamics) and the isolation mass [e.g.: @koko_ida02 infer an $a^{2.7}$ dependency]. In a recent study similar to our current work, @walsh19 report that $T_{grow}$ towards $M_{iso}$ scales as $\sim a^{3.6}$ in nominal MMSN models without collisional grinding. Since $M_{iso}$, increases with radial distance (equation \[eqn:miso\]), our measured times to reach a fixed embryo mass in different radial annuli are indicative of an even steeper scaling of $T_{grow}$ with $a$, particularly in our outermost annulus. This is, at least partially, a result of our inclusion of the giant planets in two of our 1-3 Myr simulations. While figure \[fig:powerlaw\] plots the time to reach $M=M_{moon}$ for annulus 5 (table \[table:ann\]) as the average of all 3 simulations, we note that this time $\sim$600 Kyr shorter in our NOJS run than in our other two runs. Thus, perturbations from the growing giant planets, though significantly damped in the gas disk phase, are still sufficient to moderately excite orbits and limit accretion events in the asteroid belt region.
![The same as figure \[fig:emb\_time\] for annulus 1, compared with analytical $dM/dt$ (equation \[eqn:tgrow\]) utilizing the in-situ values of $\Sigma_{pln}$, $M$ and $\bar{e}$. As $T_{grow}<< \tau_{gas}$ in annulus 1, accretion is rapid and well characterized by the runaway growth regime.[]{data-label="fig:added"}](emb_ecc.pdf){width=".5\textwidth"}
![The time required for a planetesimal to increase in mass by two orders of magnitude (top panel) and the time to accrete a Moon massed embryo (bottom panel) for each of our 5 annuli.[]{data-label="fig:powerlaw"}](powerlaw.pdf){width=".5\textwidth"}
Initial Conditions after Nebular Gas Dissipation {#sect:ics}
------------------------------------------------
Run $R$($a<$1 au) $R$(1$<a<$2 au) $R$(2$<a<$3 au)
------ --------------- ----------------- -----------------
NOJS 3.33 1.76 0.28
8JS 3.81 1.56 0.22
GROW 2.95 2.16 0.25
: Ratios of total embryo mass to total planetesimal masses ($R$) in different disk regions after 3 Myr of evolution for our three different simulations.[]{data-label="table:R"}
Our three runs (NOJS, 8JS and GROW) finish with an average total terrestrial disk mass of 4.3 $M_{\oplus}$. In general, our simulations predict $\sim$0.3 $M_{\oplus}$ embryos forming at $a<$1.0 au, Mars-massed objects accreting in the proto-Mars region, and several small, Moon-massed embryos growing in the inner asteroid belt region (see figures \[fig:mass\_comp\] and \[fig:ic\_comp\]). However, we find that the belt region is totally dominated by smaller planetesimals, rather than embryos, at this phase of evolution. Objects similar in mass to Ceres are quite prevalent throughout the asteroid belt (an average of 571 objects with $M_{Ceres}<M<10M_{Ceres}$), while each system contains only $\sim$5 larger, $\sim$Lunar-massed ($>$0.01 $M_{\oplus}$) embryos. Table \[table:R\] summarizes the ratio ($R$) of embryo ($M>$ 0.01 $M_{\oplus}$) to planetesimal mass at different locations within the terrestrial disk at $t=$ 3 Myr. The disparity between our generated $R$ values throughout the disk and assumptions of N-body studies [@chambers01; @ray09a; @clement18] is an important takeaway from our simulations because different bimodal make-ups can lead to different evolutionary outcomes in different radial regions. We address each zone individually in the following sections.
### Inner Disk
The innermost section of our disks, where $M_{iso}$ is approached rapidly, are heavily depleted of planetesimals at $t=$ 3 Myr. This can have significant implications on the follow-on evolution of Earth and Venus analogs within the giant impact phase. In general, higher $R$ values lead to less dynamical friction generated by planetesimals. As swarms of planetesimals can damp the orbits of the growing planets (discussed further in section \[sect:control\]), this can result in final systems of planets with unrealistically large eccentricities and inclinations. However, @lykawka19 analyzed terrestrial growth within various evolutionary schemes using initial values of $R=$ 1, 4 and 8. The authors noted that high-$R$ disks were typically more successful at yielding small Mars analogs, replicating late veneer accretion on the Earth, and generating Mercury-Venus pairs. Furthermore, @jacobson14 found that employing a high-$R$ disk was an effective mechanism for delaying the Moon-forming impact and thus providing an adequate match to the amount of material delivered to form the late veneer. Therefore, a high-$R$ disk would be advantageous if another mechanism were capable of limiting the eccentricities and inclinations of Earth and Venus.
### Mars Region
In the Mars region, slightly elevated $R$ values [$R\simeq$ 2.0 as opposed to $R=$ 1.0 often assumed in the literature: @chambers01; @ray09a; @clement18] can potentially alter final system outcome within an early Nice Model evolutionary scheme. @clement18 argued that instabilities timed $\sim$1-10 Myr after nebular gas dispersal are most successful at limiting Mars’ mass (compared with earlier instabilities) because the higher $R$ values achieved at more advanced evolutionary stages leads to levels of dynamical friction that are insufficient to save material from loss during the instability. Therefore, in an early instability scenario, the higher values of $R$ in the Mars region (table \[table:R\]) of our simulations might lead to greater mass loss [@clement18_ab] and improved outcomes in a scenario where the giant planet instability is the Mars mass-depletion event. Moreover, @lykawka19 found that higher-$R$ disks ($R=$ 4 or 8, as opposed to $R=$ 1) were about twice as likely to produce a Mars analog with the correct mass and orbital offset from Earth.
### Asteroid Belt
Consistent with equation \[eqn:tgrow\], the primordial asteroid belt planetesimals are relatively unprocessed at $t=$ 3 Myr. This is slightly more pronounced in our simulations that include the giant planets (however, we find this trend to be weak and inconclusive; table \[table:R\]). While we would expect our results in the inner disk to be somewhat independent of our selection of initial planetesimal mass (as $T_{grow} \lesssim \tau_{gas}$), our asteroid belt results are significantly biased by our initial conditions. Thus, as accretion events in the asteroid belt are rare in our simulations, our selected primordial asteroid sizes are strongly preserved in the final SFD. While it should be noted that we do not test this hypothesis by varying $M_{init}$, if the initial asteroids indeed formed large [@morby09_ast; @johansen15; @dermott18], our results indicate that the region would be dominated by a few Lunar-massed asteroids, and dozens of Ceres-massed objects when the nebular gas dissipated. However, it should be noted that these results are specific to our assumption of a “heavy” primordial belt and a uniform $\Sigma \propto r^{-3/2}$ disk [see, for example, @iz14; @izidoro15 for an analysis of different disk profiles].
Planetesimals are occasionally implanted into the belt from the inner disk regions in our simulations. The largest radial migration of a Ceres-massed asteroid from $t=$ 1-3 Myr in any of our simulations is $\sim$1.1 au, and each system implants an average of $\sim10^{-4}M_{\oplus}$ worth of material originating with $a<$ 1.5 au in the belt. Therefore, if Vesta had formed in the inner terrestrial region [as suggested by its composition: @bottke06_nat; @mastro17], our results imply that its implantation in the belt could have occurred during the gas disk phase of evolution [however the scarcity of such events in our simulations would indicate that this is unlikely given the subsequent depletion of the belt: @obrien07; @clement18_ab]. An example of this type of planetesimal scattering from our GROW simulation is plotted in figure \[fig:qaq\]. In that run, a small planetesimal originating at $a=$1.38 au experiences a series of close encounters with 3 large embryos in the proto-Mars region that drive it’s aphelion well in to the asteroid belt. While the planetesimal’s orbit still crosses that of one of three embryos at $t=$ 3 Myr, it is possible that it could be further scattered onto a stable orbit in the asteroid belt during the giant impact phase [e.g.: @sandine19].
![Perihelia and aphelia vs. time for three growing embryos in the Mars region (plotted in different shades of blue; each begins with a masses between that of the Moon and Mercury at $t=$ 1.0 Myr, and grow to around a Mars mass at $t=$ 3.0 Myr) interacting with and scattering a small planetesimal (pink line) into the asteroid belt. The proto-Earth embryo is plotted in orange. The horizontal red lines correspond to the modern semi-major axes of Mars and the asteroid belt’s inner edge[]{data-label="fig:qaq"}](Qaq){width=".5\textwidth"}
### Comparison with Previous Work
![Embryo (only objects with $M>0.01M_{\oplus}$ are plotted) distributions at $t=$ 2 Myr for our 3 different simulations compared with results from @carter15 and @walsh19 (grey and black circles, respectively).[]{data-label="fig:mass_comp"}](2Myr){width=".5\textwidth"}
![SFD at $t=$ 2 Myr for the inner ($a<$ 1.5 au, left panel) and outer (1.5 $<a<$ 3.0 au, right panel) regions of the terrestrial disk, compared with results from @carter15 and @walsh19 (grey and black lines, respectively).[]{data-label="fig:sfd"}](2Myr_sdf){width=".5\textwidth"}
We compare the state of our terrestrial disk at $t=$ 2 Myr with previous work by @carter15 and @walsh19 in figures \[fig:mass\_comp\] and \[fig:sfd\]. In both figures, we consider the results of each paper’s nominal, MMSN calculations [using the nomenclature of @carter15 this is the high-resolution, calm disk]. For reference, the simulation from @carter15 does not include the giant planets, while @walsh19 place 1 $M_{\oplus}$ versions of Jupiter and Saturn at $a=$ 3.5 and 6.0 au that are instantaneously moved to their modern masses and pre-instability orbits [$a=$ 5.0 and 9.2 au: @levison11] at $t=$4.0 Myr. Additionally, the simulation of @carter15 begins with $D=$ 196-1530 km planetesimals drawn from a SFD power-law of $dN = m^{-2.5}dm$ and simplifies collisions by inflating radii by a factor of 6. In contrast, @walsh19 use initial planetesimals drawn from a distribution centered around $r=$ 30 km. Because our simulations begin with $D\simeq$ 200 km, equal-mass planetesimals, it is not surprising that our final SFDs (figure \[fig:sfd\]) are relatively steep, with a tail superimposed by runaway growth.
We note significant differences between our results, and those of both previous studies. Three major factors contribute to these disparities: our wider initial planetesimal disk, our intra-annulus interpolation method, and our direct treatment of collisions and close encounters without collisional fragmentation. While @carter15 and @walsh19 study more narrow disk regions (0.5-1.5 and 0.7-3.0 au; respectively), our work considers the entire radial range of 0.48-4.0 au (table \[table:new\]). Perhaps the largest difference between our respective final embryo distributions (plotted in figure \[fig:mass\_comp\]) is the prevalence of Moon-Mars massed embryos in the 1.5-2.0 au region, and Moon massed embryos in the inner asteroid belt in our simulations. These differences are mostly a result of our larger initial planetesimals. Nominal simulations in @walsh19 begin with $r\simeq$ 15 km planetesimals placed throughout the disk. Since 2 Myr is significantly less than $T_{grow}$ at $a>$ 2.0 au, our results are more biased by the larger planetesimal sizes than they are by the outcomes of runaway growth. This is further evidenced by low $R$ values in the asteroid belt in our simulations. The difference in initial planetesimal sizes is also fossilized in our final SFDs (figure \[fig:sfd\]), thus resulting in an over abundance of $\sim$100-500 km planetesimals compared with @carter15 and @walsh19.
In general, our final SFDs significantly more indicative of runaway growth than are those of @carter15 and @walsh19. We speculate that this is a result of our treatment of collisions and close-encounters (without inflating planetary radii or utilizing tracer particles). Thus, once the total planetesimal number decreases, growing embryos in our simulations begin to both gravitationally focus smaller planetesimals onto collision courses, and heat up the local velocity dispersion through scattering events.
Our method of interpolating between annuli likely artificially accelerates growth, and we find this effect to be most consequential in the 0.5 $<a<$ 1.0 region. When we begin to interpolate at $t=$ 100 Kyr, the oligarch embryo in the annulus 1 ($r=$ 0.5 au) is about five times larger than annulus 2’s ($r=$ 1.0 au) oligarch. As we linearly interpolate between these regions when laying new oligarch embryos (rather than logarithmically), embryos in the middle of the intra-annulus regions are boosted in mass [relative to those of, say, @walsh19]. As this is also true for annuli 2 and 3, this has the cumulative effect boosting the total mass concentrated in embryos in the inner terrestrial disk relative to those of @carter15 and @walsh19.
@carter15 and @walsh19 each incorporate algorithms designed to account for the effects of collisional fragmentation by introducing new “fragment” particles (or tracers) when imperfect collisions occur [e.g.: @leinhardt12; @stewart12]. Conversely, our work treats all collisions as perfectly accretionary. It is difficult to assess the degree to which our results differ from those of models considering collisional fragmentation without a consistent set of control runs. It is also unclear whether collisional fragmentation plays a significant role in altering the final distribution of embryo masses because generated fragment particles can obviously be re-accreted later in the simulation; thus lengthening the accretion timescale while resulting in a similar final system architecture. Indeed, @deienno19 concluded that energy dissipation occurring during fragmenting, embryo-embryo collisions does not contribute to significant differences of final system structure in terrestrial planet formation (both in terms of orbital excitation and planet mass). However, other authors using different numerical implementations have reached the opposite conclusion [eg: @bonsor15; @clement18_frag; @kobayashi19]. Thus, one could argue that the results of computational investigations of fragmentation are dependent on the specific numerical approach taken (and therefore more sophisticated models are required to study the problem). While a complete analysis of imperfect accretion is beyond the scope of this work, we cannot discount collisional fragmentation as a potential contributor to the observed differences in embryo masses and planetesimal SFDs in figures \[fig:mass\_comp\] and \[fig:sfd\]. Therefore, growth towards larger embryo masses is potentially artificially accelerated in our simulations compared to those of @carter15 and @walsh19 (the degree to which is unclear).
We also note that our disk populations are significantly more bimodal in mass than @carter15. At first glance, this seems to imply that our simulations are significantly further evolved within the runaway growth phase. However, simulations in @carter15 begin with planetesimal sizes ranging from 196-1530 km in order to resemble a more advanced stage of oligarchic growth. Thus, this range of primordial sizes seems to persist in the simulations of @carter15 through the $t=$ 2 Myr point. Therefore, the differences between our respective SFD’s can be interpreted as a fossilization of the initial planetesimal population. In section \[sect:control\], we speculate further about how this fossilized initial planetesimal distribution might effect the giant impact phase.
### New Distributions for Giant Impact Studies
![Comparison of all objects with $M>$0.001$M_{\oplus}$ in our GROW simulation (red dots) with a typical distribution of 100 equal massed embryos used in N-body studies [black dots, e.g.: @chambers01; @obrien06; @clement18], and the modern terrestrial planets (open circles). The size of each point is proportional to the object’s mass.[]{data-label="fig:ic_comp"}](embryo_grow_ic_comp.pdf){width=".5\textwidth"}
In figure \[fig:ic\_comp\] we plot the state of our GROW simulation at $t=$ 3 Myr, compared with typical initial conditions used in simulations of the giant impact phase [e.g.: @chambers01]. In our simulations, the inner disk ($r\lesssim$ 1.5 au) attains quite an advanced evolutionary state prior to nebular gas dispersal, and the system already resembles some aspects of the modern terrestrial architecture. In fact, the largest two embryos in our 8JS simulation (a proto-Venus analog at $a=$ 0.6 au and a proto-Earth at $a=$ 0.9 au) each possess masses of 0.39 $M_{\oplus}$. The consequences of such a distribution of embryos for the post-gas disk phase of giant impacts is obvious. As described in @walsh19 (we expand upon this further in section \[sect:tp\_form\]), the ultimate phase of terrestrial assembly proceeds as a late instability. Thus, rather than a hundred or so embryos accreting over hundreds of giant impacts, around a dozen embryos experience just a handful of massive impacts as they continue to accrete small bodies over hundreds of Myr. Furthermore, such an evolutionary scheme might be consistent with planetary differentiation models [@rubie15] that suggest that Venus’ lack of an internally-generated magnetic dynamo implies that primordial stratification in its core was never disrupted and mixed by a late giant impact [@jacobson17b].
Fully Evolved Systems {#sect:tp_form}
---------------------
Set $N_{TP}$ $AMD/AMD_{SS}$ $RMC/RMC_{SS}$ $M_{Mars}$
--------------- ---------- ---------------- ---------------- ------------
NOJS 5.1 0.73 0.29 0.63
8JS 6.4 0.55 0.29 0.40
GROW 5.5 1.04 0.28 0.60
Classic, 1000 4.4 5.6 0.41 0.76
Classic, 2000 4.3 5.8 0.37 0.43
For a first order approximation of our $t=$ 3 Myr systems’ evolution up to $t=$ 200 Myr, we perform an additional suite of simplified, CPU integrations where the planetesimal population is approximated with 1,000 equal-mass objects. We provide a summary of important statistics for each set of runs in table \[table:tp\_form\] [commonly cited as “success criteria” for terrestrial planet formation models. See for example: @ray09a; @clement18; @izidoro18_book_review]. The evolution of these systems within the giant impact phase is strikingly different from that of the “classic” model for terrestrial planet formation [e.g.: @wetherill80; @chambers01]. Most notably, our new distributions of large, $\sim$0.1-0.4 $M_{\oplus}$ embryos struggle to combine into systems of four, larger terrestrial planets. The largest embryos accrete the remaining planetesimals (as well as the occasional smaller embryo), however the embryo systems seldom destabilize fully and experience a final series of giant impacts with one another. Thus, the resulting systems contain too many terrestrial planets that are systematically under-massed. Indeed, the mean number of planets with $a<$ 2.0 au and $m>$ 0.05 $M_{\oplus}$ among our 8JS simulations using new initial conditions is 6.4, as opposed to 4.3 in our simulations that employ “classic” initial conditions (discussed further in section \[sect:control\]). Furthermore, because our new systems evolve only slightly over 200 Myr, the final terrestrial planets largely maintain the dynamically cold orbits that were originally damped via interactions with the gas disk. To demonstrate this, we calculate the normalized angular momentum deficit [AMD: @laskar97] and radial mass concentration statistics [RMC: @chambers01] for each system:
$$AMD = \frac{\sum_{i}M_{i}\sqrt{a_{i}}[1 - \sqrt{(1 - e_{i}^2)}\cos{i_{i}}]} {\sum_{i}M_{i}\sqrt{a_{i}}}
\label{eqn:amd}$$
$$RMC = MAX\bigg(\frac{\sum_{i}m_{i}} {\sum_{i}m_{i}[\log_{10}(\frac{a}{a_{i}})]^2}\bigg)
\label{eqn:rmc}$$
Figure \[fig:amd\] plots the cumulative distribution of system AMDs for our 72 simulations that are based off the results of our GPU simulations, compared with 50 control simulations that make use of classic initial conditions. Given the limited number of large accretion events experienced in our GPU-derived simulations, the final terrestrial architectures consistently provide better matches to actual inner solar system in terms of system $AMD$. We hesitate to conclude that this result implies a potential solution to the terrestrial over-excitation problem given the poor solar system analogs produced by our integrations. Specifically, we consistently form under-massed Earth and Venus analogs that are too great in number, and over-massed Mars analogs that are also overabundant. Nevertheless, the result of final system $AMD$ being limited in systems where embryos attain a more advanced evolutionary state in the gas phase is intriguing, and an avenue for future development and study.
![Cumulative distribution of normalized AMDs (equation \[eqn:amd\]) for three separate batches of terrestrial planet formation simulations. The black lines plot simulations that employ “classic” initial conditions [e.g.: @chambers01; @obrien06; @clement18] where half of the disk mass is placed in 50 equal-massed embryos and either 1,000 (solid line) or 2,000 (dashed line) equal-massed planetesimals. The red line represents depicts the results of 72 simulations using embryo distributions generated from the GPU simulations presented in this work. The grey vertical line denotes the solar system AMD for Mercury, Venus, Earth and Mars.[]{data-label="fig:amd"}](amd.pdf){width="50.00000%"}
At first glance, it would appear that the reason for the stunted growth of our systems of larger embryos is the presence compact MMR chains that develop as a result of aerodynamic drag induced migration during the gas disk phase [a less extreme version of the “breaking of chains” model for compact systems of Super-Earths described in @izidoro17]. Indeed, many of the large embryos in our GPU simulations finish near the major first order MMRs. As an example, the 7 largest embryos with $a<$ 2.0 au in our GROW simulation lie just outside of a mutual 3:2,3:2,5:3,4:3,4:3,4:3 chain. However, on closer inspection, these proto-planets are not in resonance, nor do they fall into resonance during the giant impact phase. Instead, the dearth of massive accretion events in these simulations can be attributed to the high $R$ values in the inner disk, in conjunction with a relatively wide spacing between embryos [as opposed to, say, @koko_ida00]. Since the large embryos emerge from the gas disk on low eccentricity orbits well outside one another’s mutual Hill spheres, with very little planetesimal mass available to perturb them on to crossing orbits, these systems routinely remain stable for 200 Myr. An example of such a system from the 8JS set is plotted in figure \[fig:example\]. Given the modern eccentricities of Mercury ($e=$ 0.21) and Mars ($e=$ 0.09), the degree of orbital excitation in this system is remarkably low, and a typical outcome of our study. However, Earth and Venus only attain $\sim$70$\%$ of their modern masses, and four additional planets 2-3 times the mass of Mars are stable in the system (one in the region between Earth and Venus, and three in the Mars region).
The total mass of planets in the Mars region (as well as that of the largest Mars analog; table \[table:tp\_form\]) is clearly too large in our simulations. However, this is somewhat expected given that our systems grow from a uniform, 5 $M_{\oplus}$ disk of material. Thus, we do not account for early depletion in the Mars-forming and asteroid belt regions that might have resulted from giant planet migration [@walsh11] or a primordial gap [@ray17sci], nor do we consider the dynamical excitation of the giant planets [which is highly efficient at limiting the mass of Mars, e.g.: @ray09a; @lykawaka13; @bromley17]. Nevertheless, it is obvious that if the Earth and Venus analogs in figure \[fig:example\] each accreted one of the additional $\sim$0.2-0.3 $M_{\oplus}$ embryos, perhaps ejecting an additional Mars analog in the process, the final system would provide a much better match to the modern Venus-Earth-Mars architecture.
![An example final terrestrial system from the 8JS batch (note that this example is chosen specifically to highlight the failure of this model). In order of increasing semi-major axis, the 6 terrestrial planets in the system have masses of 0.62, 0.26, 0.69, 0.38, 0.35 and 0.25 $M_{\oplus}$, respectively.[]{data-label="fig:example"}](comp.pdf){width=".5\textwidth"}
### Implications for the Moon-forming impact
It follows naturally to speculate that, given our results, a dynamical trigger might be required to destabilize such a compact system terrestrial embryos, eject additional Mars analogs, form the Moon, and complete terrestrial planet formation. The logical trigger would be the giant planet instability [@Tsi05; @levison08; @deienno17]. Several recent authors have invoked an early instability rather than a “delayed” instability coincident with the Late Heavy Bombardment [@tera74; @gomes05]. @morb18 argued that, since the crystallization of the Moon’s mantle took longer than the Earth’s, sequestration of highly siderophile elements (HSEs) during the crystalization process can explain the observed Earth-Moon HSE disparity, and is consistent with an early instability. Indeed, the early impact chronology on the Moon and Mars appears consistent with an instability occurring within 100 Myr of the solar system’s birth [@mojzsis19; @brasser20] Additionally, @nesvorny18 showed that the instability must have occurred within 100 Myr of nebular gas dispersal in order to permit the survival of the Patroclus-Menoetius binary system of Jupiter Trojans. Furthermore, a delayed dynamical event in the outer solar system is at odds with the recently discovered asteroid families by @delbo17 and @delbo19 in the inner main belt that are inferred to be as old as the solar system [e.g.: @milani17]. Finally, an early instability is also capable of limiting Mars’ accretion [@clement18], and adequately exciting [@deienno18] and depleting [@clement18_ab] the asteroid belt. However, @clement18_frag requires a very specific timing for the instability (1-5 Myr after gas dissipation) to limit Mars’ mass and prevent the terrestrial disk from “re-spreading” and forming 3-4 equal-mass planets (however this issue might be less pronounced in a higher-$R$ disk, see discussion in section \[sect:ics\]). Imposing such a strict constraint on the instability’s occurrence is problematic in that it conflicts with recent studies of Neptune’s effect on the Kuiper Belt [@nesvorny16] that aim to explain the inclination distribution of the 3:2 MMR population [@nesvorny15a]. Specifically, @nesvorny15b requires that Neptune migrate smoothly for $\sim$20 Myr before experiencing a “jump” in semi-major axis [though recent work by @volk19 suggests that other timescales are also viable]. Thus, we propose that, assuming more realistic initial distributions of embryos in the terrestrial forming disk (as produced via our high-resolution GPU simulations) might provide greater flexibility in terms of the timing of the giant planet instability for the early instability scenario proposed in @clement18. Indeed, we find that Mars analogs already have masses of order $\sim$0.1 $M_{\oplus}$ after the gas disk phase (figure \[fig:emb\_compare\]). In several of our simulations of the giant impact phase, many Mars analogs do not grow larger beyond the $\sim$10$\%$ level over 200 Myr (31$\%$ of the planets in the region accrete no additional embryos after $t=$10 Myr). Thus, it seems reasonable, given our simulation results, that a system of 4-5 Mars-massed planets formed during the gas disk phase [consistent with the hypothesis that Mars is a “stranded embryo” given its rapid inferred accretion timescale: @Dauphas11; @kruijer17_mars] could remain stable for some tens of Myr before disrupted by the giant planet instability.
Our generated embryo configurations also imply a mass ratio between the proto-Earth and the Moon-forming impactor (Theia) closer to unity. This result is interesting given that a giant impact involving two $\sim$0.5 $M_{\oplus}$ bodies has been shown to be successful at replicating the observed isotope ratios [@canup12]. @kaibcowan15 found these conditions to be highly improbable within dynamical simulations of terrestrial planet formation that invoke classic [@chambers01] initial conditions. While beyond the scope of our present manuscript, the implications of our GPU-evolved embryo populations are nonetheless intriguing with respect to the Moon’s formation.
In summary, the results of our simplified simulations that follow the evolution of our GPU-generated embryo distributions within the giant impact phase lead us to speculate that an alternate evolutionary sequence might have ensued during the ultimate phase of terrestrial assembly in the solar system. However, given the simplicity of the numerical simulations presented in this paper, we leave the full development of this scenario to future work. In short, our results indicate that the giant impact phase might have played out as a delayed instability, as proposed in a similar study by @walsh19. This starkly contrasts the rather prolonged sequence of hundreds of giant impacts that is modeled throughout much of the literature [e.g.: @obrien06; @fischer14; @lykawka19]. Because the interior regions of our terrestrial disks achieve extremely high $R$ values and possess well-spaced orbital configurations during the nebular gas phase, we propose that a dynamical trigger such as the Nice Model instability is necessary to stimulate the destabilization of the primordial proto-planets in the inner solar system. In such a scenario, the instability would have to be responsible for both triggering the final few giant impacts on Earth and Venus [most importantly the Moon-forming impact, e.g.: @quarles15; @kaibcowan15] and evacuating the $\sim$1.3-2.0 au region of additional Mars-massed planets [@clement18].
Dependency on Planetesimal Sizes {#sect:control}
--------------------------------
Left behind after the processes of embryo formation and oligarchic growth is a remnant of the initial planetesimal size distribution. While understanding the properties of the first generation of planetesimals is still an area of active research [e.g.: @levison15; @draz16; @wallace17], we argue that the SFD of the residual planetesimals can influence the final system AMDs. As a proto-planet grows within a swarm of smaller planetesimals, it undergoes a constant series of close-encounters that tend to reduce the system’s AMD. Thus, the $e/i$ evolution of a growing planet in a planetesimal disk can be thought of as a random-walk of encounters with a net trend towards damping the planet’s orbit. It follows that, with a smaller number of larger planetesimals, it is possible to randomly walk towards lower vales of $e/i$ and smaller AMDs. We demonstrate this concept in figure \[fig:test\] with a simple numerical experiment using $Mercury$ [@chambers99]. In each simulation, we embed a 1 $M_{\oplus}$ planet at 1 au within a distribution of planetesimals with a total mass of 1 $M_{\oplus}$. We place the large planet on a moderately excited initial orbit ($e=$0.1, $i=$5.0$^{\circ}$), and in all cases the orbit is markedly damped after 1 Myr. However, simulations using a smaller number of large planetesimals experience significantly greater damping, and display a larger range of outcomes than those with a greater number of small planetesimals [despite the total planetesimal mass remaining fixed. See also: @jacobson14; @kobayashi19].
We continue to test this concept with an additional suite of 50 simulations (see table \[table:tp\_form\]) of the classic terrestrial planet formation model [e.g.: @chambers01 described in detail in section \[sect:nbody\]]. In 25 simulations, the planetesimal population is modeled using 1,000 objects, each with mass $M=$ 0.0025 $M_{\oplus}$. Our second set of 25 simulations considers 2,000 planetesimals with $M=$ 0.00125 $M_{\oplus}$. Each batch of simulations finish with nearly identical mean $AMD$ values (0.0101 and 0.0104, respectively, see figure \[fig:amd\]) that are $\sim$6 times that of the modern solar system. However, the set of simulations employing fewer, large planetesimals has a greater dispersion of $AMD$ outcomes (min$_{1000}=$ 0.0014, max $_{1000}=$ 0.024, $\sigma_{1000}=$ 0.0072, min$_{2000}=$ 0.0052, max$_{2000}=$ 0.018, $\sigma_{2000}=$ 0.0035, see figure \[fig:amd\]). Thus, a terrestrial system forming within a distribution of larger planetesimals is able to randomly walk to both lower and higher AMD values. If we are to define “success” as satisfying a constraint 50$\%$ of the time [@nesvorny12], most terrestrial formation models [with the notable exception of Grand Tack evolutionary schemes; @walsh11; @walsh16] struggle to consistently replicate the solar system’s low AMD. As we find smaller populations of more massive planetesimals to be more successful at producing low AMD terrestrial planets, we argue that a primordial terrestrial disk of $r\sim$100 km planetesimals (albeit not akin to the planetesimals used in our $N_{pln}=$1,000 simulations) formed directly via gravitational instability [e.g.: @morby09_ast; @johansen15] is worth investigating.
![Minimum eccentricity (top panel) and inclination (bottom panel) attained by a 1.0 $M_{\oplus}$ embryo embedded in a disk of $N_{pln}$ equal massed planetesimals for 1.0 Myr. $N_{pln}$ is varied in each simulation while the total mass of planetesimals is fixed at 1.0 $M_{\oplus}$.[]{data-label="fig:test"}](test.pdf){width="50.00000%"}
Conclusions
===========
We present detailed simulations of embryo formation within a decaying gas disk starting from $r\sim$100 km planetesimals. Our calculations begin by following oligarchic growth within individual radial annuli. As the total particle number decreases in each annulus, we interpolate within the intra-annulus regions, and assemble the entire terrestrial disk (0.48$<$a$<$4.0 au) in a single simulation after 1 Myr. Thus, our results are somewhat biased by our interpolation method (though we find this error term to be minor). Specifically, future work should employ a logarithmic means of interpolating between different annuli, rather than a linear one. There are several important takeaways from our work, and that of other recent high-N studies of embryo formation [@carter15; @walsh19; @wallace19].
Bimodal Makeup Depends on Radial Location
-----------------------------------------
We show that the ratio of total mass in embryos to the total mass in planetesimals ($R$) existing around the time of gas-disk dispersal varies strongly with semi-major axis. In the Earth/Venus-forming region we find $R$ values as high as $\sim$4.0, as compared with more moderate ratios ($\sim$2.0) in the proto-Mars region, and low concentrations of embryos ($R\simeq$ 0.20) in the primordial asteroid belt. We argue that the different values of $R$ in each disk region can lead to a substantial differences in system outcome during follow-on evolution. For instance, high $R$ values in the inner disk have been shown to increase the probability of forming Venus/Mercury analogs [@lykawka19], while more moderate ratios in the Mars-forming region can potentially help limit Mars’ final mass in an early Nice Model scenario [@clement18].
Few Giant Impacts in the Giant Impact phase
-------------------------------------------
Perhaps the most striking difference between our generated distributions of embryos and planetesimals [as well as those from similar studies: @carter15; @walsh19] and those supposed in classic terrestrial formation models [@chambers01; @ray09a] is the advanced evolutionary state attained in the $a\lesssim$1.0 au region during the nebular gas phase. In our simulations, only a handful of reasonably large (0.1 $\lesssim M \lesssim$ 0.4 $M_{\oplus}$) embryos grow in the Earth and Venus forming regions of the disk. Therefore, the giant impact phase of evolution ensues as a delayed instability [e.g.: @walsh19], with Earth and Venus experiencing only a few giant impacts en route to attaining their modern masses. Given the limited growth experienced by such embryos in an additional, simplified suite of simulations of the giant impact phase, we speculate that a dynamical trigger [the Nice Model instability: @Tsi05; @nesvorny12; @clement18] is required spur on the ultimate series of impacts in the inner solar system.
Large Primordial Planetesimals Generate More Extreme AMDs
---------------------------------------------------------
Finally, we perform an additional suite of test simulations to demonstrate the effects of the fossilized primordial planetesimal SFD on final terrestrial angular momentum deficit (AMD). Planetesimal-embryo encounters tend to damp the orbits of proto-planets via a random walk towards lower eccentricities, inclinations, and total system AMDs. We show that larger encounters generated from a distribution of fewer, more massive planetesimals allow a system to randomly walk towards both higher, and lower values of AMD. Thus, we speculate that a primordial generation of massive planetesimals [$r\sim$100 km, formed via gravitational instability @morby09_ast; @johansen15] might be advantageous in the ultimate giant impact phase of terrestrial assembly in terms of more consistently yielding systems with solar-system like final AMDs.
Our GPU simulations required nearly two years to complete on $NVIDIA$ GK110 (K20X) “Kepler” accelerators, and represent close to the highest contemporaneous resolution achievable with a direct N-body algorithm. We have shown that the primordial sizes of planetesimals are somewhat fossilized at the end of the gas disk phase; therefore implying that the selection of a particular initial particle mass can lead to significant differences in final system outcomes. Thus, it is imperative that future authors continue to push the limits of particle resolution as advances in computing power make such endeavors feasible.
Acknowledgments {#acknowledgments .unnumbered}
===============
M.S.C. and N.A.K. thank the National Science Foundation for support under award AST-1615975. NAK also acknowledge support under NSF CAREER award 1846388. This research is part of the Blue Waters sustained-petascale computing project, which is supported by the National Science Foundation (awards OCI-0725070 and ACI-1238993) and the state of Illinois. Blue Waters is a joint effort of the University of Illinois at Urbana-Champaign and its National Center for Supercomputing Applications [@bw1; @bw2]. Further computing for this project was performed at the OU Supercomputing Center for Education and Research (OSCER) at the University of Oklahoma (OU). This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1548562. Specifically, it used the Bridges system, which is supported by NSF award number ACI-1445606, at the Pittsburgh Supercomputing Center [PSC: @xsede]. Additional computation for the work described in this paper was supported by Carnegie Science’s Scientific Computing Committee for High-Performance Computing (hpc.carnegiescience.edu).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study quantum fidelity, the overlap between two ground states of a many-body system, focusing on the thermodynamic regime. We show how drop of fidelity near a critical point encodes universal information about a quantum phase transition. Our general scaling results are illustrated in the quantum Ising chain for which a remarkably simple expression for fidelity is found.'
author:
- 'Marek M. Rams$^{1,2}$ and Bogdan Damski$^1$'
title: Quantum fidelity in the thermodynamic limit
---
A quantum phase transition (QPT) happens when dramatic changes in the ground state properties of a quantum system can be induced by a tiny variation of an external parameter [@sachdev]. This external parameter can be a strength of a magnetic field in spin systems (e.g. Ising chains [@Coldea2010etal] and spin-1 Bose-Einstein condensates [@sadler2006etal]), intensity of a laser beam creating a lattice for cold atom emulators of Hubbard models [@Hubbard_exp], or dopant concentration in high-Tc superconductors [@Lee2006]. At the heart of the sharp transition lies non-analyticity of the ground state wave-function across the critical point separating the two phases. QPTs, traditionally associated with condensed matter physics, are nowadays intensively studied from the quantum information perspective (see e.g. [@Osterloh2002]).
Quantum fidelity – also referred to as fidelity – is a popular concept of quantum information science defined here as the overlap between two quantum states $${\cal F}(g,\delta) = |\langle g-\delta|g+\delta\rangle|,
\label{def_F}$$ where $|g\rangle$ is a ground state wave-function of a many-body Hamiltonian $\hat H(g)$ describing the system exposed to an external field $g$, and $\delta$ is a small parameter difference. It provides the most basic probe into the dramatic change of the wave-function near and at the critical point [@Zanardi2006].
The recent surge in studies of fidelity follows discovery that quantum criticality promotes decay of fidelity [@Zanardi2006]. This is in agreement with the intuitive picture of a QPT: as system properties change dramatically in the neighborhood of the critical point, ground state wave-function taken at different values of the external parameter – $|g-\delta\rangle$ and $|g+\delta\rangle$ – have little in common and so their overlap decreases.
As fidelity is given by the angle between two vectors in the Hilbert space, it is a geometric quantity [@Zanardi_geometric2]. Thus, it has been proposed as a robust geometric probe of quantum criticality applicable to all systems undergoing a QPT [*regardless*]{} of their symmetries and order parameters whose prior knowledge is required in traditional approaches to QPTs. Fidelity has been recently studied in this context in several models of condensed matter physics (see [@Gu2008] and references therein).
Besides being an efficient probe of quantum criticality, fidelity appears in numerous problems in quantum physics. Indeed, it is related to density of topological defects after a quench [@BDPRL2005; @barankov; @polkovnikov], decoherence rate of a test qubit interacting with an out-of-equilibrium environment [@BD2009], orthogonality catastrophe of condensed matter systems (see [@Anderson1967] and the references citing it). Therefore its understanding has an interdisciplinary impact.
To unravel the influence of quantum criticality on fidelity, one has to determine if its drop near the critical point encodes universal information about the transition in addition to providing the location of the critical point. This universal information is given by the critical exponents and reflects symmetries of the model rather than its microscopic details. In the “small system limit”, which broadly speaking corresponds to $\delta\to0$ at fixed system size $N$, the answer is positive. This is explored in the fidelity susceptibility approach where [@Zanardi2006; @You2007; @Gu2008] $${\cal F}(g,\delta) \simeq 1 - \delta^2\chi_F(g)/2,
\label{def_sus}$$ and $\chi_F$ stands for fidelity susceptibility. Universal information, or simply the critical exponent $\nu$, is encoded in its scaling: at the critical point $\chi_F(g_c) \sim N^{2/d \nu}$, while far away from it $\chi_F(g) \sim N|g-g_c|^{d\nu-2}$, where $d$ is system dimensionality [@ABQ2010; @barankov; @polkovnikov].
In the thermodynamic limit, which broadly speaking corresponds to $N\to\infty$ at fixed $\delta$, the answer is positive as well. This is our key result stating that $$\ln{\cal F}(g,\delta) \simeq -N|\delta|^{d\nu} A\left(\frac{g-g_c}{|\delta|} \right),
\label{near_F}$$ where $A$ is a scaling function. In particular, we see that fidelity is non-analytic in $\delta$ at the critical point, $\ln{\cal F}(g_c,\delta) \sim -N|\delta|^{d\nu}$, while away from it, i.e., for $|\delta| \ll |g-g_c| \ll 1$, we obtain $$\ln{\cal F}(g,\delta) \sim -N\delta^2 |g-g_c|^{d\nu -2},
\label{away_fid}$$ after expansion of the scaling function. These results, in particular, set firm foundations for usage of fidelity as a probe of quantum criticality in thermodynamically-large systems. In the context of theoretical studies of QPTs, the strength of the fidelity approach lies in its simplicity: all information encoded in the ground state wave-function(s) is “compactified” into a single number. A competing approach for extraction of the exponent $\nu$ – study of the asymptotic decay of correlation functions to obtain the correlation length – is considerably more complicated. Below we illustrate these predictions on the paradigmatic model of quantum phase transitions, the Ising chain, and discuss the scaling theory that leads to (\[near\_F\]) and (\[away\_fid\]).
The Hamiltonian of the one dimensional Ising chain reads [@sachdev] $$\hat H(g) =
-\sum_{i=1}^N(\sigma^x_i\sigma^x_{i+1} + g \sigma^z_i),$$ where $g$ stands for a magnetic field acting along the $z$ direction. Above the spin-spin interactions try to enforce $\pm x$ polarization of spins, while the magnetic field tries to polarize spins along its direction ($+z$ for $g>0$). This competition results in two critical points at $g_c=\pm1$: the system is in the ferromagnetic (paramagnetic) phase for $-1<g<1$ ($|g|>1$). The critical exponent $\nu=1$. This model is solved by mapping spins onto non-interacting fermions via the Jordan-Wigner transformation [@sachdev].
![(color online) Fidelity of the Ising chain near the critical point as a function of (a) the system size $N$ at fixed $\delta=10^{-4}$ and (b) parameter difference $\delta$ at the fixed system size $N=10^5$. On both panels the curves from top to bottom correspond to ${\cal F}(1,\delta)$, ${\cal F}(1+\delta, \delta)$ and ${\cal F}(1+5\delta,\delta)$. []{data-label="fig1"}](fig1new.eps){width="\columnwidth"}
Behavior of fidelity (\[def\_F\]) around the critical point, $g\approx g_c$, is summarized in Fig. \[fig1\]. In Fig. \[fig1\]a the parameter difference $\delta$ is kept fixed and the system size is increased. For small system sizes we reproduce the known result, $\ln{\cal F} \sim -N^2$ [@Zanardi2006], resulting from finite size scaling effects (see e.g. [@ABQ2010; @barankov; @polkovnikov; @Gu2008]). For large system sizes, however, we obtain $\ln{\cal F}\sim -N$ in agreement with (\[near\_F\]) and the fidelity per site approach [@Zhou2008; @Zhou2008_Vidal; @Zhou2008_Ising]. As is shown in Fig. \[fig2\], the transition between the two regimes takes place when $$N|\delta|\sim 1,
\label{Ndelta}$$ which will be discussed below.
Similarly, we observe two distinct regimes when the system size $N$ is kept fixed and the parameter difference $\delta$ is varied (Fig. \[fig1\]b). For $N|\delta|\ll1$ we observe $\ln{\cal F}\sim-\delta^2$, in agreement with (\[def\_sus\]), while for $N|\delta|\gg1$ we find $\ln{\cal F}\sim -|\delta|$ in agreement with (\[near\_F\]). In the latter fidelity [*approaches*]{} non-analytic limit (where $\partial_\delta {\cal F}$ at $\delta=0$ is undefined) reflecting singularities of the ground state wave-function resulting from the QPT [@note].
![(color online) Study of the crossover between the “small system limit” and the thermodynamic limit illustrated in Fig. \[fig1\]. As the system size is increased in Fig. \[fig1\]a, the slope of the curves changes smoothly from $2$ (corresponding to $\ln{\cal F}\sim -N^2$) to $1$ (corresponding to $\ln{\cal F}\sim -N$). The crossover region between the two limits is centered around $N=N_{3/2}$ where the slope equals $3/2$. To find it, we have calculated numerically ${\cal F}(g,\delta)$ vs. $N$ – as in Fig. \[fig1\]a – for various $\delta$’s and found that $N_{3/2}|\delta|\sim1$. This is illustrated in this figure where data sets from top to bottom correspond to results obtained for $g=1$, $1+\delta$ and $1+5\delta$, respectively (similarly as in Fig. \[fig1\]a). The power-law fits (straight lines) to numerical data (crosses) give $N_{3/2}=a|\delta|^{-b}$, where $b= 0.995\pm0.003$ for all three fits, while the prefactor $a$ changes between the fits from $3.6$ to $0.3$. Similar analysis can be done on curves shown in Fig. \[fig1\]b providing the same result. Thus we conclude that the crossover condition reads $N|\delta|\sim 1$ near the Ising critical point. []{data-label="fig2"}](fig2new.eps){width="\columnwidth"}
We also see on both panels of Fig. \[fig1\] that all curves collapse for $N|\delta|\ll1$, while they stay distinct in the opposite limit. Thus, for $N|\delta|\gg1$ sensitivity of fidelity to quantum criticality is enhanced. This can be understood if we focus on Fig. \[fig1\]a: in the large $N$ limit dramatic changes in the ground state wave-function near the critical point are expected.
As analytical results for fidelity are scarce, we find it remarkable that we can derive accurate analytical description in the complicated limit of $N|\delta|\gg1$, where the Taylor expansion (\[def\_sus\]) fails. To proceed, we calculate ${\cal F}(1+\epsilon,\delta)$, where $\epsilon$ measures distance from the critical point. For the Ising chain ${\cal F}~=~ \Pi_{k>0}f_k$, where $f_k=\cos(\theta_+(k)/2-\theta_-(k)/2)$ and $\tan(\theta_\pm(k)) = \sin k/(1+\epsilon\pm\delta-\cos k)$. We stay close to the critical point so that $0\le|\delta|,|\epsilon|\ll 1$ and introduce natural parameterization: $c=\epsilon/|\delta|$. Taking the limit of $N\to\infty$ at [*fixed*]{} $\delta$ the product $\Pi_k f_k$ can be changed into $\exp(N\int dk \ln f_k/2\pi)$, which can be further simplified to $$\ln{\cal F} ~\simeq~ -N |\delta| A(c)
\label{fid_ising}$$ in the leading order in $\delta$ and $\epsilon$. This result is in prefect agreement with our universal scaling law (\[near\_F\]): note that $\nu,d=1$ in our model and $c=(g-g_c)/|\delta|$. Moreover, it agrees well with exact numerical simulations: Fig. \[fig3\]. Above $A(c)$ is given by $$A(c)= \left \{
\begin{array}{c}
\begin{split}
& \frac{1}{4}+ \frac{|c| K(c_1)}{2\pi}+\frac{(|c|-1) {\rm Im} E(c_2)}{4\pi} ; |c|\le1 \\
& \frac{|c|}{4} - \frac{|c| K(c_1)}{2\pi}-\frac{(|c|-1) {\rm Im} E(c_2)}{4\pi} ; |c|>1.
\end{split}
\end{array}
\right.
\label{Ac}$$ where $c_1 = -4 |c|/(|c|-1)^2$, $c_2=(|c|+1)^2/(|c|-1)^2$, and $K$ and $E$ are complete elliptic integrals of the first and second kind, respectively. Agreement between (\[Ac\]) and numerics is very good: see Fig. \[fig4\] for detailed comparison of $A(c)$ to numerics. Several interesting results can be obtained now.
![(color online) Fidelity ${\cal F}(g,\delta)$ of the Ising chain near the critical point: thermodynamic limit (main plot) vs. “small system limit” (inset). Main plot: black curve is our analytic approximation (\[fid\_ising\]), while red crosses come from numerics. Both were obtained for $N= 2\times 10^5$ and $\delta=10^{-4}$ ($N|\delta|\gg1$). Inset: numerical result for $N=10^3$ and $\delta=10^{-4}$ ($N|\delta|\ll1$). In the “small system limit” fidelity stays close to unity at any distance from the critical point, while in the thermodynamic limit it can interpolate between zero and unity. []{data-label="fig3"}](fig3new.eps){width="\columnwidth"}
First, Eq. (\[fid\_ising\]) shows analytically how the so-called Anderson catastrophe – disappearance of the overlap between distinct ground states of an infinitely large many-body quantum system [@Anderson1967] – happens in the Ising chain.
Second, Eq. (\[fid\_ising\]) explains the lack of collapse of the various curves providing fidelity around the critical point in the $N|\delta|\gg1$ limit. Indeed, fidelity calculated for two ground states symmetrically around the critical point is ${\cal F}(1,\delta)\simeq\exp(-N|\delta|/4)$, but if one of the ground states is obtained at the critical point, ${\cal F}(1\pm\delta,\delta) \simeq \exp(-N|\delta|(\pi-2)/4\pi)$. In the opposite limit of $N|\delta|\ll1$, ${\cal F}\simeq 1-\delta^2 N^2/16$ in both cases explaining the collapse of all curves in this limit in Fig. \[fig1\].
Third, there is a singularity in the derivative of fidelity when one of the states is calculated at the critical point: $d{\cal F}(g\pm\delta,\delta)/dg|_{g=g_c=1}$ is divergent when $N\to\infty$ at fixed $\delta$. This reflects singularity of the wave-function at the critical point approached in the thermodynamic limit. Quantitatively, $dA(c)/dc|_{c\to1^{\pm}}= \ln|1-c|/4\pi - 3\ln2/4\pi+(1\pm1)/8 +
{\cal O}((1-c)\ln|1-c|)$, which is logarithmically divergent at $c=1$ (Fig. \[fig4\]). This divergence is a signature of a pinch point found in [@Zhou2008; @Zhou2008_Vidal; @Zhou2008_Ising] when fidelity between two distinct ground states states was studied. The logarithmic divergence in the Ising chain was numerically observed in [@Zhou2008_Ising].
Last but not least, we obtain from (\[fid\_ising\]) a compact expression for fidelity away from the critical point. Taking $|c| \gg 1$ (but still $|\epsilon| = |c\delta| \ll 1$), $A(c)\simeq 1/16 |c|$ and so $${\cal F}~\simeq~\exp(-N\delta^2/16|\epsilon|),
\label{f_away}$$ in agreement with (\[away\_fid\]). This reduces to a known result for fidelity susceptibility when the argument of the exponent is small and so ${\cal F}\simeq 1-\delta^2N/16|\epsilon|$ (see e.g. [@Gu2008]), but provides a new result in the opposite limit where lowest order of the Taylor expansion is insufficient. We notice also that (\[f\_away\]) is analytical in $\delta$ even in the limit of $N\to\infty$: there are no singularities expected when the system is far away from the critical point.
![(color online) Upper plot: scaling function $A(c)$ of the Ising chain. The solid black line provides the analytic result (\[Ac\]), while the red crosses show numerics (i.e., $\ln {\cal F} / N |\delta| $). The inset highlights singularity at $c=1$. Lower plot: logarithmic divergence of $dA/dc|_{c=1}$ discussed in the text. The solid black line is the derivative of (\[Ac\]), while the red dashed line is a numerical result: the difference between the two near the pinch point at $c=1$ is due to the finite system size $N$ [@Zhou2008_Ising]. It disappears in the limit of $N\to\infty$. In both plots numerics is done for $N=10^5$ and $\delta=\pi 10^{-3}$. []{data-label="fig4"}](fig4new.eps){width="\columnwidth"}
Below we derive general scaling results (\[near\_F\]) and (\[away\_fid\]). This can be done by studying the scaling parameter $$\tilde d(g+\delta,g-\delta) = -\lim_{N\rightarrow \infty } \ln{\cal F}(g,\delta)/N,$$ introduced in [@Zhou2008] in the context of fidelity per site approach to the thermodynamic limit. We expect that this limit is reached when $$\min[\xi(g+\delta), \xi(g-\delta)] \ll L,
\label{tlimit}$$ where $\xi(g)$ is the correlation length at magnetic field $g$ and $L$ is the linear size of the system ($N=L^d$ for a $d$-dimensional system). Indeed, the smaller of the two correlation lengths sets the scale on which the states entering fidelity “monitor” each other (\[def\_F\]). In particular, it explains our results showing that the thermodynamic limit is reached even when one of the states is calculated at the critical point and so its correlation length is infinite. Near a critical point (\[tlimit\]) is equivalent to $L|\delta|^\nu\gg1$ [@remark]. For the Ising chain studied above it reads $N|\delta|\gg1$ properly predicting the crossover condition (\[Ndelta\]) obtained from numerical simulations (Fig. \[fig2\]).
Generalizing the scaling theory of second order QPTs (Sec. 1.4 of [@ContinentinoBook]), we propose the following scaling ansatz for the universal part of the scaling parameter $$\tilde d(g_c+\epsilon+\delta,g_c+\epsilon-\delta)=b^{-d} f((\epsilon+\delta)b^{1/\nu},(\epsilon-\delta)b^{1/\nu}),$$ where $f$ is the scaling function, $b$ is the scaling factor, and $\nu$ is the critical exponent providing divergence of the coherence length $\xi\sim |g-g_c|^{-\nu}$. The scaling function depends on both $\epsilon+\delta$ and $\epsilon-\delta$ as they are renormalized simultaneously. The factor $b^{-d}$ appears for dimensional reasons. Scaling of $\epsilon+\delta$ and $\epsilon-\delta$ is given by scaling of the correlation length $\xi(\epsilon\pm\delta)=b \xi((\epsilon\pm\delta)b^{1/\nu}) $.
Taking $g=g_c+\epsilon$, introducing natural parameterization $\epsilon = c |\delta|$, and fixing the scale of renormalization through $|\delta|b^{1/\nu}=1$ we obtain $\tilde d(g+\delta,g-\delta)=|\delta|^{d\nu} f(c+1,c-1)$. It gives (\[near\_F\]) after setting $f(c+1,c-1)=A(c)$. In a general context, (\[near\_F\]) shows how universal part of the scaling parameter causes the Anderson catastrophe near a critical point.
The scaling function $A(c)$ can be simplified away from the critical point. We assume below $\epsilon,\delta>0$ for simplicity, take $\delta\ll \epsilon \ll 1$, and set $b$ through $(\epsilon+\delta)b^{1/\nu}=1$ exploring the freedom to choose the renormalization scale. Simple calculation results in $\tilde d(g+\delta,g-\delta)=(\epsilon+\delta)^{d \nu} f(1,(\epsilon-\delta)/(\epsilon+\delta))$, where the second argument of $f$ is close to unity. Expanding $f$ in it we get $\tilde d(g+\delta,g-\delta)\approx 2 \delta^2 \epsilon^{d \nu -2} f''(1,x)|_{x=1}$ as $f(1,x)$ has a minimum equal to zero at $x=1$. Thus, away from a critical point we end up with (\[away\_fid\]). When the system is small enough, $N\delta^2 |\epsilon|^{d \nu -2}\ll 1$, but still in the thermodynamic limit (\[tlimit\]), we reproduce the known result for fidelity susceptibility $1-{\cal F} \sim \delta^2 N |\epsilon|^{d \nu -2}$ [@barankov; @polkovnikov; @ABQ2010]. Otherwise, (\[away\_fid\]) provides a new result.
On general grounds, one can expect that for systems with $d\nu\ge2$ non-universal (system-specific) corrections to the above scaling relations may be significant [@polkovnikov], which requires further investigation.
Summarizing, our work characterizes fidelity – a modern probe of quantum criticality – in the thermodynamic limit. We have derived, and verified on a specific model, new universal scaling properties of fidelity. These findings should be experimentally relevant as the first experimental studies of ground state fidelity have been already done [@Zhang_all].
This work is supported by U.S. Department of Energy through the LANL/LDRD Program.
[23]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, , ).
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, (), .
, , , (), .
, ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ** (, , ).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Gravity Induced Resonant Emission (*gire*) is a new phenomenon recently reported [@Antony2017], we have analyzed the solar emissions on the basis of *gire*. We have computed the EUV solar spectrum [@Aschwanden2005; @Harrison1995] using *gire* wave length as a part of studying coronal heating problem. We find all the solar observations converge in *gire* and the *gire* wave length is sufficient to study all the solar emissions.'
address:
- 'The Astronomical Observatory, Department of Physics, University of Kerala, India - 695033 '
- 'School of Pure & Applied Physics, Mahatma Gandhi University, Kottayam, Kerala, India-686560'
- 'Center for Fundamental Research and Computational Sciences, Thiruvanathapuram, Kerala, India - 695043. '
author:
-
-
title: 'Analysis of EUV Solar Spectrum with *gire*'
---
Introduction {#S-Introduction}
============
The presence of highly ionized heavy elements in solar corona was noticed by [@Edlen1943; @Edlen1945; @Grotrian1939], which led to the problem of coronal heating, one of the most haunting problem in solar physics for the last seven decades. Corona is the source of solar wind which directly affects the earth climate and space equipments also the quest for energy from fusion, compelled several curious minds in to this strange phenomenon. Untiring search of reasons by various missions for the last 7 decades made tremendous progress in this area. Numerous journal publications and books regarding the developments over this problem could be found [@Aschwanden2005; @Kenneth2009] also in the website of NASA data system. One of the major data in the coronal heating problem is about the details of heavier elements present in the corona, detected by CDS, EIT and TRACE (CDS, Coronal Diagnostic Spectrometer; EIT, Extreme-ultraviolet Imaging Telescope; TRACE, Transition Region And Coronal Explorer (spacecraft)), compiled as EUV Solar Spectrum. As a part of studying coronal heating problem, on the basis of gravity induced resonant emission (*gire*), we are analyzing the EUV spectrum and this analysis is done in the following manner
- First we list the important characteristics of *gire* wave length
- Establishing these characteristics over the observational data about the solar emission
- Generating EUV solar spectrum with *gire* wave length
- Analyzing the variables over the the generated EUV spectrum by *gire*
Characteristics of *gire*
=========================
The effect of gravity is more on heavier ions relative to electrons induces an electric field; this electric field is very effective at high vacuum condition. Due to this electric field lower hybrid (*lh*) oscillations are induced by the propagation of *em* waves in the perpendicular direction. This *lh* oscillation leads to the resonant absorption and re-emission of *em* waves is very effective in ultra vacuum condition. This is named as Gravity Induced Resonant Emission (*gire*). Solar corona is extremely vacuum; hence *gire*may emerge as a most probable solution to Coronal Heating problem. The *gire* frequency is given as $$\omega_r^2 = \frac{c^2 \omega_{l}^2}{g} \frac{n_0^{'}}{n_0}$$ Where $\omega_l = \frac{|eB|}{\sqrt{mM}}$ is the *lh* frequency; e, charge; B, magnetic field; m, electron mass; M proton mass; $ {n^{'}_{\circ}}$, is the number density of ions; $ {n_{\circ}}$ is the number density of electrons. The *gire* frequency is deduced from the resonant absorption of *em* waves in the *lh* oscillation; to list the characteristics of *gire*, we need to know the physics of cyclotron resonant damping in the *lh* oscillations and also we should analyze this over solar coronal conditions.
The *lh* oscillation is a coupled oscillation between the oppositely charge particles bound by Coulomb force when they are in cyclotron motion due to Lorentz force. The cyclotron motion of ions and electrons are in the opposite directions hence they mutually cross over by two times in every rotation in the *lh* oscillation (see figure 1) if the frequencies are same. Here the particles are connected by the Coulomb force (attractive); when they move towards, they get accelerated and this acceleration occurs by the neutralization of their own fields (+ve & -ve). Here the field energy of the particles is converted into kinetic energy and thus the energy is conserved.
Now with the increased kinetic energy they cross over (unless the electron falls into the ion or they collide), they feel a sudden deceleration due to the deficit of fields since they are moving against the attractive force (see the figure 1). At this point, the particles are in the desperate need of field energy to compensate the lost field. At this situation, they absorb the necessary field energy from incoming *em* waves to overcome the attractive force and move forward to meet at the other point. **Thus, in the first half of the orbital path, they convert their field energy in to kinetic energy by the neutralization of their own field; and in the other half, they absorb the field energy from the *em* waves which they lost by field neutralization. Thus for a complete orbital motion, the particles convert the *em* waves in to kinetic energy. This is one of the most beautiful mechanism of auto conversion of field energy in to kinetic energy in *lh* oscillation.**
Time Factor of Cyclotron Damping process in *lh* oscillation
------------------------------------------------------------
A very important point in the *lh* oscillation is the acceleration and deceleration times. As we have discussed earlier, the particles in the *lh* oscillation; for half of the orbital length, the particles are accelerated towards by the Coulomb force and in the other half they are moving against the attractive force. We know, when the particles in the attractive force, the acceleration time ($t_a$) would be much smaller than the deceleration time ($t_d$, since when they move towards the velocities are added up). Here the electrons move in relativistic speeds, therefore; $t_d >> t_a$; this means, most of the time the particles are in the deceleration mode i.e., in the resonant absorption condition.**Thus the particles in the *lh* oscillation, continuously absorbing the field energy from the *em* waves and converting it into kinetic energy; since the solar atmosphere is sufficiently rich with all type of radiation, the gravity induced *lh* oscillation is sufficient to heat ions to any required level of temperature.**
Orbital elongation and Field deficit in *lh* oscillation
--------------------------------------------------------
Here one more thing to be noted is that the rate of absorption of the field energy is different for electrons and ions. The resonant absorption of energy by the particles due to the field deficit is directly depend on the cyclotron frequency. The electrons rotate with high frequency while the ions at lower rate, since $\omega_e > \Omega_i$, therefore the rate absorption of *em* energy by the electrons are higher than the ions. The energy of the electrons increases first; then its Larmour radius increases to conserve the angular momentum. When the orbital radii of the electrons increases; this gives more acceleration time to the coupled ions and they are get accelerated to higher energy levels and hence the orbits also elongated. These elongations of orbits are mutually in the opposite directions because the ions are controlled by electric field (gravity induced) and the electrons by magnetic field (figure 2). In this situation if there is a change in the magnetic field affects the density of electrons rather the density of ions. We know the coronal magnetic field is highly dynamical and such dynamical oscillations of magnetic field is greatly modulated in the density of electrons which leads to the deficit of field to ions. And this field deficit is the basic mechanism of ion cyclotron resoanat damping in the *lh* oscillation; this is the most important part of this paper, we shall give little more explanation on the basis of coronal oscillations.
****
Coronal Oscillation and Electron Deficit
----------------------------------------
The oscillation of the coronal loop is a regular situation in the solar atmosphere. Here we are going to prove that this coronal oscillation is the secret of the mysterious coronal heating problem. The magnetic field in the solar atmosphere is produced by the convection currents (below the photosphere), which emerges from the foot loop of the photosphere and passes through the transition layer to corona and then bend back to another foot loop. Solar surface is highly turbulent due to pressure wave, which causes Kink and Sausage oscillations in the coronal loop. These oscillations passes through the loop and finally end up in the top of the loop purely as transverse oscillations (see figure 2a). These oscillations causes sudden reduction in the magnetic field spatially which is maximum at the top of the loop (where displacement is more and field strength is less). This reduction in the magnetic field intensity leads to the heavy loss of ionized particles and this loss is more to the electrons rather than to ions in the gravity induced *lh* oscillations. This will be clear if we analyze the orbitals of the particles which undergo gravity induced *lh* oscillations in the coronal loops. To see the actual situation in the loop, we shall super impose the orbitals of gravity induced *lh* oscillations (fig.1b) over the coronal loop. The fig.1b, shows the orbitals of the particles where the direction of **B** is normal to the plane of the paper. But the direction of **B** in coronal loop is parallel to the plain of paper and therefore we rotate the orbitals $\frac{\pi}{2}$ to make B parallel to the plain of the paper and now the view of orbitals parallel to **B** is shown in figure 1c. Now we can super impose the fig.1c over the coronal loop oscillations (fig.2a) and the combined figure is shown as fig.2b (kindly note that the fig.2b is symbolic sketch and the orbital dimensions are much exaggerated to match with coronal loop, even otherwise the boundary of B is not definite in the particle dimension).
****
Now it is easy to analyze from the fig.2b which depicts the real situation of particles motion in the *lh* oscillation over the coronal loop. We find the coronal loop displacement is parallel to the elongation of the electrons and normal to the motion of ions. This shows the electrons in the gravity induced lower *lh* oscillations have minimum critical thermal velocity could escape the Coulomb potential of ions. And this loss of electrons may depend on several factors like oscillation amplitude, time, **B**, Temperature and the density (more theory is needed). Thus the variation in the magnetic field does not affect the ions much but the electrons and this loss would be high for weak field situation (solar minimum, $B\approx 10-50$G). This shows that in each oscillation, millions of coupled electrons may be escaping from the coronal loop and create a heavy reduction in the density of electrons has been confirmed by several scientists @Banerjee1998 [@Doschek1997; @Spadaro1999; @Pekeris1948; @Murawski1998; @Murawski1994; @Murawski1993; @Roberts1994] that the electron density decreases in a logarithmic scale. Thus the fluctuations in the coronal loop causes a great loss in the density of electrons, most of the magnetic field loops end in solar corona where the field deficit is very high and this is the reason that we notice the emission is along the magnetic loops. The acceleration and elongation of the orbits of the particles continued until the electrons escape from the Coulomb potential of the ions or they may under go a head on collision. If the electrons escape from the Coulomb potential of ions, this will increase the acceleration time for ions and the ions will be accelerated to high energy. Whether it is electron loss or the head on collision with ions; in both the situations, the emission is possible.
**Now we understand that in the gravity induced *lh* oscillations, the electrons and ions are confronting twice during every cycle of rotation when the frequencies are same or several times ( practically $\Omega_e >> \Omega_i$) and therefore, *gire* has the highest possibility of becoming the mother of all the atomic transitions (more work needed in the possibility of various transitions)**.\
From the above discussions we can conclude that the Ion cyclotron damping in the lower hybrid oscillations is a continuous process and the absorption is maximum when the Larmour radius of ions and electrons are same ($r_{Li} = r_{Le}$. At this situations there are number of possible transitions could take place such Couloumb’s Interaction which leads to momentary emission of radiation from the ion. The dispersion relation of a two component plasma, which has been derived by solving the momentum and continuity equations of MHD theory. The two components in the plasma are ions and electrons, therefor the ion can be any ionized heavier elements and the electron gives only the charge neutrality. For getting a generalized formula suitable to all ionized element, we replace $ \Omega_i = \frac{Z}{\mu} \Omega_i $, where $\mu$, mass number; Z, degree of ionization. The general form of *gire* wavelength is given as $${\lambda}_r = 2 \pi [(\frac{\mu}{Z}) (\frac{g}{\omega_l})(\frac{n_{\circ}}{{n_{\circ}}^{'}})] ^{\frac{1}{2}}$$ Here we find the *gire* wavelength depend on seven variables and each variable is very much sensitive to the coronal plasma and we list important characteristics of *gire* wavelength below.
1. The *gira* wave length is deduced by using Magneto Hydro Dynamic (MHD) theory
2. *gira* is a non thermal process i.e., cyclotron resonance absorption by *lh* oscillation
3. Resonant absorption in *lh* oscillation depends on field deficit [@ref16]
4. For ions the field deficit depend on the decrease in the electron density
Now we will be discussing these listed properties with the established results of coronal emissions.
Observational Convergence over *gire*
=====================================
First we shall see, how MHD theory is suitable to solar physics problems.
MHD waves in UV Emissions
-------------------------
Solar atmosphere is highly dynamical and structured in density and fields, therefore the effect of spatial and temporal oscillations of fields and density cannot be avoidable. Since the solar oscillations are macroscopic in nature, describing the macroscopic dynamical structures of magnetic field embodied with plasma, only MHD theory could help. This is the reason that the MHD oscillation has been observable in all range of wavelengths (for complete detail see Ashwanden 2005). This study is over the EUV spectrum, therefore we are mentioning the observation of EUV range only [@Antonucci1984; @Aschwanden1999; @Aschwanden2002; @Chapman1972; @Nakariakov1999; @Nakariakov2001], for many situations the EUV and soft X-rays are observed together. The second characteristic of *gire* is it is a non thermal process.
MHD with Non Thermal Emission
-----------------------------
The morphological changes in coronal magnetic field produces local heating reported by several scientists [@Tarbell1999; @Tarbell2000; @Ryutova2000; @Ryutova2001]. Micro flares in the solar atmosphere causes MHD wave propagation which is modulated in to gyro resonance and non thermal micro wave emissions [@Gopalswamy1994; @Gopalswamy1997; @White1995; @Gary1997; @Nitta1997; @Berghmans2001; @Shimizu1992; @Shimizu1994; @Shimizu1995; @Shimizu1997a; @Shimizu1997b]. An analytical model on MHD turbulent showed that the turbulent velocities can be correlated with the excess broadening of lines observed by SUMER [@Heyvaerts1992; @Inverarity1995a; @Inverarity1995b; @Milano1997]. Various numerical simulation pointed out that the nonlinear wave dissipation could be linked with the radio emission which are morphologically sensitive in nature[@Ofman1999; @Drago1974; @Lantos1975; @Furst1975; @Kundu1976; @Chiuderi1977a; @Chiuderi1977b; @Chiuderi1999; @Dulk1977; @Papagiannis1982; @Wang1987; @Kundu1989; @Bogod1997; @Gopalswamy1999; @Moran2001] and also in the study on turbulent heating, a high degree of spatial and temporal intermittent has been observed[@Einaudi1996a; @Einaudi1996b; @Dmitruk1997] and the frequency distribution of this simulation was found to be in correlation with the nano flare distribution[@Dmitruk1998]. Small scale phenomenons such as prominences and striations in the transition and coronal region exhibits various signatures of all range of non thermal emissions [@Parnell2002a; @Parnell2002b; @Shimizu2002a; @Shimizu2002b; @Berghmans2002]. Finally we look in to the figure showing the integrated Energy spectrum of flare electrons taken by RHESSI (Ramaty High Energy Solar Spectroscopic Imager) , NASA, on 20 February 2002, clearly indicate the emission spectrum is fully non thermal and the Hinode [@Zirker1993] observation confirmed the resonant absorption.
Evidence of lower hybrid Cylcotron Resonant Damping
---------------------------------------------------
*lh* waves can simultaneously heat electrons and ions [@McBride1972], such waves are excited when the ions are drifting perpendicular to the magnetic field, these distributions are found in space plasmas are arising due to diamagnetic drift associated with density gradients [@Krall1971]. In formulating a unified nonlinear theory for *lh* waves, in the study of ion & electron damping process, it is found that *lh* wave causes direct transfer of *em* energy to both ion and electrons by damping process[@Tripathi1977; @McClements1993]. There are several studies in which the *lh* waves have been invoked to explain for the acceleration of electrons causing solar radio burst type I [@Spicer1981], type II [@Lampe1977; @Thejappa1987] type IV [@Vlahos1982], also @Benz1987 proposed that *lh* waves could responsible for acceleration of flare electrons.
Evidence of Density Dependance
------------------------------
In the study of Sausage Mode oscillations, scientists found that the density variation in the electrons easily modulate in to gyro synchrotron emission [@Rosenberg1970; @Ofman1997; @Ofman2000a] and also it found that the electron density of the plasma reduces tremendously above the foot points where the heating is more by **TRACE**, [@Aschwanden2000; @Aschwanden2001a; @Chae2002]. Study of polarized brightness over the coronal holes by Time Series Analysis revealed that the average density profile can be inverted with brightness [@Fisher1975; @Guhathakurta1996] this is in correlation with the **TRACE** data of electron density reduction over foot point. In the study of soft X-ray emission by OSO-7 and Skylab showed very high density fluctuations [@Timothy1975; @Hara1994; @Hara1996; @Hara1997; @Watari1995; @Foley1997; @Aschwanden2001b] and gyro resonance emission has also been observed around a symmetric sun spots [@Alissandrakis1980; @Krueger1985; @Brosius1989; @Lee1993a; @Lee1993b; @Vourlidas1997]. A study on the large flares confirmed the non thermal emission which reflected a density modulation in the EUV brightness variation [@Lin2001; @DeForest1998]. Study over prominences by numerical simulation insisted a need of dip in the top of the horizontal magnetic field lines to explain the energy balance (see the figure 2) [@Kippenhahn1957; @Pikelner1971; @Orrall1961; @Low1975a; @Low1975b; @Lerche1977; @Heasley1976; @Milne1979], we suppose such dip could cause high electron deficit which lead to the high absorption of *em* waves by *gira*. From the above three subsections, we can arrive at the following conclusion;
1. The MHD waves are observable in all range of *em* waves from Radio wave to Hard X ray and hence it would be the most suitable theory
2. Most of the emissions are non thermal confirms and gyro emission confirms the cyclotron resonant absorption is the true mechanism which accelerate all the ionized particles of corona.
3. Fluctuations in electron density inversely correlated with spectral emission (field deficit) confirms the field deficit is behind the spectral emission
4. The *lh* waves has been used to study of radio bursts to X-ray emissions shows the *lh* waves are sufficient for the study of coronal heating
All the observations and simulations converges in *gire* which may be the ultimate mechanism behind the coronal emission. Now let us move to the next section of generating EUV solar spectrum on the basis of *gira* wavelength.
Computation EUV Spectrum
========================
There are 6 variables in the expression of *gire* wavelength (Eq.1), among these the most important variable is density ratio and the next critical variable is temperature, but it is not visible in the formula.
Line Instr. $T_{iO}$ $T_{iC}$ $T_{eC}$ $E_e (keV) $ ${\gamma}$
------------ -------- ---------- ---------- ---------- -------------- ------------ --
He II$^*$ CDS 0.08 0.074 500 43.125 1.08
O III CDS 0.08 0.074 500 43.125 1.08
O IV CDS 0.20 0.202 600 51.75 1.1
O V CDS 0.25 0.233 400 34.5 1.06
Ne III CDS 0.08 0.072 600 51.75 1.1
Ne V CDS 0.30 0.29 600 51.75 1.1
Ne VI CDS 0.40 0.37 500 43.125 1.08
Ne VII CDS 0.50 0.42 400 34.5 1.06
Ne VIII CDS 0.63 0.57 400 34.5 1.06
N IV CDS 0.10 0.11 300 25.88 1.04
Mg VI CDS 0.40 0.375 600 51.75 1.1
Mg VII CDS 0.63 0.64 700 60.38 1.118
Mg VIII CDS 0.80 0.74 600 51.75 1.1
Mg IX CDS 1.00 0.96 600 51.75 1.1
Mg X CDS 1.10 1.0 500 43.125 1.08
Si VII CDS 0.63 0.64 800 69.00 1.135
Si VIII CDS 0.80 0.75 700 60.38 1.118
Si IX CDS 1.00 0.98 700 60.38 1.118
Si X CDS 1.00 1.04 600 51.75 1.1
Si X CDS 1.30 1.23 700 60.38 1.118
Si XII CDS 1.80 1.84 700 60.38 1.118
Ca X CDS 0.63 0.6 500 43.125 1.084
Al XI CDS 1.40 1.28 700 60.38 1.118
Fe VIII CDS 0.40 0.373 700 60.38 1.118
Fe IX$^*$ CDS 1.00 0.99 1300 112.13 1.22
Fe X CDS 1.30 1.25 1300 112.13 1.22
Fe XI CDS 1.30 1.27 1100 94.88 1.19
Fe XI CDS 1.30 1.27 1100 94.88 1.19
Fe XII$^*$ CDS 1.60 1.54 1100 94.88 1.19
Fe XIII CDS 1.60 1.64 1000 86.25 1.169
Fe XIV CDS 2.00 1.92 1000 86.25 1.169
Fe XV$^*$ CDS 2.00 1.98 900 77.63 1.15
Fe XVI CDS 2.50 2.55 1000 86.25 1.169
: Comparison of Calculated and Observed Temperatures[]{data-label="T-simple"}
Temperature of Ionized Particles
--------------------------------
Even though the *gire* wavelength does not depend on temperature, but the information is contained in the relativistic factor ($\gamma$) of electron in $\omega_l$. Therefore the temperatures of the coupling electrons with the ions in the *lh* oscillations to be found out for the calculation of $\gamma$. Here the electrons and ions are bound by Coulomb’s force undergoes cyclotron motion which are in opposite directions and the resonant emission occurs when particles are much closer such that the Larmour radius of electrons and the ions are same. Implying this condition we can deduce the following relation $$T_i = \frac{Z^2}{\mu }\frac{\gamma m}{M}T_e$$ Where $T_i$ and $T_e$ are the temperatures of ion and electrons respectively but here the problem is that the temperature of the electron again depend on $\gamma$. Therefore we design the program with floating variables to tune the relativistic term. This done by finding the coupling electron temperature (set as round off value to find the relativistic term) with the ion, which has the temperature approximately equal to the observed value of the ion temperature in the EUV spectrum. Due to this reason the theoretical value of the ion temperature has a finite difference with the observed value. Finding the temperature of the coupling electron corresponding to each ionized element, the relativistic term is fixed. We have computed the $T_i$, $T_e$ and $\gamma$ for all the coronal elements detected by CDS, \*EIT and \*TRACE [@Aschwanden2005] shown in the **Table.1**. This table contains four parameters: The first and second column shows the elements and their corresponding temperature ($T_{iO}$) in the observed EUV spectrum. The third column shows the values of possible ion temperature $T_{iC}$calculated theoretically which could under go *gire*. The fourth column gives the value of the temperature of the electron $T_{eC}$ coupling with the ions in the LH oscillation. The last two column gives is the energy and relativistic factor of electron $\nu$ corresponding to the temperature $T_e$. This relativistic factor is used for the calculation of resonant wavelength shown in the table 2.\
Electron Density Reduction Factor
---------------------------------
Our formula has got the ratio of the ion to electron density, they should be fixed on the basis of the observational data. As we see in the section 2.2, that the electron density decreases due to the coronal oscillations in a logerithemic scale. The observations of **SoHO/SUMER** reveals, in the height range 20Mm to 180 Mm, the density of electron decreases approximately from ${1.1\times 10^{14} m^{-3}} $ to $1.6 \times{10^{13}m^{-3}} $[@Spadaro1990a; @Spadaro1990b; @Doyle1998; @Erdelyi1998; @Chae1998; @Esser1999; @Doschek2001]. Thus the change in electron density over a distance of of 160 Mm, is $ 15 \times 10^{14}m^{-3}$ and therefore the electron density reduction factor (EDRF) per meter is $\frac{15 \times 10^{14}}{16 \times 10^7 m^{-3}} \approx 1 \times 10^7$ (MKS units are used for computation). This shows the electron density ($n_{\circ}$) decreases logarithmically by a factor of 7. Now to arrive at the value of $\frac{n^{'}_{\circ}}{n_{\circ}}$ for the calculation and we are not certain about the value of ion density ($n^{'}_{\circ}$) (could be possible to find for a specific spectral line), we keep the value of $\frac{n^{'}_{\circ}}{n_{\circ}}$ as a floating variable in a logarithmic scale between $10^0$ to $10^7$, it could be even more these values, since the solar atmosphere is highly unpredictable.\
Line Instr. $\lambda_O (A^{\circ})$ $\lambda_C$ B(G) $(\frac{{n_{\circ}}^{'}}{n_{\circ}})$ $\theta(^{\circ})$
------------ -------- ------------------------- ------------- ------ --------------------------------------- -------------------- --
He II$^*$ CDS 303.78 303.61 12 $10^6$ 73
O III CDS 702.98 704.12 23 $10^5$ 71
O III CDS 599.59 599.07 26 $10^5 $ 61
Ne III CDS 489.50 488.91 12 $10^6$ 75
N IV CDS 765.14 765.12 16 $10^5$ 74
O IV CDS 554.52 554.32 24 $10^5$ 70
O V CDS 172.17 172.52 20 $10^6$ 61
O V CDS 629.73 628.51 17 $10^5$ 57
Ne V CDS 482.10 482.13 27 $10^5$ 72
Ne VI CDS 399.83 399.95 29 $10^5$ 74
Ne VI CDS 401.14 400.92 29 $10^5$ 75
Ne VI CDS 562.83 562.95 21 $10^5$ 87
Ne VI CDS 558.59 558.14 21 $10^5$ 79
Mg VI CDS 399.20 398.97 10 $10^6$ 68
Mg VI CDS 400.68 400.34 10 $10^6$ 69
Fe VIII CDS 168.18 168.20 10 $10^7$ 67
Fe VIII CDS 186.60 186.18 9 $10^7$ 66
Ne VII CDS 465.22 465.17 23 $10^5$ 90
Ne VIII CDS 770.40 770.08 11 $10^5$ 47
Ne VIII CDS 780.30 779.15 12 $10^5$ 63
Mg VII CDS 277.04 277.93 11 $10^6$ 40
Mg VII CDS 278.40 278.12 12 $10^6$ 50
Si VII CDS 272.60 272.80 15 $10^6$ 79
Si VII CDS 275.37 275.34 15 $10^6$ 90
Ca X CDS 557.76 557.41 22 $10^5$ 76
Mg VIII CDS 313.73 313.58 11 $10^6$ 76
Mg VIII CDS 317.01 317.16 11 $10^6$ 83
Si VIII CDS 316.22 316.51 11 $10^6$ 56
Si VIII CDS 319.83 318.65 12 $10^6$ 90
Fe IX$^*$ CDS 171.07 171.1 10 $10^7$ 86
Fe IX CDS 217.10 217.34 24 $10^6$ 68
Mg IX CDS 368.06 368.88 9 $10^6$ 64
Mg IX CDS 705.80 704.38 14 $10^5$ 65
Si IX CDS 296.12 295.54 11 $10^6$ 57
Si IX CDS 345.12 345.19 10 $10^6$ 71
Si X CDS 261.06 261.14 12 $10^6$ 63
Si X CDS 271.99 271.90 12 $10^6$ 75
Mg X CDS 624.94 625.59 14 $10^5$ 55
Si IX CDS 341.94 341.87 10 $10^6$ 66
Si X CDS 347.40 347.78 9 $10^6$ 61
Si X CDS 356.01 356.79 9 $10^6$ 67
Fe X CDS 174.53 174.5 9 $10^7$ 71
Fe X CDS 177.24 177.14 9 $10^7$ 77
Fe XI CDS 358.62 358.98 13 $10^6$ 72
Fe XI CDS 180.40 180.68 25 $10^6$ 63
Fe XI CDS 188.22 188.21 24 $10^6$ 63
Fe XI CDS 192.81 192.65 24 $10^6$ 69
Al XI CDS 568.12 568.96 17 $10^5$ 75
Al XI CDS 550.03 549.68 16 $10^5$ 53
Fe XII$^*$ CDS 195.12 195.42 21 $10^6$ 54
Fe XII CDS 193.51 193.46 23 $10^6$ 72
Fe XII CDS 364.47 364.79 12 $10^6$ 67
Fe XII CDS 346.85 346.44 13 $10^6$ 77
Fe XIII CDS 203.79 203.91 21 $10^6$ 78
Fe XIII CDS 213.77 213.69 20 $10^6$ 77
Fe XIII CDS 320.80 320.69 13 $10^6$ 68
Fe XIII CDS 359.64 359.45 12 $10^6$ 83
Fe XIII CDS 348.18 348.61 12 $10^6$ 69
Si XII CDS 520.66 520.4 18 $10^5$ 73
Fe XIV CDS 211.32 211.54 19 $10^6$ 69
Fe XIV CDS 220.08 220.09 17 $10^6$ 54
Fe XIV CDS 334.17 334.94 12 $10^6$ 69
Fe XIV CDS 353.83 353.66 11 $10^6$ 61
Fe XV$^*$ CDS 284.16 284.60 13 $10^6$ 60
Fe XV CDS 327.02 327.04 12 $10^6$ 77
Fe XVI CDS 200.80 200.77 11 $10^6$ 76
Fe XVI CDS 335.40 335.15 11 $10^6$ 65
Fe XVI CDS 360.76 360.38 10 $10^6$ 60
: Comparison of Observed and Calculated Wavelengths[]{data-label="T-simple"}
Modification of formula suitable to Corona
------------------------------------------
Before going to the computation of wavelength, there are two important corrections to be made in the formula of *gire* wavelength on the basis of coronal emissions. First one is the theoretical assumption $g \bot B$, this gives the drift velocity $v_0 = \frac{g}{\Omega_i}$. This condition is satisfied only at the top of the coronal loop, but practically the emissions are observable through out the loop. This shows that the gravitational drift is effective to all parts of the loop, since the effect of gravity depend on the geometry of the loop and this can be rectified by introducing an angle $\theta$ between g & B. Now $g\times B = gB Sin \theta$, by replacing g with $g Sin\theta$, we get the correct formula suitable to all parts of the coronal loop. The second one is that we have assumed $T_e = T_i$, but from table 1 we find $T_e >> T_i$. By imposing these two conditions, the formula will be modified as $$\lambda_r = [(\frac{\sqrt{2}\pi^2 mMg}{e^2}) (\frac{\gamma \mu}{Z}) (\frac{Sin\theta}{B^2}\frac{n_{\circ}}{{n_{\circ}}^{'}})]^{\frac{1}{2}}$$ Now the first bracket is the constant, the second bracket is fixed for particular ionized element and the third bracket has 3 variables which are B, Sin$\theta$ and $\frac{n_{\circ}}{{n_{\circ}}^{'}}$. As we have already discussed that these 3 variables are highly tunable to coronal plasma and could be studied according to the condition for solar minimum or maximum. At high magnetic field intensity (solar maximum) the spectral resolution will be good and tuning will be very easy to the maximum level of accuracy but at solar minimum, the value of B is very less and hence the spectral resolution will be affected. Therefore we choose solar minimum condition($B =10-50G$, rest all constants have usual value in SI units and $ g = 274ms^{-1}$), but we have an another variable $\theta$ and therefore it could be possible to tune it to more correct value. For each element, the mass number ($\mu$) and degree of ionization(Z) are constant; $\gamma$, relativistic factor of coupling electron corresponding to the element is taken from the table 1. The charge state of elements are taken as the value less than one of the Roman numerals of ionized elements; for example: $ Fe XV \rightarrow Z = 14$.The Computed wavelength ($\lambda_C$) by *gire* formula in comparison with the observed wavelength ($\lambda_O$) is shown in the table 2.
Analysis of Parameters over wavelength
======================================
The resonant wavelength is depend on 8 parameters, the influence of each parameter is very elaborate but we shall most relevant parameters to the corona.
Effect of $\theta$
------------------
$\theta$ is the angle between g & B. We find $\lambda_r \propto \sqrt{g Sin\theta}$; when $\theta$ is low at the bottom side of the loop where B is high. It is simple that when B is high, $r_L$ is small. Resonant wavelength, $ \lambda \propto r_L $; therefore the ions resonate with shorter wave length. This shows that at smaller angle or at the bottom side of the loop, shorter wavelengths of radiation will be absorbed and the rate of resonant absorption would be high and this is the reason that the HARD X-rays are observed more at the bottom of the coronal loop.
Magnetic field (B)
------------------
$\lambda_r \propto \frac{1}{B}$, for high field region, $\Omega$ is high and $r_L$ is very low, the wavelength will shorter. The reason is that the high magnetic intensity could hold relatively high energetic electrons ($r_L \propto \frac{v_ti}{B}$) and hence the coupling is possible highly relativistic electrons. This gives a reason for the presence of highly ionized elements are observed at high field region. As we have seen from the table, the heavier ions need high relativistic electrons to couple in the LH oscillation, which is possible in the high magnetic field regions. This situations are more realistic during solar maximum and the higher elements could be observable at the top of the loop during solar maximum; in solar minimum the higher elements would be more probable at the rear side of the loop where the magnetic field intensity is high.
Density Ratio $ \frac{{n_{\circ}}^{'}}{n_{\circ}}$
--------------------------------------------------
$\lambda_r \propto \frac{{n_{\circ}}^{'}}{n_{\circ}}$. Here the electron density, ${n_{\circ}}$ increases the wavelength. This is nothing other than what we have explained in the section EDRF, in the *lh* oscillation, the absorption of energy solely depend on the field deficit. When ${n_{\circ}}$ increases the field deficit is less, absorption rate is low and hence resonance shifts to longer wavelength. For higher ion density, the relative electron density decreases and wavelength will be shorter. More work could be done by the observational solar physicists to evaluate this factor very precisely on the basis of the elemental abundance.\
The analysis of other parameters like Z, degree of Ionization; $\gamma$, relativistic factor; $\mu$, mass number; would lead to the elemental abundances and FIP factor and that we reserve for another discussion.\
Conclusion
==========
- In corona; magnetic field pressure much greater than the particle pressure and therefore the magnetic field is controlling the motion of the particles.
- If any particles becomes neutral by collision, ans such particle will be transparent to the magnetic field descends down; and this is the reason that the solar Corona is fully ionized.
- Solar Corona is ultra vacuum and the fields are long range; since the corona is fully ionized, all the particles are in the cyclotron motion must be attracted by Coulomb force which leads to the coupling of ionized particles in the lower hybrid oscillations.
- Magnetic field in the corona is highly oscillatory since it connected with convection layer and this oscillations of magnetic field results in the loss of ionized particles lead to the field deficit.
- As per the theory of gravity induced *lh* oscillations, the particles loss is very much high for electrons compared with ions and hence Ion Cyclotron Resonant Damping is efficient and continuous in the solar corona.
**We have approached the problem with the prevailing situation of the solar atmosphere and hence the gravity induced lower hybrid Ion cyclotron resonant heating is an unavoidable solution to the coronal heating problem**.\
**Acknowledgment** This first author acknowledges the funding of UGC under Research Award Scheme (2016-2018).
[200]{}
Alissandrakis, C. E., Kundu, M. R., Lantos, P., 1980, A&A 82, 30.
Antonucci, E., Gabriel, A. H., Patchett, B. E., 1984, SP, 93, 85.
Antony, S., Sunitha, A., 2017, arXive:1704.07225v1\[physics.plasm-ph\].
Aschwanden, M. J., Newmark, J. S., Delaboudiniere, J. P., Neupert, W. M., Klimchuk, J. A., Gary, G. A., Portier-Fornazzi, F., Zucker, A., 1999, ApJ, 515, 842.
Aschwanden, M. J., Nightingale, R. W., Alexander, D., 2000, ApJ, 541, 1059.
Aschwanden, M. J., Acton, L. W., 2001a, ApJ, 550, 475.
Aschwanden, M. J., 2001b, ApJ, 559, L171.
Aschwanden, M. J. 2002, ApJ, 580, L79.
Aschwanden, M. J., 2005, Physics of Solar Corona, Springer, Praxis Publishing, UK
Banerjee, D., Teriaca, L., Doyle, J. G., 1998, A& A 339, 208.
Benz, A. O., Wentzel, D. G., 1981, A&A, 94, 100-108.
Benz, A. O., Smith, D. F., 1987, SP, 107, 299.
Berghmans, D., 2002, ESA, SP-506, 501.
Berghmans, D., McKenzie, D., Clette, F., 2001, A&A, 369, 291.
Bogod V.M. & Grebinskij A.S., 1997, SP 176, 67.
Boyd, T.J. M., Sanderson, J. J., 2005, The Physics of Plasmas, Cambridge University Press
Brosius J.W. & Holman G.D., 1989, ApJ 342, 1172.
Chae, J. C., Schuhle, U., Lemaire, P., 1998, ApJ, 505, 957.
Chae, J. C., Poland, A. I., Aschwanden, M. J., 2002, ApJ, 581, 726.
Chapman, R. D., Jordan, S. D., Neupert, W. M., 1972, ApJ, 174, L97.
Chiuderi, D. F., Avignon, Y., Thomas, R. J., 1977a, SP, 51, 143.
Chiuderi, D. F., Poletto, G., 1977b, A&A, 60, 227.
Chiuderi, D. F., Landi, E., Fludra, A., 1999, A&A, 348, 261.
DeForest, C. E., Gurman, J. B., 1998, ApJ, 501, L217.
Dmitruk, P., Gomez, D. O., 1997, ApJ, 484, L83.
Dmitruk, P., Gomez, D. O., DeLuca, E. E., 1998, ApJ, 505, 974.
Doschek, G. A., Warren, H. P., Laming, J. M., 1997, ApJ, 482, L109.
Doschek, G. A., Feldman, U., Laming, J. M., 2001, ApJ, 546, 559.
Doyle, J. G., Banerjee, D., Perez, M. E., 1998, Solar Phys., 181, 91.
Drago, F., 1974, Proc-1974-Righini, 120.
Dulk, G. A., Sheridan, K. V., Smerd, S. F., 1977, SP, 52, 349.
Edlen, B., 1943, Z. Astrophysik, 22, 30.
Edlen, B., 1945, MNRAS, 105, 323.
Einaudi, G., Velli, M., Politano, H., 1996a, ApJ, 457, L113.
Einaudi, G., Califano, F., Chiuderi, C., 1996b, ApJ, 472, 853.
Esser, R., Fineschi, S., Dobrzycka, D., 1999, ApJ, 510, L63.
Erdelyi, R., Doyle, J. G., Perez, M. E., 1998, A&A, 337, 287.
Fisher, R. R., Musman, S., 1975, ApJ, 194, 801.
Foley, C. R., Culhane, J. L., Acton, L. W., 1997, ApJ, 491, 933.
Furst, E., Hirth, W., 1975, SP, 42, 157.
Gary, D. E., Hartl, M. D., Shimizu, T., 1997, ApJ, 477, 958.
Gopalswamy, N., Payne, T. E. W., Schmahl, E. J., 1994, ApJ, 437, 522.
Gopalswamy, N., Zhang, J., Kundu, M. R., 1997, ApJ, 491, L115.
Gopalswamy, N., Shibasaki, K., Thompson, B. J., 1999, JGR, 104/A5, 9767.
Grotrian, W., 1939, Naturwissenschaften, 27, 214.
Guhathakurta, M., Fisher, R., Strong, K., 1996, ApJL, 471, L69.
Hara, H., Tsuneta, S., Acton, L. W., 1994, PASJ, 46, 493.
Hara, H., Tsuneta, S., Acton, L. W., 1996, Adv. Space Res., 17/4-5, 231.
Hara, H., 1997, PASJ, 49, 413.
Harrison, R. A., Bryans, P., Simnett, G. M., Lyons, M., 1995, SP, 162, 233
Heasley, J. N., Mihalas, D., 1976, ApJ, 205, 273.
Heyvaerts, J., Priest, E. R., 1992, ApJ, 390, 297.
Inverarity, G. W., Priest, E. R., Heyvarts, J., 1995a, A&A, 293, 913.
Inverarity, G. W., Priest, E. R., 1995b, A&A, 296, 395.
Kenneth, R. L., 2009, Springer Verlag Berlin Heidelberg.
Kippenhahn, R., Schl¨uter, A., 1957, Z.Astrophys. 43, 36.
Krall, N. A., Liewer, P. C., 1971, Phys. Rev, A4, 2094
Krueger, A., Hildebrandt, J., Fuerstenberg, F., 1985, A&A, 143, 72
Kundu, M. R., Liu, S. Y., 1976, SP, 49, 267.
Kundu, M. R., Schmahl, E. J., Gopalswamy, N., 1989, Adv. Space Res., 9, 41.
Lampe, M., Papadopoulos, K., 1977, ApJ, 212, 886.
Lantos, P., Avignon, Y., 1975, A& A 41, 137
Lee, J. W., Hurford, G. J., Gary, D. E., 1993a, SP, 144, 45.
Lee, J. W., Hurford, G. J., Gary, D. E., 1993b, SP, 144, 349.
Lerche, L., Low, B. C., 1977, SP, 53, 385.
Lin, R. P., Feffer, P. T., Schwartz, R. A., 2001, ApJ, 557, L125.
Low, B. C., 1975a, ApJ, 197, 251.
Low, B. C., 1975b, ApJ, 198, 211.
McBride, J. B., Otto, E., Boris, J. P., Oran., J. H., 1972, Phys. Fluids, 15, 2367
McClements, K. G., Bingham, R., Su, J. J., Dawson, J. M., Spicer, D. S., 1993, ApJ, 409, 465.
Milano, L. J., Gomez, D. O., Martens, P. C. H., 1997, ApJ, 490, 442.
Milne, A. M., Priest, E. R., Roberts, B., 1979, ApJ, 232, 304.
Moran, T. G., Gopalswamy, N., Dammasch, I. E., 2001, A&A, 378, 1037.
Murawski, K., Aschwanden, M. J., Smith, J. M., 1998, SP, 179, 313.
Murawski, K., Roberts, B., 1993, SP, 144, 101.
Murawski, K., Roberts, B., 1994, SP, 151, 305.
Nakariakov, V. M., Ofman, L., DeLuca, E., 1999, Science, 285, 862.
Nakariakov, V. M., Ofman, L. 2001, AA, 372, L53.
Nitta, N., 1997, ApJ, 491, 402.
Ofman, L., Romoli, M., Poletto, G., 1997, ApJ, 491, L111.
Ofman, L., Romoli, M., Poletto, G., 2000a, ApJ, 529, 529
Ofman, L., Nakariakov, V. M., DeForest, C. E., 1999, ApJ, 514, 441.
Orrall, F. Q.,Zirker, J. B., 1961, ApJ, 134, 72.
Papagiannis M.D. & Baker K.B., 1982, SP, 79, 365.
Parnell C.E., 2002a, COSPAR-CS 13, 47.
Parnell, C. E., 2002b, ESA SP-505, 231.
Pekeris, C. L., 1948, Geol.Soc.Amer.Mem., 27, 117.
Pikelner, S. B., 1971, SP, 17, 44.
Roberts, B., Joarder, P. S., 1994, Proc-1994-Belvedere, 173.
Rosenberg, H., 1970, A&A, 9, 159.
Ryutova, M., Tarbell, T. D., 2000, ApJL, 541, L29.
Ryutova, M., Habbal, S., Woo, R., 2001, SP, 200, 213.
Shimizu, T., 2002a, COSPAR-CS, 13, 29.
Shimizu, T., 2002b, ApJ, 574, 1074.
Shimizu, T., Tsuneta, S., Acton, L. W., 1992, PASJ, 44, L147.
Shimizu, T., Tsuneta, S., Acton, L. W., 1994, ApJ, 422, 906.
Shimizu, T., 1995, PASJ, 47, 251.
Shimizu, T., 1997, PhD Thesis
Shimizu, T., Tsuneta, S., 1997, ApJ, 486, 1045.
Spadaro, D., 1999, ESA SP, 448, 157.
Spadaro, D., Noci, G., Zappala, R. A., 1990a, ApJ, 355, 342.
Spadaro, D., Noci, G., Zappala, R. A., 1990b, ApJ, 362, 370.
Spicer, D. S., Benz, A. O., Huba, J. D., 1981, A&A 105, 221
Tanaka, K., Papadopoulos, K., 1983, Phys. of Fluids, 26, 1697.
Tarbell, T. D., Ryutova, M., Covington, J., 1999, ApJ, 514, L47.
Tarbell, T. D., Ryutova, M., Shine, R., 2000, SP, 193, 195.
Thejappa, G., 1987, SP, 111, 45.
Timothy, A. F., Krieger, A. S., Vaiana, G. S., 1975, SP, 42, 135.
Tripathi, V. K., Grebogi,C., Liu, C. S., 1977, Phys. Fluids, 20, 1525
Vlahos, L., Gergely, T. E., Papadopoulos, K., 1982, ApJ, 258, 812.
Vourlidas A., Bastian T.S.& Aschwanden M.J. 1997, ApJ 489, 403.
Wang, Z., Schmahl, E. J., Kundu, M. R., 1987, SP, 111, 419.
Watari, S., Kozuka, Y., Ohyama, M., 1995, J.Geomag.Geoelec., 47(11), 1063.
White, S. M., Kundu, M. R., Shimizu, T., 1995, ApJ, 450, 435.
Zirker, J. B., 1993, SP, 148, 43-60.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently, it has been proven \[R. Soc. Open Sci. 1 (2014) 140124\] that the continuous wavelet transform with non-admissible kernels (approximate wavelets) allows for an existence of the exact inverse transform. Here we consider the computational possibility for the realization of this approach. We provide modified simpler explanation of the reconstruction formula, restricted on the practical case of real valued finite (or periodic/periodized) samples and the standard (restricted) Morlet wavelet as a practically important example of an approximate wavelet. The provided examples of applications includes the test function and the non-stationary electro-physical signals arising in the problem of neuroscience.'
address:
- 'Theoretical Physics Department, Kursk State University, Radishcheva st., 33 Kursk 305000, Russia'
- 'Mathematics and Mechanics Faculty, Saint Petersburg State University, Universitetsky prospekt, 28, Peterhof, Saint Petersburg, 198504, Russia'
- 'Institute of Applied Mathematics and Mechanics, Saint Petersburg State Polytechnic University, Polytechnicheskay 29, 195251, Saint Petersburg, Russia'
- 'Institute of Chemistry and Biology, Immanuel Kant Baltic Federal University, A. Nevskogo str. 14A, Kaliningrad 236041, Russia'
author:
- 'Eugene B. Postnikov'
- 'Elena A. Lebedeva'
- 'Anastasia I. Lavrova'
title: Computational implementation of the inverse continuous wavelet transform without a requirement of the admissibility condition
---
Continuous wavelet transform ,signal processing ,Morlet wavelet
Introduction
============
The continuous wavelet transform (CWT) with the standard (restricted) Morlet wavelet $\psi(\xi)=\exp\left(i\omega_0\xi-\xi^2/2\right)$ is defined as the integral $$w(a,b)=\int\limits_{-\infty}^{+\infty}f(t)e^{-i\omega_0\frac{t-b}{a}}
e^{-\frac{(t-b)^2}{2a^2}}\frac{dt}{\sqrt{2\pi a^2}},
\label{CWTM}$$ where $a$ and $b$ are the scale and the shift correspondingly, $\omega_0$ is the central frequency.
It is one of the most powerful modern tools of signal processing especially adjusted to the extraction of instant oscillating patterns [@MallatBook; @AddisonBook] since the transform (\[CWTM\]) of the harmonic oscillation $f(t)=\exp(i\omega t)$ results in the complex function $$w(a,b)=e^{i\omega b}e^{-\frac{(a\omega-\omega_0)^2}{2}}.$$ The modulus of this function has a maximum, which allows for the determining of signal’s frequency $\omega=\omega_0/a_{max}$ and the phase coincides with signal’s one.
The modern applications of the continuous wavelet transform are focused, in particular, on a study of environmental time series [@Cazelles2008; @Galiana2014], geo- and astrophysics [@Katsavrias2012; @Soon2014; @Postnikov2009], biophysics [@Meng2013; @Worrell2012; @Suvichakorn2011] and neuroscience, see an extensive review in the recently published book [@WaveletNeuroBook].
At the same time, the actual problem is not restricted by the search of oscillating patterns localizations. It is important to extract revealed structures from the background consisting a noise, global oscillations and inhomogeneities, etc. [@Sheppard2011; @Karimi2012; @Gu2013; @Postnov2014].
However, the conventional approaches to the inversion of the wavelet transform with the standard Morlet wavelet have some principal difficulties from the point of view of functional analysis. To be applicable in a classical inversion formula, a wavelet function $\psi$ should satisfy the admissibility condition [@MallatBook] $$C_{\psi}=\int_{\mathbb{R}}\frac{|\psi^*(\omega)|^2}{|\omega|}d\omega<\infty,$$ where the asterisk denotes a complex conjugation. Then the classical inversion of the wavelet transform $W_{\psi}f=w(a,b)$ with a wavelet function $\psi$ is written as $$f(t)=\frac{1}{C_{\psi}}\int_{\mathbb{R}}\int_{\mathbb{R}} \frac{1}{a}
\psi\left(\frac{t-b}{a}\right) w(a,b) \frac{dadb}{a}.$$ However, the integral $C_{\psi}$ diverges for the standard (restricted) Morlet wavelet.
On the other hand, the alternative inversion formula for the CWT $$\frac{1}{\pi}\mathrm{v.p.}\int_{\mathbb{R}}\frac{db}{b-t} \int_{\mathbb{R}}\frac{\partial}{\partial b} w(a,b)\,da = \psi^*(0)\, f(t)
\label{rec1}$$ is proven in [@Lebedeva2014] under a mild natural conditions on a wavelet function $\psi$.
The principal aim of this paper is to show how this approach can be realized in applications. For this reason, we adapt the proof of (\[rec1\]) to the case of real-valued functions with a finite support in such a way that its line of reasoning and the result can be straightforwardly used as a practical algorithm for the reconstruction of a function from its wavelet transform with the Morlet wavelet.
The paper is organized as follows. Section 2 shortly discusses the implementation of the CWT with the Morlet wavelet in the application to real-valued functions determined within a finite interval (or periodic on $\mathbb{R}$) and then presents the simple proof the exact reconstruction formula that does not require the wavelet admissibility condition. Its formulation allows for the simple computational realization, which is presented in Section 3. The examples comprise the test one and the solution of the modern practical problem in the field of neuroscience. Neuronal electric activity is characterized by the high non-linearity and complexity, extraction of main oscillatory components and localization of them is very important for reveal of interaction dynamics between different cells in neuronal network [@Keef]. The final section gives some outlooks for the perspectives for the applications of the proposed method in computational physics.
Direct and inverse continuous wavelet transform of periodic functions
=====================================================================
CWT expansion
-------------
Let us consider a real-valued function $f(t)$ with the zero mean given by its expansion into the Fourier series $$f(t)=\sum\limits_{n=1}^{\infty}A_n\cos(\omega_nt)+B_n\sin(\omega_nt),
\label{fFourier}$$ where $\omega_n=2\pi n/T$.
This function is periodic with the period $T$. As well, it can be considered as a function determined within the interval $t\in[0,\,T]$ and periodically extended over all $\mathbb{R}$. The last point of view is applicative in practical computations since one can deal with finite samples only. The assumption of the zero mean is also not restrictive for the problems of computational physics since it can be easily achieved by the exclusion of the averaged value: $f(t)-T^{-1}\int_0^Tf(t)dt\to f(t)$.
For the wavelet analysis of spectral components, it is most convenient to operate with the complex analytical counterpart of the function $f(t)$ obtained via the Hilbert transform $f_a(t)=f(t)+iH\left[f(t)\right]$.
Since $H\left[\cos(x)\right]=\sin(x)$, $H\left[\sin(x)\right]=-\cos(x)$, the Hilbert transform applied to Eq. (\[fFourier\]), accompanied with Euler’s formula, gives $$f_a(t)=\sum\limits_{n=1}^{\infty}\left(A_n-iB_n\right)e^{i\omega_nt}=\sum\limits_{n=1}^{\infty}C_ne^{i\omega_nt}.
\label{anfun}$$
Note also that Euler’s formula applied directly to Eq. (\[fFourier\]) results in the representation $$f(t)=\frac{1}{2}\sum\limits_{n=1}^{\infty}C_ne^{i\omega_nt}+C_n^*e^{-i\omega_nt},$$ where asterisks denotes the complex conjugation.
Thus, one does not need in practice to apply any special additional procedure for the Hilbert transform of a given sample. It is enough to evaluate the discrete Fourier transform (say, via FFT algorithm) and to cut off elements corresponding to negative frequencies.
The substitution of the series (\[anfun\]) into the integral (\[CWTM\]), $$w(a,b)=\sum\limits_{n=1}^{\infty}C_n\int\limits_{-\infty}^{+\infty}e^{i\omega_nt}e^{-i\omega_0\frac{t-b}{a}}
e^{-\frac{(t-b)^2}{2a}}\frac{dt}{\sqrt{2\pi a^2}},$$ after close form computation of Poisson’s integral $\int_{-\infty}^{+\infty}\exp(-x^2)dx=\sqrt{\pi}$, results in the desired formula for the continuous wavelet transform with the Morlet wavelet $$w(a,b)=\sum\limits_{n=1}^{\infty}C_ne^{-\frac{(\omega_na-\omega_0)^2}{2}}e^{i\omega_nb}.
\label{wFourier}$$
The expression (\[wFourier\]) is known and usually applied in the numerical algorithms of CWT computations based on the intermediate FFT, e.g. realised in WaveLab package [@WaveLab]. However, the preprocessing cut-off procedure, described above, allows for the shortening of the time/memory consumption because of the twice shortened sample’s length in comparison with the standard application of Eq. (\[wFourier\]) to the initial function $f(t)$ directly. At the same time, this procedure does not loose any information about $f(t)$ since the complex coefficient $C_n=A_n-iB_n$ contains information on both real-valued ones.
As well, the representation (\[wFourier\]) provides the most clear way to obtain the exact inverse continuous wavelet transform.
CWT reconstruction
------------------
Now, we aimed to obtain the practically applicable CTW reconstruction formula basing on the series (\[wFourier\]). Its partial differentiation with respect to the shift $b$ gives the series $$\frac{\partial w(a,b)}{\partial b}=i\sum\limits_{n=1}^{\infty}C_n\omega_ne^{-\frac{(\omega_na-\omega_0)^2}{2}}e^{i\omega_nb}.$$
Since the included frequencies are strictly positive $\omega_n>0$, the subsequent integral over the non-negative scale half-line leads (using Poisson’ integral again) to $$\int\limits_0^{\infty}\frac{\partial w(a,b)}{\partial b}da=i\sqrt{2\pi}\sum\limits_{n=1}^{\infty}
C_ne^{i\omega_nb}=i\sqrt{2\pi}f_a(b).$$
Therefore, taking into account $C_n=A_n-iB_n$ and applying Euler’s formula we get the desired formula for the exact inverse CWT with the Morlet wavelet, which does not depend on the admissibility condition: $$f(t)=\frac{1}{\sqrt{2\pi}}\mathrm{Im}\left[\int\limits_0^{\infty}\frac{\partial w(a,b)}{\partial b}da\right].
\label{inwCWT}$$
Note that it reduces computational complexity in comparison with the full abstract representation (\[rec1\]) since (\[inwCWT\]) does not require the additional inverse Hilbert transform.
Moreover, Eq. (\[inwCWT\]) allows for the reconstruction of function’s instant features revealed by an exploration of the result of the direct CWT. Since the partial derivative is a local operator, it is possible to consider $w(a,b)$ with $b\in[b_{min}\,b_{max}]$, the interval, which contains the feature, which we are interested in. Thus, restricting (\[inwCWT\]) on the mentioned $b$, we reconstruct $f(t)$ within $b\in[t_{min}\equiv b_{min}\,t_{max}\equiv b_{max}]$ without boundary disturbances originated from the conventional inversion methods based on convolutions.
As well, since the Gaussian terms in (\[wFourier\]) are fast decaying functions with the maximum in $a_n=\omega_0/\omega_n$, the restriction of the integral in (\[inwCWT\]) on the interval $[a_{min}\,a_{max}]$ reconstruct the oscillating components with the frequencies from $[\omega_0/a_{max}\,\omega_0/a_{min}]$.
Combining this restriction with the choice of some interval of the shift or even considering some non-rectangular region $a=a(b)$, it is possibly to reconstruct and analyse very specific details of a studied function $f(t)$.
Tests and applications
======================
Test non-stationary oscillations
--------------------------------
As a first numerical example, let us consider the function $$f(t)=e^{-4t}\cos(20\pi t)+\chi_{\left[\frac{1}{3},\frac{2}{3}\right]}(t)\sin(40\pi t),
\label{nonstex}$$ sampled within the interval $t\in[0,\,1]$.
Here $$\chi_{[t_b,t_e]}(t)=
\left\{
\begin{array}{lr}
1,&t\in [t_b,\,t_e];\\
0,&t\notin [t_b,\,t_e].
\end{array}
\right.$$ is the indicator function. Thus, the function (\[nonstex\]) consists of the decaying oscillations over all studied interval and the stable oscillations of unit amplitude over its middle third only.
The function (\[nonstex\]) is sampled in $512=2^9$ equispaced points that allows for the direct application of FFT algorithm in further computations. It should be pointed out that the local boundary values of the function (\[nonstex\]) are sufficiently different that can result in the drastic boundary effects due to a periodization. Therefore, we used the algorithm, which expand the interval $[0, \,1]$ into $[-0.5, \,1.5]$ and function’s continuation with boundary reflections [@AddisonBook]. The MATLAB code realizing such CWT is presented in Appendix. The central frequency is chosen as $\omega_0=2\pi$, i.e. maxima lines should be located along $a=0.05$ and $a=0.1$.
![The function (\[nonstex\]) and the modulus of its CWT (dark regions correspond to the larger values of magnitude). The dashed rectangle marks the maximum line corresponding to the localized periodic components with the constant amplitude.[]{data-label="funccwt"}](ex1w){width="\textwidth"}
Then, the reconstruction formula (\[inwCWT\]) is applied to this two-dimensional wavelet function $w(a,b)$. The partial derivative with respect to the shift $b$ is realized as the three-point central difference scheme (except the left and the right boundary points, where the forward and the backward two-point scheme is exploited). The numerical integration is evaluated via the trapezoidal rule. The result of reconstruction for the full shift-scale range is presented in Fig. \[recfunc\]. One can see an accurate coincidence of the initial and the reconstructed signals.
![The inverse continuous wavelet transform via the formula (\[inwCWT\]). Upper panel: comparison of the reconstruction (solid line) and the original function (\[nonstex\]) (dashed line). Lower panel: partial reconstructions of oscillating features of the signal (\[nonstex\]).[]{data-label="recfunc"}](ex2w){width="\textwidth"}
As the next step, we process the wavelet transformed function $w(a,b)$ within intervals that bound the localized oscillating features. This should result in the decomposition of the function (\[nonstex\]) into separate instant oscillatory components.
As the first step, we extract $\partial w(a,b)/\partial b$ for the interval $b\in[0.3209,\,0.6829]$ and integrate this part of the obtained function within the limits $a\in[0.0240,\,0.052]$. This region is bounded by the dashed rectangle in Fig. \[funccwt\]. The result is drawn by solid line in Fig. \[recfunc\](lower panel). Comparing this localized oscillation with the almost everywhere constant amplitude with the corresponding component of the function (\[nonstex\]), one can conclude that the proposed methods reaches its goal. The pattern detected as a localized spot in Fig. \[recfunc\](lower panel) can be reconstructed with a sufficient accuracy using the simple differentiation and integration within the its bounding box. Certainly, it is impossible to avoid boundary disturbances completely, but they are well localized and do not corrupt majority of the reconstructed oscillation.
The rest component is reconstructed by the subtraction of this local component from the result of full inverse continuous wavelet transform. It is shown as the dashed line in Fig. \[recfunc\](lower panel). One can see that, except the short bursts originated from transient regions, the exponentially decaying oscillation is reconstructed as it should be originally.
Neuronal oscillations
---------------------
As the application of this method, it would be useful to consider an example of realistic biological oscillations, in particular, the neuronal network dynamics.
We analyse the complex rhythm, which arises prior to epileptic events, so called very fast oscillations (VFO) [@Traub]. These oscillations appear spontaneously and can be revealed after the removal of slow baseline fluctuations. Thus, there exists a problem of the reconstruction of various oscillation features from the complex signal representing experimental recordings. Moreover, the considered method should allows us to determine an approximate time localization for the switching-on and switching-off of oscillations with different frequencies and their synchronous states.
The experimental data are taken from the work [@Traub]. The potential oscillations are measured in rat’s neocortex: i) the extracellular recordings from the neocortex layer (see Fig. \[ef\], upper panel) and ii) the intracellular recordings from the pyramidal cell, so-called intrinsically bursting cell, as shown in Fig. \[1cell\], upper panel. All numerical values are presented in dimensionless units, which can be rescaled to the experimental values of the time and the potential (Fig.1 in [@Traub]) as follows: 1 time unit in Figs. \[1cell\],\[ef\] corresponds to 4.5 ms; 1 voltage unit corresponds to 0.25 mV in Fig. \[1cell\] and to 0.5 $\mu$V in Fig. \[ef\].
The visual exploration of the wavelet modulus plots, Figs \[1cell\],\[ef\] (middle panels), allows for concluding that there are at least four areas where oscillations frequencies are different. On the other hand, the corresponding maxima have not such a regular shape as in the the idealized test example considered above. For this reason, we introduce the procedure, which extracts the irregular regions as $$f(t)=\frac{1}{\sqrt{2\pi}}\mathrm{Im}\left[\int\limits_0^{\infty}C(a,b)\frac{\partial w(a,b)}{\partial b}da\right],$$ where $$C(a,b)=\theta\left(|w(a,b)|-L\cdot\mathrm{max}(|w(a,b)|)\right)$$ is the mask with the cut-off threshold $L\in[0,\,1]$. Here $$\theta(\xi)=\left\{
\begin{array}{lr}
1,\,\xi>0\\
0,\,elsewhere.
\end{array}
\right.$$ is the Heaviside function.
Thus, the reconstructed features are the results of the inverse continuous wavelet transform applied to the regions bounded to be the contours defined by the equality $|w(a,b)|=L\cdot\mathrm{max}(|w(a,b)|$.
This reconstruction within contours allows for discussing the types of dynamics, which are characterized by different frequencies (see Figs \[1cell\],\[ef\], lower panels). In the extracellular signal, as shown in Fig \[ef\], there are oscillations with two main frequencies: approximately 6 Hz (within the region marked by solid and dashed lines, red, blue and green in the colour online version) and 30 Hz (within the black dash-dotted contour). It should be noted, that the “fast signal” (thin black line in the lower panel in Figs \[ef\]) is not so strong and arises spontaneously at the time about 250 reaching the maximum of amplitude in the range of time from 450 to 650. On the other hand, the “fast dynamics” within the cell signal (thin line (green in the colour online version), Fig \[1cell\], lower panel) starts earlier at time about 220 and has a higher amplitude. It breaks abruptly and renews at the time $t$=940.
Note that our method, which explicitly extracts the main frequency components within their regions of switch on and off allows for avoiding their mixture with the subthreshold oscillations (so called spikelets), which results in the averaged frequency of “fast” signal determined as $118\pm 10$ Hz in [@Traub].
In comparison with the extracellular signal, the “slow” dynamics corresponds to the oscillations of 2 Hz (black dash-dotted line in Fig \[1cell\], lower panel) with a very small amplitude, which do not have breaks within the whole time range of time, to the oscillations of 6 Hz (thick solid (red and blue in the colour online version) lines in Fig \[1cell\], lower panel)). These oscillations exist within the time intervals from 250 to 400 and from 940 to 1024 only.
![The extracellular signal measured in the neocortex layer, its CWT and the partial reconstructions of signal’s oscillating features. The plot (upper panel) shows the signal, where VFO occurs (the frequency increases in the range of time from 250 to 650). Middle panel: Modulus of signal’s CWT, where rectangles mark the regions that contain different oscillating components; the contours bound the exact regions used for the reconstruction. Lower panel: partial reconstructions of the oscillating components contained in the signal. The total time duration corresponds to 4636 ms in the experiment, the maximal amplitude is about 0.14 mV[]{data-label="ef"}](ef){width="\textwidth"}
![The intracellular signal (intrinsic bursting cell [@Traub]), its CWT and the partial reconstructions of intracellular signal’s oscillating features. The plot (upper panel) shows the signal, where VFO occurs (the frequency increases in the range of time from 220 to 650). Middle panel: Modulus of signal’s CWT, where rectangles mark the regions that contain different oscillating components; the contours bound the exact regions used for the reconstruction. Lower panel: partial reconstructions of the oscillating components contained in the signal. The total time duration corresponds to 4636 ms in the experiment, the maximal amplitude is about 90 mV[]{data-label="1cell"}](1cell){width="\textwidth"}
Conclusion and outlook
======================
In this paper, we have proposed simple computational procedure for the inverse continuous wavelet transform that allows for processing of oscillating signals, e.g. the extraction of main frequency components. The modern methods of physical research, especially connected with electrophysiology, provides data of high complexity, and, correspondingly, their processing requires a development of new approached deeply based of methods of functional analysis [@Siddiqi2003book]. The wavelet methods play an important role among them. At the same time, they need to be adapted to needs and perceptions of computational community.
Thus, we propose new derivation of the reconstruction formula, which, in contrast to the theorem in [@Lebedeva2014] provides the direct way to a computational realization and applications.
As an example, we have considered the voltage signals measured in neuronal system. A neuronal activity is characterized by the both non-linear periodicity and complexity in time and space [@Keef; @Traub; @Buz; @Cun]. Firing patterns could be subdivided into the different frequency bands [@Buz; @Traub; @Cun] and connected with the spatial periodicity [@Keef] that provides an important information on different types of neurons at pathological processes and a normal state.
We have considered the application of this approach to the real experimental data [@Traub], which describe the spontaneous emergence of very fast oscillations at a seizure. The approach considered in the present work allows for extracting main periodical components from two signals (extracellular and intracellular). The time moments of “switching on” of fast oscillations is obtained as well. This information could be useful for the understanding of a network structure, in particular, a number of oscillating elements and the kind of interconnection between them.
However, the results are not restricted by the problems of electrophysical physiology since the proposed approach and its mathematical and computational background provide an opportunity to be applied to a much more wider class of complex oscillating patterns.
Acknowledgment {#acknowledgment .unnumbered}
==============
EBP is partially supported by grant no. 1391 of the Ministry of Education and Science of the Russian Federation within the basic part of research funding no. 2014/349 assigned to Kursk State University. EAL is supported by the Russian Foundation for Basic Research, grant No. 15-01-05796 and by the grant No. 9.38.198.2015 of Saint Petersburg State University. AIL is supported by grant no. 14.575.21.0073, code RFMEFI57514X0073 of the Ministry of Education and Science of the Russian Federation
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
S. Mallat, A wavelet tour of signal processing, Academic press, 1999.
P. Addison, The illustrated wavelet transform handbook: introductory theory and applications in science, engineering, medicine and finance, CRC Press, 2002.
B. Cazelles, M. Chavez, D. Berteaux, F. M[é]{}nard, J. O. Vik, S. Jenouvrier, N. C. Stenseth, Wavelet analysis of ecological time series, Oecologia 156 (2008) 287–304. [](http://dx.doi.org/10.1007/s00442-008-0993-2).
J. J. Galiana-Merino, C. Pla, A. Fern[á]{}ndez-Cort[é]{}s, S. Cuezva, J. Ortiz, D. Benavente, Environmentalwavelettool: Continuous and discrete wavelet analysis and filtering for environmental time series, Computer Physics Communications 185 (2014) 2758–2770. [](http://dx.doi.org/10.1016/j.cpc.2014.06.011).
C. Katsavrias, P. Preka-Papadema, X. Moussas, Wavelet analysis on solar wind parameters and geomagnetic indices, Solar Physics 280 (2) (2012) 623–640. [](http://dx.doi.org/10.1007/s11207-012-0078-6).
W. Soon, V. M. V. Herrera, K. Selvaraj, R. Traversi, I. Usoskin, C.-T. A. Chen, J.-Y. Lou, S.-J. Kao, R. M. Carter, V. Pipin, M. Severi, S. Becaglid, A review of holocene solar-linked climatic variation on centennial to millennial timescales: Physical processes, interpretative frameworks and a new multiple cross-wavelet transform algorithm, Earth-Science Reviews 134 (2014) 1–15. [](http://dx.doi.org/10.1016/j.earscirev.2014.03.003).
E. B. Postnikov, A. Loskutov, Continuous wavelet transform as an effective tool for the detecting of saturn rings’ structure, in: J. D. Denis, P. D. Aldridge (Eds.), Space Exploration Research, Nova Publishers, 2009, pp. 341–360.
T. Meng, A. T. Soliman, M.-L. Shyu, Y. Yang, S.-C. Chen, S. S. Iyengar, J. S. Yordy, P. Iyengar, Wavelet analysis in current cancer genome research: A survey, IEEE/ACM Transactions on Computational Biology and Bioinformatics 10 (2013) 1442–14359. [](http://dx.doi.org/10.1109/TCBB.2013.134).
G. A. Worrell, K. Jerbi, K. Kobayashi, J. M. Lina, R. Zelmann, M. Le Van Quyen, Recording and analysis techniques for high-frequency oscillations, Progress in neurobiology 98 (2012) 265–278. [](http://dx.doi.org/10.1016/j.pneurobio.2012.02.006).
A. Suvichakorn, C. Lemke, A. Schuck Jr, J.-P. Antoine, The continuous wavelet transform in MRS, Tutorial text, Marie Curie Research Training Network FAST, 2010.
A. E. Hramov, A. A. Koronovskii, V. A. Makarov, A. N. Pavlov, E. Sitnikova, WWavelet in neuroscience, Springer Berlin, 2015.
L. Sheppard, A. Stefanovska, P. McClintock, Detecting the harmonics of oscillations with time-variable frequencies, Phys. Rev. E 83 (2011) 016206. [](http://dx.doi.org/10.1103/PhysRevE.83.016206).
H. R. Karimi, W. Pawlus, K. G. Robbersmyr, Signal reconstruction, modeling and simulation of a vehicle full-scale crash test based on morlet wavelets, Neurocomputing 93 (2012) 88–99. [](http://dx.doi.org/10.1016/j.neucom.2012.04.010).
J. Gu, H. Xu, J. Wang, T. An, W. Chen, The application of continuous wavelet transform based foreground subtraction method in 21 cm sky surveys, Astrophysical Journal 773 (2013) 38. [](http://dx.doi.org/10.1088/0004-637X/773/1/38).
D. E. Postnov, A. Y. Neganova, D. D. Postnov, A. R. Brazhe, Monitoring of rhythms in laser speckle data, Journal of Innovative Optical Health Sciences 7 (2014) 1450015. [](http://dx.doi.org/10.1142/S1793545814500151).
E. A. Lebedeva, E. B. Postnikov, On alternative wavelet reconstruction formula: a case study of approximate wavelets, Royal Society Open Science 1 (2014) 140124. [](http://dx.doi.org/10.1098/rsos.140124).
J. Krupic, N. Burgess, J. O’Keefe, Neural representations of location composed of spatially periodic bands, Science 337 (2012) 853–857. [](http://dx.doi.org/10.1126/science.1222403).
Wavelab 850:http://statweb.stanford.edu/ wavelab/.
R. D. Traub, R. Duncan, A. J. C. Russell, T. Baldeweg, Y. Tu, M. O. Cunningham, M. A. Whittington, Spatiotemporal patterns of electrocorticographic very fast oscillations ($> 80$ hz) consistent with a network model based on electrical coupling between principal neurons, Epilepsia 51 (2010) 1587–1597. [](http://dx.doi.org/10.1111/j.1528-1167.2009.02420.x).
A. H. Siddiqi, Applied Functional Analysis: Numerical Methods, Wavelet Methods, and Image Processing, CRC Press, 2003.
M. A. Belluscio, K. Mizuseki, R. Schmidt, R. Kempter, G. Buzs[á]{}ki, Cross-frequency phase–phase coupling between theta and gamma oscillations in the hippocampus, J. Neurosci. 32 (2012) 423–435. [](http://dx.doi.org/10.1523/JNEUROSCI.4122-11.2012).
M. O. Cunningham, A. Roopun, I. S. Schofield, R. G. Whittaker, R. Duncan, A. Russell, A. Jenkins, C. Nicholson, M. A. Whittington, R. D. Traub, Glissandi: transient fast electrocorticographic oscillations of steadily increasing frequency, explained by temporally increasing gap junction conductance, Epilepsia 53 (2012) 1205–1214. [](http://dx.doi.org/10.1111/j.1528-1167.2012.03530.x).
Appendix {#appendix .unnumbered}
========
MATLAB code, which generates the example represented in Figures \[funccwt\], \[recfunc\].
% Determination of the function
N=512;
t=linspace(0,1,N);
f1=cos(20*pi*t);
f2=exp(-4*t);
f3=sin(40*pi*t).*ramp(t,t(round(N/3)),t(round(2*N/3)));
f=f1.*f2+f3;
% Boundary continuation (extension)
[te,fe]=fcontin(t,f);
%Tranform to analytical function
fa=sAnalytic(fe);
The corresponding functions, which realize extended equispaced samples via the boundary continuation
function [tp,fp]=fcontin(t,f);
n=length(t);tv(:,1)=t;fv(:,1)=f;
%Interpolation into power-two sequences:
N=2^(ceil(log2(n)));
ti(:,1)=linspace(tv(1),tv(end),N);
fi=interp1q(tv,fv,ti);
%Boundary continuation
N=2^(ceil(log2(n)));
dt=ti(2)-ti(1);
tp=zeros(2*N,1);
tp(1:N/2)=linspace(ti(1)-N*dt/2,-dt,N/2);
tp(N/2+1:1.5*N)=ti;
tp(1.5*N+1:2*N)=linspace(ti(end)+dt,ti(end)+N*dt/2,N/2);
fi=interp1q(tv,fv,ti);
% Boundary reflection
fp(1:N/2)=flipud(conj(fi(2:N/2+1)));
fp(N/2+1:1.5*N)=fi;
fp(1.5*N+1:2*N)=flipud(conj(fi(N/2:N-1)));
and by the cut off of negative frequencies that forms an analytic function:
function fa=sAnalytic(f);
N=length(f);
%Cut-off of negative frequencies
F=fft(f);
F=[2*F(1:N/2),zeros(1,N/2)];
%Output: the analytic signal
fa=ifft(F);
MATLAB function for the CWT with the Morlet wavelet in the amplitude norm, which should be applied to the formed extended sample
function w=fftMorlet(t,fp,a,omega0);
N=length(t);
%Fourier transform
F=fft(fp);
nrm=2*pi/(t(end)-t(1));
omega_=([(0:(N/2)) (((-N/2)+1):-1)])*nrm;
%Convolution
if a(1)==0
w(1,:)=fp*exp(-omega0^2/2);
k1=2;
else
k1=1;
end
for k=k1:length(a);
omega_s=a(k)*omega_;
window=exp(-(omega_s-omega0).^2/2);
cnv(k,:)=window.*F;
w(k,:)=ifft(cnv(k,:));
end
Fig. \[funccwt\](lower panel) presents `abs(w)` is obtained as
omega0=2*pi;
a=linspace(0,0.2,51);
we=fftMorlet(te,fa,a,omega0);
ti=te(0.5*N+1:1.5*N);
w=we(:,0.5*N+1:1.5*N);
The reconstruction procedure is realized in MATLAB as the function
function iw=invMorlet(t,a,w);
N=length(t);
d2t=t(3)-t(1);
dw(:,2:N-1)=(w(:,3:end)-w(:,1:end-2))/d2t;
dw(:,1)=(w(:,2)-w(:,1))/(t(2)-t(1));
dw(:,N)=(w(:,N)-w(:,N-1))/(t(N)-t(N-1));
iw=imag(trapz(a,dw))/sqrt(2*pi);
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We prove a version of Wordingham’s theorem for left regular representations in the setting of Fell bundles of inverse semigroups and use this result to discuss the various associated cross sectional $C^*$-algebras.'
address:
- 'Erik B[é]{}dos, Institute of Mathematics, University of Oslo, P.b. 1053 Blindern, 0316 Oslo, Norway.'
- 'Magnus D. Norling, Oslo, Norway.'
author:
- 'Erik B[é]{}dos'
- 'Magnus D. Norling'
bibliography:
- 'bibliography.bib'
title: On Fell bundles over inverse semigroups and their left regular representations
---
Introduction
============
Following an unpublished work of Sieben, the concept of a Fell bundle over a discrete group was generalized by Exel in [@exel11], where the notion of a Fell bundle ${\mathcal{A}}= \{A_s\}_{s\in S}$ over an inverse semigroup $S$ was introduced and the associated full cross sectional $C^*$-algebra $C^*({\mathcal{A}})$ was defined as the universal $C^*$-algebra for $C^*$-algebraic representations of ${\mathcal{A}}$. This construction may be used to present other classes of $C^*$-algebras in a unified manner. For example, Buss and Exel show in [@buss_exel11] that to each partial action $\beta$ of an inverse semigroup $S$ on a C$^*$-algebra $A$, one may associate a Fell bundle ${\mathcal{A}}_\beta$ over $S$ such that $C^*({\mathcal{A}}_\beta)$ is naturally isomorphic to the full crossed product of $A$ by $\beta$. (In fact, Buss and Exel even consider twisted partial actions.) In another direction, the same authors establish in [@buss_exel_12a] that any Fell bundle $\mathcal{B}$ over an [é]{}tale groupoid $\mathcal{G}$ gives rise to a Fell bundle $\mathcal{A}$ over $S$, where $S$ is any inverse semigroup consisting of bisections (or slices) of $\mathcal{G}$ (as defined by Renault in [@re80]), and show that, under some mild assumptions, the full cross sectional C$^*$-algebra of $\mathcal{B}$ is isomorphic to $C^*({\mathcal{A}})$.
Exel also defines in [@exel11] the reduced cross sectional $C^*$-algebra $C^*_r({\mathcal{A}})$ associated to a Fell bundle ${\mathcal{A}}$ over an inverse semigroup $S$. One drawback of his construction is that it is somewhat involved (we summarize it in section \[sec:exelreduced\]). Another approach has recently been proposed in [@buss_exel_meyer15] when ${\mathcal{A}}$ is saturated and $S$ is unital, relying on the result from [@buss_meyer14] that ${\mathcal{A}}$ may then be identified with an action of $S$ by Hilbert bimodules on the unit fibre $A$ of ${\mathcal{A}}$, thus making it possible to define $C^*_r({\mathcal{A}})$ as the reduced crossed product of $A$ by this action of $S$. We believe that it would be helpful to find a more direct construction of $C^*_r({\mathcal{A}})$, at least in some cases. For example, such a construction would make it easier to initiate a study of amenability for Fell bundles over inverse semigroups, having in mind that, as for Fell bundles over groups [@exel14; @exel97], a natural definition of amenability for ${\mathcal{A}}$ is to require that the canonical $*$-homomorphism from $C^*({\mathcal{A}})$ onto $C_r^*({\mathcal{A}})$ is injective.
Given a Fell bundle ${\mathcal{A}}$ over an inverse semigroup $S$, our main goal in the present paper is to introduce a certain $C^*$-algebra $C^*_{\rm{r, alt}}({\mathcal{A}})$ which may also be considered as a kind of reduced cross sectional C$^*$-algebra for ${\mathcal{A}}$, and to compare it with Exel’s $C_r^*({\mathcal{A}})$. Inspired by the approach used by Khoshkam and Skandalis [@khoshkam_skandalis04] in the case of an action of an inverse semigroup on a $C^*$-algebra, our first step (see Section \[Fellb\]) is to associate to ${\mathcal{A}}$ a full cross sectional $C^*$-algebra $C^*_{\rm KS}({\mathcal{A}})$ which is universal for so-called pre-representations of ${\mathcal{A}}$ in $C^*$-algebras, or, equivalently, for $C^*$-algebraic representations of the convolution $*$-algebra $C_c({\mathcal{A}})$ canonically attached to ${\mathcal{A}}$. Exel’s $C^*({\mathcal{A}})$ is then easily obtained as a quotient of $C^*_{\rm KS}({\mathcal{A}})$. Our next step (see Section \[leftregrep\]) is to show that $C_c({\mathcal{A}})$ has a natural injective left regular $C^*$-algebraic representation $\Phi_\Lambda$. The injectivity of $\Phi_\Lambda$ may be seen as an analog of Wordingham’s theorem for $\ell^1(S)$ (cf. [@paterson]), and our proof is related to his original proof, although some extra arguments are necessary. Letting $C_{\rm r, KS}({\mathcal{A}})$ denote the $C^*$-algebra generated by the range of $\Phi_\Lambda$, $C^*_{\rm{r, alt}}({\mathcal{A}})$ is then defined as the quotient of $C_{\rm r, KS}({\mathcal{A}})$ by a certain canonical ideal. From the naturality of our construction, it readily follows that there is a canonical $*$-homomorphism $\Psi_{\Lambda^{\rm alt}}$ from $C^*({\mathcal{A}})$ onto $C^*_{\rm{r, alt}}({\mathcal{A}})$.
In the case where $S$ consists only of idempotents (hence is a semilattice) and ${\mathcal{E}}$ is Fell bundle over $S$, we check in Section \[semilatt\] that $C^*_{\rm KS}({\mathcal{E}}) = C^*_{\rm r, KS}({\mathcal{E}})$ and $C^*({\mathcal{E}}) = C^*_{\rm{r, alt}}({\mathcal{E}}) \simeq C^*_{\rm{r}}({\mathcal{E}})$. Next, given a Fell bundle ${\mathcal{A}}$ over an inverse semigroup $S$ such that $S$ is $E^*$-unitary (cf. Section \[prem\]) and such that $A_0 =\{0\}$ if $S$ has a 0 element, we let ${\mathcal{E}}$ denote the Fell bundle obtained by restricting ${\mathcal{A}}$ to the semilattice $E$ of idempotents in $S$ and show (in Section \[cond-exp-diag\]) that there exists a faithful conditional expectation from $C^*_{\rm r, KS}({\mathcal{A}})$ onto $C^*_{\rm r, KS}({\mathcal{E}})$. In the final section, keeping the same assumptions, we describe $C_r^*({\mathcal{A}})$ as a quotient of $C_{\rm r, KS}({\mathcal{A}})$ and show that there exists a surjective canonical $*$-homomorphism $\Psi'$ from $C^*_{\rm{r, alt}}({\mathcal{A}})$ onto $C_r^*({\mathcal{A}})$. We also characterize when $\Psi'$ is a $*$-isomorphism and end by showing that this happens frequently when $S$ is strongly $E^*$-unitary.
Preliminaries {#prem}
=============
We recall that a semigroup is a set equipped with an associative binary operation, while a monoid is a semigroup with an identity. A commutative idempotent semigroup is called a semilattice. We also recall that an inverse semigroup is a semigroup $S$ where for each $s\in S$ there is a unique $s^* \in S$ satisfying $$ss^*s= s\mbox{ and }s^*ss^* = s^*.$$ The map $s\mapsto s^*$ is then an involution on $S$. Every inverse semigroup $S$ contains a canonical semilattice, namely $$E(S)=\{e\in S:e^2=e\}$$ satisfying $E(S) = \{e\in S:e^2=e = e^*\} = \{s^*s:s\in S\}=\{tt^*:t\in S\}.$ Throughout this article, $S$ will be a fixed inverse semigroup and $E=E(S)$ will denote its semilattice of idempotents. We refer to [@lawson] and [@paterson] for the basics of the theory of inverse semigroups. We recall below a few facts that we will need later.
There is a natural partial order relation $\leq$ on $S$ given by $s\leq t$ if and only if $s=et$ for some $e\in E$, if and only if $s=tf$ for some $f\in E$, where $e$ may be chosen to be $ss^*$, and $f$ to be $s^*s$. For $e,f \in E$, we have $e \leq f$ if and only if $e = ef$.
Many inverse semigroups have a zero, that is, an element $0$ satisfying $0s = s0 = 0$ for all $s\in S$. Such an element is necessarily unique and lies in $E$. If $S$ has a zero, we set $S^\times = S\setminus \{0\}$ and $E^\times = E\setminus \{0\}$. Otherwise, we set $S^\times = S$ and $E^\times = E$.
We will say that $S$ is *$E^*$-unitary* if the set $\{ s\in S: e\leq s\}$ is contained in $E$ for every $e \in E^\times$. If this holds for every $e \in E$, then $S$ is called *$E$-unitary*. These two concepts clearly coincide if $S$ does not have a zero. For inverse semigroups having a zero, $E$-unitarity is a too strong requirement, only satisfied by semilattices. We note that $E^*$-unitarity is usually only defined for inverse semigroups having a zero, in which case it is sometimes called $0$-$E$-unitarity (and is defined as above). Our use of terminology will allow us to unify some statements. The class of $E^*$-unitary inverse semigroups has by far been the one who has received most attention from $C^*$-algebraists, most probably because they are easier to handle.
We will also need to refer to a stronger form of $E^*$-unitarity. We recall that a map $\sigma$ from $S$ into a group with identity $1$ is called a *grading* if $\sigma(st)=\sigma(s)\sigma(t)$ whenever $s, t \in S$ and $st\in S^\times$, and that $\sigma$ is said to be *idempotent pure* if $\sigma^{-1}(\{1\})=E$. Then $S$ is said to be *strongly $E^*$-unitary* [@fleming_fountain_gould99] if there exists an idempotent pure grading from $S$ into some group. It is known that $S$ is $E^*$-unitary whenever it is strongly $E^*$-unitary.
A semigroup homomorphism from $S$ into another inverse semigroup is necessarily $*$-preserving, so this provides the natural notion of homomorphism between inverse semigroups. If $X$ is a set, then $\mathcal{I}(X)$ will denote the symmetric inverse semigroup on $X$, consisting of all partial bijections on $X$ (with composition defined on the largest possible domain). An *action of $S$ on $X$* is then a homomorphism of $S$ into $\mathcal{I}(X)$. The essence of the Wagner-Preston theorem is that there always exists an injective action of $S$ on some set.
When $A$ is a $C^*$-algebra, we will let ${\rm PAut}(A)$ denote the inverse subsemigroup of $\mathcal{I}(A)$ consisting of all partial $*$-automorphisms of $A$; so $\phi \in {\rm PAut}(A)$ if and only if $\phi$ is a $*$-isomorphism between two ideals of $A$. Here, and in the sequel, ideals in $C^*$-algebras are always assumed to be two-sided and closed, unless otherwise specified. An *action of $S$ on a $C^*$-algebra $A$* is an homomorphism $\alpha$ from $S$ into ${\rm PAut}(A)$. Khoshkam and Skandalis show in [@khoshkam_skandalis04] how to associate to such an action a full (resp. reduced) $C^*$-crossed product, which we will denote by $A\rtimes^{\rm KS}_{\alpha} S$ (resp. $A\rtimes^{\rm KS}_{\alpha, r} S$). In fact, their construction goes through for a more general kind of action of $S$ on $A$, and the interested reader should consult [@khoshkam_skandalis04] for more details, including a discussion of the relationship between their full crossed product and the crossed product construction previously introduced by Sieben in [@sieben].
Following [@buss_exel12], one may define *partial* actions of $S$ on sets and on $C^*$-algebras. As mentioned in the introduction, Buss and Exel actually consider *twisted* partial actions of $S$ in [@buss_exel12], but we will restrict ourselves to the untwisted case to avoid many technicalities. Partial actions were first introduced in the case where $S$ is a group, and the reader may consult [@exel14] for a nice introduction to this subject, including many references to the literature. We recall a few relevant definitions and facts from [@buss_exel12]. A *partial homomorphism* of $S$ in a semigroup $H$ is a map $\pi: S\mapsto H$ such that
1. $\pi(s)\pi(t)\pi(t^*) = \pi(st)\pi(t^*)$,
2. $\pi(s^*)\pi(s)\pi(t) = \pi(s^*)\pi(st)$,
3. $\pi(s)\pi(s^*)\pi(s) = \pi(s)$
hold for all $s,t\in S$. Note that if $H$ is an inverse semigroup, then (iii) implies
4. $\pi(s^*)=\pi(s)^*$
for all $s\in S$; hence, in this case, $\pi$ is a partial homomorphism if and only if (i), (ii) and (iv) hold. Moreover, still assuming that $H$ is an inverse semigroup, this is equivalent to requiring that the three conditions
- $\pi(s^*)=\pi(s)^*$,
- $\pi(s)\pi(t)\leq \pi(st)$,
- $\pi(s)\leq \pi(t)$ whenever $s\leq t$,
hold for $s,t \in S$, cf. [@buss_exel12 Proposition 3.1].
A *partial action* of $S$ on a set $X$ (resp. on a $C^*$-algebra $A$) is then defined as a partial homomorphism $\beta$ from $S$ into $\mathcal{I}(X)$ (resp. into ${\rm PAut}(A)$). As in [@buss_exel12], we will also require that a partial action $\beta$ of $S$ on a $C^*$-algebra $A$ satisfies that the union $\cup_{s\in S} J_s = \cup_{s\in S}\, {\rm im}({\beta_s})$ spans a dense subspace of $A$.
Fell bundles over inverse semigroups {#Fellb}
====================================
In [@exel11], Exel defines a Fell bundle over $S$ as a quadruple $${\mathcal{A}}= \big(\{A_s\}_{s\in S} , \{\mu_{s,t}\}_{s,t\in S} , \{*_s\}_{s\in S} , \{j_{t,s}\}_{s,t\in S, s\leq t}\big)$$ where for $s,t\in S$ we have that
1. $A_s$ is a complex Banach space,
2. $\mu_{s,t}:A_s\odot A_t\to A_{st}$ is a linear map,
3. $*_s:A_s\to A_{s^*}$ is a conjugate linear isometric map,
4. $j_{t,s}:A_s\hookrightarrow A_t$ is a linear isometric map whenever $s\leq t$.
It is moreover required that for every $r,s,t\in S$, and every $a\in A_r,b\in A_s$, and $c\in A_t$, we have
5. $\mu_{rs,t}(\mu_{r,s}(a\otimes b)\otimes c)=\mu_{r,st}(a\otimes\mu_{s,t}(b\otimes c))$,
6. $*_{rs}(\mu_{r,s}(a\otimes b))=\mu_{s^*,r^*}(*_s(b)\otimes *_r(a))$,
7. $*_{s^*}(*_s(a))=a$,
8. $\|\mu_{r,s}(a\otimes b)\|\leq\|a\| \, \|b\|$,
9. $\|\mu_{r^*,r}(*_r(a)\otimes a)\| = \|a\|^2$,
10. $\mu_{r^*,r}(*_r(a)\otimes a) \geq 0$ in $A_{r^* r}$,
11. if $r\leq s\leq t$, then $j_{t,r}=j_{t,s}\circ j_{s,r}$,
12. if $r\leq r'$ and $s\leq s'$, then $j_{r's',rs}\circ\mu_{r,s}=\mu_{r',s'}\circ (j_{r',r}\otimes j_{s',s})$ and $j_{s',s}\circ *_s = *_{s'}\circ j_{s',s}$.
As shown by Exel, axioms (i)-(ix) imply that $A_e$ is a $C^*$-algebra whenever $e \in E$, with $cd= \mu_{e,e}(c \otimes d)$ and $c^*= *_e(c)$ for $c,d \in A_e$. Hence, the requirement in axiom (x) is meaningful. Exel also shows that the following properties hold:
1. $j_{s,s}$ is the identity map ${\rm id}_{A_s}$ for every $s\in S$;
2. If $e,f\in E$ and $e\leq f$, then $j_{f,e}(A_e)$ is an ideal in $A_f$.
When no confusion is possible, we will use the simplified notation $$a\cdot b:=\mu_{s,t}(a\otimes b)\mbox{ and }a^*:=*_s(a)$$ whenever $a\in A_s$ and $b\in A_t$, and just write ${\mathcal{A}}= (\{A_s\}_{s\in S} , \{j_{t,s}\}_{s,t\in S,\, s\leq t})$, or even only ${\mathcal{A}}= \{A_s\}_{s\in S}$ in some cases. If $s \in S$ and $e:= s^*s \in E$, then one easily verifies that $A_s$ becomes a right Hilbert $A_e$-module with respect to the right action given by $(a_s, a_e)\mapsto a_s \cdot a_e \in A_{ss^*s} = A_s$ for $a_s \in A_s$ and $a_e \in A_e$, and the $A_e$-valued inner product given by $\langle a_s, b_s \rangle = a_s^* \cdot b_s$ for $a_s, b_s \in A_s$. For later use, we also note the following fact: $$\label{a-e-f}
\text{If } e, f\in E, e\leq f, \text{and }c,d \in A_e, \text{then } j_{f,e}(c)\cdot d = cd\,.$$ Indeed, using axioms (xii) and (xiii), we get $$j_{f,e}(c)\cdot d = j_{f,e}(c)\cdot j_{e,e}(d) = j_{fe,ee}(c\cdot d) = j_{e,e}(c \cdot d) = c\cdot d = cd\,.$$
We also recall that a Fell bundle ${\mathcal{A}}$ is called *saturated* when the span of $\{a_s\cdot b_t : a_s\in A_s, b_t\in A_t\}$ is dense in $A_{st}$ for all $s,t \in S$.
{#FB_PA}
An important class of examples of Fell bundles over inverse semigroups arises from (twisted) partial actions of inverse semigroups on $C^*$-algebras (cf. [@buss_exel12 Section 8]). For the ease of the reader, we sketch this construction in the untwisted case.
Let $\beta:S\to {\rm PAut}(A)$ be a partial action of $S$ on a $C^*$-algebra $A$. For each $s\in S$, set $J_{s^*}={\rm dom}(\beta_s)$ and $J_s={\rm im}(\beta_s)$, so $\beta_s$ is a $*$-isomorphism from $J_{s^*}$ onto $J_s$ and $\beta_s^{-1} = \beta_{s^*}$. One may then show (cf. [@buss_exel12 Proposition 3.4, Proposition 3.8 and Proposition 6.3]) that the family $\{J_s\}$ satisfy certain compatibility properties, such as $\beta_s(J_{s^*} \cap J_t) = J_s \cap J_{st}\,$, $J_s \subset J_{ss^*}$ (and $J_s = J_{ss^*}$ if $\beta$ is an action), $J_s \subset J_t$ whenever $s\leq t$, and $\beta_e = {\rm id}_{J_e}$ when $e\in E$.
Now, for each $s\in S$, we set $A_s= \{(a, s): a\in J_s\}$ and organize $A_s$ as a Banach space by identifying $a\in J_s$ with $(a,s)\in A_s$. Note that if $s,t \in S$, $a \in J_s$ and $b\in J_t$, then $$\beta_s^{-1}(a)\,b \in J_{s^*} \cap J_t\,, \,\text{so} \, \,\beta_s\Big(\beta_s^{-1}(a)b\Big) \in J_s \cap J_{st}\,,\,$$ $$\text{and}\,\, a^* \in J_s\,,\, \text{so} \, \,\beta_s^{-1}(a^*) \in J_{s^*}\,;$$ hence one may define $\mu_{s,t}\big((a,s)\otimes (b,t)\big)=:(a,s) \cdot (b,t)$ and $*_s(a,s)=: (a,s)^*$ by $$(a,s) \cdot (b,t) = \Big(\beta_s\big(\beta_s^{-1}(a)b\big), st\Big) \in A_{st} \, \, \text{and} \,\, (a,s)^* = \big(\beta_s^{-1}(a^*), s^*) \in A_{s^*}\,.$$ Moreover, if $s\leq t$, then $\beta_s \leq \beta_t$, so $J_s = {\rm im}(\beta_s) \subset {\rm im}(\beta_t) = J_t $, and one may then define $j_{t,s}: A_s\to A_t$ by $$j_{t,s} (a,s) = (a, t)\,, \quad \text{for all}\,\, a \in J_s\,.$$ It may then be checked that ${\mathcal{A}}= (\{A_s\}_{s\in S}, \{j_{t,s}\}_{s,t\in S,\, s\leq t})$ becomes a Fell bundle over $S$ with respect to these operations (cf. [@buss_exel11] for the case of a global (twisted) action).
{#exelfull}
Still following [@exel11], a *pre-representation* of a Fell bundle $${\mathcal{A}}=(\{A_s\}_{s\in S} , \{j_{t,s}\}_{s,t\in S,\, s\leq t})$$ in a complex $*$-algebra $B$ is a family $\Pi=\{\pi_s\}_{s\in S}$, where for each $s\in S$, $$\pi_s:A_s\to B$$ is a linear map such that for all $s,t\in S$, all $a\in A_s$, and all $b\in A_t$, one has
1. $\pi_{st}(a\cdot b)=\pi_s(a)\pi_t(b)$,
2. $\pi_{s^*}(a^*)=\pi_s(a)^*$.
If, in addition, $\Pi$ satisfies
1. $\pi_t\circ j_{t,s}=\pi_s$ whenever $s\leq t$,
then $\Pi$ is called a [*representation*]{} of ${\mathcal{A}}$ in $B$.
We recall that if $\Pi$ is a pre-representation of ${\mathcal{A}}$ in a $C^*$-algebra $B$, then for each $s\in S$ and $a\in A_s$, we have $\|\pi_s(a)\| \leq \|a\|$. Indeed, as $\pi_e:A_e\mapsto B$ is then a $*$-homomorphism between $C^*$-algebras for every $e \in E$, we get $$\|\pi_s(a)\|^2 = \|\pi_s(a)^*\pi_s(a)\| = \|\pi_{s^*}(a^*)\pi_s(a)\| = \|\pi_{s^*s}(a^*\cdot \,a)\| \leq \|a^*\cdot \,a\| =\|a\|^2.$$
Consider now the direct sum of vector spaces $${\mathcal{C}_c({\mathcal{A}})} = \bigoplus_{s\in S} A_s$$ We will often write an element $g \in {\mathcal{C}_c({\mathcal{A}})}$ as a formal sum $
g=\sum_{s\in S}a_s\delta_s
$ where $a_s \in A_s$ for $s\in S$ and all but finitely many $a_s$ are equal to $0$. Then ${\mathcal{C}_c({\mathcal{A}})}$ can be given the structure of a complex $*$-algebra by extending linearly the operations $$\begin{aligned}
(a_s\delta_s)(b_t\delta_t) &= (a_s\cdot b_t)\delta_{st}\\
(a_s\delta_s)^* &= a_s^*\delta_{s^*}\end{aligned}$$ Alternatively, if one prefers to write ${\mathcal{C}_c({\mathcal{A}})}$ as $${\mathcal{C}_c({\mathcal{A}})}=\Big\{ g\in \prod_{s\in S} A_s : g(s) = 0 \,\,\text{for all but finitely many}\, s\Big\},$$ one may define the product and the involution on ${\mathcal{C}_c({\mathcal{A}})}$ by $$(f\ast g)(r) = \sum_{s,\, t \in S,\, st \,=\, r } \,f(s)\cdot g(t)\quad \text{and} \quad f^*(r) = f(r^*)^*$$ for $f, g \in {\mathcal{C}_c({\mathcal{A}})}$ and $r\in S$.
For each $s\in S$, let $\pi^0_s:A_s \to {\mathcal{C}_c({\mathcal{A}})}$ be defined by $$\pi^0_s(a_s) = a_s \delta_s$$ for each $a_s \in A_s$. Then $\Pi^0:=\big\{\pi^0_s\big\}_{s\in S}$ is a pre-representation of ${\mathcal{A}}$ in ${\mathcal{C}_c({\mathcal{A}})}$, which satisfies the following universal property (cf. [@exel11 Proposition 3.7]):
To each pre-representation $\Pi=\{\pi_s\}_{s\in S}$ of ${\mathcal{A}}$ in a $*$-algebra $B$ one may associate a $*$-homomorphism $\Phi_\Pi:{\mathcal{C}_c({\mathcal{A}})}\to B$ given by $$\Phi_\Pi\Big(\sum_{s\in S}a_s\delta_s\Big)=\sum_{s\in S}\pi_s(a_s)\,,$$ which satisfies $\Phi_\Pi\,\circ\, \pi_s^0 = \pi_s$ for all $s\in S$. Moreover, the map $\Pi \mapsto \Phi_\Pi$ gives a bijection between pre-representations of ${\mathcal{A}}$ in $B$ and $*$-homomophisms from ${\mathcal{C}_c({\mathcal{A}})}$ into $B$.
Consider $g=\sum_{s\in S}a_s\delta_s \in {\mathcal{C}_c({\mathcal{A}})}$. If $B$ is a $C^*$-algebra and $\Pi$ is a pre-representation of ${\mathcal{A}}$ in $B$, then we have $$\|\Phi_\Pi(g)\| = \big\|\sum_{s\in S}\pi_s(a_s)\big\| \leq \sum_{s\in S}\|\pi_s(a_s)\| \leq \sum_{s\in S}\,\|a_s\|\,.$$ Hence, if we define $$\|g\|_{\rm u} := \sup_\Phi \big\{ \|\Phi(g)\| \}\,,$$ where the supremum is taken over all $*$-homomorphisms from ${\mathcal{C}_c({\mathcal{A}})}$ into any $C^*$-algebra, then $$\begin{aligned}
\|g\|_{\rm u} &= \sup\big\{ \|\Phi_\Pi(g)\| :
\Pi \ \text{is a pre-representation of ${\mathcal{A}}$ in some $C^*$-algebra}\big\} \\
&\leq \, \sum_{s\in S}\,\|a_s\| \,< \,\infty\,.\end{aligned}$$ Hence $\|\cdot\|_{\rm u}$ gives a $C^*$-seminorm on ${\mathcal{C}_c({\mathcal{A}})}$. As we will show in the next section, there always exists an injective $*$-representation of ${\mathcal{C}_c({\mathcal{A}})}$ in some $C^*$-algebra (namely the one associated to the left regular representation of ${\mathcal{C}_c({\mathcal{A}})}$). It follows that $\|\cdot\|_{\rm u}$ is in fact a $C^*$-norm, and we may therefore define the *full KS-cross sectional $C^*$-algebra of ${\mathcal{A}}$*, denoted by $C^{*}_{\rm KS}({\mathcal{A}})$, as the completion of ${\mathcal{C}_c({\mathcal{A}})}$ w.r.t. $\|\cdot\|_{\rm u}$.
We will use the same notation to denote the norm on $C^{*}_{\rm KS}({\mathcal{A}})$ and will identify ${\mathcal{C}_c({\mathcal{A}})}$ with its canonical copy in $C^{*}_{\rm KS}({\mathcal{A}})$. We may therefore regard $\Pi^0$ as a pre-representation of ${\mathcal{A}}$ in $C^{*}_{\rm KS}({\mathcal{A}})$, which is universal in the sense that given any pre-representation $\Pi$ of ${\mathcal{A}}$ in a $C^*$-algebra $B$, then there exists a unique $*$-homomorphism from $C^{*}_{\rm KS}({\mathcal{A}})$ into $B$, which we also denote by $\Phi_{\Pi}$, satisfying $\Phi_\Pi\, \circ\, \pi^0_s= \pi_s$ for all $s\in S$.
If for example ${\mathcal{A}}_\alpha$ denotes the Fell bundle associated to an action $\alpha$ of $S$ on a $C^*$-algebra $A$, then it is straightforward to verify that $C^{*}_{\rm KS}\big({\mathcal{A}}_\alpha\big)$ coincides with the full KS-crossed product $A\rtimes^{\rm KS}_\alpha S$ constructed in [@khoshkam_skandalis04]. Thus, if $\beta$ is a partial action of $S$ on a $C^*$-algebra $A$, it is natural to define the *full KS-crossed product* by $A\rtimes^{\rm KS}_\beta S:=C^{*}_{\rm KS}\big({\mathcal{A}}_\beta\big),$ where ${\mathcal{A}}_\beta$ denotes the Fell bundle over $S$ associated to $\beta$ in \[FB\_PA\].
{#sec:fullcross}
In [@exel11], Exel defines the full cross sectional $C^*$-algebra $C^{*}({\mathcal{A}})$ of a Fell bundle ${\mathcal{A}}= \big(\{A_s\}_{s\in S}, \{j_{t,s}\}_{s,t\in S,\, s\leq t}\big)$. This algebra may be described as a quotient of $C^{*}_{\rm KS}({\mathcal{A}})$. To explain this, we first have to review Exel’s construction. Let ${\mathcal{N}}_{\mathcal{A}}$ denote the subspace of ${\mathcal{C}_c({\mathcal{A}})}$ spanned by the set $$\Big\{a_s\delta_s-j_{t,s}(a_s)\delta_t:s,t\in S, s\leq t, a_s\in A_s\Big\}.$$ Exel shows in [@exel11 Proposition 3.9] that ${\mathcal{N}}_{\mathcal{A}}$ is a two-sided selfadjoint ideal of ${\mathcal{C}_c({\mathcal{A}})}$. It follows that ${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$ becomes a complex $*$-algebra in the obvious way. Moreover, [@exel11 Proposition 3.10] says that if $\Pi$ is a pre-representation of ${\mathcal{A}}$ in a $*$-algebra $B$, then $\Pi$ is a representation of ${\mathcal{A}}$ if and only if $\Phi_\Pi$ vanishes on ${\mathcal{N}}_{\mathcal{A}}$, in which case we will denote the associated $*$-homomorphism from ${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$ into $B$ by $\widetilde\Phi_\Pi$. The map $\Pi\mapsto \widetilde\Phi_\Pi$ gives then a bijection between representations of ${\mathcal{A}}$ in $B$ and $*$-homomorphisms from ${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$ into $B$.
Now, for any $g=\sum_{s\in S}a_s\delta_s \in {\mathcal{C}_c({\mathcal{A}})}$ and any representation $\Pi$ of ${\mathcal{A}}$ in a $C^*$-algebra, we have $$\|\widetilde\Phi_\Pi(g+ {\mathcal{N}}_{\mathcal{A}})\| = \|\Phi_\Pi(g)\| \leq \|g\|_{\rm u} \,.$$ It follows that if we define $$\|g+ {\mathcal{N}}_{\mathcal{A}}\|_{*}:=\sup_\Psi \big\{ \|\Psi\big(g+ {\mathcal{N}}_{\mathcal{A}}\big)\|\}$$ where the supremum is taken over all $*$-homomorphisms $\Psi$ from ${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$ into a $C^*$-algebra, we get $$\begin{aligned}
\|g+ {\mathcal{N}}_{\mathcal{A}}\|_{*}&= \sup\big\{ \|\Phi_\Pi(g)\| :
\Pi \ \text{is a representation of ${\mathcal{A}}$ in some $C^*$-algebra}\big\} \\
&\leq \|g\|_{\rm u} \,,\end{aligned}$$ so $\|\cdot\|_{*}$ gives a $C^*$-seminorm on ${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$. The *full $($Exel$)$ cross sectional $C^*$-algebra $C^*({\mathcal{A}})$* is then defined as the Hausdorff completion of ${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$ w.r.t. to this seminorm.
Letting $Q_{{\mathcal{A}}}:{\mathcal{C}_c({\mathcal{A}})}\to {\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$ denote the quotient map and $R_{{\mathcal{A}}}:{\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}\to C^*({\mathcal{A}})$ denote the canonical map, we get that $\iota_{{\mathcal{A}}}:=R_{{\mathcal{A}}}\,\circ \,Q_{{\mathcal{A}}}$ is a contractive $*$-homomorphism from ${\mathcal{C}_c({\mathcal{A}})}$ into $C^*({\mathcal{A}})$ having dense range. Thus, $\iota_{{\mathcal{A}}}$ extends to a $*$-homomorphism $q_{{\mathcal{A}}}$ from $C^{*}_{\rm KS}({\mathcal{A}})$ onto $C^*({\mathcal{A}})$ such that $$\label{fullcross}
C^*({\mathcal{A}}) \simeq \,C^{*}_{\rm KS}({\mathcal{A}})\, / \,{\rm Ker} \, q_{{\mathcal{A}}}\,.$$ Now, for each $s\in S$, define $\pi^{{\mathcal{A}}}_s: A_s \to C^*({\mathcal{A}})$ by $$\pi^{{\mathcal{A}}}_s = q_{{\mathcal{A}}}\, \circ \, \pi^{0}_s \ (=\iota_{{\mathcal{A}}}\, \circ \, \pi^{0}_s)\,.$$ Then one checks (cf. [@exel11 Proposition 3.12] and the proof of [@exel11 Proposition 3.13]) that $$\Pi^{{\mathcal{A}}}:=\big\{\pi^{{\mathcal{A}}}_s\big\}_{s\in S}$$ is a representation of ${\mathcal{A}}$ in $C^*({\mathcal{A}})$ satisfying the following universal property: given any representation $\Pi=\{\pi_s\}_{s\in S}$ of ${\mathcal{A}}$ in a $C^*$-algebra $B$, there exists a unique $*$-homomorphism $\Psi_\Pi: C^*({\mathcal{A}})\to B$ such that $\Psi_\Pi \circ\, \pi^{{\mathcal{A}}}_s = \pi_s$ for all $s\in S$. It follows immediately that $\Phi_\Pi = \Psi_\Pi \, \circ \,q_{{\mathcal{A}}}$ for every such representation $\Pi$.
The ideal $\,{\rm Ker} \, q_{{\mathcal{A}}}$ has a natural description in terms of ${\mathcal{N}}_{\mathcal{A}}$. Indeed, letting ${\mathcal{M}}_{\mathcal{A}}$ denote the ideal of $C^{*}_{\rm KS}({\mathcal{A}})$ given by ${\mathcal{M}}_{\mathcal{A}}:=\overline{{\mathcal{N}}_{\mathcal{A}}}^{\,\|\cdot\|_{\rm u}}$, we have $$\,{\rm Ker} \, q_{{\mathcal{A}}} = {\mathcal{M}}_{\mathcal{A}}.$$ To prove this, we first note that since $q_{{\mathcal{A}}}({\mathcal{N}}_{\mathcal{A}}) = \iota_{{\mathcal{A}}}({\mathcal{N}}_{\mathcal{A}}) = \{0\} $, we have $${\mathcal{M}}_{\mathcal{A}}\subset \,{\rm Ker} \, q_{{\mathcal{A}}}\,.$$ Next, let $s\in S$ and define $\omega_s:A_s \to C^{*}_{\rm KS}({\mathcal{A}})/{\mathcal{M}}_{\mathcal{A}}$ by $$\omega_s (a_s) := \pi^0_s(a_s) + {\mathcal{M}}_{\mathcal{A}}$$ for every $a_s \in A_s$. As ${\mathcal{M}}_{\mathcal{A}}$ contains ${\mathcal{N}}_{\mathcal{A}}$, one easily verifies that $\Omega=\{\omega_s\}_{s\in S}$ is a representation of ${\mathcal{A}}$ in $C^{*}_{\rm KS}({\mathcal{A}})/{\mathcal{M}}_{\mathcal{A}}$. The associated $*$-homomorphism $\Phi_\Omega$ from $C^{*}_{\rm KS}({\mathcal{A}})$ into $C^{*}_{\rm KS}({\mathcal{A}})/{\mathcal{M}}_{\mathcal{A}}$ is then nothing but the quotient map. Since $\Phi_\Omega = \Psi_\Omega \, \circ \,q_{{\mathcal{A}}}$, it follows that $${\rm Ker} \, q_{{\mathcal{A}}} \subset {\rm Ker}\, \Phi_\Omega = {\mathcal{M}}_{\mathcal{A}}\,.$$ Thus we get $\,{\rm Ker} \, q_{{\mathcal{A}}} = {\mathcal{M}}_{\mathcal{A}}$, as desired. It follows that $$\label{fullcross2}
C^*({\mathcal{A}}) \simeq \,C^{*}_{\rm KS}({\mathcal{A}})\, / {\mathcal{M}}_{\mathcal{A}}\,.$$
{#sec:exelreduced}
In [@exel11], Exel also constructs the reduced cross sectional $C^*$-algebra $C_r^{*}({\mathcal{A}})$ of a Fell bundle ${\mathcal{A}}= (\{A_s\}_{s\in S}, \{j_{t,s}\}_{s,t\in S,\, s\leq t})$. His construction, which is somewhat involved, may be summarized as follows.
Consider first $e\in E$ and $s\in S$ such that $e\leq s$. Then $j_{s,e}$ gives an isometric embedding of $A_e$ into $A_s$, so one may view $A_e$ as a subspace of $A_s$. Let $\varphi_e$ be a continuous linear functional on $A_e$. Exel shows in [@exel11 Proposition 6.1] that $\varphi_e$ extends to a continuous linear functional $\tilde{\varphi}_e^s$ on $A_s$ satisfying $\|\tilde{\varphi}_e^s\|=\|\varphi_e\|$ and $$\tilde{\varphi}_e^s(x) = \lim_i \varphi_e(xu_i) = \lim_i \varphi_e(u_ix)=\lim_i \varphi_e(u_ixu_i)$$ for every approximate unit $\{u_i\}_i$ for $A_e$ and every $x \in A_s$.
Next, let $e\in E$ and let $\varphi_e$ be a state on $A_e$. Define $\tilde{\varphi}_e$ on ${\mathcal{C}_c({\mathcal{A}})}$ by $$\tilde{\varphi}_e\left(\sum_{s\in S}a_s\delta_s\right)=\sum_{s\in S, \,s\geq e}\tilde{\varphi}_e^s(a_s)\,.$$ Then, as shown in [@exel11 Proposition 6.9], $\tilde{\varphi}_e$ is a state on ${\mathcal{C}_c({\mathcal{A}})}$ when ${\mathcal{C}_c({\mathcal{A}})}$ is considered as a normed $*$-algebra with respect to the norm $\|g\|_1= \sum_{s\in S}\|a_s\|$ for $g =\sum_s a_s\delta_s \in {\mathcal{C}_c({\mathcal{A}})}$.
Now, let ${\mathcal{E}}=\{A_e\}_{e\in E}$ denote the restriction of ${\mathcal{A}}$ to the semilattice $E$, let $\Pi^{{\mathcal{E}}}= \{\pi_e^{{\mathcal{E}}}\}_{e\in E}$ denote the universal representation of ${\mathcal{E}}$ in $C^*({\mathcal{E}})$, and fix a pure state $\varphi$ on $C^*({\mathcal{E}})$. For each $e \in E$, one has that $\pi^{{\mathcal{E}}}_e(A_e)$ is an ideal of $C^*({\mathcal{E}})$, and $\varphi_e:=\varphi\circ\pi_e^{{\mathcal{E}}}$ is a state on $A_e$ as long as $\varphi_e \neq 0$, that is, whenever $\varphi$ does not vanish on $\pi^{{\mathcal{E}}}_e(A_e)$. Moreover, [@exel11 Proposition 7.4] says that there exists a positive linear functional $
\tilde{\varphi}$ on ${\mathcal{C}_c({\mathcal{A}})}$ such that:
1. For every $s\in S$ and $a_s\in A_s$, one has that $$\tilde{\varphi}(a_s\delta_s)=\begin{cases}
\tilde{\varphi}^s_e(a_s) & \text{if there exists
$e\in E$ such that $\varphi_e\neq 0$ and $e\leq s$,}\\
\hspace{2.5ex} 0 & \text {otherwise,}
\end{cases}$$
2. For every $e\in E$ and every $a_e\in A_e$ one has that $$\tilde{\varphi}(a_e\delta_e)= \varphi_e(a_e) = \varphi(\pi^{{\mathcal{E}}}_e(a_e)),$$
3. $\|\tilde{\varphi}\|\leq \|\varphi\|$,
4. $\tilde{\varphi}$ vanishes on the ideal ${\mathcal{N}}_{\mathcal{A}}$.
For later use we note that if $S$ is $E^*$-unitary, then (i) and (ii) together just say that $$\label{eq:stateextension}
\tilde{\varphi}(a_s\delta_s)=\begin{cases}
\varphi(\pi^{{\mathcal{E}}}_s(a_s)) & \text{if $s\in E$,}\\
\hspace{4ex} 0 & \text{ otherwise.}
\end{cases}$$ Let $H_{\tilde{\varphi}}$ be the Hilbert space completion of ${\mathcal{C}_c({\mathcal{A}})}$ with respect to the pre-inner-product given by $$\langle g,h \rangle_{\tilde{\varphi}}=\tilde{\varphi}(h^*g) \quad \text{for} \, \, g,h\in{\mathcal{C}_c({\mathcal{A}})},$$ and let $h\mapsto\widehat{h}$ denote the canonical map ${\mathcal{C}_c({\mathcal{A}})}\to H_{\tilde{\varphi}}$. The GNS representation of $\tilde{\varphi}$, which is defined in the usual way by $$\Upsilon_{\tilde{\varphi}}(g)\widehat{h}=\widehat{gh} \quad \text{for} \, \, g,h\in{\mathcal{C}_c({\mathcal{A}})},$$ gives a $*$-representation of ${\mathcal{C}_c({\mathcal{A}})}$ on $H_{\tilde{\varphi}}$.
Exel’s reduced cross sectional $C^*$-algebra $C^*_r({\mathcal{A}})$ is then defined as the Hausdorff completion of ${\mathcal{C}_c({\mathcal{A}})}$ with respect to the $C^*$-seminorm given by $$\|g\|'_{\rm r}=
\sup_{\varphi}\|\Upsilon_{\tilde{\varphi}}(g)\|$$ where the supremum is taken over the set $\mathcal{P}(C^*({\mathcal{E}}))$ consisting of all pure states of $C^*({\mathcal{E}})$. Note that the kernel of $ \Upsilon_{\tilde{\varphi}}$ is given by $${\rm Ker} \Upsilon_{\tilde{\varphi}}=\{g\in{\mathcal{C}_c({\mathcal{A}})}: \tilde{\varphi}(h^*gh') = 0 \text{ for all } h, h'\in {\mathcal{C}_c({\mathcal{A}})} \}.$$ So if $\mathcal{K}_{\mathcal{A}}:= \{ g\in{\mathcal{C}_c({\mathcal{A}})}: \|g\|'_{\rm r}=0\}$, then $$\mathcal{K}_{\mathcal{A}}= \bigcap_{\varphi\in \mathcal{P}(C^*({\mathcal{E}}))} \, {\rm Ker} \Upsilon_{\tilde{\varphi}}\,,$$ and $C^*_r({\mathcal{A}})$ is the completion of ${\mathcal{C}_c({\mathcal{A}})}/\mathcal{K_{\mathcal{A}}}$ with respect to the norm $$\|g+ \mathcal{K}_{\mathcal{A}}\|''_{\rm r} := \|g\|'_{\rm r}.$$ Letting $\iota^{\rm red}_{\mathcal{A}}: {\mathcal{C}_c({\mathcal{A}})} \to C^*_r({\mathcal{A}})$ denote the canonical $*$-homomorphism, one gets the left regular representation $\Pi^{\rm red} = \{ \pi_s^{\rm red}\}_{s\in S}$ of ${\mathcal{A}}$ in $C^*_r({\mathcal{A}})$ by setting $$\pi_s^{\rm red}
= \iota^{\rm red}_{\mathcal{A}}\circ \pi^0_s$$ for each $s\in S$. The associated $*$-homomorphism $\Psi_{\Pi^{\rm red}}:
C^*({\mathcal{A}}) \to C^*_r({\mathcal{A}})$ is then surjective (cf. [@exel11 Proposition 8.6]).
The left regular representation of $C_{\rm KS}^{*}({\mathcal{A}})$ {#leftregrep}
==================================================================
Let ${\mathcal{A}}$ be a Fell-bundle over $S$. In this section we will describe how one may define the left regular representation $\Phi_\Lambda$ of ${\mathcal{C}_c({\mathcal{A}})}$ in a certain $C^*$-algebra $B$ naturally associated with ${\mathcal{A}}$, and show that $\Phi_\Lambda$ is injective. We will first construct the left regular pre-representation $\Lambda$ of ${\mathcal{A}}$ in $B$. The associated $*$-homomorphim $\Phi_\Lambda$ from $C_{\rm KS}^{*}({\mathcal{A}})$ into $B$ will then give the left regular representation of $C_{\rm KS}^{*}({\mathcal{A}})$.
{#WagnerP}
We begin by recalling some notation and a few facts that will be useful in our construction. For each $u \in S$, we set $$D(u) = \{s\in S:ss^*\leq u^*u\}\,,$$ so $D(u^*) = \{v\in S:vv^*\leq uu^*\}$, and for each $e\in E$, we set $$S_e=\{s\in S: s^*s=e\}\,.$$ The Wagner-Preston theorem (and its proof), see for example [@paterson Proposition 2.1.3], says that for each $u\in S$, the map $\gamma_u: D(u) \to D(u^*)$ given by $\gamma_u(s) = us$ is a bijection, with inverse given by $\gamma_{u^*}:D(u^*)\to D(u)$. Moreover, it says that the map $\gamma: u\mapsto \gamma_u$ is an injective homomorphism from $S$ into $\mathcal{I}(S)$. A part of the last statement is that for $u_1, u_2, s \in S$, we have $$\label{domain}
s \in D(u_1u_2)\ \text{if and only if} \ s \in D(u_2) \ \text {and} \ u_2s \in D(u_1)\,.$$
Consider $u\in S$ and assume $s \in S_e \cap D(u)$ for some $e \in E$. Then we have $$(us)^*us = s^*u^*us = s^*u^*u\,ss^*s = s^*ss^*s = s^*s = e\,$$ so $us \in S_e \cap D(u^*)$. Hence, if $v\in S_e \cap D(u^*)$, then $u^*v \in S_e \cap D(u)$. It follows that the map $s\mapsto us $ gives a bijection from $S_e \cap D(u) $ onto $ S_e \cap D(u^*)$, with inverse given by $v \mapsto u^*v$ for $ v\in S_e \cap D(u^*)$.
{#section}
Let now ${\mathcal{A}}=\{A_s\}_{s\in S}$ be a Fell bundle over $S$. Given $e\in E$, set $$\begin{aligned}
X_e=\Big\{\xi \in \prod_{s\in S_e}A_s
:\sum_{s\in S_e}\xi(s)^*\cdot\,\xi(s)\mbox{ is norm convergent in} \, \, A_e\Big\}\,.\end{aligned}$$ Note that the sum $\sum_{s\in S_e}\xi(s)^*\cdot\,\xi(s)$ makes sense since $$\xi(s)^*\cdot\xi(s)\in A_{s^*s}=A_e$$ for each $s\in S_e$. Proceeding in the same way as for the direct sum of a family of right Hilbert $C^*$-modules over the same $C^*$-algebra [@lance95], it is not difficult to check that $X_e$ is a subspace of the product vector space $\prod_{s\in S_e}A_s$, which becomes a right Hilbert $A_e$-module with respect to the operations $$\begin{aligned}
(\xi\cdot a)(s)&=\xi(s)\cdot a \, \in A_{se}=A_s, \\
\langle \xi,\eta \rangle_e&=\sum_{s\in S_e}\xi(s)^*\cdot\eta(s) \, \in A_e, \end{aligned}$$ for $\xi, \eta \in X_e, \, a\in A_e$ and $ s\in S_e$.
Consider $e\in E$ and $u\in S$. For $a_u\in A_u$, we let $$\lambda_{e,u}(a_u):X_e\to X_e$$ be the linear operator defined by $$\big(\lambda_{e,u}(a_u)\xi\big)(v)=\begin{cases}
a_u\cdot\xi(u^*v) & \mbox{ if } v\in D(u^*),\\
\hspace{4ex} 0 & \mbox{ otherwise.}
\end{cases}$$ for $\xi \in X_e$ and $v\in S_e$. To see that $\lambda_{e,u}(a_u)$ is well defined, let $\xi \in X_e$. If $v\in S_e \cap D(u^*)$, then $u^*v\in S_e$, and $\xi(u^*v) \in A_{u^*v}$, so we get $$a_u\cdot\xi(u^*v) \in A_u\cdot A_{u^*v}\subset A_{uu^*v}=A_{uu^*vv^*v}=A_{vv^*v}=A_v\,.$$ Thus we see that $\lambda_{e,u}(a_u)\xi$ lies in $\prod_{v\in S_e} A_v$. Moreover, if $v \in \,S_e \,\cap \,D(u^*)$, then one readily verifies that the map $b \mapsto a_u \cdot b$ is an adjointable linear map from $A_{u^*v}$ into $A_v$ (with adjoint map $c \mapsto a_u^*\cdot c$); thus, using [@lance95 Proposition 1.2], we get $$\big (a_u\cdot \xi(u^*v)\big)^*\cdot \big(a_u\cdot \xi(u^*v)\big)\, \leq \, \|a_u\|^2\ \xi(u^*v)^*\cdot \xi(u^*v)\,.$$ Now, since $\xi \in X_e$, the sum $$\sum_{v \in \,S_e \,\cap \,D(u^*)} \xi(u^*v)^*\cdot \xi(u^*v)$$ is norm-convergent in $A_e$, and it follows that $$\sum_{v\in S_e}\big(\lambda_{e,u}(a_u)\xi\big)(v)^*\cdot \big(\lambda_{e,u}(a_u)\xi\big)(v)
\, = \sum_{v \in \,S_e \,\cap \,D(u^*)} \big (a_u\cdot \xi(u^*v)\big)^*\cdot \big(a_u\cdot \xi(u^*v)\big)$$ is also norm-convergent in $A_e$. Thus, $\lambda_{e,u}(a_u)\xi \in X_e$, as desired.
Next, we show that $\lambda_{e,u}(a_u)\in{\mathcal{L}(X_e)}$. For $\xi, \eta \in X_e$, we have $$\begin{aligned}
\langle \lambda_{e,u}(a_u)\xi,\eta \rangle_e &= \sum_{v\in S_e}(\lambda_{e,u}(a_u)\xi)(v)^*\cdot\eta(v) \\
&= \sum_{v\in S_e \cap D(u^*)}(a_u\cdot\xi(u^*v))^*\cdot\eta(v)\\
&= \sum_{v\in S_e \cap D(u^*) }\xi(u^*v)^*\cdot a_u^*\cdot\eta(v)\\
&= \sum_{s\in S_e \cap D(u) }\xi(s)^*\cdot a_u^*\cdot\eta(us)\\
&= \sum_{s\in S_e}\xi(s)^*\cdot(\lambda_{e,u^*}(a_u^*)\eta)(s) \\
&=\langle \xi,\lambda_{e,u^*}(a_u^*)\eta \rangle_e\,,\end{aligned}$$ where we have used that the map $v \mapsto u^*v$ is a bijection from $S_e \cap D(u^*)$ onto $S_e \cap D(u)$. This shows that $\lambda_{e,u}(a_u)$ is an adjointable operator on $X_e$, with adjoint given by $$\label{adjoint}
\lambda_{e,u}(a_u)^*=\lambda_{e,u^*}(a_u^*).$$
Thus we get a map $\lambda_{e,u}:A_u\to{\mathcal{L}(X_e)}$ for each $e\in E$ and each $u\in S$. For each $e \in E$ we set $$\Lambda^e :=\{\lambda_{e,u}\}_{u\in S}\,.$$ To show that $\Lambda^e$ is a pre-representation of ${\mathcal{A}}$ in ${\mathcal{L}(X_e)}$, in view of (\[adjoint\]), we only have to show that for $u,u'\in S$, $a\in A_u$ and $a'\in A_{u'}$, we have $$\label{rep-prop}
\lambda_{e,uu'}(a\cdot a')=\lambda_{e,u}(a)\,\lambda_{e,u'}(a').$$
To prove this, consider $\xi \in X_e$ and $v\in S_e$. Then $$\begin{aligned}
\big(\lambda_{e,uu'}(a\cdot a')\xi\big)(v)&=\begin{cases}
a\cdot a'\cdot \xi((uu')^*v) & \mbox{ if } v\in D\big((uu')^*\big),\\
\hspace{6ex}0 & \mbox{ otherwise,}
\end{cases}\\
&=\begin{cases}
a\cdot a'\cdot \xi(u'^*u^*v) & \mbox{ if } v\in D\big(u'^*u^*\big),\\
\hspace{6ex}0 & \mbox{ otherwise,}
\end{cases}\end{aligned}$$ while $$\big(\lambda_{e,u}(a)\lambda_{e,u'}(a')\xi\big)(v)=\begin{cases}
a\cdot(\lambda_{e,u'}(a')\xi)(u^*v) & \mbox{ if } v\in D(u^*),\\
\hspace{8ex}0 & \mbox{ otherwise}
\end{cases}$$ $$=\begin{cases}
a\cdot a'\cdot\xi(u'^*u^*v) & \mbox{ if } v\in D(u^*) \mbox{ and } u^*v \in D(u'^*),\\
\hspace{8ex}0 & \mbox{ otherwise.}
\end{cases}$$ Now, using (\[domain\]) with $u_1 = u'^* $ and $u_2= u^*$ gives that $v\in D\big(u'^*u^*\big)$ if and only if $v\in D(u^*) \mbox{ and } u^*v \in D(u'^*)$, so we see that $$\big(\lambda_{e,uu'}(a\cdot a')\xi\big)(v) = \big(\lambda_{e,u}(a)\lambda_{e,u'}(a')\xi\big)(v)\,.$$ It follows that (\[rep-prop\]) holds, as desired.
We can now form the product pre-representation $\Lambda = \prod_{e\in E} \Lambda^e$ of ${\mathcal{A}}$ in the product $C^*$-algebra $B:=\prod_{e\in E}{\mathcal{L}(X_e)}$. It is natural to call $\Lambda$ the *left regular pre-representation* of ${\mathcal{A}}$ in $B$. It is given by $\Lambda =\{\lambda_u\}_{u\in S}$, where $\lambda_u:A_u\to B$ is defined by $$\lambda_u(a_u)=\Big(\lambda_{e,u}(a_u)\Big)_{e\in E}$$ for $u\in S$ and $a_u \in S$. The associated $*$-homomorphism $\Phi_\Lambda:{\mathcal{C}_c({\mathcal{A}})}\to B$ (resp. $C^*_{\rm KS}({\mathcal{A}})\to B$), which satisfies $$\Phi_\Lambda\Big(\sum_{u\in S}a_u\delta_u\Big)=
\ \left(\sum_{u\in S}\lambda_{e,u}(a_u)\right)_{e\in E},$$ will be called the *left regular representation of* ${\mathcal{C}_c({\mathcal{A}})}$ (resp. $C^*_{\rm KS}({\mathcal{A}})$) in $B$. Note that $\Phi_\Lambda=\prod_{e\in E}\Phi_{\Lambda^e}$, since $$\big(\prod_{e\in E}\Phi_{\Lambda^e}\big)\Big(\sum_{u\in S}a_u\delta_u\Big)= \left(\Phi_{\Lambda^e}\Big(\sum_{u\in S}a_u\delta_u\Big) \right)_{e\in E} = \ \left(\sum_{u\in S}\lambda_{e,u}(a_u)\right)_{e\in E}.$$
{#section-1}
Our aim is to show that $\Phi_\Lambda$ is injective on ${\mathcal{C}_c({\mathcal{A}})}$ (cf. Theorem \[theo:wording\]). The following lemma will be crucial.
\[lem:wordinghamsub\] Assume that $g=\sum_{u\in S}a_u\delta_u\in{\mathcal{C}_c({\mathcal{A}})}$ satisfies $\Phi_\Lambda(g)=0$ and let $e,f\in E$ with $e\leq f$.
Then, for each $t\in S$ and each $b\in A_e$, we have $$\sum_{u\in S, \,f\leq u^*u, \,ue=t } a_u\cdot b = 0\,.$$
Note that here (and elsewhere), we use the convention that a sum over an empty index set is equal to $0$.
For each $s\in S$ and $a\in A_s$, we will let $a\odot\varepsilon_{s}$ denote the element of $X_{s^*s}$ given for each $t\in S_{s^*s}$ by $$(a\odot\varepsilon_{s})(t) = \begin{cases} a & \ \text{if} \ t=s,\\
0 & \ \text{if} \ t \neq s\,.
\end{cases}$$ For every $v\in E$, we set $g_v=\sum_{u\in S_v} a_u\delta_u$. Then $g_v\in{\mathcal{C}_c({\mathcal{A}})}$, $g_v=0$ for all but finitely many $v$ in $E$, and $$g=\sum_{v\in E}g_v\,.$$ Moreover, $$\label{eq:wordingham1}
0=\Phi_\Lambda(g)=\sum_{v\in E}\Phi_\Lambda(g_v)=\left(\sum_{v\in E}\sum_{u\in S_v}\lambda_{p,u}(a_u)\right)_{p\in E}$$
Now, consider $v\in E$, $u\in S_v$, $a\in A_u$ and $a'\in A_f$.
Note first that $a'\odot\varepsilon_f \in X_{f^*f}=X_f$. Moreover, if $f\leq v$, then $$(uf)^*uf = fu^*uf= fvf = f\,,$$ so $uf \in S_f$ and $(a\cdot a')\odot\varepsilon_{uf} \in X_{(uf)^*uf} = X_f$. We claim that $$\label{regfund}
\lambda_{f,u}(a)(a'\odot\varepsilon_f)=\begin{cases}
(a\cdot a')\odot\varepsilon_{uf}&\mbox{if }f\leq v\,,\\
\hspace{6ex}0&\mbox{otherwise.}
\end{cases}$$ To prove this claim, let $t\in S_{f}$. Then we have $t \in D(u^*) $, that is, $tt^*\leq uu^*$, if and only if $f=t^*t\leq u^*u=v$. As $$\big(\lambda_{f,u}(a)(a'\odot\varepsilon_f)\big)(t)=\begin{cases}
a\cdot (a'\odot\varepsilon_{f})(u^*t)&\mbox{if } t\in D(u^*)\\
\hspace{6ex} 0&\mbox{otherwise,}
\end{cases}$$ we see that $\lambda_{f,u}(a)(a'\odot\varepsilon_f)=0 $ when $f\not\leq v$.
If $f\leq v$, thus $t\in S_f \cap D(u^*)$, then we have $u^*t \in S_f \cap D(u)$, with $u^*t = f$ if and only $uf =t$ (cf. \[WagnerP\]), so we get $$\begin{aligned}
\big(\lambda_{f,u}(a)(a'\odot\varepsilon_f)\big)(t) &= a\cdot (a'\odot\varepsilon_{f})(u^*t)\\
&=\begin{cases}
a\cdot a' &\mbox{if } t= uf \\
\hspace{2ex} 0&\mbox{otherwise}
\end{cases}
\,\, =
\big((a\cdot a')\odot\varepsilon_{uf}\big)(t)\,.\end{aligned}$$ We have thus shown that $\lambda_{f,u}(a)(a'\odot\varepsilon_f)= (a\cdot a')\odot\varepsilon_{uf}$ whenever $f\leq v$, and this finishes the proof of (\[regfund\]).
Let now $b \in A_e$. By the Cohen-Hewitt factorization theorem [@hewitt_ross70 Theorem 32.22] we can write $b$ as a product $b=cd$ where $c,d\in A_e$. As $e\leq f$, we get from (\[a-e-f\]) that $$\label{c-d}
j_{f,e}(c)\cdot d= cd =b\,.$$
For each $v\in E$ we get from (\[regfund\]) that $$\sum_{u\in S_v}\lambda_{f,u}(a_u)(j_{f,e}(c)\odot\varepsilon_f)=\begin{cases}
\sum_{u\in S_v}(a_u\cdot j_{f,e}(c))\odot\varepsilon_{uf}&\mbox{if }f\leq v\,,\\
\hspace{12ex}0&\mbox{otherwise.}
\end{cases}$$ Using it then follows that $$0=\sum_{v\in E}\sum_{u\in S_v}\lambda_{f,u}(a_u)(j_{f,e}(c)\odot\varepsilon_f)=\sum_{\{v\in E:f\leq v\}}\sum_{u\in S_v}(a_u\cdot j_{f,e}(c))\odot\varepsilon_{uf}\,.$$ By looking at individual coefficients we can then conclude that in ${\mathcal{C}_c({\mathcal{A}})}$, $$\label{eq:wordingham2}
0=\sum_{\{v\in E:f\leq v\}}\sum_{u\in S_v}(a_u\cdot j_{f,e}(c))\,\delta_{uf}$$ Since $e\leq f$ we get from and (\[c-d\]) that $$0=\left(\sum_{\{v\in E:f\leq v\}}\sum_{u\in S_v}(a_u\cdot j_{f,e}(c))\,\delta_{uf}\right)\big(d\delta_e\big)=\sum_{\{v\in E:f\leq v\}}\sum_{u\in S_v}(a_u\cdot b)\,\delta_{ue}\,.$$ We see that given $t\in S$, the $t$-coefficient of the sum on the right hand side of the above equation is $$\sum_{u\in S, \,f\leq u^*u, ue=t}\, a_u\cdot b\,,$$ which must then be equal to $0$.
We will need another lemma. Let $F$ be a semilattice and $A$ be a Banach space. As usual, we will denote the dual space of $A$, consisting of all continuous linear functionals on $A$, by $A^*$. We let ${\mathcal{C}_c(F,A)}$ denote the vector space of all finitely supported functions from $F$ to $A$. We will describe an element of ${\mathcal{C}_c(F,A)}$ as a formal sum $\sum_{f\in F}a_f\delta_f$ where each $a_f\in A$ and $a_f=0$ for all but finitely many $f$ in $F$. Given $\psi\in A^*$ and $e\in F$, we define $\theta_{\psi,e}:{\mathcal{C}_c(F,A)}\to\mathbb{C}$ to be the linear functional given by $$\theta_{\psi,e}\Big(\sum_{f\in F}a_f\delta_f\Big)= \sum_{f \in F, \,f\geq e} \psi(a_f)$$
\[lem:separation\] Let $A$ be a Banach space and $F$ be a semilattice. Then the set $\{\theta_{\psi,e}:e\in F,\psi\in A^*\}$ separates the elements of ${\mathcal{C}_c(F,A)}$.
Suppose $\sum_{f\in F}a_f\delta_f\neq 0$. Since $a_f=0$ for all but finitely many $f$ in $F$, we can choose $e\in F$ such that $a_e\neq 0$ and $a_f=0$ for all $f\in F\setminus \{e\}$ satisfying $f\geq e$. We may then pick $\psi\in A^*$ such that $\psi(a_e)\neq 0$, and this gives $$\theta_{\psi,e}\left(\sum_{f\in F} a_f\delta_f\right)=\sum_{f\in F,\,f\geq e}\psi(a_f)=\psi(a_e)\neq 0.$$
The following theorem is a generalization of Wordingham’s theorem [@paterson Theorem 2.1.1], and our proof follows the pattern of Wordingham’s original proof.
\[theo:wording\] Let ${\mathcal{A}}=\{A_s\}_{s\in S}$ be a Fell bundle over an inverse semigroup $S$. Then the left regular representation of ${\mathcal{C}_c({\mathcal{A}})}$ is injective.
Let $g\in{\mathcal{C}_c({\mathcal{A}})}$ and express $g$ as a sum $g=\sum_{u\in S}a_u\delta_u$, where $\mathrm{supp}(g)=\{u\in S : a_u\neq 0\}$ is finite. Assume $\Phi_\Lambda(g)=0$. We want to show that $a_t=0$ for each $t\in S$. Since $g=\sum_{e\in E}\sum_{u\in S_e}a_u\delta_u$ it is sufficient to show that for any $e\in E$, $a_t=0$ when $t\in S_e$.
Fix $e\in E$ and consider $t\in S$. Let $F$ be the subsemilattice of $E$ given by $$F=\{v\in E:e\leq v\}\,.$$ Also, let $f\in F$ and $b\in A_e$. For each $v\in F$ set $$\begin{aligned}
\beta^t_v&=\sum_{u\in S, \, v=u^*u, \,ue=t} a_u\cdot b \,\, \, \in A_t\,,\\
\beta^t&=\sum_{v\in F}\,\beta^t_v\, \delta_v \,\, \in {\mathcal{C}_c(F,A_t)}\,.\end{aligned}$$ Note that $\beta^t$ has finite support since $\beta^t_v=0$ if $v\notin\{u^*u:u\in\mathrm{supp}(g)\}$, and $\mathrm{supp}(g)$ is finite. Now, for each $\psi\in (A_t)^*$, we get from Lemma \[lem:wordinghamsub\] that $$\begin{aligned}
\theta_{\psi,f}\big(\beta^t\big)&=\sum_{v\in F,\,f\leq v}\,\psi\big(\beta^t_v\big)\\
&=\psi\left(\sum_{u\in S,\, f\leq u^*u,\,ue=t} a_u\cdot b\right)=0\,.\end{aligned}$$ Then $\beta^t=0$ by Lemma \[lem:separation\], so $\beta^t_v=0$ for each $v\in F$. In particular, since $e\in F$, we get $$\label{beta-te}
\sum_{u\in S_e, \, ue=t} a_u\cdot b = \beta^t_e = 0\,.$$ Assume now that $t\in S_e$. If $u\in S_e$ satisfies that $ ue=t$, then we have $ut^*t=t$ and $u^*u=t^*t$, which together imply that $u=t$. So (\[beta-te\]) gives that $a_t\cdot b=0$. Choosing $b=a_t^*\cdot a_t\in A_{t^*t}=A_e$, we get $a_t\cdot a_t^*\cdot a_t=0$, so $(a_t^*\cdot a_t)^2 = a_t^*\cdot a_t\cdot a_t^*\cdot a_t=0$, hence $a_t^*\cdot a_t=0$, and axiom (x) in the definition of a Fell bundle gives that $a_t=0$, as desired.
{#section-2}
We define the *reduced KS-cross sectional $C^*$-algebra ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ of ${\mathcal{A}}$* as the completion of ${\mathcal{C}_c({\mathcal{A}})}$ with respect to the norm $\|\cdot\|_r$ given by $$\|g\|_{\rm r} :=\|\Phi_\Lambda(g)\|$$ for $g \in {\mathcal{C}_c({\mathcal{A}})}$. Alternatively, we may consider ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ to be given as the norm-closure of $\Phi_\Lambda({\mathcal{C}_c({\mathcal{A}})})$ in $B$, or, equivalently, as $\Phi_\Lambda({C^*_{\mathrm{KS}}({\mathcal{A}})})$.
Recall that $\mathcal{N}_{\mathcal{A}}$ denotes the two-sided selfadjoint ideal of ${\mathcal{C}_c({\mathcal{A}})}$ spanned by the set $$\{a_s\delta_s-j_{t,s}(a_s)\delta_t:s,t\in S, s\leq t, a_s\in A_s\}\,.$$ We define $\mathcal{I}_{\mathcal{A}}$ to be the closure of $\mathcal{N}_{\mathcal{A}}$ inside ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$. In other words, we set $$\mathcal{I}_{\mathcal{A}}= \overline{\Phi_\Lambda(\mathcal{N}_{\mathcal{A}})}\,.$$ It is easy to check that $\mathcal{I}_{\mathcal{A}}$ is an ideal of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$. Hence we may form the quotient $C^*$-algebra $${C^*_{r,\mathrm{alt}}({\mathcal{A}})}:={C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{I}_{\mathcal{A}}\,,$$ which provides an alternative version of the reduced cross sectional $C^*$-algebra of ${\mathcal{A}}$. We will let $q_{\mathcal{A}}^r:{C^*_{r,\mathrm{KS}}({\mathcal{A}})}\to {C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ denote the quotient map. It is not clear whether ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ is isomorphic to Exel’s reduced $C^*$-algebra $C^*_r({\mathcal{A}})$ (cf. \[sec:exelreduced\]). We will show in Section \[sec:comparison\] that this is true under certain assumptions.
For $u\in S$ let $\lambda^{\rm alt}_u:A_u \to {C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ be defined by $$\lambda^{\rm alt}_u(a_u) = \lambda_u(a_u) + \mathcal{I}_{\mathcal{A}}$$ for all $a_u \in A_u$. It is then almost immediate that $\Lambda^{\rm alt}:=\{\lambda^{\rm alt}_u\}_{u\in S}$ is a representation of ${\mathcal{A}}$ in ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$. Using the universal property of $C^*({\mathcal{A}})$ we get a surjective $*$-homomorphism $\Psi_{\Lambda^{\rm alt}}$ from $C^*({\mathcal{A}})$ onto ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ satisfying $$\Psi_{\Lambda^{\rm alt}}\big(\pi_u^{\mathcal{A}}(a_u)\big) = \lambda^{\rm alt}_u(a_u) = \lambda_u(a_u) + \mathcal{I}_{\mathcal{A}}$$ for all $u\in S$ and $a_u \in A_u$. Similarly, we get a surjective $*$-homomorphism $\Phi_{\Lambda^{\rm alt}}$ from ${C^*_{\mathrm{KS}}({\mathcal{A}})}$ onto ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$, which satisfies $$\Phi_{\Lambda^{\rm alt}} = q^r_{\mathcal{A}}\circ \Phi_{\Lambda} = \Psi_{\Lambda^{\rm alt}}\circ q_{\mathcal{A}}\,.$$
The following commutative diagram sums up the relationship between the various algebras and some of the $*$-homomorphisms defined so far.
(CcA) at (0,2) [${\mathcal{C}_c({\mathcal{A}})}$]{}; (CKSA) at (1,2) [${C^*_{\mathrm{KS}}({\mathcal{A}})}$]{}; (CrKSA) at (2,2) [${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$]{}; (CcANA) at (0,1) [${\mathcal{C}_c({\mathcal{A}})}/{\mathcal{N}}_{\mathcal{A}}$]{}; (CA) at (1,1) [$C^*({\mathcal{A}})$]{}; (CraltA) at (2,1) [${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$]{}; (CrA) at (1,0) [$C^*_r({\mathcal{A}})$]{}; (CcA) edge (CKSA); (CcA) edge node\[auto\] [$Q_{\mathcal{A}}$]{} (CcANA) (CKSA) edge node\[auto\] [$q_{\mathcal{A}}$]{} (CA) (CrKSA) edge node\[auto\] [$q_{\mathcal{A}}^{\rm r}$]{} (CraltA) (CKSA) edge node\[above\] [$\Phi_\Lambda$]{} (CrKSA) (CKSA) edge node\[sloped, above, yshift=1pt\] [$\Phi_{\Lambda^{\rm alt}}$]{} (CraltA) (CA) edge node\[above, yshift=1pt\] [$\Psi_{\Lambda^{\rm alt}}$]{} (CraltA) (CA) edge node\[auto\] [$\Psi_{\Pi^{\rm red}}$]{} (CrA); (CcA) edge node\[sloped, above\] [$\iota_{\mathcal{A}}$]{} (CA) (CcANA) edge node\[above\] [$R_{\mathcal{A}}$]{} (CA) (CcA) edge\[bend right=80\] node\[above, sloped, near end\] [$\iota_{\mathcal{A}}^{\rm red}$]{} (CrA);
Fell bundles over semilattices {#semilatt}
==============================
In this section we look at the case where $S=E$ is a semilattice, and consider a Fell bundle ${\mathcal{E}}=\{A_e\}_{e\in E}$. Since $E_e=\{f\in E:f^*f=e\}=\{e\}$ for each $e\in E$, the Hilbert $A_e$-module $X_e$ that occurs in the definition of the pre-representation $\Lambda^e=\big(\lambda_{e,f}\big)_{f\in E}$ of ${\mathcal{E}}$ in ${\mathcal{L}(X_e)}$ is nothing but $A_e$ itself (with its standard structure). Thus ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$ can be viewed as a $C^*$-subalgebra of $\prod_{e\in E}{\mathcal{L}(A_e)}$, and for $e, f\in E$ and $a_f\in A_f$, $\lambda_{e,f}(a_f):A_e\to A_e$ is given by $$\label{eq:lamb-e-f}
\lambda_{e,f}(a_f)b = \begin{cases}
a_f\cdot b & \text{ if } e\leq f,\\
0 & \text{ otherwise, }
\end{cases} \quad \quad \text{for all}\ b\in A_e.$$ As before, let $\Phi_{\Lambda^e}:{\mathcal{C}_c({\mathcal{E}})}\to{\mathcal{L}(A_e)}$ be the corresponding $*$-homomorphism given by $$\Phi_{\Lambda^e}\Big(\sum_{f\in E}a_f\delta_f\Big)=\sum_{f\in E}\lambda_{e,f}(a_f).$$ Let $(a^i)$ be an approximate unit for $A_e$, let $f\in E$ be such that $f\geq e$, and let $a_f\in A_f$. Then for all $b\in A_e$ we have $$\lim_i \,(a^i \cdot a_f)\cdot b = \lim_i \,a^i \cdot (a_f \cdot b) = a_f\cdot b,$$ and $\lim_i \, a_f \cdot (a^i\cdot b)=a_f\cdot b$. Thus, $$\label{eq:reducedrepformula}
\lambda_{e,f}(a_f)=\lim_i a^i\cdot a_f=\lim_i a_f\cdot a^i$$ where the limits in equation are taken in the strict topology of ${\mathcal{L}(A_e)}$.
We recall that a *character* (sometimes called a semicharacter) on $E$ is a non-zero homomorphism from $E$ into the semilattice $ \{0,1\}$, see e.g. [@paterson].
\[semichar\] Let $\pi:{C^*_{\mathrm{KS}}({\mathcal{E}})}\to\mathcal{B}(H)$ be a non-zero irreducible representation of ${C^*_{\mathrm{KS}}({\mathcal{E}})}$ on a Hilbert space $H$. For $e \in E$, let $p_e$ denote the orthogonal projection of $H$ onto the norm-closure of $\pi(A_e\delta_e)H$ in $H$. Then $p_e\in\{0,I_H\}$. Moreover, the map $\widehat\pi:E\to \{0,1\}$ defined by $$\widehat\pi(e) = \begin{cases} 1& \quad \text{if} \ p_e = I_H, \\ 0& \quad \text{if} \ p_e = 0,
\end{cases}$$ is a character on $E$.
Let $(a^i)$ be an approximate unit for $A_e$. It is straightforward to check that $\pi(a^i\delta_e)$ converges strongly to $p_e$.
Let $f\in E$ be such that $f\leq e$ and let $a\in A_f$. Since $j_{e,f}(a^i\cdot a)=a^i\cdot j_{e,f}(a)$, and $a^i\cdot j_{e,f}(a)$ converges to $j_{e,f}(a)$ in norm, it follows, using that $j_{e,f}$ is isometric, that $a^i\cdot a$ converges to $a$ in norm. Hence, for any $\xi\in H$, we have $$p_e\pi(a\delta_f)\xi = \lim_i \pi(a^i\delta_e)\pi(a\delta_f)\xi =\lim_i \pi(a^i\cdot a\delta_f)\xi=\pi(a\delta_f)\xi.$$ It follows that $p_fH\subset p_eH$, that is, $p_f \leq p_e$.
Consider now $e'\in E$. Then $e'e \leq e$, so $p_{e'e}H\subset p_eH$.
Hence for $a\in A_{e'}$ we get $$\begin{aligned}
\pi(a\delta_{e'})p_eH&=\pi(a\delta_{e'})\overline{\pi(A_e\delta_e)H}\\
&\subset\overline{\pi(A_{e'}\cdot A_e\delta_{e'e})H}\\&\subset \overline{\pi(A_{e'e}\delta_{e'e})H}\\&=p_{e'e}H\subset p_eH.
\end{aligned}$$ This implies that $p_eH$ is a closed invariant subspace for $\pi({C^*_{\mathrm{KS}}({\mathcal{E}})})$, hence that $p_e\in\{0,I_H\}$ since $\pi$ is irreducible. Moreover, for any $e,e'\in E$, $p_{e'}p_e=p_{e}p_{e'}$ is then a projection, and, as seen above, we have $p_{e'e}\leq p_e$, and similarly $p_{e'e}=p_{ee'} \leq p_{e'}$, so $p_{e'e}\leq p_{e'}p_{e}$. On the other hand, for $a\in A_{e'}$, we know that $\pi(a\delta_{e'}) p_{e}H\subset p_{e'e}H$. Hence, using an approximate unit for $A_{e'}$, one easily deduces that $p_{e'}p_{e}\leq p_{e'e}$. Thus we get $p_{e'e} = p_{e'}p_{e}$. Since $\pi$ is non-zero, it clearly follows that $\widehat{\pi}$ is a character on $E$.
Given a character $\psi$ on $E$, we set $F_\psi = \{e\in E:\psi(e)=1\}$. Then $F_\psi$ is an example of a filter in $E$ (and every filter on $E$ can be obtained this way), cf. [@paterson]. We recall that a filter in $E$ is a nonempty subsemilattice $F$ of $E$ such that if $f\in E$ and $f\geq e$ for some $e\in F$, then $f\in F$.
\[prop:universalKS\] We have ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}={C^*_{\mathrm{KS}}({\mathcal{E}})}$.
We only have to check that every irreducible representation of ${C^*_{\mathrm{KS}}({\mathcal{E}})}$ is dominated in norm by the left regular representation $\Phi_\Lambda$. More precisely, it suffices to show that given a Hilbert space $H$ and a non-zero irreducible representation $\pi:{C^*_{\mathrm{KS}}({\mathcal{E}})}\to \mathcal{B}(H)$, we have $$\|\pi(g)\|\leq \|\Phi_\Lambda(g)\|$$ for all $g\in {\mathcal{C}_c({\mathcal{E}})}$. Let $\widehat{\pi}$ be the character on $E$ described in Lemma \[semichar\], and let $F=F_{\widehat{\pi}}$ be the corresponding filter in $E$. Note that if $p_e$ is defined as in Lemma \[semichar\], then $p_e=I_H$ when $e\in F$, while $p_e = 0$ when $e\in E\setminus F$. Hence for all $e \in E\setminus F$ and all $a_e\in A_e$ we have $\pi(a_e\delta_e)=0$. Indeed, letting $(a^i)$ be an approximate unit for $A_e$, we then have $$\pi(a_e\delta_e)\xi = \lim_i \pi(a_ea^i \delta_e)\xi = \lim_i \pi(a_e\delta_e)\pi(a^i\delta_e)\xi = \pi(a_e\delta_e)p_e\xi = 0$$ for all $\xi \in H$. Consider now $g =\sum_{u\in E} a_u\delta_u \in{\mathcal{C}_c({\mathcal{E}})}$. Using the observation we just made, we get $$\pi(g) = \sum_{f\in K\cap F}\pi(a_f\delta_f),$$ where $K:=\{u\in E:a_u\neq 0\}$ is finite. Since $F$ is a semilattice and $K\cap F$ is a finite subset of $F$, there exists some $e\in F$ such that $e\leq f$ for all $f\in K\cap F$. Let $(a^i)$ be an approximate unit for $A_e$. Since the restriction $\pi_e$ of $\pi$ to $A_e\delta_e\simeq A_e$ is a non-degenerate representation of $A_e\delta_e$ on $H$, it may be extended to a representation $\overline{\pi_e}:{\mathcal{L}(A_e)}\to\mathcal{B}(H)$ (see for instance [@blackadar Theorem II.7.3.9]). Moreover, if $(b^i\delta_e)$ is a net in $A_e\delta_e$ converging strictly to some $x\in{\mathcal{L}(A_e)}$, then $\pi_e(b^i\delta_e)$ converges strongly to $\overline{\pi_e}(x)$ in $\mathcal{B}(H)$. Thus for any $\xi\in H$, using equations and , we get $$\begin{aligned}
\pi(g)\xi&=p_e\pi(g)\xi= \lim_i\pi(a^i\delta_e)\,\sum_{f\in K\cap F}\pi(a_f\delta_f)\xi\\
&=\lim_i \sum_{f\in K \cap F} \pi\Big((a^i\cdot a_f)\delta_e\Big)\xi\\
&=\sum_{f\in K\cap F} \overline{\pi_e}(\lambda_{e,f}(a_f))\xi\\
&=\overline{\pi_e}(\Phi_{\Lambda^e}(g))\xi\,.
\end{aligned}$$ It follows that $\pi(g)=\overline{\pi_e}(\Phi_{\Lambda^e}(g))$, so $\|\pi(g)\|\leq\|\Phi_{\Lambda^e}(g)\|\leq \|\Phi_\Lambda(g)\|$.
Proposition \[prop:universalKS\] implies that ${\mathcal{M}}_{\mathcal{E}}=\mathcal{I}_{\mathcal{E}}$, hence that $C^*({\mathcal{E}})={C^*_{r,\mathrm{alt}}({\mathcal{E}})}$. Since it follows from [@exel11 Corollary 8.10] that $C^*_r({\mathcal{E}})\simeq C^*({\mathcal{E}})$, we get:
$ {C^*_{r,\mathrm{alt}}({\mathcal{E}})}=C^*({\mathcal{E}})\simeq C^*_r({\mathcal{E}}).$
Conditional Expectations onto the Diagonal {#cond-exp-diag}
==========================================
Let ${\mathcal{A}}=\{A_s\}_{s\in S}$ be a Fell bundle over $S$ and let ${\mathcal{E}}=\{A_e\}_{e\in E}$ denote the Fell bundle obtained by restricting ${\mathcal{A}}$ to the semilattice $E$ of idempotents in $S$. Recall that ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ can be viewed as a $C^*$-subalgebra of $\prod_{e\in E}{\mathcal{L}(X_e)}$ and that ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$ can be viewed as a $C^*$-subalgebra of $\prod_{e\in E}{\mathcal{L}(A_e)}$. When it is necessary to distinguish them, we will denote by $\Phi^{\mathcal{A}}_{\Lambda}$ the left regular representation of ${\mathcal{C}_c({\mathcal{A}})}$ and by $\Phi^{\mathcal{E}}_\Lambda$ the left regular representation of ${\mathcal{C}_c({\mathcal{E}})}$. Similarly, we will write $\{\lambda^{\mathcal{A}}_{e,s}\}$ and $\{\lambda^{\mathcal{E}}_{e,f}\}$ for the respective pre-representations.
Let $e\in E$ and $s\in S_e$ (so that $s^*s=e$). Recalling that $A_s$ is a (right) $A_e$-module, we define an $A_e$-module map $\gamma_s:A_s\to X_e$ by $\gamma_s(a)=a\odot\varepsilon_s$, i.e. $$\gamma_s(a)(t)=\begin{cases}
a & \mbox{ if }s=t\\
0 & \mbox{ otherwise.}
\end{cases}$$ One readily checks that $\gamma_s$ is adjointable with adjoint given by $\gamma_s^*(\xi)=\xi(s)$ for every $\xi \in X_e$. Then $\gamma_s^*\gamma_s$ is clearly the identity map on $A_s$, so $\gamma_s$ is isometric. Moreover, we have $\gamma_s\gamma_s^*\xi = \xi(s)\odot \varepsilon_s$ for every $\xi \in X_e$, and it follows that $\sum_{s\in S_e}\gamma_s\gamma_s^*\xi=\xi$ for every $\xi\in X_e$, where the sum converges in the norm topology on $X_e$.
\[lem:conditional0\] Let $e\in E$, $s,t\in S_e$, $u\in S$ and $a\in A_u$. Then the map $\gamma_s^*\lambda_{e,u}(a)\gamma_t: A_t\to A_s$ is given by $$\big(\gamma_s^*\lambda_{e,u}(a)\gamma_t\big)(b)=\begin{cases}
a\cdot b & \text{ if }u \geq st^*,\\
0 & \text{ otherwise}
\end{cases}$$ for all $b\in A_t$.
For $b\in A_t$ we have $$\begin{aligned}
\gamma_s^*\lambda_{e,u}(a)\gamma_t(b)&=\big(\lambda_{e,u}(a)\gamma_t(b)\big)(s)\\
&=\begin{cases}
a\cdot\gamma_t(b)(u^*s) & \text{ if }s \in D(u^*),\\
\hspace{6ex} 0 & \text{ otherwise}
\end{cases}\\
&=\begin{cases}
a\cdot b & \text{ if }u^*s=t\text{ and }ss^*\leq uu^*,\\
\hspace{2ex} 0 & \text{ otherwise.}
\end{cases}
\end{aligned}$$ Suppose first that $u\geq st^*$. Then $uu^*\geq (st^*)(st^*)^*=ss^*$ since $t^*t=s^*s$. Moreover, $u^*\geq ts^*$, so $ts^*=u^*(ts^*)^*(ts^*)=u^*ss^*$, hence $u^*s=ts^*s=tt^*t=t$. Conversely, if $ss^*\leq uu^*$ and $u^*s=t$, then $$u^*(ts^*)^*(ts^*)=u^*st^*ts^* = u^* uu^*ss^*ss^* =u^*ss^*=ts^*,$$ so $u^*\geq ts^*$, hence $u\geq st^*$.
\[lem:factorize\] Let $t\in S$ and $b\in A_t$. Then there exist $c\in A_{tt^*}$ and $d\in A_t$ such that $b=c\cdot d$.
Let $(u^i)$ be an approximate unit for $A_{tt^*}$. Since $||u^i||\leq 1$ for all $i$, we get $$\begin{aligned}
||u^i\cdot b-b||^2&=||(u_i\cdot b-b)(u_i\cdot b-b)^*||\\
&=||u_i\cdot b\cdot b^*-b\cdot b^*+(u_i\cdot b\cdot b^*-b\cdot b^*)\cdot u_i||\\
&\leq 2||u_i\cdot b\cdot b^*-b\cdot b^*||.
\end{aligned}$$ So $u^i\cdot b$ converges to $b$. Regarding $A_t$ as a left $A_{tt^*}\,$-module in the obvious way, we may then apply the Cohen-Hewitt factorization theorem [@hewitt_ross70 Theorem 32.22] to $b$ and deduce that $b=c\cdot d$ for some $c\in A_{tt^*}$ and $d\in A_t$.
\[lem:fourierfactorize\] Let $t\in S$ and $b\in A_t$. Let $b=c\cdot d$ be any factorization of $b$ with $c\in A_{tt^*}$ and $d\in A_t$. Let $(T_e)_{e\in E}\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}$, and let $s\in S$ be such that $s^*s=t^*t$. Then $$(\gamma_s^*T_{t^*t}\gamma_t)(b)= (\gamma_{st^*}^*T_{tt^*}\gamma_{tt^*})(c)\cdot d.$$
Note first that the expression on the right-hand side is well-defined since $(st^*)^*(st^*)= t s^*st^*= tt^*$ as $s^*s =t^*t$. By linearity and continuity, it suffices to prove that for any $u\in S$ and $a\in A_u$, we have $$(\gamma_s^*\lambda_{t^*t,u}(a)\gamma_t)(b)= (\gamma_{st^*}^*\lambda_{tt^*,u}(a)\gamma_{tt^*})(c)\cdot d.$$ This follows immediately by applying Lemma \[lem:conditional0\] to both sides, and using that $st^*(tt^*)^*=st^*$.
\[lem:conditional1\] For any $e,f\in E$, and $a_f \in A_f$ we have $$\gamma_e^*\lambda^{\mathcal{A}}_{e,f}(a_f)\gamma_e = \lambda^{\mathcal{E}}_{e,f}(a_f).$$ Moreover, if $S$ is $E^*$-unitary and $A_0=\{0\}$ (if $S$ has a $0$-element), then for any $e\in E$, $u\in S$ and $a_u\in A_u$, we have $$\label{gamma-e-u}
\gamma_e^*\lambda^{\mathcal{A}}_{e,u}(a_u)\gamma_e = \begin{cases}
\lambda^{\mathcal{E}}_{e,u}(a_u)\quad&\mbox{ if } u\in E,\\
\hspace{3ex}0 & \mbox{ otherwise.}
\end{cases}$$
We prove the second statement. The proof of the first statement follows from a small adjustment to the argument and is left to the reader. Assume that $S$ is $E^*$-unitary and $A_0=\{0\}$ (if $S$ has a $0$-element). Let $b\in A_e$. Lemma \[lem:conditional0\] gives that $$\label{gamma-e}
\big(\gamma_e^*\lambda^{\mathcal{A}}_{e,u}(a_u)\gamma_e\big)(b)=
\begin{cases}
a_u\cdot b & \text{if } e\leq u,\\
\hspace{2ex}0 & \text{otherwise. }
\end{cases}$$ Since $S$ is $E^*$-unitary, $e\leq u$ implies that $u$ is idempotent or $e=0$. If $e\neq 0$, the right hand side of (\[gamma-e\]) is equal to $$\begin{cases}
a_u \cdot b & \text{if }e\leq u\text{ and }u\in E,\\
\hspace{2ex}0 & \text{ otherwise, }
\end{cases}\\
\ = \
\begin{cases}
\lambda^{\mathcal{E}}_{e,u}(a_u)b & \text{ if } u\in E,\\
\hspace{3ex} 0 & \text{ otherwise,}
\end{cases}$$ so we see that (\[gamma-e-u\]) holds in this case. If $e=0$ (so $S$ has a $0$-element), then both sides of (\[gamma-e-u\]) are equal to $0$ since $A_0=\{0\}$ by assumption.
\[lem:diagembedding\] There is an embedding of ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$ into ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ extending the inclusion ${\mathcal{C}_c({\mathcal{E}})}\subset{\mathcal{C}_c({\mathcal{A}})}$.
Let $\sum_{f\in E}a_f\delta_f\in{\mathcal{C}_c({\mathcal{E}})}$. We need to prove that $$\label{Phi-E}
\big\|\Phi^{\mathcal{E}}_\Lambda\Big(\sum_{f\in E}a_f\delta_f\Big)\big\|\, =\, \big\|\Phi^{\mathcal{A}}_\Lambda\Big(\sum_{f\in E}a_f\delta_f\Big)\big\|.$$ Since ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}={C^*_{\mathrm{KS}}({\mathcal{E}})}$, cf. Proposition \[prop:universalKS\], the expression on the left-hand side of (\[Phi-E\]) is the same as the universal norm of $\sum_{f\in E}a_f\delta_f$ in ${C^*_{\mathrm{KS}}({\mathcal{E}})}$. As $\Phi^{\mathcal{A}}_\Lambda$ restricts to a $*$-homomorphism of $C_c({\mathcal{E}})$, we see that the $\geq$ inequality in (\[Phi-E\]) must hold. On the other hand, since $\gamma_e$ is an isometry for each $e\in E$, Lemma \[lem:conditional1\] implies that $$\sup_{e\in E}\big\|\sum_{f\in E}\lambda^{\mathcal{E}}_{e,f}(a_f)\big\|\,\leq\,\sup_{e\in E}\big\| \sum_{f\in E}\lambda^{\mathcal{A}}_{e,f}(a_f)\big\|.$$ This shows that $\leq$ inequality in (\[Phi-E\]) holds.
We will identify ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$ with its canonical image in ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$, and call it the *diagonal* ($C^*$-subalgebra) of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$. Define ${\mathfrak{E}}:{\mathcal{C}_c({\mathcal{A}})}\to{\mathcal{C}_c({\mathcal{E}})}$ by $${\mathfrak{E}}\Big(\sum_{u\in S}a_u\delta_u\Big)=\sum_{e\in E}a_e\delta_e$$ for all $\sum_{u\in S}a_u\delta_u\in {\mathcal{C}_c({\mathcal{A}})}$. Moreover, define a positive linear map ${\mathfrak{E}_{\rm KS}}$ from ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ into $\prod_{e\in E}{\mathcal{L}(A_e)}$ by $${\mathfrak{E}_{\rm KS}}\big((T_e)_{e\in E}\big)=(\gamma_e^*T_e\gamma_e)_{e\in E}$$ for all $(T_e)_{e\in E} \in {C^*_{r,\mathrm{KS}}({\mathcal{A}})}$.
\[condi-exp\] The map ${\mathfrak{E}_{\rm KS}}:{C^*_{r,\mathrm{KS}}({\mathcal{A}})}\to\prod_{e\in E}{\mathcal{L}(A_e)}$ is faithful.
If $S$ is $E^*$-unitary and $A_0=\{0\}$ $($if $S$ has a $0$-element$)$, then ${\mathfrak{E}_{\rm KS}}$ satisfies $$\label{eq:conditionalformula}
{\mathfrak{E}_{\rm KS}}\Big(\Phi^{{\mathcal{A}}}_\Lambda(g)\Big)=\Phi^{\mathcal{E}}_\Lambda\Big({\mathfrak{E}}(g)\Big)$$ for all $g\in {\mathcal{C}_c({\mathcal{A}})}$. Moreover, in this case, ${\mathfrak{E}_{\rm KS}}$ is a faithful conditional expectation from ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ onto ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$.
Let $e\in E$, $T_e\in{\mathcal{L}(X_e)}$ and $a\in A_e$. For each $s\in S_e$, we have $$(\gamma_s^*T_e\gamma_e)(a) = \big(T_e\gamma_e(a)\big)(s),$$ so we get $$\begin{aligned}
\Big\langle (\gamma_e^*T_e^*T_e\gamma_e)(a),a\Big\rangle_{A_e} &= \Big\langle T_e\gamma_e(a),T_e\gamma_e(a)\Big\rangle_{X_e}\\
&=\sum_{s\in S_e} \big(T_e\gamma_e(a)\big)(s)^*\big(T_e\gamma_e(a)\big)(s)\\
&=\sum_{s\in S_e} \big(\gamma_s^*T_e\gamma_e\big)(a)^*\big(\gamma_s^*T_e\gamma_e\big)(a).\end{aligned}$$ So we see that if $\gamma_e^*T_e^*T_e\gamma_e=0$, then $\gamma_s^*T_e\gamma_e=0$ for each $s\in S_e$.
Consider $T=(T_e)_{e\in E}\in {C^*_{r,\mathrm{KS}}({\mathcal{A}})}$. If $\gamma_s^*T_e\gamma_e=0$ for all $e\in E$ and $s\in S_e$, then for any $e\in E$ and $s,t\in S_e$, we have in particular that $\gamma_{st^*}^*T_{tt^*}\gamma_{tt^*}=0$, so Lemma \[lem:factorize\] and Lemma \[lem:fourierfactorize\] imply that $\gamma_s^*T_e\gamma_t=0$. Combining this with our first observation, we get that if $\gamma_e^*T_e^*T_e\gamma_e=0$ for each $e\in E$, then $\gamma_s^*T_e\gamma_t=0$ for each $e\in E$ and $s,t\in S_e$.
Assume now that ${\mathfrak{E}_{\rm KS}}(T^*T) = 0$. This means that $\gamma_e^*T_e^*T_e\gamma_e=0$ for all $e\in E$. Hence, for $e\in E$ and $\xi\in X_e$, we get $$T_e\xi = \sum_{s\in S_e} \gamma_s\gamma_s^*T_e\xi = \sum_{s\in S_e}\sum_{t\in S_e}\gamma_s\gamma_s^*T_e\gamma_t\gamma_t^*\xi=0.$$ Thus $T_e=0$ for every $e\in E$, so $T=0$. This proves that ${\mathfrak{E}_{\rm KS}}$ is faithful.
Next, assume that $S$ is $E^*$-unitary and $A_0=\{0\}$ (if $S$ has a $0$-element). To show that holds amounts to show that for any $e\in E$, we have $$\gamma_e^*\left(\sum_{u\in S}\lambda_{e,u}(a_u)\right)\gamma_e=\sum_{f\in E}\lambda_{e,f}(a_f).$$ for all $\sum_{u\in S}a_u\delta_u\in {\mathcal{C}_c({\mathcal{A}})}$. This follows readily from Lemma \[lem:conditional1\]. It is then clear that the image of ${\mathfrak{E}_{\rm KS}}$ is ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$. Note also that ${\mathfrak{E}_{\rm KS}}$ is contractive since $\gamma_e$ is an isometry for each $e\in E$. Moreover, it is immediate from that ${\mathfrak{E}_{\rm KS}}$ is a projection map. Hence, Tomiyama’s theorem (see for instance [@blackadar Theorem II.6.10.2]) gives that ${\mathfrak{E}_{\rm KS}}$ is a conditional expectation.
\[groupbundle\] Suppose that $S$ is strongly $E^*$-unitary and and $A_0=\{0\}$ (if $S$ has a zero). Let $\sigma$ be an idempotent pure grading from $S^\times$ into a group $G$. Then for each $g\in G$ one can form the Banach space $$B_g := \overline{\bigoplus_{s\in S,\,\sigma(s)=g}\Phi_\Lambda(A_s\delta_s)}\subset {C^*_{r,\mathrm{KS}}({\mathcal{A}})}.$$ It is straightforward to check that $\mathcal{B}:=\{B_g\}_{g\in G}$ is a Fell bundle over $G$, giving a $G$-grading for ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ in the sense of [@exel14 Definition 16.2]. Moreover, since $\sigma$ is idempotent pure, we have $\{s\in S: \sigma(s) =1_G\}=E$, so $B_{1_G}={C^*_{r,\mathrm{KS}}({\mathcal{E}})}$. Since ${\mathfrak{E}_{\rm KS}}$ is faithful by the previous proposition, it then follows from [@exel14 Proposition 19.8] that ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ is naturally isomorphic to the reduced cross sectional $C^*$-algebra $C^*_r(\mathcal{B})$ associated with $\mathcal{B}$.
The following covariance property of ${\mathfrak{E}_{\rm KS}}$ will be useful later.
\[covar\] Suppose $S$ is $E^*$-unitary and $A_0=\{0\}$ (if $S$ has a $0$ element). Then for all $s\in S$, $b\in A_s$ and $T\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ we have $$\label{eq:covar}
{\mathfrak{E}_{\rm KS}}\big(\lambda_s(b)^*T\lambda_s(b)\big)=\lambda_s(b)^*{\mathfrak{E}_{\rm KS}}(T)\lambda_s(b).$$
Let $s\in S$ and $b\in A_s$. Consider $\sum_{t\in S}a_t\delta_t\in {\mathcal{C}_c({\mathcal{A}})}$. Then for any $t\in S$ we have that $s^*ts=0$ if and only if $ss^*tss^*=0$. Moreover, $s^*ts\in E$ if and only if $ss^*tss^*\in E$; thus, since $S$ is $E^*$-unitary and $t\geq ss^*tss^*$, we get that $s^*ts\in E$ if and only if $t\in E$ or $s^*ts=0$. Hence, using that $A_0=\{0\}$ (if $S$ has a $0$ element), we get $$\begin{aligned}
{\mathfrak{E}}\left((b\delta_s)^*\Big(\sum_{t\in S}a_t\delta_t\Big)(b\delta_s)\right)
&={\mathfrak{E}}\left(\sum_
{t\in S,\, s^*ts\neq 0}(b^*\cdot a_t\cdot b)\delta_{s^*ts}\right)\\
&=\sum_
{t\in S, \, s^*ts\in E^\times}
(b^*\cdot a_t\cdot b)\delta_{s^*ts}\\
&=(b\delta_s)^*{\mathfrak{E}}\left(\sum_{t\in S}a_t\delta_t\right)(b\delta_s).\end{aligned}$$ Using equation (\[eq:conditionalformula\]), we then see that equation (\[eq:covar\]) holds whenever $T= \Phi_\Lambda^{\mathcal{A}}(g)$ for some $g \in {\mathcal{C}_c({\mathcal{A}})}$. By linearity and continuity of ${\mathfrak{E}_{\rm KS}}$ and density of $\Phi_\Lambda^{\mathcal{A}}({\mathcal{C}_c({\mathcal{A}})})$ in ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$, it then holds for all $T\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}$.
Comparison with Exel’s reduced cross sectional $C^*$-algebras {#sec:comparison}
=============================================================
[*Throughout this section we consider a Fell bundle ${\mathcal{A}}=\{A_s\}_{s\in S}$ over an $E^*$-unitary inverse semigroup $S$ and assume that $A_0=\{0\}$ $($if $S$ has a $0$ element$)$.*]{} Our aim is to show that Exel’s $C^*_r({\mathcal{A}})$ is a quotient of ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ and that these $C^*$-algebras are canonically isomorphic under certain assumptions.
As in the previous section, we let ${\mathcal{E}}=\{A_e\}_{e\in E}$ denote the Fell bundle obtained by restricting ${\mathcal{A}}$ to the semilattice $E=E(S)$. From Proposition \[condi-exp\], we see that ${\mathfrak{E}_{\rm KS}}\big(\Phi^{\mathcal{A}}_\Lambda({\mathcal{N}}_{\mathcal{A}})\big)=\Phi^{\mathcal{E}}_\Lambda({\mathcal{N}}_{\mathcal{E}})$, and it easily follows that ${\mathfrak{E}_{\rm KS}}(\mathcal{I}_{\mathcal{A}})=\mathcal{I}_{\mathcal{E}}$.
For any ideal $\mathcal{K}$ of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ satisfying ${\mathfrak{E}_{\rm KS}}(\mathcal{K})=\mathcal{I}_{\mathcal{E}}$ we can define a surjective linear map ${\mathfrak{E}}_\mathcal{K}:{C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{K}\to{C^*_{r,\mathrm{alt}}({\mathcal{E}})}$ by $${\mathfrak{E}}_\mathcal{K}(T+\mathcal{K})={\mathfrak{E}_{\rm KS}}(T)+\mathcal{I}_{\mathcal{E}}.$$ It is straightforward to check that ${\mathfrak{E}}_\mathcal{K}$ is contractive. Note also that for each $T\in\Phi^{\mathcal{A}}_\Lambda({\mathcal{C}_c({\mathcal{E}})})$ we have $$\|T+\mathcal{I}_{\mathcal{E}}\|=\|{\mathfrak{E}}_\mathcal{K}(T+\mathcal{K})\|\leq\|T+\mathcal{K}\|\leq\|T+\mathcal{I}_{\mathcal{E}}\|$$ where the last inequality uses that the map $$g\mapsto\Phi^{\mathcal{A}}_\Lambda(g) +\, \mathcal{K}$$ is a representation of ${\mathcal{C}_c({\mathcal{E}})}$ in ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{K}$ and that ${C^*_{r,\mathrm{alt}}({\mathcal{E}})}=C^*({\mathcal{E}})$. It follows that $\|T + \mathcal{I}_{\mathcal{E}}\| = \|T+\mathcal{K}\|$ for each $T \in {C^*_{r,\mathrm{KS}}({\mathcal{E}})}$, so we can identify ${C^*_{r,\mathrm{alt}}({\mathcal{E}})}$ with the image of ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$ in the quotient ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{K}$. Using Tomiyama’s theorem (see for instance [@blackadar Theorem II.6.10.2]), we get that ${\mathfrak{E}}_\mathcal{K}$ is a conditional expectation; in particular it is completely positive.
Define $$\mathcal{J}_{\mathcal{A}}=\{T\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}:{\mathfrak{E}_{\rm KS}}(T^*T)\in \mathcal{I}_{\mathcal{E}}\}.$$ Then we have $$\label{tempideal}
\mathcal{J}_{\mathcal{A}}=\{T\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}\,:\, {\mathfrak{E}_{\rm KS}}(QTR)\in\mathcal{I}_{\mathcal{E}}\, \text{ for all } Q,R\in {C^*_{r,\mathrm{KS}}({\mathcal{A}})}\}\,.$$ Thus $\mathcal{J}_{\mathcal{A}}$ is an ideal of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$, satisfying $\mathcal{I}_{\mathcal{A}}\subset\mathcal{J}_{\mathcal{A}}$ and ${\mathfrak{E}_{\rm KS}}(\mathcal{J}_{\mathcal{A}})=\mathcal{I}_{\mathcal{E}}$. Moreover, the conditional expectation ${\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}$ from ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}$ onto ${C^*_{r,\mathrm{alt}}({\mathcal{E}})}$ is faithful.
Let $\mathcal{K}$ be the ideal of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ defined by the right hand side of equation . Then by using an approximate unit for ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ one easily deduce that ${\mathfrak{E}_{\rm KS}}(\mathcal{K})=\mathcal{I}_{\mathcal{E}}$ and $\mathcal{K}\subset\mathcal{J}_{\mathcal{A}}$.
Let $T\in\mathcal{J}_{\mathcal{A}}$. Then we have $$\label{norm-zero}
\|{\mathfrak{E}}_\mathcal{K}(T^*T+\mathcal{K})\| = \|{\mathfrak{E}_{\rm KS}}(T^*T)+\mathcal{I}_{\mathcal{E}}\| = 0$$ since ${\mathfrak{E}_{\rm KS}}(T^*T)\in \mathcal{I}_{\mathcal{E}}$. Consider now $Q\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}$. Then, by using the Cauchy-Schwarz inequality (cf. [@lance95]) and equation (\[norm-zero\]), we get $$\begin{aligned}
\|{\mathfrak{E}_{\rm KS}}\big((QT)^*(QT)\big)+\mathcal{I}_{\mathcal{E}}\|^2&=
\|{\mathfrak{E}}_\mathcal{K}\big((QT)^*(QT)+\mathcal{K}\big)\|^2\\
&=\|{\mathfrak{E}}_\mathcal{K}\big((T+\mathcal{K})^*(Q^*QT+\mathcal{K})\big)\|^2 \\
&\leq \|{\mathfrak{E}}_\mathcal{K}(T^*T+\mathcal{K})\|\|{\mathfrak{E}}_\mathcal{K}((Q^*QT)^*(Q^*QT)+\mathcal{K})\|\\
&=0\end{aligned}$$ So ${\mathfrak{E}_{\rm KS}}((QT)^*(QT))\in \mathcal{I}_{\mathcal{E}}$, hence $QT\in\mathcal{J}_{\mathcal{A}}$.
Next, consider $R=\lambda_s(b)$ for $s\in S$ and $b\in A_s$. Lemma \[covar\] gives that ${\mathfrak{E}_{\rm KS}}(R^*T^*TR)=R^*{\mathfrak{E}_{\rm KS}}(T^*T)R$. So ${\mathfrak{E}_{\rm KS}}(R^*T^*TR)\in\mathcal{I}_{\mathcal{A}}$ since ${\mathfrak{E}_{\rm KS}}(T^*T)\in \mathcal{I}_{\mathcal{E}}\subset\mathcal{I}_{\mathcal{A}}$ and $\mathcal{I}_{\mathcal{A}}$ is an ideal. As the range of ${\mathfrak{E}_{\rm KS}}$ is ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$ we also get that ${\mathfrak{E}_{\rm KS}}(R^*T^*TR)\in{C^*_{r,\mathrm{KS}}({\mathcal{E}})}$, so ${\mathfrak{E}_{\rm KS}}(R^*T^*TR)\in\mathcal{I}_{\mathcal{E}}$. By the Schwarz inequality (sometimes called the Kadison inequality), see for instance [@blackadar Proposition II.6.9.14], we have $${\mathfrak{E}_{\rm KS}}(TR)^*{\mathfrak{E}_{\rm KS}}(TR)\leq{\mathfrak{E}_{\rm KS}}((TR)^*TR)= {\mathfrak{E}_{\rm KS}}(R^*T^*TR)\in\mathcal{I}_{\mathcal{E}}.$$ Hence ${\mathfrak{E}_{\rm KS}}(TR)^*{\mathfrak{E}_{\rm KS}}(TR) \in \mathcal{I}_{\mathcal{E}}$ since $\mathcal{I}_{\mathcal{E}}$ (being an ideal) is a hereditary subalgebra of ${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$, and it therefore follows that ${\mathfrak{E}_{\rm KS}}(TR)\in\mathcal{I}_{\mathcal{E}}$ (cf. [@blackadar Proposition II.5.1.1]). By linearity and continuity of ${\mathfrak{E}_{\rm KS}}$ and density of $\Phi_\Lambda^{\mathcal{A}}({\mathcal{C}_c({\mathcal{A}})})$ in ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$, we get that ${\mathfrak{E}_{\rm KS}}(TR)\in\mathcal{I}_{\mathcal{E}}$ for all $R\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}$.
If now $T\in\mathcal{J}_{\mathcal{A}}$ and $Q, R\in{C^*_{r,\mathrm{KS}}({\mathcal{A}})}$, then we get that $T':=QT\in\mathcal{J}_{\mathcal{A}}$, and this implies that ${\mathfrak{E}_{\rm KS}}(QTR)={\mathfrak{E}_{\rm KS}}(T'R)\in\mathcal{I}_{\mathcal{E}}$. This shows that $\mathcal{J}_{\mathcal{A}}\subset\mathcal{K}$, hence that $\mathcal{J}_{\mathcal{A}}=\mathcal{K}$.
Since we now have shown that $ \mathcal{J}_{\mathcal{A}}$ is an ideal of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ satisfying ${\mathfrak{E}_{\rm KS}}(\mathcal{J}_{\mathcal{A}})= \mathcal{I}_{\mathcal{E}}$, the canonical conditional expectation ${\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}$ from ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}$ onto ${C^*_{r,\mathrm{alt}}({\mathcal{E}})}$ is well defined. Showing that ${\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}$ is faithful amounts to verify that $T^*T\in\mathcal{J}_{\mathcal{A}}$ whenever ${\mathfrak{E}_{\rm KS}}(T^*T)\in\mathcal{I}_{\mathcal{E}}$. This readily follows from the definition of $\mathcal{J}_{\mathcal{A}}$ and the fact that $\mathcal{J}_{\mathcal{A}}$ is an ideal of ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$.
\[Psi-fEalt\] Let $q_{\mathcal{J}_{\mathcal{A}}} $ denote the quotient map from ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ onto ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}$. Then there exists a canonical isomorphism $$\Psi:{C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}\to C^*_r({\mathcal{A}})$$ satisfying $\Psi\circ q_{\mathcal{J}_{\mathcal{A}}}\circ \Phi_\Lambda^{\mathcal{A}}= \iota^{\rm red}_{\mathcal{A}}$.
The strategy for proving the proposition is to show that there exists a $*$-homomorphism $\Psi:{C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}\to C^*_r({\mathcal{A}})$ and a linear map ${\mathfrak{E}_{\rm r}}:C^*_r({\mathcal{A}})\to C^*({\mathcal{E}})$ making the following diagram commute:
(CrA) at (0,3) [$C^*_r({\mathcal{A}})$]{}; (CrJA) at (3,3) [${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}$]{}; (CE) at (0,0) [$C^*({\mathcal{E}})$]{}; (CraltE) at (3,0) [${C^*_{r,\mathrm{alt}}({\mathcal{E}})}$]{}; (CcA) at (1,2) [${\mathcal{C}_c({\mathcal{A}})}$]{}; (CrKSA) at (2,2) [${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$]{}; (CcE) at (1,1) [${\mathcal{C}_c({\mathcal{E}})}$]{}; (CrKSE) at (2,1) [${C^*_{r,\mathrm{KS}}({\mathcal{E}})}$]{}; (CcA) edge node\[above\] [$\Phi^{\mathcal{A}}_\Lambda$]{} (CrKSA) (CcE) edge node\[above\] [$\Phi^{\mathcal{E}}_\Lambda$]{} (CrKSE); (CrKSA) edge node\[above, sloped\] [$q_{\mathcal{J}_{\mathcal{A}}}$]{} (CrJA) (CrKSE) edge (CraltE); (CcA) edge node\[auto\] [${\mathfrak{E}}$]{} (CcE) (CrKSA) edge node\[auto\] [${\mathfrak{E}_{\rm KS}}$]{} (CrKSE) (CrJA) edge node\[auto\] [${\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}$]{} (CraltE) (CraltE) edge node\[above\] [$=$]{} (CE) (CcE) edge node\[above, sloped\] [$\iota_{\mathcal{E}}$]{} (CE) (CcA) edge node\[above, sloped\] [$\iota^{\rm red}_{\mathcal{A}}$]{}(CrA); (CrA) edge node\[auto\] [${\mathfrak{E}_{\rm r}}$]{} (CE) (CrJA) edge node\[above\] [$\Psi$]{} (CrA);
It will then follow that $\Psi$ is an isomorphism by considering the outer square in this diagram and using that ${\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}$ is faithful, as shown in the previous proposition.
Let $\varphi$ be a pure state on $C^*({\mathcal{E}})$. It is easy to deduce from equation (\[eq:stateextension\]) that the functional $\tilde{\varphi}$ on ${\mathcal{C}_c({\mathcal{A}})}$ defined in section \[sec:exelreduced\] is given by $$\tilde{\varphi}=\varphi\circ \iota_{{\mathcal{E}}}\circ{\mathfrak{E}}$$ where $\iota_{\mathcal{E}}$ denotes the canonical map from $\mathcal{C}_c({\mathcal{E}})$ to $C^*({\mathcal{E}})$. Moreover, it is straightforward to see that we have $\iota_{{\mathcal{E}}}\circ{\mathfrak{E}}= {\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}\circ q_{\mathcal{J}_{\mathcal{A}}}\circ\Phi_\Lambda^{\mathcal{A}}$, so we get $$\tilde{\varphi}=\varphi\circ{\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}\circ q_{\mathcal{J}_{\mathcal{A}}}\circ\Phi_\Lambda^{\mathcal{A}}\,.$$ Let $\varphi'=\varphi\circ{\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}\circ q_{\mathcal{J}_{\mathcal{A}}}$. Then $\varphi'$ is a state on ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$. As before, let $(\Upsilon_{\tilde{\varphi}}, H_{\tilde{\varphi}})$ be the GNS representation associated to $\tilde{\varphi}$, with $x\mapsto\hat{x}$ denoting the canonical map ${\mathcal{C}_c({\mathcal{A}})}\to H_{\tilde{\varphi}}$. Form also the GNS-representation $(\pi_{\varphi'}, H_{\varphi'})$ associated to $\varphi'$, with $T\mapsto \widehat{T}$ denoting the canonical map from ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}$ into $H_{\varphi'}$. For any $x\in{\mathcal{C}_c({\mathcal{A}})}$, we obtain $$\|\hat{x}\|^2=\tilde{\varphi}(x^*x)=\varphi'\big(\Phi_\Lambda^{\mathcal{A}}(x^*x)\big)=\|\widehat{\Phi_\Lambda^{\mathcal{A}}(x)}\|^2.$$ Since $\{\hat{x}:x\in{\mathcal{C}_c({\mathcal{A}})}\}$ is dense in $H_{\tilde{\varphi}}$, the assignment $\hat{x}\mapsto\widehat{\Phi_\Lambda^{\mathcal{A}}(x)}$ extends to an isometry $V:H_{\tilde{\varphi}}\to H_{\varphi'}$.
Consider now $g\in{\mathcal{C}_c({\mathcal{A}})}$. For any $x,y\in{\mathcal{C}_c({\mathcal{A}})}$ we get $$\begin{aligned}
\Big\langle V^*\pi_{\varphi'}\big(\Phi_\Lambda^{\mathcal{A}}(g)\big) V\hat{x},\hat{y}\Big\rangle&=\Big\langle \pi_{\varphi'}\big(\Phi_\Lambda^{\mathcal{A}}(g)\big)\widehat{\Phi_\Lambda^{\mathcal{A}}(x)},\widehat{\Phi_\Lambda^{\mathcal{A}}(y)} \Big\rangle\\
&=\varphi'\big(\Phi_\Lambda^{\mathcal{A}}(y^*gx)\big)
=\tilde{\varphi}(y^*gx)\\
&=\Big\langle \Upsilon_{\tilde{\varphi}}(g)\hat{x},\hat{y}\Big\rangle.\end{aligned}$$ So $\Upsilon_{\tilde{\varphi}}(g)=V^*\pi_{\varphi'}\big(\Phi_\Lambda^{\mathcal{A}}(g)\big)V$, and it follows that $\|\Upsilon_{\tilde{\varphi}}(g)\|\leq \|\pi_{\varphi'}\big(\Phi_\Lambda^{\mathcal{A}}(g)\big)\|$ since $V$ is an isometry. Moreover, as $\varphi'$ annihilates $\mathcal{J}_{\mathcal{A}}$, the kernel of $\pi_{\varphi'}$ contains $\mathcal{J}_{\mathcal{A}}$, so we get $\|\pi_{\varphi'}\big(\Phi_\Lambda^{\mathcal{A}}(g)\big)\|\leq\|q_{\mathcal{J}_{\mathcal{A}}}(\Phi_\Lambda^{\mathcal{A}}(g))\|$. Hence we conclude that $$\|\Upsilon_{\tilde{\varphi}}(g)\|\leq \|q_{\mathcal{J}_{\mathcal{A}}}(\Phi_\Lambda^{\mathcal{A}}(g))\|.$$ Since this holds for all $\varphi \in \mathcal{P}(C^*({\mathcal{E}}))$ we get $\|\iota^{\rm red}_{\mathcal{A}}(g)\|\leq \|q_{\mathcal{J}_{\mathcal{A}}}(\Phi_\Lambda^{\mathcal{A}}(g))\|$. It follows that there exists a $*$-homomorphism $\Psi:{C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}\to C^*_r({\mathcal{A}})$ satisfying $\Psi\big(q_{\mathcal{J}_{\mathcal{A}}}(\Phi_\Lambda^{\mathcal{A}}(g))\big) = \iota^{\rm red}_{\mathcal{A}}(g)$ for all $g \in{\mathcal{C}_c({\mathcal{A}})}$, as desired.
Next, we will show that the map ${\mathfrak{E}_{\rm r}}':\iota_{\mathcal{A}}^{\rm red}({\mathcal{C}_c({\mathcal{A}})})\to C^*({\mathcal{E}})$ given by ${\mathfrak{E}_{\rm r}}'(\iota_{\mathcal{A}}^{\rm red}(g))=\iota_{\mathcal{E}}({\mathfrak{E}}(g))$ is well defined, linear and contractive. By density, it will then extend to a (contractive) linear map ${\mathfrak{E}_{\rm r}}:C^*_r({\mathcal{A}})\to C^*({\mathcal{E}})$, as desired.
To see that ${\mathfrak{E}_{\rm r}}'$ is well defined, note that if $g\in{\mathcal{C}_c({\mathcal{A}})}$ and $\iota_{\mathcal{A}}^{\rm red}(g)=0$, then $0=\tilde{\varphi}(x^*gy)=\varphi(\iota_{\mathcal{E}}({\mathfrak{E}}(x^*gy)))$ for all $x,y\in{\mathcal{C}_c({\mathcal{A}})}$ and all $\varphi \in \mathcal{P}(C^*({\mathcal{E}}))$. Letting $x$ and $y$ range over ${\mathcal{C}_c({\mathcal{E}})}$ so that ${\mathfrak{E}}(x^*gy)=x^*{\mathfrak{E}}(g)y$ (which follows from Proposition \[condi-exp\] since ${\mathfrak{E}_{\rm KS}}$ is a conditional expectation), and using the density of $\iota_{\mathcal{E}}({\mathcal{C}_c({\mathcal{E}})})$ in $C^*({\mathcal{E}})$ we get $\iota_{\mathcal{E}}({\mathfrak{E}}(g))=0$. It readily follows that ${\mathfrak{E}_{\rm r}}'$ is well defined, and its linearity is then obvious.
Further, consider $g\in{\mathcal{C}_c({\mathcal{A}})}$. For $x,y\in{\mathcal{C}_c({\mathcal{E}})}$ with $\|\iota_{\mathcal{E}}(x)\|,\|\iota_{\mathcal{E}}(y)\|\leq 1$, we have $$\begin{aligned}
\|\iota_{\mathcal{E}}(x^*{\mathfrak{E}}(g)y)\|&=\|\iota_{\mathcal{E}}({\mathfrak{E}}(x^*gy))\|=\sup_{\varphi\in\mathcal{P}(C^*({\mathcal{E}}))}\varphi(\iota_{\mathcal{E}}({\mathfrak{E}}(x^*gy)))\\
&\leq \sup_{\varphi\in\mathcal{P}(C^*({\mathcal{E}}))}\|\Upsilon_{\tilde{\varphi}}(g)\|=\|\iota^{\rm red}_{\mathcal{A}}(g)\|\,.
\end{aligned}$$ For every $\varepsilon>0$ it is not difficult to see that we can find $x$ and $y$ as above such that $$\|\iota_{\mathcal{E}}({\mathfrak{E}}(g))\|\leq\|\iota_{\mathcal{E}}(x^*{\mathfrak{E}}(g)y)\|+\varepsilon,$$ so we get $$\|{\mathfrak{E}_{\rm r}}'(\iota_{\mathcal{A}}^{\rm red}(g))\| =\|\iota_{\mathcal{E}}({\mathfrak{E}}(g))\|\, \leq\,\|\iota_{\mathcal{E}}(x^*{\mathfrak{E}}(g)y)\|+\varepsilon \,\leq\, \|\iota^{\rm red}_{\mathcal{A}}(g)\| + \varepsilon\,.$$ Thus we conclude that ${\mathfrak{E}_{\rm r}}'$ is contractive.
The reader will have no problem to check that the maps $\Psi$ and ${\mathfrak{E}_{\rm r}}$ we have constructed make the above diagram commutative, thus finishing the proof.
From Proposition \[Psi-fEalt\] and its proof we get the following result which may be worthy of being stated separately. It is proved for saturated Fell bundles over unital inverse semigroups in [@buss_exel_meyer15] using a different approach.
\[faith-cond\] There exists a faithful conditional expectation $${\mathfrak{E}_{\rm r}}:C^*_r({\mathcal{A}})\to C^*({\mathcal{E}})$$ satisfying ${\mathfrak{E}_{\rm r}}\circ\iota_{\mathcal{A}}^{\rm red} = \iota_{\mathcal{A}}\circ{\mathfrak{E}}$.
Since $\mathcal{I}_{\mathcal{A}}\subset\mathcal{J}_{\mathcal{A}}$, there is a natural surjective $*$-homomorphism $p_{\mathcal{A}}$ from ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}= {C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{I}_{\mathcal{A}}$ onto ${C^*_{r,\mathrm{KS}}({\mathcal{A}})}/\mathcal{J}_{\mathcal{A}}$. Using Propositions \[Psi-fEalt\] and \[faith-cond\], the relationship between the reduced $C^*$-algebra ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ introduced in the present article and Exel’s $C^*_r({\mathcal{A}})$ can be described as follows:
\[Alt-Exel\] There exists a surjective canonical $*$-homomorphism $$\Psi':{C^*_{r,\mathrm{alt}}({\mathcal{A}})}\to C^*_r({\mathcal{A}})$$ satisfying $\Psi'\circ \Psi_{\Lambda^{\rm alt}} = \Psi_{\Pi^{\rm red}}$.
Moreover, the conditional expectation ${\mathfrak{E}_{\rm r}}^{\rm alt}:{C^*_{r,\mathrm{alt}}({\mathcal{A}})}\to{C^*_{r,\mathrm{alt}}({\mathcal{E}})}=C^*({\mathcal{E}})$ given by ${\mathfrak{E}_{\rm r}}^{\rm alt} = {\mathfrak{E}_{\rm r}}\circ \Psi'$ is canonical in the sense that $${\mathfrak{E}_{\rm r}}^{\rm alt}\big( \Phi^{\mathcal{A}}_{\Lambda^{\rm alt}}(g)\big)
= \Phi^{\mathcal{E}}_{\Lambda^{\rm alt}}({\mathfrak{E}}(g))$$ for all $g \in C_c({\mathcal{A}})$, and the following conditions are equivalent:
- $\Psi'$ is an isomorphism;
- $\mathcal{I}_{\mathcal{A}}=\mathcal{J}_{\mathcal{A}}$;
- ${\mathfrak{E}_{\rm r}}^{\rm alt}$ is faithful.
It suffices to set $\Psi' = \Psi \circ p_{\mathcal{A}}$ and observe that ${\mathfrak{E}_{\rm r}}^{\rm alt} = {\mathfrak{E}}_{\mathcal{J}_{\mathcal{A}}}\circ p_{\mathcal{A}}$.
We don’t know whether ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ is isomorphic to $C^*_r({\mathcal{A}})$ in general. When $S$ is strongly $E^*$-unitary this happens quite often.
Assume $S$ is strongly $E^*$-unitary and let $\sigma:S^\times\to G$ be an idempotent pure grading into a group $G$. Let $\mathcal{B}$ be the associated Fell bundle over $G$ defined in Remark \[groupbundle\]. If $G$ is exact [@brown_ozawa08], or if $\mathcal{B}$ satisfies Exel’s approximation property [@exel97], then ${C^*_{r,\mathrm{alt}}({\mathcal{A}})}$ is canonically isomorphic to $ C^*_r({\mathcal{A}})$.
By using [@exel14 Theorem 23.7] if $G$ is exact, or [@exel14 Proposition 23.6] if $\mathcal{B}$ satisfies Exel’s approximation property, one deduces easily that $\mathcal{I}_{\mathcal{A}}=\mathcal{J}_{\mathcal{A}}$ after making appropriate identifications of these ideals in $C^*_r(\mathcal{B})$. Hence, the result follows from Theorem \[Alt-Exel\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the first parallax and luminosity measurements for an L subdwarf, the sdL7 [[2MASS J05325346+8246465]{}]{}. Observations conducted over three years by the USNO infrared astrometry program yield an astrometric distance of 26.7$\pm$1.2 pc and a proper motion of 2.6241$\pm$0.0018$\arcsec$ yr$^{-1}$. Combined with broadband spectral and photometric measurements, we determine a luminosity of $\log{L_{bol}/L_{\sun}}$ = -4.24$\pm$0.06 and [[T$_{eff}$]{}]{} = 1730$\pm$90 K (the latter assuming an age of 5–10 Gyr), comparable to mid-type L field dwarfs. Comparison of the luminosity of [[2MASS J05325346+8246465]{}]{} to theoretical evolutionary models indicates that its mass is just below the sustained hydrogen burning limit, and is therefore a brown dwarf. Its kinematics indicate a $\sim$110 Myr, retrograde Galactic orbit which is both eccentric (3 $\lesssim$ R $\lesssim$ 8.5 kpc) and extends well away from the plane ($\Delta$Z = $\pm$2 kpc), consistent with membership in the inner halo population. The relatively bright $J$-band magnitude of [[2MASS J05325346+8246465]{}]{} implies significantly reduced opacity in the 1.2 $\micron$ region, consistent with inhibited condensate formation as previously proposed. Its as yet unknown subsolar metallicity remains the primary limitation in constraining its mass; determination of both parameters would provide a powerful test of interior and evolutionary models for low-mass stars and brown dwarfs.'
author:
- 'Adam J. Burgasser, Frederick J. Vrba, Sébastien Lépine, Jeffrey A. Munn, Christian B. Luginbuhl, Arne A. Henden, Harry H. Guetter, and Blaise C. Canzian'
title: Parallax and Luminosity Measurements of an L Subdwarf
---
Introduction
============
The lowest luminosity stars and brown dwarfs are among the most useful probes of planetary, stellar and Galactic processes. With hydrogen burning lifetimes far in excess of a Hubble time (e.g., @lau97) and space densities that exceed those of hotter stars, low-mass dwarfs are ubiquitous in the disk, thick disk and halo populations (e.g., @dah95 [@rei02; @dig03; @cru07]), and may host the bulk of terrestrial planets in the Galaxy (e.g., @bos06 [@tar07]). The steady cooling of brown dwarfs over time makes them useful chronometers for coeval clusters [@bil97; @sta98], and these sources probe star formation down to its lowest mass limit [@luh07 and references therein]. Observations of cool brown dwarfs provide empirical constraints on atmospheric chemical models, opacities and dynamics (e.g., @ack01 [@lod02; @hel04]), and facilitate studies of hot exoplanetary atmospheres (e.g., @bar03).
Over the past decade, hundreds of low-mass stars and brown dwarfs have been identified by wide-field optical and near-infrared surveys such as the Two Micron All Sky Survey (2MASS; @skr06), the Sloan Digital Sky Survey (SDSS; @yor00) and the Deep Near Infrared Survey of the Southern Sky (DENIS; @epc97). These discoveries include members of two new spectral classes, the L dwarfs and T dwarfs [@kir05 and references therein][^1]. As the number of low-luminosity dwarfs grows, distinct populations of “peculiar” sources are being found which exhibit unusual surface gravities (e.g., @kir06), metallicities (e.g., @me0532 [@lep1610]), or atmospheric structures (e.g., @cru03 [@cru07; @kna04]). One such population is the L subdwarf class [@mecs13], the metal-poor counterparts to solar-metallicity field L dwarfs and the low-temperature extension of the M subdwarf sequence [@giz97]. L subdwarfs are distinguished from L dwarfs by the presence of relatively enhanced metal-hydride absorption bands (CaH, FeH, CrH) and metal lines (, , ) relative to reduced metal-oxide absorption (TiO, VO; @mou76). They also exhibit exceptionally blue near-infrared colors ($J-K_s \lesssim 0$ compared to $J-K_s \approx 1.5-2.5$) caused by strong collision-induced [[H$_2$]{}]{} absorption [@lin69; @sau94; @bor97]. Their halo or thick disk kinematics [@me0532; @rei06] indicate an origin early in the Galaxy’s history. L subdwarfs are useful for studying metallicity effects on cool atmospheric chemistry, particularly at temperatures in which photospheric condensates first become an important source of opacity [@ack01; @all01; @bur06]. Like L dwarfs, they may also span the metal-dependent hydrogen-burning mass limit [@me0532], and are therefore probes of low-mass star formation in the metal-poor halo. However, L subdwarfs are also exceptionally rare. Only four are currently known, identified serendipitously in the 2MASS [@me0532; @me1626], SDSS [@siv07] and SUPERBLINK [@lep1610] surveys.
Understanding the physical characteristics of metal-poor low-mass stars and brown dwarfs requires the characterization of basic observational properties — distance, luminosity and effective temperature ([[T$_{eff}$]{}]{}) — which can be facilitated by parallax measurements. However, while astrometric studies of late-type field dwarfs have provided robust absolute magnitude, [[T$_{eff}$]{}]{} and luminosity scales down to the lowest luminosity brown dwarfs known [@dah02; @tin03; @vrb04], parallax measurements for late-type subdwarfs are rare [@mon92], and there are no such measurements for any L subdwarfs.
To address this deficiency, we report the first parallax measurement of an L subdwarf, [[2MASS J05325346+8246465]{}]{} (hereafter [[2MASS J0532+8246]{}]{}; @me0532), one of the first and latest-type L subdwarfs to be identified. This source is tentatively classified sdL7 due to the similarity of its optical spectrum to those of solar-metallicity L7 dwarfs [@megmos]. Its late spectral type, low estimated temperature ([[T$_{eff}$]{}]{} $\approx$ 1400–2000 K) and halo kinematics all make [[2MASS J0532+8246]{}]{} a strong candidate halo brown dwarf.
In $\S$ 2 we describe astrometric and photometric measurements of [[2MASS J0532+8246]{}]{}, conducted as part of the United States Naval Observatory (USNO) infrared astrometry program [@vrb04], and compare its absolute photometry to field late-type M, L and T dwarfs. In $\S$ 3 we calculate the bolometric luminosity and [[T$_{eff}$]{}]{} of [[2MASS J0532+8246]{}]{} based on our astrometric measurements and broad-band spectral and photometric measurements reported in the literature. In $\S$ 4 we analyze these results, determining estimates of the physical properties of [[2MASS J0532+8246]{}]{} using evolutionary models from @bar97 [@bar98] and @bur01, and examine the kinematics of this source and its Galactic orbit. Results are summarized in $\S$ 5.
Observations
============
Astrometric Measurements
------------------------
Astrometric observations of [[2MASS J0532+8246]{}]{} were obtained with the ASTROCAM near-infrared imager [@fis03] at the USNO Flagstaff Station 61-inch Kaj Straand Astrometric Reflector on 37 nights spanning a $\sim$3 yr period beginning in February 2003 and ending in February 2006.[^2] Each of the observations was made in the $H$-band (1.7 $\micron$), with three dithered exposures of 450 to 900 s each, depending on seeing conditions (always less than 2$\farcs$5), or 1350 to 2700 s total integration per visit. Data acquisition and astrometric reduction procedures are discussed in detail in @vrb04. Note that observations for [[2MASS J0532+8246]{}]{} span a longer baseline than the L and T dwarfs reported in @vrb04, so that stable parallaxes in both right ascension and declination could be determined and combined. Twelve stars were employed in the reference frame, and photometric parallaxes for these sources were determined using 2MASS photometry transformed to the CIT system. The correction from relative to absolute parallax was found to be 1.14$\pm$0.13 mas.
Final values for the parallax (relative and absolute) and proper motion solutions for [[2MASS J0532+8246]{}]{} are listed in Table \[tab:usno\]. The astrometric distance of this source is 26.7$\pm$1.2 pc, within the 10-30 pc range estimated by @me0532. The proper motion measurement, 2.6241$\pm$0.0018 $\arcsec$ yr$^{-1}$, is also consistent with but substantially more accurate than the previous determination (2.60$\pm$0.15 $\arcsec$ yr$^{-1}$). The USNO measurements confirm the high tangential space velocity of [[2MASS J0532+8246]{}]{}, $V_{tan}$ = 332$\pm$15 [[km s$^{-1}$]{}]{}, over three times larger than any of the known “disk” field L dwarfs [@vrb04; @sch07]. Combined with its radial velocity ($V_{rad}$ = $-172{\pm}1$ [[km s$^{-1}$]{}]{}; @rei06), we determine local standard of rest (LSR) velocities of $(U,V,W)_{LSR}$ = $(-70,-354,78)$ $\pm$ $(9,13,7)$ [[km s$^{-1}$]{}]{}, assuming a LSR solar velocity of (10.00, 5.25, 7.17) [[km s$^{-1}$]{}]{} [@deh98]. The substantial negative $V_{LSR}$ velocity implies that [[2MASS J0532+8246]{}]{} is orbiting in a retrograde motion about the Galactic center; i.e., in the opposite sense of the Galactic disk (assuming that $V_{disk}$ = 220 [[km s$^{-1}$]{}]{} in galactocentric coordinates; @ker86). This would appear to rule out membership in the thick disk population [@chi00], so [[2MASS J0532+8246]{}]{} is likely to be a halo low-mass object. The Galactic orbit of this source is discussed further in $\S$ 4.2.
Photometric Measurements
------------------------
Near-infrared photometric measurements of [[2MASS J0532+8246]{}]{} were also obtained during the course of the astrometric observations in an attempt to improve upon existing 2MASS photometry (uncertainties of 0.06, 0.09, and 0.15 mag in $J$-, $H$-, and $K_s$-bands, respectively). $J$-, $H$- and $K$-band observations on the CIT system [@gue03] were obtained with ASTROCAM on 2003 October 12 (UT). Conditions were clear with $\sim$0.9$\arcsec$ seeing. Three dithered exposures were obtained in each filter with total integration times of 720, 720, and 1500 s, respectively. The ASTROCAM field of view contained 9 stars significantly brighter than [[2MASS J0532+8246]{}]{}, with mean 2MASS $J$, $H$, and $K_s$ uncertainties of 0.03, 0.04, and 0.07 mag, respectively. 2MASS photometry for these sources were converted to the CIT system using the transformations of @car01, which were then employed as “local photometric standards” on the ASTROCAM frames. Aperture photometry of these stars and [[2MASS J0532+8246]{}]{} was then carried out using DAOPHOT in the IRAF[^3] environment on each of the individual dithers, and the combined instrumental magnitudes and colors were used to determine transformations to the CIT standard system. Final results for [[2MASS J0532+8246]{}]{} are listed in Table \[tab:usno\]. We note that along with reduced errors, the USNO photometry of this source is brighter and somewhat bluer than the 2MASS photometry, albeit within the 3$\sigma$ uncertainties of the latter.
Examination of Absolute Magnitudes
----------------------------------
[[2MASS J0532+8246]{}]{} is the first L subdwarf to have a measured distance; hence, examination of its absolute magnitudes (Table \[tab:properties\]) is of some importance in understanding the atmospheric properties of low-temperature metal-poor dwarfs. Figures \[fig:absmvscolor\] and \[fig:absmvsspt\] display a sample of optical and infrared color/magnitude and spectral type/magnitude diagrams for late-type M, L and T dwarfs with parallax measurements, including [[2MASS J0532+8246]{}]{}. Photometric data were culled from @dah02 and @me0532 for $I_c$-band (0.8 $\micron$); 2MASS for near-infrared $JHK_s$, and @pat06 for mid-infrared measurements made with the [*Spitzer*]{} IRAC instrument [@faz04]. Astrometric data are from @pry97 [@dah02; @tin03]; and @vrb04. For the spectral type/magnitude plots, we used published optical spectral types for late-type M and L dwarfs (e.g., @kir99) and near-infrared spectral types for T dwarfs (e.g., @meclass2). Sources were constrained to have color and magnitude uncertainties no greater than 0.2 mag, and not known to be multiple [@meppv and references therein]. We also include absolute magnitudes and colors for the sdM7 LHS 377 [@mon92; @giz97], until now the latest-type subdwarf with a reported parallax measurement.
In general, [[2MASS J0532+8246]{}]{} is somewhat overluminous for both its color and spectral type. This is particularly the case at $J$-band, where it is 1–2 mag brighter than the disk L dwarf/T dwarf sequence based on both spectral type and color. Indeed, [[2MASS J0532+8246]{}]{} lies in a relatively unpopulated region in the $M_J$ versus $(J-K_s)$ color/magnitude diagram. $J$-band opacity in L dwarf photospheres is dominated by absorption from condensates, which gives rise to their red $J-K_s$ colors [@tsu96; @tsu99; @bur99; @ack01; @all01]. @me0532, @rei06 and @megmos have all speculated that condensate formation may be inhibited in L subdwarf photospheres based on the unexpected strength of gaseous TiO, and features. Reduced condensate opacity allows $J$-band light to escape from deeper and hotter layers, resulting in an overall brightening at these wavelengths [@ack01]. In contrast, $K$-band opacity is dominated by collision-induced [[H$_2$]{}]{} opacity in late-type dwarfs, a species whose absolute abundance is not modified by metallicity. The higher pressure photospheres of old, high surface gravity subdwarfs will in fact increase [[H$_2$]{}]{} opacity. This trend explains why the $M_{K_s}$ magnitude of [[2MASS J0532+8246]{}]{} is consistent with, and perhaps slightly fainter than, those of solar-metallicity L7 field dwarfs. Given that the primary deviation in the near-infrared brightness of [[2MASS J0532+8246]{}]{} occurs in the spectral region in which condensate opacity should be dominant, these results provide further support for the idea that condensate formation is inhibited in low-temperature metal-poor atmospheres.
The location of [[2MASS J0532+8246]{}]{} redward and/or above the L dwarf sequence in the $M_{I_c}$ versus $(I_c-J)$ and $M_{K_s}$ versus $(K_s-[4.5])$ color/magnitude diagrams can also be explained by metallicity effects. Reduced $J$-band condensate opacity coupled with residual (albeit reduced) TiO and pressure-broadened absorption at $I_c$-band result in slightly redder $(I_c-J)$ colors than comparably classified field dwarfs. This is in contrast to the sdM7 LHS 377 which is brighter at $I_c$-band rather than $J$-band as compared to M7 field dwarfs, and hence bluer in $(I_c-J)$. The difference in color peculiarity between these two sources can be attributed to the fact that late M dwarfs are too warm to have significant condensates in their photospheres [@ack01], while variations in metal oxide absorption have a greater effect. The $(K_s-[4.5])$ color of [[2MASS J0532+8246]{}]{} is slightly redder than field dwarfs with similar $M_{K_s}$ magnitudes because of enhanced 4.5 $\micron$ flux. This region is dominated by metallicity-sensitive molecular CO and [[H$_2$O]{}]{} opacity, in contrast to the to strong [[H$_2$]{}]{} absorption at $K_s$. Both [[2MASS J0532+8246]{}]{} and LHS 377 have redder $([3.6]-[4.5])$ colors than the dwarf sequence; similar metallicity-induced color effects have also been noted amongst field T dwarfs [@lie07; @leg07].[^4] The red $([3.6]-[4.5])$ color of [[2MASS J0532+8246]{}]{} may be muted, however, if [[CH$_4$]{}]{} absorption at 3.3 $\micron$, present in mid- and late-type L dwarfs [@nol00; @cus05] is also weakened by metallicity effects.
Bolometric Luminosity and Effective Temperature Determinations
==============================================================
Our distance measurement for [[2MASS J0532+8246]{}]{} enables calculation of its bolometric luminosity and [[T$_{eff}$]{}]{}, through the use of existing spectroscopic and photometric data spanning 0.6–15 $\micron$. Our baseline calculation, illustrated in Figure \[fig:sed\], was constructed as follows. First, near-infrared spectral data for [[2MASS J0532+8246]{}]{} from @me0532 were piece-wise scaled to absolute 2MASS $J$, $H$ and USNO $K$-band magnitudes (Table \[tab:properties\]). Gaps in the [[H$_2$O]{}]{} bands and the 1 $\micron$ peak were substituted by a NEXTGEN spectral model [@hau99] with parameters [[T$_{eff}$]{}]{} = 2000 K, [[$\log{g}$]{}]{} = 5.5, \[M/H\] = -1.0 dex, scaled to match the flux-calibrated spectral data in the near-infrared peaks. Red optical spectral data for [[2MASS J0532+8246]{}]{} were scaled to the absolute $I_c$ magnitude of this source [@me0532]. As there are no reported spectral data for [[2MASS J0532+8246]{}]{} at longer wavelengths, we used a spectral template based on data for the L5 dwarf 2MASS J15074769-1627386 [@rei00; @cus06 hereafter 2MASS J1507-1627], spanning the 2.9–4.1 and 5.2–15.4 $\micron$ regions and piece-wise scaled to match the absolute 3.6, 5.8 and 8.0 $\micron$ [*Spitzer*]{} IRAC photometry of [[2MASS J0532+8246]{}]{} [@pat06]. Flux calibration of these data was done using the appropriate correction factor for IRAC photometry as discussed in @cus06. The 2.4–2.9 and 4.1–5.2 $\micron$ gaps were again replaced with theoretical model spectra, scaled to overlap with the template spectra and to match absolute 4.5 $\micron$ photometry of [[2MASS J0532+8246]{}]{}. Short (0.1 $<$ $\lambda < 0.6$ ) and long (15 $<$ $\lambda < 1000$ ) wavelength regions were calculated by scaling the NEXTGEN spectral model to the observed and template spectra, respectively.
The resulting broad spectral energy distribution was then integrated to determine the total spectral flux, with nominal uncertainties based on the absolute photometry, including astrometric uncertainties. To examine systematic effects, we repeated our analysis replacing the NEXTGEN model with a [[T$_{eff}$]{}]{} = 1700 K, [[$\log{g}$]{}]{} = 5.5., \[M/H\] = -0.5 condensate cloud model from @bur06 [hereafter, Tucson model], as well as linearly interpolating over gaps in the observed and template spectral data.
Table \[tab:sed\] provides a breakdown of the total fluxes of [[2MASS J0532+8246]{}]{} in five spectral regions, with comparisons between the different computational methods. As expected, the bulk of the spectral flux for this late-type dwarf arises at infrared wavelengths ($\sim$93%), with the 1.0–2.9 $\micron$ region encompassing almost 70% of the total light. Differences between the computational methods are generally less than the formal uncertainties, except in the shortest and longest wavelength regions which fortuitously contribute negligibly to the aggregate flux ($<$0.25%). Bolometric flux values computed using the two spectral models and linear interpolation are within 3$\sigma$ of each other, with our baseline calculation providing the median value. We therefore use this value as the measured bolometic flux, and propagate the estimated systematic error ($\sim$14%) into our final uncertainties.
We determine an absolute bolometric flux of (18.2$\pm$2.6)$\times$10$^{-12}$ erg cm$^{-2}$ s$^{-1}$ for [[2MASS J0532+8246]{}]{}, which translates into a bolometric luminosity of $\log{L/L_{bol}}$ = -4.24$\pm$0.06, or $M_{bol}$ = 15.35$\pm$0.16. This is comparable to mid-type L dwarfs like 2MASS J1507-1627 ($M_{bol}$ = 15.41$\pm$0.13; @vrb04), and is roughly half a magnitude more luminous than the average L7 dwarf. [[2MASS J0532+8246]{}]{} is nevertheless over 10 times less luminous than the next coolest subdwarf with a parallax measurement, sdM7 LHS 377 ($\log{L/L_{bol}}$ = -3.11$\pm$0.02; @leg00). Bolometric corrections ($BC_b \equiv M_{bol} - M_b$) at $J$ and $K$ were determined as $BC_J$ = 2.30$\pm$0.08 and $BC_K$ = 2.68$\pm$0.09. The latter is $\sim$0.5 mag smaller than comparable values computed for L6–L8 field dwarfs, further illustrating the substantial redistribution of flux from this metal-poor source.
An effective temperature for [[2MASS J0532+8246]{}]{} can be estimated assuming that this source is likely older than $\sim$5 Gyr based on its kinematics, and has a mass near the hydrogen burning limit (0.05–0.09 M$_{\sun}$, depending on metallicity; see $\S$ 4.1). These assumptions imply a radius of 0.096$\pm$0.015 R$_{\sun}$ based on the solar-metallicity evolutionary models of @bur01 and @bar03, and yields [[T$_{eff}$]{}]{} = 1600$\pm$300 K, where the uncertainty is entirely dominated by the uncertainty in the radius. If we include the measured luminosity (including 3$\sigma$ uncertainty) as an additional constraint on these models, a more refined radius estimate of 0.084$\pm$0.003 R$_{\sun}$ is obtained, corresponding to [[T$_{eff}$]{}]{} = 1730$\pm$90 K. Again, this is comparable to temperatures for mid-type L dwarfs [@vrb04; @gol04] and is nearly 1200 K cooler than LHS 377 [@leg00]
Discussion
==========
The Substellar Nature of [[2MASS J0532+8246]{}]{}
-------------------------------------------------
While parallax and bolometric luminosity measurements are able to constrain some of the atmospheric properties of [[2MASS J0532+8246]{}]{}, they do not directly address the question of whether this source is substellar. For this we require comparison to evolutionary models; in particular, models incorporating subsolar metallicities. The reduced opacity of a metal-poor atmosphere results in more rapid cooling; hence, the luminosity of a metal-poor brown dwarf will be lower than that of a solar metallicity brown dwarf with the same mass and age, while the hydrogen burning minimum mass limit increases for lower metallicities (e.g., @dan85). Figure \[fig:evol\] illustrates these trends by comparing theoretical mass/luminosity relations from @bur01 and @bar97 [@bar98], spanning the hydrogen burning limit for ages of 5 and 10 Gyr and metallicities \[M/H\] = 0, -1 and -2 dex. The limits at which 50% and 99% of the luminosity is generated from core hydrogen fusion at 10 Gyr are indicated; note how these limits shift to higher masses and luminosities for lower metallicity models. The luminosity of [[2MASS J0532+8246]{}]{} falls below the 99% limits for all models shown, as well as for intermediate and lower metallicities. This appears to confirm [[2MASS J0532+8246]{}]{} as a brown dwarf, incapable of sustaining its luminosity by core hydrogen fusion alone. However, note that for \[M/H\] $\gtrsim$ -1, hydrogen fusion provides over half of the energy emitted, so [[2MASS J0532+8246]{}]{} is fairly close to the stellar/substellar boundary. Indeed, if its metallicity is sufficiently high, this object may eventually cool to the point at which its lower luminosity is sustained by core fusion, changing this low-mass brown dwarf into a star.
The luminosity of [[2MASS J0532+8246]{}]{} falls on a particularly steep section of the mass/luminosity relations shown, enabling relatively tight (1%), albeit model-dependent, constraints on the mass of this source for a given metallicity. However, the variation in the derived mass for the metallicity range shown is considerably larger, ranging from 0.0744 to 0.0835 M$_{\sun}$ for \[M/H\] = 0 to -2 for the @bur01 models (these values are consistent with mass estimates based on the @bar97 [@bar98] models). Our best guess for the metallicity of this source, \[M/H\] = -1 (@megmos; see also @sch1444), implies a mass of 0.0783$\pm$0.0013 M$_{\sun}$ (0.0788$\pm$0.0006 M$_{\sun}$ for the @bar97 [@bar98] models). However, it is clear that the unknown metallicity of [[2MASS J0532+8246]{}]{} remains the largest source of uncertainty in characterizing its physical properties.
The Galactic Orbit of [[2MASS J0532+8246]{}]{}
----------------------------------------------
The kinematics of [[2MASS J0532+8246]{}]{} argue strongly for membership in the Galactic halo, but was this source formed early in the Galaxy or was it tidally stripped from one of the Galaxy’s dwarf satellites during a merging event? To examine this question, we calculated a probable orbit for [[2MASS J0532+8246]{}]{} based on its measured kinematics, distance and 2MASS coordinates (epoch 1999 March 1 UT; Table \[tab:properties\]). We used the Galactic mass model of @dau95 and a Runge-Kutta fourth-order integrator; see @lep1826 for details. Figure \[fig:orbit\] displays the results of this calculation, projecting the $\sim$110 Myr Galactocentric orbit in cylindrical coordinates, as well as the evolution of Galactic radius and vertical scaleheight over time. [[2MASS J0532+8246]{}]{} appears to have a relatively eccentric orbit, moving between roughly 3 and 8.5 kpc of the Galactic center ($e$ $\approx$ 0.5). The vertical scaleheight of its orbit extends $\sim$2 kpc above and below the Galactic plane.
The orbital characteristics of [[2MASS J0532+8246]{}]{} appear to be inconsistent with tidal capture from an external system, as this source generally remains interior to the Solar Galactic orbit. Rather, it is likely to be a member of the Galaxy’s inner flattened halo [@som90], indicating formation from the early building blocks of the Galaxy (e.g., @nor94 [@car96]). The substantial retrograde motion of [[2MASS J0532+8246]{}]{} also argues against formation and (violent) ejection from the inner disk. In other words, [[2MASS J0532+8246]{}]{} is likely to be quite old. However, this does not provide a robust constraint on its metallicity. Mean eccentricity/metallicity relations by @car96 and @chi00 suggest \[Fe/H\] $\lesssim$ -1, but the latter study also finds no clear correlation between orbital eccentricity and metallicity for \[Fe/H\] $<$ -0.8. The chemical abundances present in the atmosphere of [[2MASS J0532+8246]{}]{}, a necessary constraint for mass estimation, must be determined through other means, most likely spectral modeling.
Summary
=======
We have presented the first parallax and luminosity measurements for an L subdwarf, the sdL7 [[2MASS J0532+8246]{}]{}. The derived parameters, summarized in Table \[tab:properties\], confirm the low-luminosity, low-temperature nature of this source, similar in both respects to mid-type field L dwarfs. Examination of absolute photometry indicates that this source is substantially brighter at $J$-band than comparable solar-metallicity field dwarfs, consistent with reduced condensate opacity as previously suggested by optical spectroscopy. The low luminosity of [[2MASS J0532+8246]{}]{}, $\log{L_{bol}/L_{\sun}}$ = -4.24$\pm$0.06, compared to theoretical evolutionary models also confirms the substellar nature of this source (M $\approx$ 0.078 M$_{\sun}$, assuming \[M/H\] = -1), although it is probably still fusing hydrogen at a low level in its core. Its kinematics are consistent with membership in the inner halo population, making this source the first [*bona-fide*]{} halo brown dwarf.
Further investigation of [[2MASS J0532+8246]{}]{} should be directed toward determining its metallicity and chemical abundances. As discussed in $\S$ 4.1, uncertainty in the estimated mass of this source is dominated by its unknown metallicity. High resolution spectroscopy (e.g., @rei06) and accurate spectral modeling can improve mass constraints, although such analysis must also provide a consistent luminosity determination (cf., @smi03). Independent determination of both metallicity and mass could also provide an important empirical test on the evolutionary models themselves, and thus one of the few constraints on interior brown dwarf physics (in addition to radii measurements; e.g., @sta06). Studies of metallicity effects in the physical and observational properties of low-mass stars and brown dwarfs will also benefit from the measurement of parallaxes for additional late-type subdwarfs, to fill the gap between LHS 377 (sdM7) and [[2MASS J0532+8246]{}]{} (sdL7). Such observations should be of high priority, as detailed studies of metallicity effects in the atmospheric properties and evolution of the lowest-luminosity stars and brown dwarfs are currently limited by the absence of these data.
The authors would like to thank I. Baraffe and A. Burrows for making available electronic versions of their evolutionary models for our analysis, and M. Cushing for making available his spectral data for 2MASS J1507-1627.. We also thank J. Bochanski for additional comments to the manuscript and J. Gizis for his prompt review. This publication makes use of data from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center, and funded by the National Aeronautics and Space Administration and the National Science Foundation. 2MASS data were obtained from the NASA/IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This research has benefitted from the M, L, and T dwarf compendium housed at DwarfArchives.org and maintained by Chris Gelino, Davy Kirkpatrick, and Adam Burgasser; and the VLM Binary Archive maintained by N. Siegler at [http://paperclip.as.arizona.edu/\$\\sim\$nsiegler/VLM\_binaries/](http://paperclip.as.arizona.edu/$\sim$nsiegler/VLM_binaries/).
Facilities:
Ackerman, A. S., & Marley, M. S. 2001, , 556, 872
Allard, F., Hauschildt, P. H., Alexander, D. R., Tamanai, A., & Schweitzer, A. 2001, , 556, 357
Baraffe, I., Chabrier, G., Allard, F., & Hauschildt, P. H. 1997, , 327, 1054
Baraffe, I., Chabrier, G., Allard, F., & Hauschildt, P. H. 1998, , 337, 403
Baraffe, I., Chabrier, G., Barman, T., Allard, F., & Hauschildt, P. H. 2003, , 382, 563
Bildsten, L., Brown, E. F., Matzner, C. D., & Ushomirsky, G. 1997, , 482, 442
Borysow, A., J[ø]{}rgensen, U. G., & Zheng, C. 1997, , 324, 185
Boss, A. P. 2006, , 644, L79
Burgasser, A. J. 2004, , 614, L73
Burgasser, A. J., Cruz, K. L., & Kirkpatrick, J. D. 2007, , 657, 494
Burgasser, A. J., Geballe, T. R., Leggett, S. K., Kirkpatrick, J. D., & Golimowski, D. A. 2006, , 637, 1067
Burgasser, A. J., Kirkpatrick, J. D., Burrows, A., Liebert, J., Reid, I. N., Gizis, J. E., McGovern, M. R., Prato, L., & McLean, I. S. 2003, , 592, 1186
Burgasser, A. J., Kirkpatrick, J. D., & Lépine, S. 2005 in The 13th Cambridge Workshop on Cool Stars, Stellar Systems, and the Sun (ESA-SP-560), ed. F. Favata, G. A. J. Hussain & B. Battrick (Noordwijk: ESA), p. 237
Burgasser, A. J., Reid, I. N., Siegler, N., Close, L. M., Allen, P., Lowrance, P. J., & Gizis, J. E. 2007, in Planets and Protostars V, eds. B. Reipurth, D. Jewitt and K. Keil (Univ. Arizona Press: Tucson), p. 427
Burrows, A., Hubbard, W. B., Lunine, J. I., & Liebert, J. 2001, Rev. of Modern Physics, 73, 719
Burrows, A., & Sharp, C. M. 1999, , 512, 843
Burrows, A., Sudarsky, D., & Hubeny, I. 2006, , 640, 1063
Carney, B. W., Laird, J. B., Latham, D. W., & Aguilar, L. A. 1996, , 112, 668
Carpenter, J. M. 2001, , 121, 2851
Chiba, M., & Beers, T. C. 2000, , 119, 2843
Cruz, K. L., Reid, I. N., Liebert, J., Kirkpatrick, J. D., & Lowrance, P. J. 2003, AJ, 126, 2421
Cruz, K. L., et al. 2007, , 133, 439
Cushing, M. C., Rayner, J. T., & Vacca, W. D. 2005, , 623, 1115
Cushing, M. C., et al. 2006, , 648, 614
Dahn, C. C., Liebert, J., Harris, H. C., & Guetter, H. H. 1995 in The Bottom of the Main Sequence - and Beyond, ed. C. G. Tinney (Springer-Verlag: Heidelberg), p.239
Dahn, C. C., et al. 2002, , 124, 1170
D’Antona, F., & Mazzitelli, I. 1985, , 296, 502
Dauphole, B., & Colin, J. 1995, , 300, 117
Dehnen, W., & Binney, J. J. 1998, , 298, 387
Digby, A. P., Hambly, N. C., Cooke, J. A., Reid, I. N., & Cannon, R. D. 2003, , 344, 583
Epchtein, N., et al. 1997, The Messenger, 87, 27
Fazio, G. G., et al. 2004, , 154, 10
Fischer, J., et al. 2003, Proc. SPIE, 4841, 564
Gizis, J. E. 1997, , 113, 806
Golimowski, D. A., et al. 2004, , 127, 3516
Guetter, H. H., Vrba, F. J., Henden, A. A., & Luginbuhl, C. B. 2003, , 125, 3344
Hauschildt, P. H., Allard, F., & Baron, E. 1999, , 512, 377
Helling, Ch., Klein, R., Woitke, R., Nowak, U., & Sedlmayr, E. 2004, , 423, 657
Kerr, F. J., & Lynden-Bell, D. 1986, , 221, 1023
Kirkpatrick, J. D. 2005, , 43, 195
Kirkpatrick, J. D., Barman, T. S., Burgasser, A. J., McGovern, M. R., McLean, I. S., Tinney, C. G., & Lowrance, P. J. 2006, , 639, 1120
Kirkpatrick, J. D., et al. 1999, , 519, 802
Knapp, G., et al. 2004, , 127, 3553
Kuiper, G. P. 1939, , 89, 549
Laughlin, G., Bodenheimer, P., & Adams, F. C. 1997, , 482, 420
Leggett, S. K., Allard, F., Dahn, C., Hauschildt, P. H., Kerr, T. H., & Rayner, J. 2000, , 535, 965
Leggett, S. K., Saumon, D., Marley, M. S., Geballe, T. R., Golimowski, D. A., Stephens, D., & Fan, X. 2007, , 655, 1079
Lépine, S., Rich, R. M., Neill, J. D., Caulet, A., & Shara, M. M. 2002, , 581, L47
Lépine, S., Rich, R. M., & Shara, M. M. 2003, , 591, L49
Liebert, J., & Burgasser, A. J. 2007, , 655, 522
Linsky, J. L. 1969, , 156, 989
Lodders, K. 2002, , 577, 974
Luhman, K. L., Joergens, V., Lada, C., Muzerolle, J., Pascucci, I., & White, R. 2007, in Planets and Protostars V, eds. B. Reipurth, D. Jewitt and K. Keil (Univ. Arizona Press: Tucson), p. 443
Monet, D. G., Dahn, C. C., Vrba, F. J., Harris, H. C., Pier, J. R., Luginbuhl, C. B., & Ables, H. D. 1992, , 103, 638
Mould, J. R. 1976, , 207, 535
Noll, K. S., Geballe, T. R., Leggett, S. K., & Marley, M. S. 2000, , 541, L75
Norris, J. E. 1994, , 431, 645
Patten, B. M., et al. 2006, , 651, 502
Perryman, M. A. C., et al. 1997, , 323, L49
Reid, I. N., Kirkpatrick, J. D., Gizis, J. E., Dahn, C. C., Monet, D. G., Williams, R. J., Liebert, J., & Burgasser, A. J. 2000, , 119, 369
Reid, I. N., Kirkpatrick, J. D., Liebert, J., Gizis, J. E., Dahn, C. C., & Monet, D. G. 2002, , 124, 519
Reiners, A., & Basri, G. 2006, , 131, 1806
Saumon, D., Bergeron, P., Lunine, J. I., Hubbard, W. B., & Burrows, A. 1994, , 424, 333
Schmidt, S. J., Cruz, K. L., Bongiorno, B. J., Liebert, J., & Reid, I. N. 2007, , in press
Scholz, R.-D., Lodieu, N., Ibata, R., Bienaymé, O., Irwin, M., McCaughrean, M. J., & Schwope, A. 2004, , 347, 685
Sivarani, T., Kembhavi, A. K., & Gupchup, J. 2007, , in preparation
Skrutskie, M. F., et al. 2006, , 131, 1163
Smith, V. V., Tsuji, T., Hinkle, K. H., Cunha, K., Blum, R. D., Valenti, J. A., Ridgway, S. T., Joyce, R. R., & Bernath, P. 2003, , 599, L107
Sommer-Larsen, J., & Zhen, C. 1990, , 242, 10
Stassun, K., Mathieu, R. D., Vaz, L. P. R., Valenti, J. A., & Gomez, Y. 2006, Nature, 440, 311
Stauffer, J. R., Schultz, G., & Kirkpatrick, J. D. 1998, , 499, 199
Tarter, J. C., et al. 2007, AsBio, 7, 30
Tinney, C. G., Burgasser, A. J., & Kirkpatrick, J. D. 2003, , 126, 975
Tsuji, T., Ohnaka, K., & Aoki, W. 1996, , 305, L1
—. 1999, , 520, L119
Vrba, F. J., et al. 2004, , 127, 2948
York, D. G., et al. 2000, AJ, 120, 1579
[ll]{} $\pi_{rel}$ & 36.3$\pm$1.6 mas\
$\pi_{abs}$& 37.5$\pm$1.7 mas\
$\mu$ & 2.6241$\pm$0.0018$\arcsec$ yr$^{-1}$\
$\theta$ & 128.91$\degr\pm$0.02$\degr$\
$V_{\tan}$ & 332$\pm$15 [[km s$^{-1}$]{}]{}\
$(J-H)$ & 0.121$\pm$0.017 mag\
$(J-K)$ & 0.17$\pm$0.07 mag\
$K$ & 14.80$\pm$0.07 mag\
[lccl]{} 0.01–0.64 $\micron$ & [**0.031$\pm$0.003**]{} & 0.17 & NEXTGEN Model, 2MASS photometry\
& 0.096$\pm$0.009 & & Linear extrapolation, 2MASS photometry\
0.64–1.0 $\micron$ & [**1.24$\pm$0.11**]{} & 6.81 & Observed spectrum, 2MASS photometry\
1.0-2.9 $\micron$ & [**12.6$\pm$0.5**]{} & 69.45 & Observed spectrum, NEXTGEN model, 2MASS/USNO photometry\
& 12.4$\pm$0.4 & & Observed spectrum, Tucson model, 2MASS/USNO photometry\
& 12.7$\pm$0.4 & & Observed spectrum, spectral template, linear interpolation, 2MASS/USNO photometry\
2.9–15 $\micron$ & [**4.27$\pm$0.19**]{} & 23.52 & Spectral template, NEXTGEN model, IRAC photometry\
& 4.4$\pm$0.5 & & Spectral template, Tucson model, IRAC photometry\
& 4.39$\pm$0.12 & & Spectral template, linear interpolation, IRAC photometry\
15–1000 $\micron$ & [**0.0096$\pm$0.0009**]{} & 0.05 & NEXTGEN model, IRAC photometry\
& 0.077$\pm$0.007 & & Raleigh-Jeans tail, IRAC photometry\
0.1–1000 $\micron$ & [**18.2$\pm$1.7**]{} & 100.00 & Observations + NEXTGEN model\
& 18.0$\pm$1.8 & & Observations + Tucson model\
& 18.5$\pm$1.7 & & Observations + linear interpolation\
[lll]{} $\alpha$ & 05$^h$32$^m$53$\fs$46 & 1\
$\delta$ & +82$\degr$46$\arcmin$46$\farcs$5 & 1\
$\mu$ & 2.6241$\pm$0.0018$\arcsec$ yr$^{-1}$ & 2\
$\theta$ & 128.91$\degr\pm$0.02$\degr$ & 2\
Spectral Type & sdL7 & 3,4\
$d$ & 26.7$\pm$1.2 pc & 2\
$M-m$ & $-2.13{\pm}0.10$ mag & 2\
$(U,V,W)_{LSR}$ & $(-70,-354,78){\pm}(9,13,7)$ [[km s$^{-1}$]{}]{} & 2,5\
Kinematic Pop. & Halo & 2\
$\log{L_{bol}/L_{\sun}}$ & $-4.24{\pm}0.06$ & 1\
$M_{bol}$ & 15.35$\pm$0.16 mag & 2\
$M_{I_c}$ & 17.07$\pm$0.14 mag & 2,3\
$M_J$ & 13.05$\pm$0.12 mag & 1,2\
$M_H$ & 12.77$\pm$0.13 mag & 1,2\
$M_{K_s}$ & 12.79$\pm$0.18 mag & 1,2\
$M_K$ & 12.67$\pm$0.12 mag & 2\
$M_{[3.6]}$ & 11.24$\pm$0.10 mag & 2,6\
$M_{[4.5]}$ & 11.09$\pm$0.10 mag & 2,6\
$M_{[5.8]}$ & 11.10$\pm$0.14 mag & 2,6\
$M_{[8.0]}$ & 10.90$\pm$0.14 mag & 2,6\
$BC_J$ & 2.30$\pm$0.08 mag & 1,2\
$BC_K$ & 2.68$\pm$0.09 mag & 2\
[[T$_{eff}$]{}]{} & 1730$\pm$90 K & 2,7,8\
Mass & 0.0744$\pm$0.0009 M$_{\sun}$ & 2,7\
& 0.0783$\pm$0.0013 M$_{\sun}$ & 2,7\
& 0.0825$\pm$0.0008 M$_{\sun}$ & 2,7\
[^1]: A current list of known L and T dwarfs is maintained at <http://dwarfarchives.org>.
[^2]: Astrometric observations of [[2MASS J0532+8246]{}]{} were terminated only after a June 2006 cryogenic explosion seriously damaged ASTROCAM. When this instrument is restored to operational status (expected mid-2008), observations will resume for this source and other L and T dwarfs in the USNO infrared astrometric program.
[^3]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^4]: Late-type esdMs also have redder $B-V$ colors at a given $M_V$ magnitude due to similar metal opacity effects [@giz97].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In 1979, in the course of the proof of the irrationality of $\zeta(2)$ Robert Apéry introduced numbers $b_n = \sum_{k=0}^n {n \choose k}^2{n+k \choose k}$ that are, surprisingly, integral solutions of recursive relations $$(n+1)^2 u_{n+1} - (11n^2+11n+3)u_n-n^2u_{n-1} = 0.$$ Zagier performed a computer search on first 100 million triples $(A,B,C)\in {\mathbb{Z}}^3$ and found that the recursive relation generalizing $b_n$ $$(n+1)u_{n+1} - (An^2+An+B)u_n + C n ^2 u_{n-1}=0,$$ with the initial conditions $u_{-1}=0$ and $u_0=1$ has (non-degenerate i.e. $C(A^2-4C)\ne 0$) integral solution for only six more triples (whose solutions are so called sporadic sequences) .
Stienstra and Beukers showed that for the prime $p\ge 5$ $$b_{(p-1)/2} \equiv \begin{cases} 4a^2-2p \pmod{p} \textrm{ if } p = a^2+b^2,\textrm{ a odd}\\ 0 \pmod{p} \textrm{ if } p\equiv 3 \pmod{4}.\end{cases}$$
Recently, Osburn and Straub proved similar congruences for all but one of the six Zagier’s sporadic sequences (three cases were already known to be true by the work of Stienstra and Beukers) and conjectured the congruence for the sixth sequence (which is a solution of recursion determined by triple $(17,6,72)$.
In this paper we prove that remaining congruence by studying Atkin and Swinnerton-Dyer congruences between Fourier coefficients of certain cusp form for non-congurence subgroup.
address: 'Department of Mathematics, University of Zagreb, Bijenička cesta 30, 10000 Zagreb, Croatia'
author:
- Matija Kazalicki
bibliography:
- 'bibl.bib'
title: 'Congruences for sporadic sequences and modular forms for non-congruence subgroups'
---
Introduction
============
In 1979, in the course of his famous proof of the irrationality of $\zeta(3)$ and $\zeta(2)$ Robert Apéry [@Ap] introduced numbers $a_n = \sum_{k=0}^n {n \choose k}^2{n+k \choose k}^2$ and $b_n = \sum_{k=0}^n {n \choose k}^2{n+k \choose k}$. These numbers, which was important for the proof, are integral solutions of recursive relations $$(n+1)^3 u_{n+1}-(34n^3+51n^2+27n+5)u_n + n^3 u_{n-1}=0 \quad \textrm{ and }$$ $$(n+1)^2 u_{n+1} - (11n^2+11n+3)u_n-n^2u_{n-1} = 0$$ respectively. The integrality came as a big surprise since to calculate $a_n$ (or $b_n$) in each step one has to divide by $n^3$ (or $n^2$) so a priori one would expect that these numbers have denominators of the size $n!^3$ (or $n!^2)$. Inspired by Beukers [@B1], Zagier [@Zag1] performed a computer search on first 100 million triples $(A,B,C)\in {\mathbb{Z}}^3$ and found that the recursive relation generalizing $b_n$ $$(n+1)u_{n+1} - (An^2+An+B)u_n + C n ^2 u_{n-1}=0,$$ with the initial conditions $u_{-1}=0$ and $u_0=1$ has (non-degenerate i.e. $C(A^2-4C)\ne 0$) integral solution for only six more triples (whose solutions are so called sporadic sequences) $$(0,0,-16), (7,2,-8), (9,3,27), (10,3,9), (12,4,32) \textrm{ and } (17,6,72).$$
Interestingly, Stienstra and Beukers [@SB] showed that the generating function of Apéry’s numbers $b_n$ is a holomorphic solution of Picard-Fuchs differential equation of elliptic K3-surface $\mathcal{S}:X(Y-Z)(Z-X)-t(X-Y)YZ=0$ (other sporadic sequences are related in this way to K3 surfaces as well, see [@Zag1]). Using this connection they also proved that for prime $p\ge 5$ $$b_{(p-1)/2} \equiv \begin{cases} 4a^2-2p \pmod{p} \textrm{ if } p = a^2+b^2,\textrm{ a odd}\\ 0 \pmod{p} \textrm{ if } p\equiv 3 \pmod{4}.\end{cases}$$ Here one can interpret the right-hand side of the congruences as a $p$-th Fourier coefficient of a certain $CM$ modular form of weight 3 whose $L$-function is a factor of the zeta function of $\mathcal{S}$. (Later Beukers [@B1] proved a similar result for the numbers $a_n$ - this time relating them to the coefficients of Hecke eigenform of weight 4.) For a beautiful survey of these results see [@Zag2].
Recently, Osburn and Straub [@OS] proved similar congruences for all but one of the six Zagier’s sporadic sequences (three cases were already known to be true by the work of Stienstra and Beukers) and conjectured the congruence for the sixth sequence $F(n)$ (which is a solution of recursion determined by triple $(17,6,72)$. In this paper we prove that remaining congruence.
Denote by $$F(n) = \sum_{k=0}^n (-1)^k 8^{n-k}{n \choose k}\sum_{j=0}^k {k \choose j}^3,$$ the sporadic sequence corresponding to triple $(17,6,72)$. For $\tau \in \mathbb{H}$ and $q=e^{2\pi i \tau}$ let $$f(\tau)=\sum_{n=0}^\infty=q-2q^2+3q^3+\cdots=\sum_{n=0}^\infty \gamma(n)q^n \in S_3\left(\Gamma_0(24),\left( \frac{-6}{\cdot}\right)\right)$$ be a newform. Our main result is the following theorem.
\[thm:1\] For all primes $p>2$ we have $$F\left(\frac{p-1}{2}\right) \equiv \gamma(p) \pmod{p}.$$
One can check that $f(\tau)$ is CM form such that for prime $p$ $$\gamma(p) \equiv \begin{cases} 2(a^2-6b^2) \textrm{ if } p = a^2+6b^2\\ 0 \pmod{p} \textrm{ if } p \equiv 5, 11, 13, 17, 19, 23 \pmod{24}.\end{cases}$$
In Section \[sec:2\] using the method of Beukers [@B1 Proposition 3.] and Verrill [@V Theorem 1.1] we reduce Theorem \[thm:1\] to showing that the weight three cusp form (for non-congruence subgroup $\Gamma_2$ of $\Gamma_1(6)$) $$g(\tau) =q^{1/2} + \frac{3}{2} q^{3/2} - \frac{9}{8} q^{5/2} - \frac{85}{16} q^{7/2} - \frac{981}{128} q^{9/2}+\cdots \in S_3(\Gamma_2),$$ satisfies a three-term Atkin and Swinnerton-Dyer congruence relation with respect to $f(\tau)$ for all primes $p>3$ (see Proposition \[prop:2\]). The similar idea was used previously by the author [@K] in proving three term congruence relations for some multinomial sums by employing Atkin and Swinnerton-Dyer congruence relations satisfied by the Fourier coefficients of certain weakly holomorphic modular forms (but for congruence subgroups).
In Section \[sec:ASD\] we explain how using Scholl’s theory [@Scholl] we can reduce Proposition \[prop:2\] to the equivalence of two strictly compatible families of $\ell$-adic Galois representations: $\tilde{\rho}_\ell$ isomorphic to $\ell$-adic realization of the motive associated to the space of cusp forms $S_3(\Gamma_2)$, and $\rho'_\ell$ attached to the newform $f(\tau)\otimes \left( \frac{-1}{\cdot}\right)$ by Deligne’s work.
In Section \[sec:4\] and Section \[sec:Serre\] we prove that these two $\ell$-adic families are isomorphic by showing that they are isomorphic to the third $\ell$-adic family $\rho_\ell$ which is constructed from the explicit model of the universal family of elliptic curves over modular curve of $\Gamma_2$.
Elliptic surfaces, modular forms and the proof of Theorem \[thm:1\] {#sec:2}
===================================================================
Consider modular rational elliptic surface attached to $\Gamma_1(6)$ (see third example in [@V Section 4.2.2.]) $$\mathcal{W}: (x+y)(x+z)(y+z) - 8xyz =\frac{1}{t} xyz,$$ with fibration $\phi:\mathcal{W} \rightarrow P^1$, $(x,y,z,t) \mapsto t$. For $t \notin \{\infty, 0, -\frac{1}{9}, -\frac{1}{8} \}$ the preimage $\phi^{-1}(t)$ is an elliptic curve with a distinguished point of order $6$. Picard-Fuchs differential equation associated to this elliptic surface $$(8t + 1)(9t + 1)P(t)''+ t(144t + 17)P(t)' + 6t(12t + 1)P(t) = 0,$$ has a holomorphic solution around $t=0$ $$P(t) = \sum_{n=0}^\infty (-1)^n F(n)t^n.$$ (Our notation is slightly different from [@V Section 4.2.2.] since $F(n)=(-1)^n c_n$, with $c_n$ defined in [@V]) If we identify $t$ with a modular function (for $\Gamma_0(6)$) $$t(\tau) = \frac{\eta(2\tau)\eta(6\tau)^5}{\eta(\tau)^5\eta(3\tau)}, \quad \tau \in \mathbb{H}$$ then $P(\tau) := \sum_{n=0}^\infty (-1)^n F(n) t(\tau)^n$ is a weight one modular form for $\Gamma_1(6)$.
Now consider a two cover $\mathcal{S}$ of $\mathcal{W}$, a K3-surface given by the equation $$\mathcal{S}: (x+y)(x+z)(y+z) - 8xyz = \frac{1}{s^2}xyz,$$ where $t = s^2$. Then $s(\tau)=\sqrt{\frac{\eta(2\tau)\eta(6\tau)^5}{\eta(\tau)^5\eta(3\tau)}}$ is a corresponding modular function for index two genus zero subgroup $\Gamma_2 \subset \Gamma_1(6)$.
By identifying $s$-line with the modular curve $X(\Gamma_2)$, we can identify singular fibers of K3-surface $\mathcal{S}$ with cusps of modular curve $X(\Gamma_2)$. More precisely, using Tate’s algorithm one finds that Kodaira types of singular fibers at $s=\infty, 0, \pm \frac{i}{2\sqrt{2}}$ and $\pm\frac{i}{3}$ are $I_2, I_{12}, I_3, I_3, I_2$ and $I_2$ respectively. Hence all the cusps of $X(\Gamma_2)$ are regular.
In general, for a finite index subgroup $\Gamma$ of ${{\text {\rm SL}}}_2({\mathbb{Z}})$ of genus $g$ such that $-I\notin \Gamma$ and $k$ odd, [@Shi Theorem 2.25] gives the formula for the dimension of $S_k(\Gamma)$ $$\dim S_k(\Gamma) = (k-1)(g-1)+\frac{1}{2}(k-2)r_1+\frac{1}{2}(k-1)r_2+\sum_{i=1}^{j}\frac{e_i-1}{2e_i},$$ where $r_1$ is the number of regular cusps, $r_2$ is the number of irregular cusps, and $e_i's$ are the orders of elliptic points. Since $\Gamma_2$ has no elliptic points ($\Gamma_1(6)$ is a free group), we have that $\dim S_3(\Gamma_2)=1$.
Our starting point for studying congruences involving $F(n)$ is the following proposition of Beukers [@B1].
\[prop:Beukers\] Let $p$ be a prime and $$\omega(t)=\sum_{n=1}^\infty b_nt^{n-1}dt$$ a differential form with $b_n\in {\mathbb{Z}}_p$. Let $t(q)=\sum_{n=1}^\infty A_nq^n$,$A_n\in {\mathbb{Z}}_p$, and suppose $$\omega(t(q))=\sum_{n=1}^\infty c_n q^{n-1}dq.$$ Suppose there exist $\alpha_p,\beta_p\in {\mathbb{Z}}_p$ with $p|\beta_p$ such that $$b_{mp^r}-\alpha_p b_{mp^{r-1}}+\beta_pb_{mp^{r-2}}\equiv 0 \pmod{p^r}, \quad \forall m,r\in {\mathbb{N}}.$$ Then $$c_{mp^r}-\alpha_p c_{mp^{r-1}}+\beta_pc_{mp^{r-2}}\equiv 0 \pmod{p^r}, \quad \forall m,r\in {\mathbb{N}}.$$ Moreover, if $A_1$ is $p$-adic unit then the second congruence implies the first, and we have that $b_p \equiv \alpha_p b_1 \pmod{p}$.
Given prime $p>2$, if we apply the previous proposition to a differential form $$\omega(s) = \sum_{n=0}^\infty (-1)^nF(n)s^{2n}ds,$$ and $s(q)$ - the $q$-expansion of modular function $s(\tau)$ (where $q=e^{\pi i \tau}$), we obtain that $\omega(s(q))=\sum_{n=1}^\infty c_n q^{n-1}dq$, where $c_n$ are Fourier coefficients of weight $3$ cusp form $g(\tau)\in S_3(\Gamma_2)$ $$g(q) = P(q)q\frac{d}{dq}s(q)=q + \frac{3}{2} q^3 - \frac{9}{8} q^5 - \frac{85}{16} q^7 - \frac{981}{128} q^9+\cdots=\sum_{n=1}^\infty c_n q^n.$$
- For $p=2$ the Fourier coefficients of $s(q)$ are not $p$-integral so we can not use Proposition \[prop:Beukers\].
- It is well known that a differential operator $q\frac{d}{dq}$ maps modular functions to meromorphic modular forms of weight $2$. Holomorphicity and cuspidality of $g(\tau)$ then follow since zeros of $P(\tau)$ cancel out the poles of $s(\tau)$.
- Since Fourier coefficients of $g(\tau)$ have unbounded denominators, it follows that $\Gamma_2$ is non-congruence subgroup of $\Gamma_1(6)$ (for congruence subgroups the Hecke eigenforms (which form the basis for the space of cuspforms) have Fourier coefficients that are algebraic integers).
We will show that, for all primes $p>3$, the cusp form $g(\tau)$ satisfies a three term Atkin and Swinnerton-Dyer congruence relation with respect to the quadratic twist of the newform $f(\tau)=\sum_{n=1}^\infty \gamma(n) e^{2\pi i \tau}$ by quadratic character $\left(\frac{-1}{\cdot} \right)$. Hence Theorem \[thm:1\] follows from Proposition \[prop:Beukers\] and the following proposition.
\[prop:2\] Let $p>3$ be a prime. Then for all $m,r \in {\mathbb{N}}$, we have that $$c_{mp^r}-\left(\frac{-1}{p} \right)\gamma(p) c_{mp^{r-1}}+\left( \frac{-6}{p}\right)p^2 c_{mp^{r-2}}\equiv 0 \pmod{p^{2r}}.$$
Atkin and Swinnerton-Dyer congruences for $S_3(\Gamma_2)$ {#sec:ASD}
=========================================================
For a finite index non-congruence subgroup $\Gamma \subset {{\text {\rm SL}}}_2({\mathbb{Z}})$ and a prime $p$, we say that weight $k$ cusp form $f(\tau)=\sum_{n=0}^\infty a_n q^n \in S_k(\Gamma, \overline{{\mathbb{Z}}_p})$ satisfy Atkin and Swinnerton-Dyer (ASD) congruence at $p$ if there exist an algebraic integer $A_p$ and a root of unity $\mu_p$ such that for all non-negative integers $m$ and $r$ we have $$\label{eq:1}
a_{mp^r}-A_p a_{mp^{r-1}} + \mu_p p^{k-1}a_{mp^{r-2}}\equiv 0 \pmod{p^{(k-1)r}}.$$ (In our example ${a_{n}}'s$ and ${A_{p}}'s$ are rational integers, and $\mu_p=\pm 1$.)
In the absence of the useful theory of Hecke operators for non-congruence subgroups, such $f(\tau)$ can be regarded as Hecke eigenfunction at prime $p$. A discovery of these congruences by Atkin and Swinnerton-Dyer [@ASD] initiated a systematic study of modular forms for non-congruence subgroups. For more information see a survey article by Li, Long and Yang[@LL].
In the case when the space of cusp forms is one dimensional and generated by $f(\tau)$ (which is the case for $S_3(\Gamma_2)$ and $g(\tau)$), Scholl [@Scholl] proved that the ASD congruence holds for all but finitely many $p$. The congruences were obtained by embedding the module of cusp forms into certain de Rham cohomology group $DR(\Gamma,k)$ which is the de Rham realization of the motive associated to the relevant space of modular forms. At a good prime $p$, crystalline theory endows $DR(\Gamma,k)\otimes {\mathbb{Z}}_p$ with a Frobenius endomorphism whose action on $q$-expansion gives rise to Atkin and Swinnerton-Dyer congruences, i.e. if $T^2-A_p T+\mu_p p^2$ is a characteristic polynomial of Frobenius acting on $DR(\Gamma,k)\otimes {\mathbb{Z}}_p$ then congruence holds ($A_p$ is the trace of Frobenius). See [@KS Section 2] for the summary of these results.
To calculate the trace of Frobenius $A_p$, following Scholl [@Scholl Sections 4 and 5], we associate to the subgroup $\Gamma_2$ a strictly compatible family of $\ell$-adic Galois representations of ${\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}$, $\tilde{\rho}_\ell$, that is isomorphic to $\ell$-adic realization of the motive associated to the space of cusp forms $S_3(\Gamma_2)$. From [@Scholl2 2.7. Proposition] and algebraic relation between $s(\tau)$ and modular $j$-invariant $j(\tau)$ $$(s^2-\frac{1}{6})^3(s^6-\frac{7}{2}s^4+\frac{3}{4}s^2-\frac{1}{24})^3+\frac{1}{72}(s-\frac{1}{3})^2(s+\frac{1}{3})^2 s^{12} (s^2-\frac{1}{8})^3 j = 0,$$ it follows that $\tilde\rho_\ell$ is unramified outside $2, 3$ and $\ell$.
In particular, for $\ell=2$ and prime $p>3$ we have that [@Scholl Theorem 5.4.] $$\label{eq:2}
A_p=trace(\tilde\rho_2({\textrm{Frob}}_p)) \textrm{ and } \mu_p=\det(\tilde\rho_2(Frob_p)).$$
Compatible families of $\ell$-adic Galois representations of ${\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}$ {#sec:4}
===================================================================================================================
Denote by $\rho_\ell'$ a strictly compatible family of two dimensional $\ell$-adic Galois representation of ${\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}$ attached to the newform $f(\tau)\otimes \left( \frac{-1}{\cdot}\right)$ by the work of Deligne [@D]. Hence, $$\label{eq:3}
\textrm{trace}(\rho_\ell'(\textrm{{\textrm{Frob}}}_p))=\left( \frac{-1}{p}\right)\gamma(p) \textrm{ and } \det(\rho_\ell'({\textrm{Frob}}_p))=\left(\frac{-24}{p}\right)p^2,$$ for prime $p \ne 2,3$ and $\ell$.
We will prove that representations $\rho_\ell'$ and $\tilde{\rho_\ell}$ are isomorphic by showing that both of them are isomorphic to the representation $\rho_\ell$ which we define now. Proposition \[prop:2\] then follows from and .
Let $X(\Gamma_2)^{0}$ be the complement in $X(\Gamma_2)$ of the cusps. Denote by $i$ the inclusion of $X(\Gamma_2)^0$ into $X(\Gamma_2)$, and by $h':\mathcal{S} \rightarrow X(\Gamma_2)^0$ the restriction of elliptic surface $h:\mathcal{S}\rightarrow X(\Gamma_2)$ to $X(\Gamma_2)^0$. For a prime $\ell$ we obtain a sheaf $$\mathcal{F}_\ell = R^1h'_{*}{\mathbb{Q}}_\ell$$ on $X(\Gamma_2)^0$, and also sheaf $i_{*} \mathcal{F}_\ell$ on $X(\Gamma_2)$ (here ${\mathbb{Q}}_\ell$ is the constant sheaf on the elliptic surface $\mathcal{S}$, and $R^1$ is derived functor). The action of ${\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}$ on the ${\mathbb{Q}}_\ell$-vector space $$W = H_{et}^1(X(\Gamma_2)\otimes \overline{{\mathbb{Q}}}, i_* \mathcal{F}_\ell)$$ defines $\ell$-adic representation $\rho_\ell$. Representation is unramified outside $2,3$ and $\ell$. By the argument similar to [@LLY Proposition 5.1.], $\rho_\ell$ is isomorphic to $\tilde\rho_\ell$ up to a twist by quadratic character.
Using explicit equation for $\mathcal{S}$, we can calculate $trace(\rho_l({\textrm{Frob}}_p))$ and $$\det(\rho_l({\textrm{Frob}}_p))=\frac{1}{2}((trace(\rho_l({\textrm{Frob}}_p))^2-trace(\rho_l({\textrm{Frob}}_p^2)))=\frac{1}{2}((trace(\rho_l({\textrm{Frob}}_p))^2-trace(\rho_l({\textrm{Frob}}_{p^2})))$$ for $p \ne 2,3,\ell$ using the following theorem.
\[thm:trace\] Let $q=p^s$ be a power of prime $p\ne 2,3, \ell$. The following are true:
- We have that $${\textrm{trace}}({\textrm{Frob}}_q|W)=-\sum_{t\in X(\Gamma_j)(\mathbb{F}_q)} {\textrm{trace}}({\textrm{Frob}}_q|(i_*\mathcal{F}_\ell)_t).$$
- If the fiber $E_t := h^{-1}(t)$ is smooth, then $${\textrm{trace}}({\textrm{Frob}}_q|(i_*\mathcal{F}_\ell)_t)={\textrm{trace}}({\textrm{Frob}}_q|H^1(E_t,
{\mathbb{Q}}_\ell))=q+1-\#E_t(\mathbb{F}_q).$$
- If the fiber $E^j_t$ is singular, then $${\textrm{trace}}({\textrm{Frob}}_q|(i_*\mathcal{F}_\ell)_t)=
\begin{cases}
1 & \text{if the fiber is split multiplicative}, \\
-1 & \text{if the fiber is nonsplit multiplicative},\\
0 & \text{if the fiber is additive}.
\end{cases}$$
Serre-Faltings method and proof of Proposition \[prop:2\] {#sec:Serre}
=========================================================
We will prove the following proposition.
\[prop:3\] For every prime $\ell$ the representations $\rho_\ell$ and $\rho_\ell'$ are isomorphic.
Since the families are strictly compatible, by Chebotarev density theorem it is enough to prove that $\rho_2$ and $\rho_2'$ are isomorphic. We apply the method of Serre and Faltings as formulated in [@Scholl2 Section 5].
\[thm:3\] For a finite set of primes $S$ of ${\mathbb{Q}}$, let $\chi_1, \ldots, \chi_r$ be a maximal independent set of quadratic characters of ${\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}$ unramified outside $S$, and $G$ a subset of ${\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}$ such that the map $(\chi_1, \ldots, \chi_r):G \rightarrow ({\mathbb{Z}}/2{\mathbb{Z}})^r$ is surjective.
Let $\sigma, \sigma':{\textrm{Gal}(\overline{{\mathbb{Q}}}/{\mathbb{Q}})}\rightarrow {\mathbb{GL}}_2({\mathbb{Q}}_2)$ be continuous semisimple representation unramified away from $S$, whose images are pro-2-groups. If for every $g\in G$ $${\textrm{trace}}(\sigma(g))={\textrm{trace}}(\sigma'(g)) \textrm{ and } \det(\sigma(g))=\det(\sigma'(g)),$$ then $\sigma$ and $\sigma'$ are isomorphic.
Images of representations $\rho_2$ and $\rho_2'$ are pro-2-groups.
We can assume that the images of both representations are contained in ${\mathbb{GL}}_2({\mathbb{Z}}_2)$. It is enough to prove that the images of their mod $2$ reductions have order $2$ (since the kernel of the natural homomorphism ${\mathbb{GL}}_2({\mathbb{Z}}/2^{k+1}{\mathbb{Z}}) \rightarrow {\mathbb{GL}}_2({\mathbb{Z}}/2^{k}{\mathbb{Z}})$ is a $2$-group). For primes $p\in \{7,11,13\}$ using Theorem \[thm:trace\] and an explicit model for surface $\mathcal{S}$, we compute that $$\textrm{trace}(\rho_2(\textrm{{\textrm{Frob}}}_p))=\left( \frac{-1}{p}\right)\gamma(p) \textrm{ and } \det(\rho_2({\textrm{Frob}}_p))=\left(\frac{-24}{p}\right)p^2.$$
Moreover, if $\left(\frac{-6}{p}\right)=-1$, we find that $\gamma(p)=0$ and the eigenvalues of $\rho_2({\textrm{Frob}}_p)$ are $\pm p\sqrt{-1}$ from which it follows that mod $2$ reduction of $\rho_2({\textrm{Frob}}_p)$ has order $2$. If $\left(\frac{-6}{p}\right)=1$, then the eigenvalues mod $2$ are equal, and mod $2$ reduction of $\rho_2({\textrm{Frob}}_p)$ is trivial.
Since the group ${\mathbb{GL}}_2({\mathbb{Z}}/2{\mathbb{Z}})$ is isomorphic to the symmetric group $S_3$, if we assume that the mod $2$ image is not of order two, then it must be the whole group. In that case, denote by $L$ a $S_3$ Galois extension of ${\mathbb{Q}}$ cut out by mod $2$ reduction of $\rho_2$ (i.e. $L$ is the fixed field of the kernel of the mod $2$ reduction of $\rho_2$). Then $L$ contains a unique quadratic field $K$ which is unramified outside $2$ and $3$ in which $7$ and $11$ split and $13$ is inert. It follows that $K={\mathbb{Q}}(\sqrt{-6})$. We know by the Hermite-Minkowski theorem that there are finitely many $S_3$ extensions of ${\mathbb{Q}}$ unramified outside $2$ and $3$, and using LMFDB [@lmfdb] we find that there is only one such field ${\mathbb{Q}}(x)$, where $x^6-3x^2+6=0$, whose Galois group contains ${\mathbb{Q}}(\sqrt{-6})$. This field contains a cubic field $F = {\mathbb{Q}}(s)$, where $s^3+3s-2=0$. One finds that $7$ is inert in $F$, hence $\rho_2({\textrm{Frob}}_7)$ has order $3$. This is impossible since ${\textrm{trace}}(\rho_2({\textrm{Frob}}_7))=10$ is an even number which implies that mod $2$ reduction of $\rho_2({\textrm{Frob}}_7)$ has order $1$ or $2$.
To apply Theorem \[thm:3\] for $S=\{2,3\}$ we choose characters $$\chi_1=\left( \frac{-1}{p}\right), \chi_2=\left( \frac{2}{p}\right), \chi_3=\left( \frac{3}{p}\right),$$ and $G=\{{\textrm{Frob}}_p:31 \le p \le 73, \textrm{ for } p \textrm{ prime}\}$. Using Theorem \[thm:trace\] and we can check that $${\textrm{trace}}(\rho_2(g))={\textrm{trace}}(\rho'_2(g)) \textrm{ and } \det(\rho_2(g))=\det(\rho'_2(g)),$$ for all $g \in G$, hence Proposition \[prop:3\] follows.
To prove Proposition \[prop:2\] (and consequently Theorem \[thm:1\]), we need to show that representations $\rho_\ell$ and $\tilde{\rho_\ell}$ are isomorphic. In particular, it is enough to prove this claim for $\ell=2$. By the argument similar to [@LLY Proposition 5.1.], it follows that $\rho_2$ is isomorphic to $\tilde\rho_2$ up to a twist by a quadratic character. Since both representations are unramified outside $2$ and $3$, this character is an element of the group generated by characters $\chi_1$, $\chi_2$ and $\chi_3$. For every nontrivial $\chi$ from that group, we can find a prime $p>3$ such that $\chi(p)=-1$, and numerically check that ASD congruence relation for the Fourier coefficients of $g(\tau)$ $$c_{mp^r}-\chi(p)\left(\frac{-1}{p} \right)\gamma(p) c_{mp^{r-1}}+\left( \frac{-6}{p}\right)p^2 c_{mp^{r-2}}\equiv 0 \pmod{p^{2r}},$$ does not hold for some choice of $m$ and $r$. The claim follows.
All the computations in this paper were done in SageMath [@sagemath] and Magma [@magma].
Future work
===========
It is natural to ask do similar mod $p$ congruences exist for the numbers $F(\frac{p-1}{n})$, where $n>2$ and $p \equiv 1 \pmod{n}$? E.g. when $n=3$, by considering the $3$-cover (defined by $t=s^3$) of the elliptic surface $\mathcal{W}$, one can show that for $p\equiv 1 \pmod{3}$ we have $F\left(\frac{p-1}{3}\right) \equiv A_p \pmod{p}$, where $A_p$ is the trace of ${\textrm{Frob}}_p$ under the Galois representation defined analogously to $\rho_\ell$ (in this situation the representation is four-dimensional).
In the paper under the preparation, we are going to investigate this phenomena for sequence $F$ and other Apéry numbers.
Acknowledgments
===============
The author would like to thank Robert Osburn and Armin Straub for bringing this problem to his attention.
The author was supported by the QuantiXLie Centre of Excellence, a project co-financed by the Croatian Government and European Union through the European Regional Development Fund - the Competitiveness and Cohesion Operational Programme (Grant KK.01.1.1.01.0004), and by the Croatian Science Foundation under the project no. IP-2018-01-1313.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We examine phases of the Shraiman-Siggia model of lightly-doped, square lattice quantum antiferromagnets in a self-consistent, two-loop, interacting magnon analysis. We find magnetically-ordered and quantum-disordered phases both with and without incommensurate spin correlations. The quantum disordered phases have a pseudo-gap in the spin excitation spectrum. The quantum transition between the magnetically ordered and commensurate quantum-disordered phases is argued to have the dynamic critical exponent $z=1$ and the same leading critical behavior as the disordering transition in the pure $O(3)$ sigma model. The relationship to experiments on the doped cuprates is discussed.'
address: |
Departments of Physics and Applied Physics, P.O. Box 6666,\
Yale University, New Haven, CT 06511
author:
- Subir Sachdev
title: 'Quantum phases of the Shraiman-Siggia model'
---
INTRODUCTION {#intro}
============
The last few years have seen numerous experiments examining the magnetic properties of the doped copper-oxide compounds in some detail. However, our theoretical understanding has not kept in step, partly due to the numerous competing effects and ensuing complexity of these materials: at low temperatures the effects of disorder are paramount, and at larger doping there is the onset of superconductivity.
On the theoretical side, a popular model for investigating the interplay of doping and antiferromagnetic spin correlations has been the $t$-$J$ model. Numerical and high temperature series studies of this model have been especially valuable as a testing ground for various theoretical ideas. In an important advance, Shraiman and Siggia [@boris] (SS) proposed a phenomenological description of the long-wavelength interplay between spin and charge transport in this model. There are numerous theoretical reasons for believing that their long-wavelength model is a correct description of the $t$-$J$ model at low temperatures. It is expected that the SS model is quite robust, especially at low doping concentrations, and will describe a whole class of doped antiferromagnets, not just the $t$-$J$ model. However, at sufficiently large doping, an approach based upon long-wavelength distortions of the background antiferromagnetic order must eventually become invalid; the mechanism of this break-down of the SS model is not understood and remains an important open problem. This paper shall examine the SS model in a new approach which is designed to explicate the nature of the quantum-disordered phases [*i.e.*]{} the phases with no long-range magnetic order. These phases appear at a reasonably small doping concentration, where it is expected, though not certain, that the mapping between the SS and $t$-$J$ models is still valid. We will also discuss the relationship of our results to other theoretical work and some experiments in Section \[conc\].
We begin by writing down a simplified version of the action $S$ of the Shraiman-Siggia model: $$S = \int_0^{\beta \hbar} d \tau \int d^2 r \left[ S_n + S_f + S_c \right],$$ with $\tau$ the Matsubara time, $r = (x,y)$ the spatial co-ordinates, $\beta = 1/(k_B T)$, and $T$ the absolute temperature, The first part, $S_n$ describes fluctuations of the antiferromagnetic order parameter $n_{\ell}$. Here $n_{\ell}$ is a 3-component vector and is taken to have unit length $$\sum_{\ell=1}^{3} n_{\ell}^2 = 1.
\label{const}$$ We then have $$S_n = \frac{\rho_s^0}{2\hbar} \left( (\partial_x n_{\ell} )^2 +
(\partial_y n_{\ell} )^2 + \frac{1}{c_0^2}
\left(\frac{\partial n_{\ell}}{\partial \tau} \right)^2 \right)$$ with $\rho_s^0$ the bare spin stiffness and $c_0$ the bare spin-wave velocity. The momentum of the $n_{\ell}$ field is restricted to be smaller than an ultraviolet cutoff $\Lambda$, which is also the scale at which the coupling constant are defined.
The action of the fermionic dopant holes is given by $S_f$. The holes are describes by fermionic spinor fields $\Phi_{v\alpha}$ where $\alpha=\uparrow,
\downarrow$ is the spin index, and $v$ is ‘valley’ index. The valleys are regions in the Brillouin zone around the minima of the fermion dispersion spectrum. For the square-lattice $t$-$J$ model there is a great deal of evidence that there are two vallleys centered at the points $(\pi/2 , \pm \pi/2)$ in the reduced Brillouin zone [@veit; @variational]. We will rotate our co-ordinate system from the conventional one, so that our $x$-axis is at an angle of $45$ degrees to the axes of the square lattice, and the principal axes of the valleys are along the new $x$ and $y$ axes; the index $v$ will therefore take the values $v=x,y$. Further, we will measure momenta from the center of the valleys. With these conventions, the fermionic action $S_f$ is $$\begin{aligned}
S_f = &&
\Phi_{x\alpha}^{\dagger}
\left( \frac{\partial}{\partial\tau} - \frac{\hbar^2}{2m_l} \frac{\partial^2}{
\partial x^2} - \frac{\hbar^2}{2 m_h} \frac{\partial^2}{\partial y^2}
\right) \Phi_{x\alpha} \nonumber \\
&&~~~~ + \Phi_{y\alpha}^{\dagger}
\left( \frac{\partial}{\partial\tau} - \frac{\hbar^2}{2m_h} \frac{\partial^2}{
\partial x^2} - \frac{\hbar^2}{2 m_l} \frac{\partial^2}{\partial y^2}
\right) \Phi_{y\alpha},\end{aligned}$$ where $m_l$, $m_h$ are the light and heavy masses of the hole.
It is important to realize that there is no simple relationship between the bare fermionic field $c_{i\alpha}^{\dagger}$ of the $t$-$J$ model ($i$ is the site index) and the continuum fields $(n_{\ell}, \Phi_{v \alpha} )$ of the hydrodynamic SS model. Crudely speaking, one may consider this field transmutation as a form of spin-charge separation in which the spin-1/2, charge $e$ $c_{i\alpha}^{\dagger}$ fermionic field has separated into the spin-1, charge 0, bosonic $n_{\ell}$ quanta and the spin-1/2, charge $e$, $\Phi_{v\alpha}$ fermions.
Finally, we will consider only the leading term in the coupling, $S_c$ between the holes and the $n_{\ell}$ field, which is responsible for inducing local incommensurate spin correlations. It is expected that the remaining terms in the SS model are innocuous and do not change the results of this paper qualitatively. We have $$S_c = \kappa \left( \Phi_{v\alpha}^{\dagger}
\sigma^{\ell}_{\alpha\beta} \Phi_{v\beta} \right)
\left(\epsilon_{\ell m p} n_{m} \partial_v n_{p} \right)
\label{sc}$$ where $\sigma^{\ell}$ are three Pauli matrices, and $\kappa$ is a coupling constant. A crucial feature of the SS model is that there is no three-body coupling like $n_{\ell} \Phi_{v\alpha}^{\dagger}
\sigma^{\ell}_{\alpha\beta} \Phi_{v\beta}$ between the fermions and $n_{\ell}$ quanta [@comment]: such a term is forbidden by a sublattice interchange symmetry of the SS model [@boris] under which $n_{\ell}$ change sign while the $\Phi_{v\alpha}$ remain invariant. Other models of nearly antiferromagnetic fermi liquids [@hertz; @Andy] posses such a term and as a result, have quantum disordered phases with rather different properties.
We wish to clarify a crucial point about our particular form of $S$ at the outset. SS have shown that there is a mapping, in principle exact, to an alternative form of $S$ in which the degrees of freedom are a spin-1/2, charge 0, complex scalar $z_{\alpha}$ with $\alpha = \uparrow,\downarrow$ and [*spinless*]{}, charge $e$ fermions $\Psi_{va}$ where $a=A,B$ is a sublattice index. The $(n_{\ell}, \Phi_{v\alpha})$ fields are related to the $(z_{\alpha}, \Psi_{va})$ fields by $$n_{\ell} = z_{\alpha}^{\ast} \sigma^{\ell}_{\alpha\beta} z_{\beta}
~~;~~ \left( \begin{array}{c} \Phi_{v\uparrow} \\ \Phi_{v\downarrow}
\end{array} \right) = \left( \begin{array}{cc} z_{\uparrow}, &
-z_{\downarrow}^{\ast} \\ z_{\downarrow} & z_{\uparrow}^{\ast} \end{array}
\right) \left( \begin{array}{c} \Psi_{vA} \\ \Psi_{vB} \end{array} \right)$$ Note that there is a $U(1)$ gauge transformation on the $(z_{\alpha}, \Psi_{va})$ fields which leaves the $(n_{\ell}, \Phi_{v\alpha})$ fields invariant: $$z_{\alpha} \rightarrow z_{\alpha} e^{i \phi} ~~~;~~~\Psi_{vA}
\rightarrow \Psi_{vA} e^{-i\phi}~~~;~~~\Psi_{vB} e^{i \phi},$$ where $\phi$ has an arbitrary dependence on spacetime. The action in terms of the $(z_{\alpha}, \Psi_{va})$ will therefore also have to be invariant under the gauge transformation; there is no such restriction on the action in the $(n_{\ell}, \Phi_{v\alpha})$ variables. This latter absence of gauge restrictions, and associated long-range gauge forces will be quite useful to us in our analysis.
SS go on to state that equivalent results are obtained in computations using either choice of fields. In principle, this statement is correct. In practice, however, one is usually restricted in the analysis to perturbation theory, in which the quantum numbers of the low-lying excitations are essentially identical to those of the degrees of freedom in the action. Equivalent results in both theories have been obtained in the vicinity of magnetically ordered states, where one is performing a small fluctuation, spin-wave analysis. On the other hand, the choice of fields is expected to have dramatic consequences in a quantum-disordered state. Kane [*et. al.*]{} and others [@kane] made a choice of fields equivalent to the $(z_{\alpha} , \Psi_{va})$ formulation of the SS model. Thus, not surprisingly, their Schwinger-boson mean-field theory yielded a quantum disordered phase with deconfined, spin-1/2, charge 0, bosonic spinons (the $z_{\alpha}$) and spinless, charge $e$ fermions ($\Psi_{va}$).
In this paper, we use another approach and shall examine the phases of the SS model which are obtained naturally in the $(n_{\ell} , \Phi_{v\alpha})$ formulation. Some of the results of our calculations were noted some time ago [@jinwu]. While, in the end, we have no formal justification for claiming that the $(n_{\ell}, \Phi_{v\alpha})$ formulation is more accurate than the $(z_{\alpha}, \Psi_{va})$ approach, we can offer the following motivations. The $n_{\ell}$ field formulation has been quite successful in describing the undoped, frustrated antiferromagnet. The long wavelength action of the undoped antiferromagnet [@CHN] is simply the $O(3)$ sigma model, $S_n$, and it is expected to display a quantum-disordered phase in which massive $n$ quanta form the lowest excitations and carry spin 1 [@polyakov]. Consistent with this, recent investigations of the quantum-critical behavior in these systems have argued for the superiority of the $n$-field based approach, and have successfully explained a number of experimental and numerical computations on the square lattice antiferromagnet [@jinwu; @CSY]. Further, careful analysis of fluctuations [@Sach-Read1] in the $z_{\alpha}$-based theories of the undoped antiferromagnet has yielded quantum disordered phases in which the quantum numbers of the low-lying excitations are identical to those $n$-field, which means that fluctuations in fact confine bosonic spinons into $S=1$ particles. It is then natural to explore the consequences of doping in a model in which the correct physics in the limit of zero doping is captured most directly [*i.e.*]{} in the $(n_{\ell}, \Phi_{v\alpha})$ approach. We shall argue later in this paper, that the results of such an investigation are consistent with the available numerical and experimental data on doped antiferromagnets.
Summary of Results
------------------
We now discuss the main results of our calculations. We will distinguish the quantum phases by the properties of the equal-time $n_{\ell}$ correlator $$\tilde{S} (q) = \langle | n_{\ell} ( q) |^2 \rangle$$ as $T \rightarrow 0$, as a function of the momentum $q$. We emphasize that $\tilde{S} (q)$ is [*not*]{} the full structure factor $S(q)$ measured in neutron scattering experiments; $S(q)$ will contain additional terms involving the contribution of the spin-1/2 $\Phi_{v\alpha}$ fields. Phases with magnetic long-range order have a delta function term in $\tilde{S} (q)$ at $T=0$; this delta function is at $q=0$ in the commensurate long-range-ordered phase (hereafter referred to as CLRO) and at $q\neq 0$ for the case of incommensurate long-range-order (ILRO) ($q$ is measured from ($\pi,\pi$)). The quantum disordered phases have no delta-function terms at $T=0$, but only a peak of finite, though possibly small, width. This peak is at $q=0$ in the commensurate quantum-disordered phase (CQD) and at $q\neq 0$ in the incommensurate case (IQD). In agreement with the SS analysis [@boris] and experimental results [@aeppli], the peak in the ILRO and IQD phases was found to occur along the conventional $(1,0)$ and $(0,1)$ axes of the square lattice (these are the $(1,1)$ and $(1,-1)$ axes in our rotated co-ordinate system).
We will have little to add here to existing studies [@boris; @frenkel; @borislast; @gorkov] of the properties of the magnetically-ordered phases (CLRO and IRLO): their low-lying excitations are spin-waves involving long-wavelength deformations of the ordered state. Our focus will mainly be upon the new quantum-disordered phases (CQD and IQD) and their unusual properties. The $n_{\ell}$ quanta in both phases were found to be fully gapped. The low-lying excitations in the $n_{\ell}$ sector consist of a triply-degenerate spin-1 particle with a finite energy. However, the spin-1/2 $\Phi_{v\alpha}$ particles continue to form a Fermi sea which possesses gapless fermionic excitations with charge $e$ and spin-1/2. Despite the presence of these gapless excitations, the $n_{\ell}$ gap is robust as there is no term in the SS model which permits the decay of a $n_{\ell}$ quantum to a fermion particle-hole pair. The importance of the absence of the three-body term noted above, is now evident. Taken as a whole, the model thus only has a pseudo-gap to spin excitations in the CQD and IQD phases. One of the consequences of the presence of the gapless spin-1/2 fermions is that the uniform spin susceptibility of the CQD and IQD phases will be finite at $T=0$ due to the Pauli contribution [@comment]. We also note, that our calculation has completely neglected the effect of Berry phases; in the context of undoped antiferromagnets it has been argued [@Sach-Read1] that Berry phases should induce spin-Peierls long-range order in the CQD phases. It is possible that such spin-Peierls order will also exist in the CQD phase of the doped antiferromagnet.
We undertook a partial numerical survey of the phase diagram of the SS model as a function of $\rho_s^0$, $\kappa$ and the hole density. Parameters were always chosen so that the zero doping state was CLRO. This CLRO state was always found to be stable over a small, but finite, doping concentration. Over some of the regime examined, the sequence of phases with increasing doping was CLRO - CQD - IQD. We studied the $T=0$ quantum transition between the CLRO and CQD phases and will present evidence indicating that it has dynamic critical exponent $z=1$ and the same leading critical behavior as the transition in the pure $O(3)$ sigma model; however, the corrections to scaling in the two models were found to be quite different. The boundary between the CQD and IQD phases is an example of a disorder line [@disorder]: our calculation only found a non-analyticity in the dependence of the structure factor on the bare coupling constants at the disorder line, but no strong long-wavelength fluctuations.
In a region of the phase diagram with $\kappa$ large, we found the sequence CLRO - ILRO with increasing doping. In principle, there should eventually be a ILRO to IQD transition, but for the parameters examined, we did not find one before a doping level where the incommensuration wavevector was almost as large as the momentum upper cutoff.
We will argue from the above numerical results, and from theoretical considerations, that there is a Lifshitz point in the $\rho_s^0$, $\kappa$ plane where all the four phases - CLRO, ILRO, CQD, IQD - meet. Some properties of this multicritical point will be discussed.
We also have extensive results on the temperature dependence of equal-time correlation functions in the various phases. In particular, the temperature dependences in the spin correlation length and the structure factor are quite instructive, and will be described later.
CALCULATIONS {#calcs}
============
For the case of the undoped antiferromagnet, the $1/N$ expansion on the $O(N)$ non-linear sigma model offers a convenient and accurate method for exploring properties in the vicinity of $T=0$ quantum transitions [@CSY]. The extension of the $1/N$ expansion to the doped antiferromagnet is however not straightforward because of the presence of the third-rank $\epsilon_{\ell m p}$ tensor in $S_c$ (Eqn. (\[sc\])), which is special to the case $N=3$. Even with this complication, it is still possible to justify perturbative 1/N calculations, although in a rather inelegant way: after the fermions have been integrated out, the coupling $\kappa$ in the effective action of the $n$ field has be to scaled by $1/N^{\mu}$ where $0 < \mu < 1/2$. Not much is learned from this extension to general $N$, and we will therefore spare the reader the details. We will be satisfied, instead, in restricting our discussion to the special case of $N=3$, and viewing our $1/N$ calculation as a physically motivated, self-consistent, interacting magnon approximation. The magnon-magnon interactions are computed in a manner which is directly inspired by the $1/N$ expansion of the undoped antiferromagnet [@CSY].
An important property of our approach is that spin-rotation invariance is explicitly preserved at all stages of the calculation. This is crucial for a proper study of the quantum disordered phases of the model, especially when the $n$ quanta acquire a gap. We thus expect our approximations to work best in the quantum-disordered phase, at the $T=0$ quantum transition, and in the intermediate temperature quantum-critical region [@CHN; @CSY]. At the same time, the low temperature properties in a region with magnetic long-range order (the renormalized classical region [@CHN]) in the ground state, may not be well described. Even in the undoped antiferromagnet, the $1/N$ expansion is singular in the renormalized-classical region, and a careful interpretation of the results is required [@CSY]. We therefore will focus below here will be mainly on the quantum-disordered phases phases.
We begin by expressing the action in suitable dimensionless parameters. We rescale lengths such that the upper-cutoff in momentum space for the $n$ field is 1; thus $$r \rightarrow \frac{r}{\Lambda}$$ Similarly, the times $\tau$ are rescaled so that the bare spin-wave velocity of the $n$ field is unity: $$\tau \rightarrow \frac{\tau}{c_0 \Lambda}$$ After rescaling the temperature $$T \rightarrow T \frac{\hbar c_0 \Lambda }{k_B}$$ we then have the following modified form of $S_n$ $$S_n = \frac{1}{2g} \left( (\partial_{x} n_{\ell} )^2 +
(\partial_{y} n_{\ell} )^2 + (\partial_{\tau} n_{\ell})^2 \right)$$ where the dimensionless coupling constant $g$ is given by $$g = \frac{\hbar c_0 \Lambda}{\rho_s^0}$$ The fermionic action will retain its form after rescaling the field $\Phi \rightarrow \Lambda \Phi$ and rescaling to the dimensionless effective masses $$m_{h,l} \rightarrow m_{h,l} \frac{\Lambda \hbar}{c_0}$$ Finally in $S_c$ we replace $\kappa \rightarrow \kappa c_0$.
We will impose the constraint (\[const\]) by a Lagrange multiplier field $\lambda$. Thus we need to evaluate the functional integral $$\begin{aligned}
Z = \int {\cal D} n_{\ell} {\cal D} \Phi_{v\alpha} && {\cal D} \lambda
\exp \Biggl(
- S \nonumber \\
&& - i \left. \int_0^{1/T} d\tau \int d^2 x \frac{\lambda}{2g} ( n_{\ell}^2 - 1)
\right)\end{aligned}$$
It is now possible to set up a rotationally-invariant, diagrammatic expansion of all observables associated with $Z$. We will work at finite $T$, and so no breaking of spin rotation invariance can occur; the properties of the ground state will be elucidated by taking the $T \rightarrow 0$ limit. We treat the $\lambda$ field in much the same way as in the undoped system [@polyakov]. We assume that fluctuations of $i\lambda$ occur about a saddle point value $\overline{\lambda}$; we therefore write $$i \lambda = \overline{\lambda} + i \tilde{\lambda}$$ where $\tilde{\lambda}$ is the fluctuating part of $\lambda$. The value of $\overline{\lambda}$ is to be determined at the end of the calculation to satisfy the constraint (\[const\]). The diagrammatic expansion now has three bare propagators: the conventional Green’s function of the fermions $\Phi_{va}$, the propagator, $G^0$, of the $n_{\ell}$ field $$G^0 ( q, i\omega_n ) = \frac{1}{q^2 + \omega_n^2 + \overline{\lambda}},
\label{g0}$$ ($q$ is the wavevector and $\omega_n$ is a Matsubara frequency) and the propagator $1/\Pi$ of $\tilde{\lambda}$, with [@polyakov] $$\Pi ( q, i\omega_n) =
T \sum_{\epsilon_n} \int \frac{d^2 k }{4 \pi^2}
G^0 ( k+q, i\epsilon_n+i\omega_n ) G^0 ( k , i\epsilon_n ).
\label{Pi}$$ There are two interaction vertices: the four-body $(n-n-\Phi-\Phi)$ coupling in $S_c$ and a three-body $(\tilde{\lambda}-n-n)$ vertex with the value $i/(2 g)$. Finally, there is a rule to prevent over-counting: no $\tilde{\lambda}$ propagator can be followed by a bubble consisting just of two $G^0$ propagators.
We may now write the fully-renormalized correlator of the $n$ field in the form $$G ( q, i\omega_n ) = \frac{1}{q^2 + \omega_n^2 + m^2 +
\Sigma ( q, i\omega_n ) - \Sigma (0, 0)}
\label{Gfunc}$$ where $\Sigma$ is the self energy and the ‘mass’ $m$ is given by $$m^2 = \overline{\lambda} + \Sigma (0, 0)$$ The lowest order contributions to $\Sigma$ from magnon-magnon ($\Sigma_n$) and magnon-fermion ($\Sigma_f$) interactions are shown in Fig \[feyn\], and their values are $$\begin{aligned}
\Sigma &=& \Sigma_n + \Sigma_f \nonumber \\
\Sigma_n ( q, i\omega_n ) &=& \frac{2}{3} T \sum_{\epsilon_n}
\int \frac{d^2 k }{4 \pi^2}
\frac{G^0 ( k+q , i\epsilon_n + i\omega_n )}{\Pi (k , i\epsilon_n )} \nonumber \\
\Sigma_f ( q, i\omega_n ) &=& -4 g^2 \kappa^2 T \sum_{\epsilon_n}
\int \frac{d^2 k }{4 \pi^2} (2 q_v + k_v )^2 \chi_v ( k , i\epsilon_n )
G^0 ( k+q , i\epsilon_n + i\omega_n )
\label{Sigma}\end{aligned}$$ where $\chi_v$ is polarization of the fermion in valley $v$. We used the following expression, appropriate for an elliptical Fermi surface: $$\chi_x ( q, i\epsilon_n ) = \chi_0 (T) \left( 1 - \frac{|\epsilon_n |}{
\left[ \epsilon_n^2 + v_F^2 \left( q_x^2 (m_h/m_l)^{1/2}
+ q_y^2 (m_l / m_h )^{1/2} \right) \right]^{1/2}} \right)$$ and similarly for $\chi_y$. The Fermi velocity $v_F$, the Fermi wavevector $k_F$ and the doping concentration $\delta$ are related by the equations $$v_F = \frac{k_F}{\sqrt{m_l m_h}}~~~~~~~;~~~~~~~~\delta = \frac{k_F^2}{\pi}.$$ and the polarization is taken to vanish unless $$\left( q_x^2 (m_h/m_l)^{1/2}
+ q_y^2 (m_l / m_h )^{1/2}\right)^{1/2} < 2 k_F$$ The prefactor $\chi_0$ we chose as the compressibility of a free Fermi gas at a temperature $T$: $$\chi_0 (T) = \frac{\sqrt{m_l m_h}}{2 \pi} \left( 1 - e^{-k_F v_F /
(2 T)} \right)
\label{chival}$$
Our approach consisted of using the above approximation for $\Sigma$ and then solving the constraint equation (\[const\]), or $$3 g T \sum_{\omega_n} \int \frac{d^2 q}{4 \pi^2} G (q, i\omega_n) = 1
\label{constg}$$ for the value of $m^2$ (or equivalently $\overline{\lambda}$). The dependence of $\Sigma$ on $G^{0}$ was made partially self consistent by replacing $\overline{\lambda}$ by $m^2$ in (\[g0\]), thus using $G^0 (q, i\omega_n )= 1/(q^2 + \omega_n^2 + m^2)$ in (\[Pi\]) and (\[Sigma\]). A fully self-consistent approach would require we replace $G^0$ by $G$ in these equations; this is computationally much more difficult and was not numerically implemented. Our approximation thus amounts to replacing $G^0$ by $G$, but then ignoring the momentum and frequency dependence of the self energy in $G$. For the most part, this omission is not expected to be serious, as corrections can be organized order by order in $\kappa$. However, we cannot rule out the possibility, especially in the quantum-disordered phases, that there is some entirely different, possibly gapless, solution of the fully self-consistent equations; such a solution will clearly be non-perturbative in $\kappa$. We also note here that in our analytical considerations below of the boundaries between the phases, we will include the full $G$ in (\[Sigma\]).
The numerical determination $m^2$ as a function of $T, \delta$ was carried out on a HP-RISC workstation. A meaningful solution always existed at all finite $T$, with no phase transitions as a function of $T$ or $\delta$. Phase transitions are however present at $T=0$, and were examined by studying the $T \rightarrow 0$ limit of our solutions. The computations required about 3 weeks of computer time.
RESULTS
=======
We now describe the results of our numerical calculations. The nature of the ground state can be determined from the values and $T$ dependences of $q_c$ and $\overline{m}$ where $q=q_c$ is the location of the maximum of $G(q, i\omega_n = 0)$, and $$\overline{m}^2 = m^2 + \Sigma (q_c , i\omega_n = 0) - \Sigma(0,0)
\label{defmb}$$ It is easy to see from (\[Gfunc\]) that $\overline{m}$ is roughly the inverse correlation length (‘roughly’ because this neglects $\partial \Sigma/ \partial q^2$; including this term yields corrections of the order of unity which are not strongly $T$ dependent). The various phases can be identified by studying the $T$ dependence of $\overline{m}$ as $T \rightarrow 0$, as will be described below. The values of $q_c$ were approximately $T$ independent and distinguish between commensurate and incommensurate phases.
Two samples of our results are contained in Figs. \[mbar1\] and \[mbar2\] which plot the $T$ and doping dependence of $\overline{m}$ for two sets of coupling constants. For completeness we also show in Figs \[m1\] and \[m2\] the values of $m$ for the same samples. We will now describe the properties of the phases in these figures and follow that up with some general discussion on the nature of the quantum transitions between them.
Long-range-ordered states
-------------------------
The states with magnetic order are expected to have $\overline{m} \rightarrow 0$ as $T \rightarrow 0$. In particular, the low $T$ dependence should be [@brezin] $$\overline{m} \sim \exp\left( - \frac{2 \pi \rho_s}{k_B T} \right)$$ where $\rho_s$ is the fully renormalized spin stiffness. Numerical solutions at very lower $T$ took longer times to converge, so it was difficult to see this exponential behavior in some of the doped samples. We simply identified the samples in which $\overline{m}$ vanished with an upward curvature as $T \rightarrow
0$, as possessing magnetic long-range order. Further states with $q_c = 0$ ($q_c \neq 0$) were identified as CLRO (ILRO).
Quantum disordered states
-------------------------
These states have $\overline{m}$ saturating at a finite value as $T \rightarrow 0$, which is roughly the gap, $\Delta \sim \overline{m}
(T \rightarrow 0)$, in the $n$ sector. Again, the value of $q_c$ distinguishes between the CQD and IQD states.
We examine the nature of the spin correlations at a point in the IQD phase by plotting the $n$ field contribution to the structure factor, $\tilde{S}(q)$, $$\tilde{S}(q) = g T \sum_{\omega_n} G(q, \omega_n)$$ in Fig. \[sfac\]. Notice that there is strong overlap between the peaks at high temperature. Upon lowering the temperature, the peaks first sharpen considerably, but then their width saturates.
Let us discuss the form of the $n$ spectrum at $T=0$ in the CQD phase. We will focus on real frequencies, $\omega$ just above the gap $\Delta$, and small momenta $q$. The magnon contribution to the self energy, $\Sigma_n$ in (\[Sigma\]) does not acquire an imaginary part until $\omega = 3 \Delta$ and can therefore be completely ignored. The damping from the fermion particle-hole continuum, $\Sigma_f$ is however not so innocuous. We find [@square] $$\mbox{Im} \Sigma_f ( q, \omega ) \sim \left\{
\begin{array}{ll}
0 & ~~ 0< \omega < \Delta \\
|q| (\omega - \Delta)^2 &
~~0 < \omega - \Delta \ll {\displaystyle \frac{q^2}{2 \Delta}} \\
\Delta^{1/2} (\omega - \Delta)^{5/2} &
~~\omega - \Delta \gg {\displaystyle \frac{q^2}{2 \Delta}}
\end{array}
\right.$$ and $\mbox{Im} \Sigma_f ( q, - \omega ) = - \mbox{Im} \Sigma_f ( q,\omega )$. The $n$ spectral weight is then given by $$G ( q, \omega ) \sim \frac{1}{\Delta^2 + q^2 - \omega^2 + \Sigma_f (q, \omega )}$$ From the above results it follows that at $q=0$ we have $$\mbox{Im} G ( q=0, \omega \geq \Delta )
= \frac{a_1}{\Delta} \delta ( \omega - \Delta ) + \frac{a_2}{
\Delta^{3/2}} ( \omega - \Delta)^{1/2}$$ for some constants $a_1 , a_2$. Thus there is a sharp spin-1 quasiparticle peak, and a second background term which is a direct consequence of the coupling of the $n$ quanta to the particle-hole continuum. At small, but finite $q$, the sharp peak moves to $\omega \sim \Delta + q^2 /(2 \Delta)$ and acquires a finite width; there is absorption at all frequencies greater than $\Delta$.
The spectral properties of the IQD phase are essentially identical except that the role of the point $q=0$ is replaced by $q=q_c$; in obtaining this result it is, of course, necessary to replace $G^0$ by $G$ in (\[Sigma\]).
Quantum transitions
-------------------
Our results in Figs. \[mbar1\] and \[mbar2\] show two sequences of quantum transitions with increasing doping: CLRO-CQD-IQD and CLRO-ILRO. In the second case there should eventually be a ILRO to IQD transition, but for the parameters examined, we did not find one before a doping level where the incommensuration wavevector was almost as large as the momentum upper cutoff. We now present a theoretical analysis of some issues raised by the existence of these quantum transitions.
### CLRO to CQD quantum transition
An important ingredient in determining the universality class of this transition is the analytic structure of the $n$ field self-energy $\Sigma_f$ as $T\rightarrow 0$ in the Néel phase and the quantum-critical point. We consider (\[Sigma\]) in the limit $m\rightarrow 0$, and $T\rightarrow 0$ when the frequencies become continuous variables, and the Matsubara summations can be converted to integrations. It is evident that $\Sigma_f ( q, i\omega )$ is an even function of $\omega$. Moreover, it is not difficult to show that there are no infra-red divergences in either $\partial \Sigma_f / \partial q^2 |_{q=\omega=0}$ or $\partial \Sigma_f / \partial \omega^2 |_{q=0, \omega \searrow 0}$. This implies that for $q, \omega$ small we have $$\Sigma_f ( q, i\omega ) = \Sigma_f (0,0) + b_1 q^2 + b_2 \omega^2 + \ldots
\label{sigexpand}$$ Thus the gapless fermion particle-hole sea has not induced any non-analyticities in $\Sigma_f$ to this order. There are indeed non-analytic terms present at higher order in $\Sigma_f$ which are signaled by infra-red divergences in higher derivatives of $\Sigma_f$; we will discuss the form of such terms below. For our purposes, it is sufficient to note here that all such higher gradient terms are expected to be irrelevant at the CLRO to CQD transition. Thus the gapless fermion particle-hole excitations have had a relatively innocuous effect: they have mainly lead to renormalizations of the spin-wave velocity and spin stiffness. The universality class of the CLRO-CQD transition is thus expected to be the same as that in the undoped sigma model. This is a transition with dynamic critical exponent $z=1$ and its leading universal properties have been discussed in some detail by Chubukov [*et. al.*]{} [@CSY]. All of the scaling functions of Chubukov [*et. al.*]{} [@CSY] should therefore also apply to the present doped antiferromagnet. The main effect of the fermions has been to change the value of the effective coupling constant and renormalize the spin-wave velocity. Consistent with this identification, observe the linear dependence of $\overline{m}$ with $T$ in Fig. \[mbar1\] at $k_F =0.2$ over a wide temperature region. This value of $k_F$ places the system quite close to the quantum-critical point as the value of $\Delta$ is very small. At the quantum-critical point of the sigma model, it is predicted that $\xi^{-1} = C_Q k_B T / \hbar c$ with $C_Q \approx 1$ a universal number. The slope of $m$ versus $T$ at $k_F =0.2$ in Fig. \[mbar1\] is about $0.65$ - this matches with the expected result if there is renormalization of spin-wave velocity $c/c_0 \approx 1/0.65$. A renormalization of the spin-wave velocity of order unity is to be expected, as the fermionic polarization $\chi$ in (\[chival\]) is not suppressed by a factor that vanishes as $\delta \rightarrow 0$. Differences between the quantum transition in the doped and undoped antiferromagnet do however show up at the correction to scaling level. The higher-order non-analytic terms in $\Sigma_f$ will have a form which is quite specific to the doped model. One such term can be obtained by analytically continuing to real frequencies and computing $\mbox{Im} \Sigma_f$ at the critical point. We find $$\mbox{Im} \Sigma_f ( q , \omega )
\sim q \omega^2 ~~~~\mbox{at the quantum-critical point}$$
### Lifshitz point
The existence of a direct CLRO to ILRO transition in Fig. \[mbar2\] has a strong consequence for the phase diagram of the SS model. As both phases can be transformed into their quantum-disordered partners simply by increasing the value of $g$, we conclude that there must be a point in the phase diagram where all the four phases - CLRO, CQD, IRLO, and IQD - meet. Such a point is called a Lifshitz point [@horn]. Lifshitz points have so far been studied primarily in the context of thermal transitions in classical spin systems [@selke]. An important result is that such points can exist only above a lower critical dimension determined as follows [@grest]: a system in $D$ dimensions, with incommensurate instabilities in $m$ of those dimensions, has lower critical dimension $2 + m/2$. This result appears to be in conflict with our results here for the doped antiferromagnet. For we have incommensuration in $m=2$ spatial dimensions, giving a lower critical dimensionality equal to the spacetime dimension $D=3$. So how can a Lifshitz point exist ?
The answer to this apparent inconsistency lies in the form of $\Sigma_f$. At the Lifshitz point we clearly have $\partial G^{-1} /
\partial q^2 |_{q=0,\omega=0 } = 0$. Thus the leading $q$ dependence of $G^{-1}$ at small $q$ will come from higher-order terms in $\Sigma_f$. Let us assume that $G^{-1} \sim \omega^2
+ q^p$ at the Lifshitz point. Inserting this fully renormalized $G$ in the result (\[Sigma\]) for $\Sigma_f$ we find by power counting $$\Sigma_f ( q, 0) = \Sigma_f (0,0) + b_1 q^2 + b_3 q^{4-p/2} + \ldots$$ Consistency now demands that $p=4-p/2$ which yields $p=8/3$. This differs from the value $p=4$ use in classical spin systems [@selke]. With this modified form of $G$ we may repeat the calculation of Grest and Sak [@grest] and verify that spacetime dimension $D=3$ is above the lower critical dimension which is $D=7/3$. Thus it is possible to have a Lifshitz point in $D=3$.
Finally we note that a point where CLRO, ILRO, CQD, and IQD phases meet was also found in the large $N$, $Sp(N)$ theory of frustrated, two-dimensional quantum Heisenberg antiferromagnets [@Sach-Read2]. The nature of this point appears to be quite different from the Lifshitz point in the present theory. In particular, the IQD phase of the $Sp(N)$ frustrated antiferromagnets [@Sach-Read2] contains deconfined, bosonic, spin-1/2 spinons, while here we have found massive, triply-degenerate $n$ quanta. Perhaps related to this difference is the fact that the large $N$ $Sp(N)$ theory [@Sach-Read2] finds no softening in the spinon spectrum at the CQD-IQD boundary. Instead, at the boundary the parabolic spinon spectrum splits into two parabola with minima at incommensurate points; the curvature at the minima of the parabola always remains finite. Contrast this with the behavior of the $n$ spectrum found here: the curvature at the minimum of the $n$ spectrum vanishes at the CQD-IQD boundary leading to double minima at incommensurate points in the IQD phase.
CONCLUSIONS {#conc}
===========
The most notable features of our results on lightly-doped antiferromagnets are the quantum disordered phases with a spin pseudo-gap. These phases posses fully-gapped, triply-degenerate, spin 1 magnons, and gapless, spin-1/2, charge $e$ fermions. Several other investigators [@nagaosa; @bza; @spn; @sokol2; @monien] have also recently explored models of the normal state of the lightly-doped cuprates which have spin gaps/pseudo-gaps. This interest in pseudo-gaps is of course motivated by numerous experiments on the underdoped cuprates showing gap-like features in the normal state. [@walstedt; @bucher; @Hardy] There are also interesting trends in the doping and temperature dependence of uniform spin susceptibility of the cuprates [@millis].
For completeness we review some of the previous theoretical results, and point out the differences to our results. A number of the models [@nagaosa] are related to resonating-valence-bond type mean-field theories [@bza]; there is a BCS-like pairing of spin-1/2, neutral spinons in the normal state at low-doping, leading to a gap-like feature in the spectrum at finite temperature. However, unlike our results, this state extrapolates to a true gap at $T=0$. An extension of these models to a three-band, $CuO_2$ layer model [@spn] did posses gapless, spin-1/2, charge $e$ fermionic excitations on the oxygen sites. However the above-gap spectrum in all of these models [@nagaosa; @spn] consists of unbound spin-1/2, neutral fermions; in contrast, the spin spectral weight above the gap of our model is dominated by a spin-1, bosonic magnon. Millis and Monien have attributed the gaps to interlayer couplings in the Yttrium based cuprates [@monien]. The recent work of Sokol, Pines and collaborators [@sokol2] is perhaps closest in spirit to ours, although their scenario for the mixing between the $n$ quanta and the fermions appears to be somewhat different.
Also relevant to our result is the recent high temperature series analysis of the CLRO to CQD transition in the $t-J$ model [@sokol1]. This work provides some evidence in support of our result that $z=1$ at this transition.
Finally, if our model is to provide a complete picture of the cuprates, it should also explain the nature of the photo-emission spectrum [@photo]. For this one needs to understand better the connection between the bare electron ejected in the photo-emission and the $(n_{\ell}, \Phi_{va})$ fields. This problem is currently under investigation.
The research was supported by NSF Grant No. DMR-9224290. I am indebted to A. Chubukov for valuable discussions on every aspect of this work, for his comments on the manuscript, and for our previous collaborations on related subjects [@comment; @CSY]. Jinwu Ye participated in some of the initial stages of this work. I thank N. Read, and B. Shraiman for useful discussions, and ITP Santa Barbara for hospitality while part of this work was completed.
B.I. Shraiman and E.D. Siggia, Phys. Rev. Lett. [**61**]{}, 467 (1988); Phys. Rev. B [**42**]{}, 2485 (1990).
V. Elser, D. Huse, B.I. Shraiman, and E.D. Siggia, Phys. Rev. B [**41**]{}, 6715 (1990).
S. Sachdev, Phys. Rev. B [**39**]{}, 12232, (1989).
A.V. Chubukov and S. Sachdev, reply to comment by A.J. Millis, Phys. Rev. Lett. to be published.
J.A. Hertz, Phys. Rev. B [**14**]{}, 525 (1976).
A.J. Millis, preprint
C. Kane, P.A. Lee, T.K. Ng, B. Chakraborty, and N. Read, Phys. Rev. [**B**]{} 41, 2610 (1990); C. Jayaprakash, H.R. Krishnamurthy, and S. Sarkar, Phys. Rev. [**B**]{}, 40, 2610 (1989); D. Yoshioka, J. Phys. Soc. Jpn. [**58**]{}, 1516 (1989).
S. Sachdev and Jinwu Ye, Phys. Rev. Lett. [**69**]{}, 2411 (1992).
S. Chakravarty, B.I. Halperin and D.R. Nelson, Phys. Rev. B [**[39]{}**]{}, 2344 (1989).
A.M. Polyakov, [*Gauge Fields and Strings*]{}, Harwood, New York (1987).
A.V. Chubukov and S. Sachdev, Phys. Rev. Lett. [**71**]{}, 169 (1993); A.V. Chubukov, S. Sachdev and J. Ye, Phys. Rev. B, submitted;
N. Read and S. Sachdev, Phys. Rev. Lett. [**62**]{}, 1694 (1989); Phys. Rev. B [**42**]{}, 4568 (1990).
S.-W. Cheong [*et. al.*]{}, Phys. Rev. Lett. [**67**]{}, 1791 (1991).
A.V. Chubukov and D. Frenkel, Phys. Rev. B [**46**]{}, 11884 (1992).
B.I. Shraiman and E.D. Siggia, Phys. Rev. B [**46**]{}, 8305 (1992).
L.P. Gor’kov, V.N. Nicopoulos, and P. Kumar, preprint.
J. Stephenson, Can. J. Phys. [**48**]{}, 2118 (1970).
A.M. Polyakov, Phys. Lett. B [**59**]{}, 79 (1975); E. Brezin and J. Zinn-Justin, Phys. Rev. B [**14**]{}, 3110 (1976).
The system does not have full rotational invariance, only the symmetry of the square lattice: the notation $q^p$ is a schematic for a function of $q_x$, $q_y$ which is homogeneous of degree $p$ and is invariant under the square lattice space group.
R.M. Hornreich, M. Luban, and S. Shtrikman, Phys. Rev. Lett. [**35**]{}, 1678 (1975); R.M. Hornreich and A.D. Bruce, J. Phys. A [**11**]{}, 595 (1978).
W. Selke in [*Phase Transitions and Critical Phenomena*]{}, C. Domb and J.L. Lebowitz eds., v. 15, pg 57, Academic Press, New York, 1992.
G.S. Grest and J. Sak, Phys. Rev. B [**17**]{}, 3607 (1978).
N. Read and S. Sachdev, Phys. Rev. Lett. [**66**]{}, 1773 (1991); Int. J. Mod. Phys. [**B5**]{}, 219 (1991).
P.A. Lee and N. Nagaosa, Phys. Rev. B [**46**]{}, 5621 (1992); G. Kotliar and J. Liu, Phys. Rev. B [**38**]{}, 5142 (1988); N. Andrei and P. Coleman, Phys. Rev. Lett. [**62**]{}, 595 (1989); H. Fukuyama, Prog. Theor. Phys. Suppl. [**108**]{}, 287 (1992).
G. Baskaran, Z. Zou, and P.W. Anderson, Sol. State Comm. [**63**]{}, 973 (1987); G. Baskaran and P.W. Anderson , 580 (1987); A.E. Ruckenstein, P.J. Hirschfeld, and J. Appel, Phys. Rev. B [**36**]{}, 857 (1987).
J. Ye and S. Sachdev, Phys. Rev. B [**44**]{}, 10173 (1991); S. Sachdev, Phys. Rev. B [**45**]{}, 389 (1992).
A. Sokol and D. Pines, preprint; V. Barzykin, D. Pines, A. Sokol and D. Thelen, preprint.
A.J. Millis and H. Monien, Phys. Rev. Lett. [**70**]{}, 2810 (1993).
R.E. Walstedt [*et. al.*]{}, Phys. Rev. B [**41**]{}, 9574 (1990).
B. Bucher [*et. al.*]{} Phys. Rev. Lett. [**70**]{}, 2012 (1993).
C.C. Homes, T. Timusk, R. Liang, D.A. Bonn, and W.N. Hardy, Phys. Rev. Lett. [**71**]{}, 1645 (1993).
A.J. Millis, Phys. Rev. Lett., to be published.
A. Sokol, R.L. Glenister, and R.R.P. Singh, preprint.
C.G. Olson [*et. al.*]{}, Phys. Rev. B [**42**]{}, 381 (1990).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the signature of a cosmic string wake in 21cm redshift surveys. Since 21cm surveys probe higher redshifts than optical large-scale structure surveys, the signatures of cosmic strings are more manifest in 21cm maps than they are in optical galaxy surveys. We find that, provided the tension of the cosmic string exceeds a critical value (which depends on both the redshift when the string wake is created and the redshift of observation), a cosmic string wake will generate an emission signal with a brightness temperature which approaches a limiting value which at a redshift of $z + 1 = 30$ is close to 400 mK in the limit of large string tension. The signal will have a specific signature in position space: the excess 21cm radiation will be confined to a wedge-shaped region whose tip corresponds to the position of the string, whose planar dimensions are set by the planar dimensions of the string wake, and whose thickness (in redshift direction) depends on the string tension. For wakes created at $z_i + 1 = 10^3$, then at a redshift of $z + 1 = 30$ the critical value of the string tension $\mu$ is $G \mu = 6 \times 10^{-7}$, and it decreases linearly with redshift (for wakes created at the time of equal matter and radiation, the critical value is a factor of two lower at the same redshift). For smaller tensions, cosmic strings lead to an observable absorption signal with the same wedge geometry.'
author:
- 'Robert H. Brandenberger$^{1}$, Rebecca J. Danos$^{1}$, Oscar F. Hernández$^{2,1}$ and Gilbert P. Holder$^{1}$'
title: The 21 cm Signature of Cosmic String Wakes
---
Introduction
============
Cosmic strings (see [@Zel; @Vil; @Kibble; @TB; @Sato; @Stebbins] for initial work on cosmic strings and structure formation) cannot be [@Albrecht; @Turok] the dominant source of the primordial fluctuations, however they can still provide a secondary source of fluctuations. Over the past decade, the realization has grown that many inflationary scenarios constructed in the context of supergravity models lead to the formation of gauge theory cosmic strings at the end of the inflationary phase [@Rachel]. Also, in a large class of brane inflation models the formation of cosmic superstrings [@Witten] at the end of inflation is generic [@CS-BI], and in some cases (see [@Pol1]) these strings are stable (see also [@recentCS] for reviews on fundamental cosmic strings). Cosmic superstrings are also a possible remnant of an early Hagedorn phase of string gas cosmology [@SGrev]. In all of these contexts, both a scale-invariant spectrum of adiabatic coherent perturbations and a sub-dominant contribution of cosmic strings is predicted. Hence, it is important to search for the existence of cosmic strings.
In this paper we study the signature of cosmic strings in 21cm radiation maps (see e.g. [@Furlanetto] for an in-depth review of 21cm cosmology). Observing the intensity of the cosmological background radiation at wavelengths corresponding to the red-shifted 21cm transition line of neutral hydrogen has several potential advantages compared to the currently explored windows . First of all, it probes the distribution of the dominant form of baryonic matter and is thus not sensitive to our incomplete understanding of star formation and non-linear evolution, which is a problem when interpreting the results of optical redshift surveys. Secondly, it probes the universe at higher redshifts and allows us to explore the “dark ages" (the epoch before star formation and non-linear clustering set in). Related to this, it explores the distribution of matter in a regime when the amplitude of the fluctuations is smaller and linear theory is a better approximation. Finally, 21cm surveys provide three-dimensional maps, a significant potential advantage over CMB anisotropy maps.
Cosmic strings are known to give rise to distinctive signatures in CMB temperature anisotropy maps [@KS], CMB polarization maps [@Danos3] and large-scale structure (LSS) maps [@Silk; @Rees; @Vach; @SVBST; @Charlton; @Hara]. These distinctive signatures come from moving long (compared to the Hubble radius) strings (see Section 2). In the CMB temperature maps, these strings lead to line discontinuities, in the polarization maps to (roughly) rectangular regions with extra polarization, and in LSS maps to thin planar regions of enhanced density. These signals are manifest in position space maps, but they become obscured when calculating power spectra. Hence, the lesson is to study the maps in position versus momentum space.
Here, we compute the signature of a single cosmic string wake in 21cm emission. We find that strings with tensions $\mu$ somewhat below the current limits of $G \mu = 3 \times 10^{-7}$ could be detected in 21cm maps where they would appear as wedges in 21cm maps with either extra emission or extra absorption, depending on the tension and on the specific redshift. The planar dimensions are set by the direction of the string and its velocity vector, and the width of the wedges is proportional to the string tension. This signal must be searched for in position space maps. In Fourier space the distinctive phase information would be washed out.
The outline of this paper is as follows: In Section 2 we discuss how string wakes are generated and how these lead to distinctive signals for observations. We review gravitational accretion of matter onto cosmic string wakes and compute the temperature of the HI gas inside the wake. In Section 3 we then study the 21cm emission signal from a single cosmic string wake. The generalization to the case of a network of string wakes will be addressed in a future publication. We compute the brightness temperature and describe the geometrical structure of the signal, a structure very characteristic for cosmic strings. In the final section we summarize our results and put them in the context of other work on the possible detection of cosmic strings. When computing the magnitude of the 21cm signal, we use the same WMAP concordance values [@WMAP] for the cosmological parameters as used in the review [@Furlanetto].
Cosmic Strings and Large-Scale Structure
========================================
In many field theory models, the formation of a network of cosmic strings is an inevitable consequence of a phase transition in the early universe (for reviews on cosmic strings see e.g. [@Kibble2; @VilShell; @HK; @RHBrev4]). This network of cosmic strings approaches a “scaling solution" which means that the statistical properties of the string network are the same at all times if distances are scaled to the Hubble radius [@RHBrev4] . The detailed form of the scaling solution must be obtained from numerical simulations [@CSsimuls] of the evolution of a network of cosmic strings (see also [@Pol2] for some recent analytical work). There are two components of the string network - firstly a network of “long" strings with a mean curvature radius $\zeta = c_1 t$, where $c_1$ is a constant of order unity, and secondly a distribution of string loops with radii smaller than the horizon which results from the “cutting up" of the long string network as a consequence of string intersections. According to more recent cosmic string evolution simulations, the long string component is more important for cosmological structure formation.
Numerical simulations by different groups have clearly verified the scaling solution for the long string network. There is, however, still a large uncertainty concerning the distribution of string loops (the only agreement seems to be that the loops are less important than the long strings for cosmological structure formation). Hence, in order to obtain constraints in models with cosmic strings which are robust against the uncertainties in the numerical simulations, it is important to focus on signatures of long strings as opposed to signatures of string loops. The tightest current constraints on cosmic strings come from the shape of the angular power spectrum of CMB anisotropies, yielding a constraint of [@Wyman] (see also [@others]) \[limit\] G < 3 10\^[-7]{} , where $\mu$ is the string tension and $G$ is Newton’s constant. However, it is important to keep in mind that this limit is based on an analytical description of the scaling solution which contains a number of parameters which can only be determined from comparisons with numerical simulations and whose values thus have a substantial uncertainty. Work based on a field theory simulation of strings [@Hind] gives a limit of twice the above value (see also [@recent] for a very recent analysis of the different bounds). Note that limits on the cosmic string tension which come from gravitational radiation from string loops [@CSgrav] are sensitive both to the large uncertainties in the distribution of string loops, and also to back-reaction effects on cosmic string loops (see e.g. [@RHBloop]), and are hence not robust.
Direct limits on the cosmic string tension $\mu$ can be obtained by looking for specific signatures of individual long strings. These are limits which are insensitive to the parameters in the cosmic string scaling solution. Limits obtained from searching for the Kaiser-Stebbins signature of a long string in CMB temperature maps were derived in [@Lo; @Smoot] using WMAP data. The limits obtained were weaker than the ones from (\[limit\]). However, the angular scale of the WMAP experiment is too large to be able to effectively search for sharp features in position space maps such as those predicted by cosmic strings. It was pointed out [@Fraisse] that the string signatures for values of $G \mu$ somewhat smaller than the limiting value of (\[limit\]) should be clearly visible in smaller angular scale CMB anisotropy maps such as those provided by the ACT [@ACT] and SPT [@SPT] experiments. In recent work [@Amsel; @Stewart; @Rebecca1] it was shown that limits up to an order of magnitude tighter than (\[limit\]) might be achievable using SPT data. In the following, we will discuss direct signals of wakes created by long strings in 21cm surveys.
As first pointed out in [@Silk] and then further discussed in [@Vach; @Rees; @SVBST; @Charlton], long strings moving perpendicular to the tangent vector along the string give rise to “wakes" behind the string, i.e. in the plane spanned by the tangent vector to the string and the velocity vector. The wake arises as a consequence of the geometry of space behind a long straight string [@Vil2; @Ruth] - space perpendicular to the string is conical with a deficit angle given by \[deficit\] = 8 G . From the point of view of an observer behind the string (relative to the string velocity vector), it appears that matter streaming by the string (from the point of view of the observer traveling with the string) obtains a velocity kick of magnitude v = 4 G v\_s \_s \~ 4 [km/s]{} (G )\_6 v\_s \_s towards the plane behind the string. In the above, $v_s$ is the velocity of the string (in units of the speed of light), $\gamma_s$ is the corresponding relativistic gamma factor, and $(G \mu)_6$ is the value of $G \mu$ in units of $10^{-6}$. This leads to a wedge-shaped region behind the string with twice the background density (see Figure 1).
![Sketch of the mechanism by which a wake behind a moving string is generated. Consider a string perpendicular to the plane of the graph moving straight downward. From the point of view of the frame in which the string is at rest, matter is moving upwards, as indicated with the arrows in the left panel. From the point of view of an observer sitting behind the string (relative to the string motion) matter flowing past the string receives a velocity kick towards the plane determined by the direction of the string and the velocity vector (right panel). This velocity kick towards the plane leads to a wedge-shaped region behind the string with twice the background density (the shaded region in the right panel).[]{data-label="fig:1"}](wakewedge.eps){height="5cm"}
Since the strings are relativistic, they generally move with a velocity of the order of the speed of light. There will be frequent intersections of strings. The long strings will chop off loops, and this leads to the conclusion that the string distribution will be statistically independent on time scales larger than the Hubble radius. Hence, to model the effects of strings we will make use of a toy model introduced in [@Periv] and used in most analytical work on cosmic strings and structure formation since then: we divide the time interval between the time of equal matter and radiation and the current time $t_0$ into Hubble expansion time steps. In each time step, we lay down a distribution of straight string segments moving in randomly chosen directions with velocities chosen at random between $0$ and $1$ (in units of the speed of light). The centers and directions of these string segments are also chosen randomly, and the string density corresponds to $N$ strings per Hubble volume, where $N$ is an integer which is of the order $1$ according to the scaling of the string network. The distribution of string segments is uncorrelated at different Hubble times. Each string segment will generate a wake, and it is the signal of one of these wakes which we will study in the following.
A string segment laid down at time $t_i$ will generate a wake whose dimensions at that time are the following: \[size\] c\_1 t\_i t\_i v\_s \_s 4 G t\_i v\_s \_s , where $c_1$ is a constant of order one. In the above, the first dimension is the length in direction of the string, the second is the depth. The fact that the string wake has a finite depth is due to the causality constraints on density fluctuations produced during a phase transition [@Traschen]: The information about the formation of a string (and hence the information about the existence of a deficit angle) cannot have propagated farther than the horizon from the point of nucleation of the defect. It was shown in [@Joao] that the deficit angle quite rapidly tends to zero as the horizon is approached.
The overdense region in the wake will lead to the gravitational accretion of matter above and below the wake towards the center of the wake. In this way, the wake will grow in thickness. The accretion of matter onto a cosmic string wake was studied in [@SVBST; @Leandros1] in the case of the dark matter being cold, and in [@Leandros1; @Leandros2] in the case of the dark matter being hot. We are interested in the case of cold dark matter. As an aside, we mention that if the strings have lots of small-scale structure then they will have an effective tension which is less than the effective energy density [@wiggly]. This will lead to a local gravitational attraction of matter towards the string, a smaller transverse velocity, and hence to string filaments instead of wakes. The gravitational accretion onto string filaments was studied in [@Aguirre].
We will now review the computation of the width of the wake. We are interested in the distribution of baryons in the vicinity of the string. However, for $t > t_{rec}$ the baryons and cold dark matter feel the same gravitational attraction and will thus behave in the same way - modulo thermal velocity effects which will be discussed later - until the baryons undergo a shock. Thus, for the moment we focus on the onset of clustering of the cold dark matter. We are interested in wakes created after the time $t_{eq}$ of equal matter and radiation (there is no gravitational clustering of cold dark matter before that time). For times between $t_{eq}$ and recombination ($t_{rec}$), the baryons are coupled to the radiation. However, for $t > t_{rec}$ the baryons will rapidly fall into the potential wells created by the cold dark matter and thus, once again, it is legitimate to focus attention on the clustering of the cold dark matter. Note that wakes produced at the earliest time are the most numerous.
We consider a mass shell located at a comoving distance $q$ above the wake. Its physical distance above the wake at time $t$ is \[height1\] w(q, t) = a(t) ( q - ) , where $\psi$ is the co-moving perturbation induced by gravitational accretion. If the wake is laid down at the initial time $t_i$ (with corresponding redshift $z_i $), then the initial conditions for the cold dark matter fluctuation are \[IC\] (t\_i) = (t\_i) = 0 . As a consequence of the initial wake planar overdensity $\sigma(t_i)$, the co-moving displacement $\psi$ will begin to increase. The clustering dynamics can be studied making use of the Zel’dovich approximation [@Zelap] in which the gravitational force is treated in the Newtonian limit. As reviewed recently in [@Danos3], we obtain \[height2\] (t) = (t\_i) ( )\^[2/3]{} ( )\^[2/3]{} t\_0\^2 , where $t_0$ is the present time and $\sigma(t_i)$ is the initial wake planar density excess.
The mass shell with initial comoving distance $q$ above the wake “turns around" when ${\dot w}(q, t) = 0$. At time $t$, this occurs for a value $q = q_{nl}(t)$ given by \[qnl\] q\_[nl]{}(t) = G v\_s \_s (z\_i + 1)\^[-1/2]{} t\_0 ( )\^[2/3]{} . It is easy to check that at the point of turnaround (q\_[nl]{}, t) = q , and that hence the density inside the turnaround surface is twice the background density.
Once a matter shell reaches its maximal distance $w_{max}$ from the center of the wake, it will start to collapse onto the wake. The infall of matter will halt as the shell hits other streams of matter. This will lead to a shock. In analogy with what can be shown analytically in the context of the spherical collapse model (see e.g. [@Peebles]) we assume that the shock occurs at approximately half the maximal distance. The hydrodynamical simulations of [@Sorn1] confirm the applicability of this assumption. Since the distance at turnaround is half the width the shell would have without gravitational accretion onto the wake, the shock occurs at a distance $1/4$ of that which the matter shell would have under unperturbed Hubble expansion. Hence, the average density inside the wake is four times the background density, a result which will be used several times in the computations of the following section.
The evolution of the mass shell between turnaround and shock can be followed using (\[height1\]) and (\[height2\]). The shock will occur when $w(q, t) = (1/2) w_{max}(q)$. A straightforward computation shows that at this point the velocity is given by \[hresult\] (q, t) = - G v\_s \_s ( )\^[1/2]{} . It is this velocity which then determines the temperature inside the shocked region.
The shocks will lead to thermalization of the wake. The gas temperature will be given by \[temp\] k\_B T = m v\^2 , where $m$ is the mass of a HI atom and $v$ is the velocity of the in-falling particles when they hit the shock, which is given by \[vel\] v = [w]{}(q(t), t) , where $q(t)$ is the comoving distance of the shell which starts to collapse at the time $t$.
Inserting the result (\[hresult\]) into (\[vel\]) and then into (\[temp\]) we obtain the following result for the temperature $T_K$ of the HI atoms inside the wake \[wakeT\] T\_K &=& (G )\^2 (v\_s \_s)\^2 k\_B\^[-1]{} m\
& & \[20 [K]{}\] (G )\_6\^2 (v\_s \_s)\^2 , where in the second line we have written the result in degrees K and expressed $G \mu$ in units of $10^{-6}$ (and kept only one significant figure in reporting the final number). The wake temperature is largest for wakes produced at the earliest times since the initial wake density is then the highest, and increases as time increases because more matter has time to accrete.
Assuming values of $G \mu = 3 \times 10^{-7}$ (the current upper bound on the string tension), $z_i = 10^3$, (close to the time of recombination), $z + 1 = 30$ and $(v_s \gamma_s)^2 = 1/3$, Eq. (\[wakeT\]) yields $T_K \, \sim \, 20 K$. This temperature is smaller than the CMB temperature $T_{\gamma} \simeq 82 K$ at this redshift. As we will see in the following section, this leads to an absorption signal in the 21cm radiation.
The formation of a cosmic string wake and the thermalization which takes place inside the shocked region has been studied in [@Sorn1] using an Eulerian hydro code [@Sorn2] optimized to resolve shocks. The numerical simulations of [@Sorn1] show that the density and temperature inside the shocked region are indeed roughly uniform and that the temperature obtained agrees with the values obtained here using analytical approximations.
Looking for signals from cosmic strings in 21cm surveys is potentially more powerful than looking in large-scale optical redshift surveys. This is because firstly the non-Gaussian signatures from strings are more pronounced at higher redshifts, secondly because we are directly looking at the distribution of the baryons, and not just the distribution of stars, the latter being affected by non-linear and gas dynamics, and thirdly because the distribution of matter is more linear at higher redshifts and hence easier to follow analytically. Compared to CMB and CMB polarization maps, 21cm surveys have the advantage of yielding three- rather than just two-dimensional maps, maps thus containing much more information. To our knowledge, there has been little previous work on 21cm emission from strings. Two exceptions are [@Wandelt] in which the angular power spectrum of 21cm emission from a network of cosmic strings was considered, and [@Aaron], a recent study in which the correlation between 21cm emission and CMB signals from cosmic strings was studied. In contrast to these works, we are looking for direct string signals in position space 21cm maps.
In the following section we will briefly summarize some key features of 21cm cosmology and apply the equations for the 21cm brightness temperature to the case of emission from a string wake.
Cosmic Strings and 21cm Maps
============================
Neutral hydrogen is the dominant form of baryonic matter in the “dark ages", i.e. before star formation. Neutral hydogen has a 21cm hyperfine transition line which is excited if the hydrogen gas is at finite temperature. Hence, we expect redshifted 21cm radiation to reach us from all directions in the sky, and the intensity of this radiation can tell us about the distribution of neutral hydrogen in the universe, as a function of both angular coordinates in the sky and redshift. Thus, in contrast to the CMB, 21cm surveys can probe the three-dimensional distribution of matter in the universe (see [@early] for pioneering papers on the cosmology of the 21cm line and [@Furlanetto] for an in-depth review).
Let us now consider the equation of radiative transfer along the line of sight through a hydrogen gas cloud of uniform temperature- in the case of interest to us this gas cloud is the cosmic string wake. The brightness temperature $T_b(\nu)$ at an observed frequency $\nu$ due to 21cm emission is given by \[three1\] T\_b() = T\_S ( 1 - e\^[- \_]{} ) + T\_() e\^[- \_]{} , where $T_S$ is the spin temperature, $T_{\gamma}$ is the microwave radiation temperature, and $\tau_{\nu}$ is the optical depth obtained by integrating the absorption coefficient along the light ray through the gas cloud. The frequency $\nu$ is the blue-shifted frequency at the position of the cloud corresponding to the observed frequency $\nu_o$. The term proportional to $T_S$ is due to spontaneous emission, while the term proportional to $T_{\gamma}$ is due to absorption and stimulated emission.
As explained in [@Furlanetto], what is of observational interest is the comparison of the temperature coming from the hydrogen cloud with the “clear view” of the 21 cm radiation from the CMB. \[three15\] T\_b() = \_ . Note that the “clear view” of the CMB is hypothetical since even without a string wake’s gas cloud the intergalactic medium is partly a less dense hydrogen gas. In the second part of the equation above we have expanded the exponential factor to linear order in the optical depth.
The spin temperature $T_S$ is defined as the relative number density of atoms in the hyperfine energy states through $
{n_1}/{n_0} \, = \, 3 \exp({-T_{\star} / T_S} )\, .
$ Here $n_1$ and $n_0$ are the number densities of atoms in the two hyperfine states, and $T_{\star}=E_{10}/k_B=0.068 {\rm K}$ is the temperature corresponding to the energy splitting $E_{10}$ between these states.
The spin temperature is determined solely by the temperature $T_K$ of the gas in the wake, as long as UV scattering is negligible (which is true in our case). The relationship between spin and kinetic gas temperatures is expressed through the collision coefficients $x_c$ which describe the rate of scattering among hydrogen atoms and electrons: \[three2\] T\_S = T\_ . We will shortly discuss the numerical values of the $x_c$ for a particular case of interest.
Combining (\[three15\]) and (\[three2\]) we find for the difference $\delta T_b(\nu)$ in brightness temperature induced by the interaction of the radiation with neutral hydrogen is \[three3\] T\_b() T\_S ( 1 - ) \_ (1 + z)\^[-1]{} , where the last factor represents the red-shifting of the temperature between the time of emission and the present time. It is important to keep in mind that the brightness temperature from a region of space without density perturbations is nonzero. It is negative since the temperature of gas after redshift 200 is lower than that of the CMB photons, the former red-shifting as $(z(t) + 1)^2$, the latter as $z(t) + 1$. The optical depth of a cloud of hydrogen is \[three5\] \_ = ( ) () , where $N_{HI}$ is the column density of HI.
Up to this point, the analysis has been general. Let us now specialize to the case where the gas cloud is the gas inside the cosmic string wake. In this case, the column density is the hydrogen number density $n_{HI}^{wake}$ times the length that the light ray traversed in the cloud. This length will depend on the width $w$ of the wake and the angle $\theta$ that the light ray makes relative to the vertical to the wake so that N\_[HI]{} = . The factor of $2$ results since $w$ is the width of the wake from the center.
The line profile $\phi(\nu)$ is due to broadening of the emission line. This broadening reflects the fact that not all photons resulting from the hyperfine transition will leave the gas cloud at the same frequency. Frequency differences are in general due to thermal motion, bulk motion and pressure effects. Since the pressures we are considering are small by astrophysical standards, pressure broadening is negligible. Since the gas temperature inside the wake is not much larger than the CMB temperature, thermal broadening will not be important, either. This leaves us with the effects of bulk motion, more specifically the expansion of the wake in planar directions. The line profile is normalized such that the integral of it over frequency is unity.
The origin of the line broadening due to the expansion of the wake is illustrated in Figure 2. Let us consider a point on the wake for which the photons travel to us at an angle which is not orthogonal to the plane of the wake. Then, relative to photons emitted at the central point on the photon path, photons emitted from the highest point and the lowest point obtain a relative Doppler shift of \[three7\] = 2 [[sin]{}]{}() , where $H$ is the expansion rate of space and $w$ is the wake width computed in the previous section, both evaluated at the redshift $z$ when the photons are emitted. The angle $\theta$ is indicated in Figure 2. It is the angle relative to the vertical to the wake. As a consequence of the normalization of $\phi(\nu)$ we hence find \[three8\] () = [for]{} \[\_[10]{} - , \_[10]{} + \] , and $\phi(\nu) = 0$ otherwise.
In addition to its role in determining the brightness temperature, the frequency shift $\delta \nu$ is important since it determines the width of the 21cm signal of strings in the redshift direction and is hence central to the issue of observability of the signal. From (\[three7\]) and (\[qnl\]) (and setting the $2 sin{\theta} tan{\theta}$ factor to one) we find that \[freqwidth\] &=& G v\_s \_s ( z\_i + 1 )\^[1/2]{} ( z(t) + 1 )\^[-1/2]{}\
& & 3 10\^[-5]{} (G )\_6 (v\_s \_s) , where in the second line we have used the cosmological parameters mentioned at the end of the introductory section and inserted the representative redshifts $z_i + 1 = 10^3$ and $z(t) + 1 = 30$.
![Photons reaching us from a particular direction given by the angle $\theta$ to the vertical to the wake are emitted at different points in the wake. Three such points are indicated in the Figure - the central point and two points on the bottom and top of the wake, respectively. Since the wake is undergoing Hubble expansion in its planar directions, there is a relative Doppler frequency shift $\delta \nu$ between the 21cm photons from the top and the center of the wake.[]{data-label="fig:2"}](21cmwake.eps){height="5cm"}
Taking the formula (\[three3\]) for the brightness temperature, inserting the results (\[three5\]) for the optical depth and (\[three7\]) and (\[three8\]) for the line profile, we get \[result1\] T\_b() &=& 2 ( 1 - )\
& & n\_[HI]{}\^[bg]{}(t\_0)\
& & (2 [[sin]{}]{}\^2())\^[-1]{} (1 + z)\^[1/2]{} , where $\Omega_m$ is the fraction of the critical energy density which is in the form of matter. Here the ratio of the density inside the wake to the background density, $n_{HI}^{wake}/n_{HI}^{bg}$, is approximately 4. We have rescaled the Hubble constant and the background HI density to its current values $H_0$ and $n_{HI}^{bg}(t_0)$, respectively, and made use of $H(z) = H_0 \Omega_m^{1/2} (1 + z)^{3/2}$ in the re-scaling. Note that the width of the wake has cancelled out between the HI column density and the line profile. The width of the wake, however, has not disappeared completely from the calculation since it determines the wake temperature $T_K$, and since it yields the width of the signal in redshift direction.
Taking the values $A_{10} = 2.85 \times 10^{-15}~{\rm s}^{-1}$, $T_{\star} = 0.068~{\rm K}$, $H_0=73~{\rm km~s}^{-1}~{\rm Mpc}^{-1}$, $\nu_{10}=1420$ MHz, $\Omega_b=0.042$, $\Omega_m=0.26$, $2\sin^2\theta=1$, eq. (\[result1\]) becomes \[result2\] T\_b() = \[0.07 [K]{}\] ( 1 - ) (1 + z)\^[1/2]{} . The collision coefficient $x_c$ is dominated by the hydrogen-hydrogen collisions because of the very low fraction of free electrons. It is given by [@Zygelman] x\_c = . where $\kappa_{10}^{HH}$ is the de-excitation cross section and is approximately $2.7 \times 10^{-11} \rm{cm^3 s^{-1}}$ at a wake gas temperature of $T_K = 20 K$ [@Zygelman]. This temperature corresponds to the parameter values we took at the end of Section 2, i.e. $(G \mu)_6 = 0.3$, $(v_s \gamma_s)^2 = 1/3$, $z_i = 10^3$ and $z + 1 = 30$. Remembering that the density $n$ inside the wake is four times the background density $n_{bg} = 1.9 \times 10^{-7} {\rm{cm}}^{-3} (1 + z)^3$ we find for a redshift $1 + z = 30$ that $x_c \, \simeq \, 0.16$ and hence (from (\[result2\])) $\delta T_b(\nu) \, \simeq \, - 160~{\rm mK}$. For a formation redshift $z_i = z_{eq} = 3200$ corresponding to the time of equal matter and radiation the corresponding temperature is $\delta T_b(\mu) \, \simeq \, -15~{\rm mK}$. Note that the dependence of the above result on the cosmic string tension $\mu$ enters only through the wake temperature. Note also that in the large $G \mu$ limit ($G \mu \gg 10^{-6}$), the brightness temperature approaches a constant value of close to $400 {\rm mK}$.
There is a critical value of the string tension (which depends on the redshift) at which the cosmic string signal changes from emission to absorption. This value is determined by $T_K = T_{\gamma}$ which yields \[crit1\] (G )\_6\^2 0.1 (v\_s \_s)\^[-2]{} 0.4 , for the value $(v_s \gamma_s)^2 = 1/3$ which we are using throughout, and this corresponds to $(G \mu)_6 \, \simeq \, 0.6$ if we use $(z_i + 1) = 10^3$ and insert our reference redshift of $z + 1 = 30$. At a redshift of $z + 1 = 20$ the critical value is $(G \mu)_6 \, \simeq \,
0.4$. For a formation time corresponding to the redshift of equal matter and radiation the critical value of $G \mu$ is a factor of two lower.
For lower values of the tension the cosmic string signal in the 21cm maps shifts from an emission signal to an absorption signal. However, two effects must be taken into account when discussing the visibility of the string signal. The first is the fact that at very small values of $G \mu$ our calculation of the wake temperature breaks down. This occurs when the resulting wake temperature computed in (\[wakeT\]) is smaller than the gas temperature obtained by taking the background gas temperature and computing its increase under adiabatic compression to the overdensity of the wake. In this case it is no longer justified to take the initial conditions (\[IC\]). At redshifts $z$ below 150 the cosmic gas is cooling adiabatically as \[gasT\] T\_g = 0.02 [K]{} (1+z)\^2 because at this point Compton heating of the gas by the CMB is negligible [@SSS]. Setting $T_K = 2.5 \times T_g$ (the factor $2.5$ being due to the fact that for adiabatic compression of a mono-atomic gas by a factor of $4$ in volume, one expects a temperature increase by a factor of $4^{2/3} \sim 2.5$) yields (for our standard values of $1 + z_i = 10^3$ and $(v_s \gamma_s)^2 = 1/3$) \[crit2\] (G )\_6 3 10\^[-3]{} (1 + z)\^[3/2]{} which equals $(G \mu)_6 = 0.5$ at a redshift of $1 + z = 30$. For a formation redshift $z_i = z_{eq}$ the factor $3$ in (\[crit2\]) is replaced by $2$.
For values of $G \mu$ smaller than the one given in (\[crit2\]), the effects of thermal pressure start to become important, and this will effect both the density and temperature structure in the wake.
Our results are illustrated in Figure 3 which shows a comparison of the temperatures relevant to the above discussion and their dependence on redshift. The vertical axis is the temperature axis, the horizontal axis is inverse redshift. The magenta (dashed) line shows the CMB temperature $T_{\gamma}$, the orange (dotted) line is the gas temperature in a wake $2.5 T_g$ after adiabatic contraction. The two solid lines with the positive slope represent $T_K$ for two different values of $G \mu$. The upper curve is for $(G \mu)_6 = 1$, the lower curve for $(G \mu)_6 = 0.3$. Along a fixed $G \mu$ curve, the 21cm signal of a cosmic string is an emission signal above the $T_{\gamma}$ curve and an absorption signal below it. Above the $2.5 T_g$ curve, the temperature of the gas inside the wake is well approximated by the equation (\[wakeT\]), below it the initial thermal gas temperature effects dominate and the wake temperature will follow the $2.5 T_g$ curve. The wake temperature curves are for a formation redshift of $z_i + 1 = 10^3$.
![The redshift dependence of the relevant temperatures. The majenta dashed line is the CMB temperature $T_{\gamma}$, the orange dotted line is $2.5 T_g$, the gas temperature after adiabatic contraction by a factor 4 in volume, and the green (dot-dashed) line is the $T_g$ curve. The two solid curves with positive slope represent $T_K$ for two different values of $G \mu$, $(G \mu)_6 = 1$ in the case of the upper curve and $(G \mu)_6 = 0.3$ in the case of the lower curve (in both cases for a formation redshift $z_i + 1 = 10^3$). Following one of the solid curves, we see that the 21cm string signal is an emission signal above the majenta curve and an absorption signal below it. The wake temperature is given by $T_K$ only above the orange curve. For higher redshifts, the initial gas temperature dominates the final temperature of the gas inside the wake, and the latter follows the orange curve. Once $T_K$ drops below the green curve, there will no longer be any shock and our effect disappears.[]{data-label="fig:3"}](21cmtemprange3.eps){height="5cm"}
The second issue is that the string signal must be compared to what would be seen if the wake region were replaced by unclustered neutral gas. In this case, the absorption signal has a temperature given by (\[result2\]) with $T_K$ replaced by $T_g$ and the collision coefficient computed with $T_g$ instead of $T_K$. Due to the overdensity of the wake, the signal of a string wake should persist. The overdensity in the string wake by a factor of 4 will lead to an enhancement in the magnitude of the brightness temperature by a factor of 16 compared to what would be seen if the wake region were replaced by unclustered gas. One factor of 4 comes from the fact that the collision coefficient is proportional to the gas density, the second from the fact that the optical depth is also proportional to the density. In addition, the temperature ratios in (\[result2\]) are different.
However, for values of $G \mu$ so small that $T_K \ll T_g$ there will no longer be any shock and hence no well-defined wake region with overdensity 4. This will occur at a value of $G \mu$ which is smaller than the limit in (\[crit2\]) by a factor of $\sqrt{2.5}$. Note also that the thickness in redshift space (discussed below) of the signal decreases as $G \mu$ decreases, and hence improved sensitivity will be required to detect the signal.
The bottom line of the above analysis is that at a redshift of $z + 1 = 30$, strings with tensions below the current observational limit are predicted to be visible in absorption in 21cm surveys. Strings with larger tensions would yield an emission signal. The critical value (\[crit1\]) of $G \mu$ at which the emission signal turns into an absorption signal decreases as $z$ decreases, so that strings with tensions at the current observational limit would become visible in emission at a redshift of below 20. However, the value of $z$ cannot be smaller than that corresponding to reionization since we have not considered how UV scattering affects the spin temperature.
The 21cm signature of a cosmic string wake has a distinctive shape in redshift space. The string signal will be a wedge-like region of extra 21cm emission. The signal will be wide in the two angular directions, and narrow in redshift direction [^1]. The projection of this region onto the plane corresponding to the two angles in the sky (we are working in the limit of small angles and thus can use the flat sky approximation) corresponds to the projection of the wake onto the observer’s past light cone. The orientation of the wedge in redshift space is given by the orientation of the wake relative to the observer’s light cone. This is illustrated in Figure 4. The left panel is a space-time sketch of the geometry (with two spatial directions suppressed). The wake is created by a string segment which starts out at time $t_i$ at the position $x_1$ and which at time $2 t_i$ (roughly one Hubble time step later) has moved to the position $x_2$ (the arrow on the segment connecting the two events gives the direction of the velocity of the string segment. For the orientation of the string chosen, the past light cone intersects the “back” of the string wake at a later time than the front. Hence, the 21cm signal has a larger red-shift for photons from the tip of the wake than from the back side of the wake. The width of the wake vanishes at the tip and increases towards the back. Hence, in redshift space the region of extra 21cm emission has the orientation sketched on the right panel of Figure 4. The planar dimensions of the region of extra emission are in the angular directions, and their sizes are the angles corresponding to the comoving area (see (\[size\])) c\_1 t\_i (z\_i + 1) t\_i v\_s \_s (z\_i + 1) . The amplitude of the emission signal depends on the orientation of the wake relative to us (through the redshift when our past light cone intersects the various parts of the wake). This is not correlated with the direction of string motion. The width in redshift space, on the other hand, depends on the emission point on the wake - the width is larger in the back and approaches zero at the tip, as given by $\delta \nu$ (see \[three7\]).
![Geometry of the 21cm signal of a cosmic string wake. The left panel is a sketch of the geometry of the wake in space-time - vertical axis denoting time, the horizontal axis one direction of space. The string segment producing the wake is born at time $t_i$ and travels in the direction of the arrow, ending at the position $x_2$ at the time $2 t_i$. The past light cone of the observer at the time $t_0$ intersects the tip of the wake at the time $s_2$, the back of the wake at time $s_1$. These times are in general different. Hence, the 21cm radiation from different parts of the wake is observed at different red-shifts. The resulting angle-redshift signal of the string wake shown in the left panel is illustrated in the right panel, where the horizontal axis is the same spatial coordinate as in the left panel, but the vertical axis is the redshift of the 21cm radiation signal. The wedge in 21cm has vanishing thickness at the tip of the wedge, and thickness given by $\delta \nu$ at the back side.[]{data-label="fig:4"}](21cmgeometry2.eps){height="4.7cm"}
Conclusions and Discussion
==========================
In this paper we have calculated the 21cm signal of a single cosmic string wake. We found that wakes leave a characteristic signal in 21cm surveys - wedge shaped regions of extra emission or absorption, depending on the value of the string tension. There is an emission signal as long as the string tension exceeds the critical value given by (\[crit1\]) and as the tension increases the temperature approaches the value given by $200 {\rm mK}$, a value larger by almost two orders of magnitude compared to the backgrounds from standard cosmology sources which the Square Kilometer Array (SKA) telescope project is designed to be sensitive to [@Carilli]. For values of $G \mu$ below the critical value given by (\[crit1\]), the cosmic string wake signal changes to an absorption signal. This will persist down to a value of $G \mu$ a factor of $\sqrt{2.5}$ smaller than the value given in (\[crit2\]), at which point the material which is gravitationally attracted to the wake will no longer undergo a shock and hence our analysis ceases to be applicable. The width of the wedge in redshift space depends on the string tension via (\[three7\]). The width increases from zero at the tip of the string to the value given by (\[three7\]) at the midpoint of the wake. The sensitivity of 21cm surveys to cosmic strings is best at the lowest redshifts sufficiently higher than the redshift of reionization. This is due to the fact that the string wakes keep increasing in width as a function of time, leading to an increasing temperature of the gas in the wake and hence to a higher 21cm signal. At the same time, the wakes become more stable towards thermal disruption.
As mentioned above, the brightness temperature shift which we predict will be large compared to that from the surrounding intergalactic medium. The fractional frequency width of the signal is given by (\[freqwidth\]) and is near the limit of the frequency resolution of radio telescopes for values of $G \mu$ close to the current upper bound. For example, the Square Kilometer Array will have a fractional spectral resolution of the order [@Carilli] of $10^{-4}$. In both angular directions, the signal will cover a rectangle whose side length is given by the comoving Hubble radius at the redshift $z_i $, which for redshifts close to recombination corresponds to a scale of about one degree. Thus, the signal of cosmic strings which we predict is clearly a possible target for future radio telescopes. For example, the angular resolution of the SKA is designed to be smaller than $0.1$ arcsec, and the frequency range of the SKA will allow the detection of the 21cm signal up to a redshift of 20. We have focused on the signal of a single wake laid down at a time $t > t_{eq}$. The reader may worry that the signal of such a wake is masked by the “noise” due to wakes laid down at earlier times. Even though the baryons will stream out of these wakes due to their coupling with the photons, the dark matter wakes persist. For $t < t_{eq}$ the dark matter wakes do not grow in thickness until $t = t_{eq}$. After that, the thickness will begin to grow, and at $t > t_{rec}$ the baryons will start to fall into the dark matter potential wells. However, the width of such wakes (averaged over the length of the wake) is smaller than that of wakes at $t = t_{eq}$. In addition, due to their large number, on length scales of wakes laid down at $t \geq t_{eq}$, the earlier wakes will act like Gaussian noise. The coherent signature in position space of such a “late” wake can be picked out of the Gaussian noise even if the amplitude of the Gaussian noise is comparable to or slightly larger than the amplitude of the signal (as measured in terms of the contribution to the power spectrum). This has been studied in the context of picking out the signatures of late strings in CMB temperature maps in [@Amsel; @Stewart; @Rebecca1], and we expect similar conclusions to hold here.
Let us end with a brief comparison of our work with that of [@Wandelt] and [@Aaron] who also considered 21cm signals of cosmic strings. What sets our work apart is that we focus on the specific position space signature of wakes rather than on the (Fourier space) power spectrum. In computing a power spectrum, one loses the information about the non-Gaussianities in the distribution of strings. It is these non-Gaussianities which most clearly distinguish the predictions of a string model from a model with only Gaussian fluctuations. Therefore, the sensitivity to the presence of cosmic strings will be much higher in a study like ours compared to what can be achieved by only computing power spectra.
This work is supported in part by a NSERC Discovery Grant, by funds from the CRC Program, by the FQRNT Programme de recherche pour les enseignants de collège, and by a Killam Research Fellowship awarded to R.B. We wish to thank U.-L. Pen, J. Magueijo, R. Rutledge and in particular Y. Wang for useful discussions.
[99]{}
Y. B. Zeldovich, “Cosmological fluctuations produced near a singularity,” Mon. Not. Roy. Astron. Soc. [**192**]{}, 663 (1980). A. Vilenkin, “Cosmological Density Fluctuations Produced By Vacuum Strings,” Phys. Rev. Lett. [**46**]{}, 1169 (1981) \[Erratum-ibid. [**46**]{}, 1496 (1981)\]. T. W. B. Kibble, “Phase Transitions In The Early Universe,” Acta Phys. Polon. B [**13**]{}, 723 (1982);\
T. W. B. Kibble, “Some Implications Of A Cosmological Phase Transition,” Phys. Rept. [**67**]{}, 183 (1980). N. Turok and R. H. Brandenberger, “Cosmic Strings And The Formation Of Galaxies And Clusters Of Galaxies,” Phys. Rev. D [**33**]{}, 2175 (1986). H. Sato, “Galaxy Formation by Cosmic Strings,” Prog. Theor. Phys. [**75**]{}, 1342 (1986).
A. Stebbins, “Cosmic Strings and Cold Matter”, Ap. J. (Lett.) [**303**]{}, L21 (1986).
J. Magueijo, A. Albrecht, D. Coulson and P. Ferreira, “Doppler peaks from active perturbations,” Phys. Rev. Lett. [**76**]{}, 2617 (1996) \[arXiv:astro-ph/9511042\]. U. L. Pen, U. Seljak and N. Turok, “Power spectra in global defect theories of cosmic structure formation,” Phys. Rev. Lett. [**79**]{}, 1611 (1997) \[arXiv:astro-ph/9704165\]. R. Jeannerot, “A Supersymmetric SO(10) Model with Inflation and Cosmic Strings,” Phys. Rev. D [**53**]{}, 5426 (1996) \[arXiv:hep-ph/9509365\];\
R. Jeannerot, J. Rocher and M. Sakellariadou, “How generic is cosmic string formation in SUSY GUTs,” Phys. Rev. D [**68**]{}, 103514 (2003) \[arXiv:hep-ph/0308134\]. E. Witten, “Cosmic Superstrings,” Phys. Lett. B [**153**]{}, 243 (1985). S. Sarangi and S. H. H. Tye, “Cosmic string production towards the end of brane inflation,” Phys. Lett. B [**536**]{}, 185 (2002) \[arXiv:hep-th/0204074\]. E. J. Copeland, R. C. Myers and J. Polchinski, “Cosmic F- and D-strings,” JHEP [**0406**]{}, 013 (2004) \[arXiv:hep-th/0312067\]. A. C. Davis and T. W. B. Kibble, “Fundamental cosmic strings,” Contemp. Phys. [**46**]{}, 313 (2005) \[arXiv:hep-th/0505050\];\
M. Sakellariadou, “Cosmic Superstrings,” arXiv:0802.3379 \[hep-th\]. R. H. Brandenberger and C. Vafa, “Superstrings In The Early Universe,” Nucl. Phys. B [**316**]{}, 391 (1989).;\
A. Nayeri, R. H. Brandenberger and C. Vafa, “Producing a scale-invariant spectrum of perturbations in a Hagedorn phase of string cosmology,” Phys. Rev. Lett. [**97**]{}, 021302 (2006) \[arXiv:hep-th/0511140\];\
R. H. Brandenberger, A. Nayeri, S. P. Patil and C. Vafa, “String gas cosmology and structure formation,” arXiv:hep-th/0608121;\
R. H. Brandenberger, “String Gas Cosmology,” arXiv:0808.0746 \[hep-th\]. S. Furlanetto, S. P. Oh and F. Briggs, “Cosmology at Low Frequencies: The 21 cm Transition and the High-Redshift Universe,” Phys. Rept. [**433**]{}, 181 (2006) \[arXiv:astro-ph/0608032\]. N. Kaiser and A. Stebbins, “Microwave Anisotropy Due To Cosmic Strings,” Nature [**310**]{}, 391 (1984). R. J. Danos, R. H. Brandenberger and G. Holder, “A Signature of Cosmic Strings Wakes in the CMB Polarization,” arXiv:1003.0905 \[astro-ph.CO\].
J. Silk and A. Vilenkin, “Cosmic Strings And Galaxy Formation,” Phys. Rev. Lett. [**53**]{}, 1700 (1984). M. Rees, “Baryon concentrations in string wakes at $z \geq 200$: implications for galaxy formation and large-scale structure," Mon. Not. R. astr. Soc. [**[222]{}**]{}, 27p (1986).
T. Vachaspati, “Cosmic Strings and the Large-Scale Structure of the Universe,” Phys. Rev. Lett. [**57**]{}, 1655 (1986). A. Stebbins, S. Veeraraghavan, R. H. Brandenberger, J. Silk and N. Turok, “Cosmic String Wakes,” Astrophys. J. [**322**]{}, 1 (1987). J. C. Charlton, “Cosmic String Wakes and Large Scale Structure,” Astrophys. J. [**325**]{}, 52 (1988). T. Hara and S. Miyoshi, “Formation of the First Systems in the Wakes of Moving Cosmic Strings,” Prog. Theor. Phys. [**77**]{}, 1152 (1987);\
T. Hara and S. Miyoshi, “Flareup of the Universe After Z Appproximately 10\*\*2 for Cosmic String Model,” Prog. Theor. Phys. [**78**]{}, 1081 (1987);\
T. Hara and S. Miyoshi, “Large Scale Structures and Streaming Velocities Due to Open Cosmic Strings,” Prog. Theor. Phys. [**81**]{}, 1187 (1989). D. N. Spergel [*et al.*]{} \[WMAP Collaboration\], “Wilkinson Microwave Anisotropy Probe (WMAP) three year results: Implications for cosmology,” Astrophys. J. Suppl. [**170**]{}, 377 (2007) \[arXiv:astro-ph/0603449\]. T. W. B. Kibble, “Phase Transitions In The Early Universe,” Acta Phys. Polon. B [**13**]{}, 723 (1982);\
T. W. B. Kibble, “Topology Of Cosmic Domains And Strings,” J. Phys. A [**9**]{}, 1387 (1976). A. Vilenkin and E.P.S. Shellard, *Cosmic Strings and other Topological Defects* (Cambridge Univ. Press, Cambridge, 1994).
M. B. Hindmarsh and T. W. B. Kibble, “Cosmic strings,” Rept. Prog. Phys. [**58**]{}, 477 (1995) \[arXiv:hep-ph/9411342\]. R. H. Brandenberger, “Topological defects and structure formation,” Int. J. Mod. Phys. A [**9**]{}, 2117 (1994) \[arXiv:astro-ph/9310041\]. A. Albrecht and N. Turok, “Evolution Of Cosmic Strings,” Phys. Rev. Lett. [**54**]{}, 1868 (1985);\
D. P. Bennett and F. R. Bouchet, “Evidence For A Scaling Solution In Cosmic String Evolution,” Phys. Rev. Lett. [**60**]{}, 257 (1988);\
B. Allen and E. P. S. Shellard, “Cosmic String Evolution: A Numerical Simulation,” Phys. Rev. Lett. [**64**]{}, 119 (1990);\
C. Ringeval, M. Sakellariadou and F. Bouchet, “Cosmological evolution of cosmic string loops,” JCAP [**0702**]{}, 023 (2007) \[arXiv:astro-ph/0511646\];\
V. Vanchurin, K. D. Olum and A. Vilenkin, “Scaling of cosmic string loops,” Phys. Rev. D [**74**]{}, 063527 (2006) \[arXiv:gr-qc/0511159\]. J. Polchinski and J. V. Rocha, “Analytic Study of Small Scale Structure on Cosmic Strings,” Phys. Rev. D [**74**]{}, 083504 (2006) \[arXiv:hep-ph/0606205\]. L. Pogosian, S. H. H. Tye, I. Wasserman and M. Wyman, “Observational constraints on cosmic string production during brane inflation,” Phys. Rev. D [**68**]{}, 023506 (2003) \[Erratum-ibid. D [**73**]{}, 089904 (2006)\] \[arXiv:hep-th/0304188\];\
M. Wyman, L. Pogosian and I. Wasserman, “Bounds on cosmic strings from WMAP and SDSS,” Phys. Rev. D [**72**]{}, 023513 (2005) \[Erratum-ibid. D [**73**]{}, 089905 (2006)\] \[arXiv:astro-ph/0503364\]. A. A. Fraisse, “Limits on Defects Formation and Hybrid Inflationary Models with Three-Year WMAP Observations,” JCAP [**0703**]{}, 008 (2007) \[arXiv:astro-ph/0603589\];\
U. Seljak, A. Slosar and P. McDonald, “Cosmological parameters from combining the Lyman-alpha forest with CMB, galaxy clustering and SN constraints,” JCAP [**0610**]{}, 014 (2006) \[arXiv:astro-ph/0604335\];\
R. A. Battye, B. Garbrecht and A. Moss, “Constraints on supersymmetric models of hybrid inflation,” JCAP [**0609**]{}, 007 (2006) \[arXiv:astro-ph/0607339\];\
R. A. Battye, B. Garbrecht, A. Moss and H. Stoica, “Constraints on Brane Inflation and Cosmic Strings,” JCAP [**0801**]{}, 020 (2008) \[arXiv:0710.1541 \[astro-ph\]\]. N. Bevis, M. Hindmarsh, M. Kunz and J. Urrestilla, “CMB power spectrum contribution from cosmic strings using field-evolution simulations of the Abelian Higgs model,” Phys. Rev. D [**75**]{}, 065015 (2007) \[arXiv:astro-ph/0605018\];\
N. Bevis, M. Hindmarsh, M. Kunz and J. Urrestilla, “Fitting CMB data with cosmic strings and inflation,” arXiv:astro-ph/0702223. R. Battye and A. Moss, “Updated constraints on the cosmic string tension,” arXiv:1005.0479 \[astro-ph.CO\]. T. Damour and A. Vilenkin, “Gravitational wave bursts from cosmic strings,” Phys. Rev. Lett. [**85**]{}, 3761 (2000) \[arXiv:gr-qc/0004075\];\
T. Damour and A. Vilenkin, “Gravitational wave bursts from cusps and kinks on cosmic strings,” Phys. Rev. D [**64**]{}, 064008 (2001) \[arXiv:gr-qc/0104026\]. R. H. Brandenberger, “On the Decay of Cosmic String Loops,” Nucl. Phys. B [**293**]{}, 812 (1987). A. S. Lo and E. L. Wright, “Signatures of cosmic strings in the cosmic microwave background,” arXiv:astro-ph/0503120. E. Jeong and G. F. Smoot, “Search for cosmic strings in CMB anisotropies,” Astrophys. J. [**624**]{}, 21 (2005) \[arXiv:astro-ph/0406432\];\
E. Jeong and G. F. Smoot, “The Validity of the Cosmic String Pattern Search with the Cosmic Microwave Background,” arXiv:astro-ph/0612706. A. A. Fraisse, C. Ringeval, D. N. Spergel and F. R. Bouchet, “Small-Angle CMB Temperature Anisotropies Induced by Cosmic Strings,” arXiv:0708.1162 \[astro-ph\]. A. Kosowsky \[the ACT Collaboration\], “The Atacama Cosmology Telescope Project: A Progress Report,” New Astron. Rev. [**50**]{}, 969 (2006) \[arXiv:astro-ph/0608549\]. J. E. Ruhl [*et al.*]{} \[The SPT Collaboration\], “The South Pole Telescope,” Proc. SPIE Int. Soc. Opt. Eng. [**5498**]{}, 11 (2004) \[arXiv:astro-ph/0411122\]. S. Amsel, J. Berger and R. H. Brandenberger, “Detecting Cosmic Strings in the CMB with the Canny Algorithm,” JCAP [**0804**]{}, 015 (2008) \[arXiv:0709.0982 \[astro-ph\]\]. A. Stewart and R. Brandenberger, “Edge Detection, Cosmic Strings and the South Pole Telescope,” arXiv:0809.0865 \[astro-ph\]. R. J. Danos and R. H. Brandenberger, “Canny Algorithm, Cosmic Strings and the Cosmic Microwave Background,” arXiv:0811.2004 \[astro-ph\];\
R. J. Danos and R. H. Brandenberger, “Searching for Signatures of Cosmic Superstrings in the CMB,” arXiv:0910.5722 \[astro-ph.CO\]. A. Vilenkin, “Gravitational Field Of Vacuum Domain Walls And Strings,” Phys. Rev. D [**23**]{}, 852 (1981). R. Gregory, “Gravitational Stability of Local Strings,” Phys. Rev. Lett. [**59**]{}, 740 (1987). L. Perivolaropoulos, “COBE versus cosmic strings: An Analytical model,” Phys. Lett. B [**298**]{}, 305 (1993) \[arXiv:hep-ph/9208247\];\
L. Perivolaropoulos, “Statistics of microwave fluctuations induced by topological defects,” Phys. Rev. D [**48**]{}, 1530 (1993) \[arXiv:hep-ph/9212228\]. J. H. Traschen, “Causal Cosmological Perturbations And Implications For The Sachs-Wolfe Effect,” Phys. Rev. D [**29**]{}, 1563 (1984);\
J. H. Traschen, “Constraints On Stress Energy Perturbations In General Relativity,” Phys. Rev. D [**31**]{}, 283 (1985). J. C. R. Magueijo, “Inborn metric of cosmic strings,” Phys. Rev. D [**46**]{}, 1368 (1992). R. H. Brandenberger, L. Perivolaropoulos and A. Stebbins, “Cosmic Strings, Hot Dark Matter and the Large-Scale Structure of the Universe," Int. J. Mod. Phys. A [**5**]{}, 1633 (1990). L. Perivolaropoulos, R. H. Brandenberger and A. Stebbins, “Dissipationless Clustering Of Neutrinos In Cosmic String Induced Wakes,” Phys. Rev. D [**41**]{}, 1764 (1990). B. Carter, “Integrable equation of state for noisy cosmic string,” Phys. Rev. D [**41**]{}, 3869 (1990);\
A. Vilenkin, “Effect of Small Scale Structure on the Dynamics of Cosmic Strings," Phys. Rev. D [**41**]{}, 3038 (1990). A. N. Aguirre and R. H. Brandenberger, “Accretion of hot dark matter onto slowly moving cosmic strings,” Int. J. Mod. Phys. D [**4**]{}, 711 (1995) \[arXiv:astro-ph/9505031\]. Y. B. Zeldovich, “Gravitational instability: An Approximate theory for large density perturbations,” Astron. Astrophys. [**5**]{}, 84 (1970). P.J.E. Peebles, “The Large-Scale Structure of the Universe" (Princeton Univ. Press, Princeton, 1980).
A. Sornborger, R. H. Brandenberger, B. Fryxell and K. Olson, “The structure of cosmic string wakes,” Astrophys. J. [**482**]{}, 22 (1997) \[arXiv:astro-ph/9608020\]. A. Sornborger, B. Fryxell, K. Olson and P. MacNeice, “An Eulerian PPM & PIC Code for Cosmological Hydrodynamics,” arXiv:astro-ph/9608019. R. Khatri and B. D. Wandelt, “Cosmic (super)string constraints from 21 cm radiation,” Phys. Rev. Lett. [**100**]{}, 091302 (2008) \[arXiv:0801.4406 \[astro-ph\]\]. A. Berndsen, L. Pogosian and M. Wyman, “Correlations between 21 cm Radiation and the CMB from Active Sources,” arXiv:1003.2214 \[astro-ph.CO\]. C. Hogan and M. Rees, Mon. Not. R. Astron. Soc. [**188**]{}, 791 (1979);\
D. Scott and M. Rees, “The 21-cm line at high redshift: a diagnostic for the origin of large-scale structure", Mon. Not. R. Astron. Soc. [**247**]{}, 510 (1990);\
P. Madau, A. Meiksin and M. Rees, “21-cm Tomography of the Intergalactic Medium at High Redshift,” Astrophys. J. [**475**]{}, 429 (1997) \[arXiv:astro-ph/9608010\]. B. Zygelman, “Hyperfine Level-changing Collisions of Hydrogen Atoms and Tomography of the Dark Age Universe", Astrophys. J. [**622**]{}, 1356 (2005);\
S. Furlanetto and M. Furlanetto, “Spin Exchange Rates in Electron-Hydrogen Collisions,” Mon. Not. Roy. Astron. Soc. [**374**]{}, 547 (2007) \[arXiv:astro-ph/0608067\]. S. Seager, D. D. Sasselov and D. Scott, “A New Calculation of the Recombination Epoch,” Astrophys. J. [**523**]{}, L1 (1999) \[arXiv:astro-ph/9909275\];\
S. Seager, D. D. Sasselov and D. Scott, “How exactly did the Universe become neutral?,” Astrophys. J. Suppl. [**128**]{}, 407 (2000) \[arXiv:astro-ph/9912182\]. C. L. Carilli and S. Rawlings, “Science with the Square Kilometer Array - Editorial”, New Astronomy Reviews [**48**]{}, 979 (2004);\
C. L. Carilli et al, “Probing the Dark Ages with the Square Kilometer Array”, New Astronomy Reviews [**48**]{}, 1029 (2004).
[^1]: The signal at a fixed redshift would be extended only in one angular direction.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider the general form of the linear transformation for point rotation coordinate frames. The frames have the rotation axis at every point. In the transformation the frequency of one frame relative to another is not equivalent to the reverse frequency. Using symmetry of the direct and reverse transformation as well as symmetry of the frame coordinates we show that two different type of the transformation are possible. The first type is a generalization of the Lorentz transformation. This case can not be checked in optical measurements. In contrast to that the second unusual type allows us to observe consequences of the transformation in an optical experiment even though the characteristic constant inherent in the transformation is less than nuclear time of the order of 10$^{-23}$ sec. We describe the experiment.'
author:
- 'Boris V. Gisin'
title: The strange transformation for point rotation coordinate frames and its experimental verification
---
Introduction
============
The concept of a point rotation framearises in crystallooptics. Distinctive feature of the frame, in contrast to Cartesian one, is existence of the rotation axis at every point. In such a frame the axes are constructed on field amplitudes and only the axis direction is essential. Similar (non-rotating) frames have been used in quantum field theory for a long time. An example of the frame is the rotating optical indicatrix (index ellipsoid). One more difference in comparison with the Cartesian frame is the absence of centrifugal forces in the point rotation frame. The frame coordinates is an angle (phase) and time, the frequency of rotation is a parameter. In electro-optical crystals the rotation is stimulated by an applied rotating electric field. In crystals with the linear (Pockels) effect the frequency equals half (and even quarter [@rot2]) of that of the electric field. In the Kerr crystals this frequency is doubled [@Kam; @Nye]. The sense of rotation of the plane circularly polarized light wave moving through the electro-optical crystal with rotating optical indicatrix is reversed and the optical frequency is shifted if the amplitude of the applied electric field equals the half-wave value. The device for the shift by means of electro-optical crystals is the single-sideband modulator [@ssm; @jpc; @camp]. Note that optically the rotating phase plate is equivalent to the modulator but physically they are different as the plate has only one axis of rotation.
It is convenient to use for the description of circularly polarized plane light wave in the single-sideband modulator the transition to a frame with the resting optical indicatrix. Apparently, for the first time, convenience of that had been described in initial works on single-sideband modulation of light [@ssm]. Such a transition results in change of the optical frequency. The change equals the frequency of the optical indicatrix. After the polarization reversal and returning back to the initial frame the frequency deviation is doubled. Emphasize that in the frame with the resting indicatrix the modulating electric field is also at rest in spite the fact that both the indicatrix and field rotate at different frequencies relative to the initial frame?! This is one further unusual and strange property of the point rotation frames.
The transition to the rotating frame always is connected with the question what is the frequency superposition law, is it linear or not. The nonlinear law always corresponds to an extra frequency shift. Emphasize that the consideration in the framework of the Maxwell equation can not give such an extra shift. The situation is analogous to a comparison between results obtained for rectilinear move with help of the Lorentz transformation and the Newton mechanics.
In Ref. [@pra] the question was considered in assumption that the combined frequency may be presented in terms of power series of two other frequencies. It was also assumed that the frequency of one frame relative to another equals to the reverse frequency. The only difference is the sign. The negative sign corresponds to the rotation in opposite direction. In this condition the extra shift in the first approximation is proportional to the product of the optical frequency squared and the modulation frequency. The characteristic constant in the extra shift has dimension of time. In Ref. [@pra] a optical experiment was also proposed for measurement of the term and it was shown that a lower limit for measurements of the characteristic constant with such a form of the shift is about $10^{-17}$sec.
Shortly after it was shown that an analogy exists between the light propagation in medium with the rotating optical indicatrix and the motion particle in the rotating magnetic field and both the phenomena can be described in the framework of the Pauli equation [@job]. In other words a plane circularly polarized light wave propagating along the optical axis of 3-fold electro-optical crystal under the action of an applied electric field possesses properties of two-component spinor. It means that measurements of the optical frequency shift in the single-sideband modulator is similar to measurements of the magnetic moment in the magnetic resonance and anomalous magnetic moment may be associated with the nonlinear frequency shift. It was understood that the probable value of the characteristic constant is, as maximum, of the order of nuclear time $\sim 10^{-23}$ sec. [@job; @rot]. Such a small value excludes possibility to observe in optical experiments the term calculated in Ref. [@pra]
Meanwhile the immediate way to determine the frequency superposition law is the transformation for the point rotation frames. Emphasize that the Maxwell equation do not contain any information about the transformation. The transformation *must be postulated*.
In the given paper we consider general linear transformation for the point rotation frames. We use symmetry of frame coordinates and assume that the reverse frequency is a function of the direct frequency with the same function in vice versa. We show that two different types of the transformation exist. The first type is a generalization of the Lorentz transformation. The type in an experimental sense corresponds to the case of Ref. [@pra]. The second type is principally different. The type give us a chance to measure the extra term. We describe an optical experiment for the measurement of the term. The experiment keeps the main features of that in Ref. [@pra].
General linear transformation
=============================
The general form of the linear transformation for the transition from one frame to another can be written as follows $$\tilde{\varphi}=q(\varphi -\nu t),\;\;\;\tilde{t}=\frac{\tilde{q}q-1}{\tilde{%
q}\tilde{\nu}}\varphi -q\frac{\nu }{\tilde{\nu}}t, \label{trp}$$where $\varphi $ and $t$ is an angle (phase) and time, tilde corresponds to the reverse transformation $$\varphi =\tilde{q}(\tilde{\varphi}-\tilde{\nu}\tilde{t}),\;\;\;t=\frac{%
\tilde{q}q-1}{q\nu }\tilde{\varphi}-\tilde{q}\frac{\tilde{\nu}}{\nu }\tilde{t%
}, \label{trn}$$$\nu $ is the frequency of second frame relative to first one. It is obvious that Eq. (\[trp\]) turns out into Eq. (\[trn\]) if variables with tilde change to variables without tilde and vice versa.
First of all we exclude from the consideration the Galilean transformation, i.e., the case $q\equiv 1$. This case with its infinite frequencies seems unbelievable from the viewpoint of contemporary physics.
Making normalization$$\begin{aligned}
\varphi &\rightarrow &\varphi \sqrt{|\tilde{\nu}/\nu |},\text{ \ }%
t\rightarrow t,\text{ }q\rightarrow q|\nu /\tilde{\nu}|,\text{ }\nu
\rightarrow \sqrt{|\tilde{\nu}\nu |}\text{ \ \ or } \\
\varphi &\rightarrow &\varphi \sqrt{|\tilde{\nu}|},\text{ \ }t\rightarrow t%
\sqrt{|\nu |},\text{ }q\rightarrow q\sqrt{|\nu /\tilde{\nu}|},\text{ }\nu
\rightarrow \sqrt{|\tilde{\nu}\nu |},\end{aligned}$$etc., we would arrive to the case $\nu =-\tilde{\nu}$. However we can not carry out arbitrary normalization since usually $t$ and $\varphi $ already determined and connected with the space Cartesian coordinates. Generally the point rotation frames is not compatible with the Cartesian frame except when the frames are at rest. In this case if the rotation axis coincides with some Cartesian axis we may associate $\varphi $ and $t$ with the Cartesian cylindric angle and time. The above normalization would result, in particular, in the change of the speed of light. Therefore we must consider here the general case assuming that $\tilde{\nu}$ is a function of $\nu $. It means that rotations to the right and left are not equivalent in the approach.
For the point rotation frame we have not a general principle like the relativity principle for the Cartesian frames, however we use principle of symmetry instead. It means, in particular, that if $\ \tilde{\nu}(\nu )=$ $%
f(\nu )$ then $\nu (\tilde{\nu})=f(\tilde{\nu}).$
The function $\tilde{\nu}(\nu )$ and $q(\nu )\ $remain indeterminate except the condition at small $\nu $, namely, $\tilde{\nu}\rightarrow -\nu
,\;q\rightarrow 1$ if $\nu \rightarrow 0$. If the characteristic constant $%
\tau $ is of the order $\sim 10^{-23}$ sec. then the normalized frequency $%
\tau \nu $ even in microwave range is about $10^{-12}$. Therefore we assume that the function $\tilde{\nu}(\nu )$ may be expanded in the powers series in $\nu $ $$\tilde{\nu}(\nu )=-\nu +a_{2}\nu ^{2}+a_{3}\nu ^{3}+a_{4}\nu ^{4}+...,\text{
\ } \label{un}$$This expansion is compatible with the reverse expansion at certain conditions for the coefficients $a_{n},$ namely, the expansion can not contain only odd powers of $\nu $ and up to term $\nu ^{2n}$ has only $n$ independent coefficients $a_{n}$. Obviously that any expansion (\[un\])together with the reverse expansion can be written in the symmetric form with help of finite or infinite series
$$\tilde{\nu}+\nu =\sum_{n=1}b_{n}(\tilde{\nu}\nu )^{n}\equiv F(\tilde{\nu}\nu
), \label{nugen}$$
where $F$ is a function. Results below will be also valid for arbitrary $F(%
\tilde{\nu}\nu )$ with the only condition $F(0)=0$.
The main problem in the given approach is the nonlinear frequency shift. However since it is defined by product $\tilde{q}q$ we do not need to know the explicit form of function $q(\nu )$. For finding the form of $\ \tilde{q}%
q$ we use symmetry between $\varphi $ and $t$. Transformation (\[trp\]) can be written as $$\tilde{t}=Q(t-\Lambda \varphi ),\;\;\;\tilde{\varphi}=\frac{\tilde{Q}Q-1}{%
\tilde{Q}\tilde{\Lambda}}t-Q\frac{\Lambda }{\tilde{\Lambda}}\varphi ,
\label{prt}$$where role of $(\varphi ,t,q,\nu )$ is played by $(t,\varphi ,Q,\Lambda )$ respectively and
$$Q=-q\frac{\nu }{\tilde{\nu}},\text{ \ \ }\Lambda =(1-\frac{1}{q\tilde{q}})%
\frac{1}{\nu }. \label{lambda}$$
It is naturally to assume that the equality in Eq. (\[nugen\]) would keep if $\Lambda /\sigma ,\tilde{\Lambda}/\sigma $ is substituted for $\nu ,%
\tilde{\nu}$. Here $\sigma $ is a dimensional constant. Making use of the substitution and excluding $(\tilde{\nu}+\nu )$ we obtain the equation for $%
\tilde{q}q$ $$F\left( \frac{\Theta ^{2}}{\tilde{\nu}\nu }\right) -\frac{\Theta }{\tilde{\nu%
}\nu }F(\tilde{\nu}\nu )=0, \label{eqnu}$$where $\Theta =(1-1/q\tilde{q})/\sigma $. In the given case $\sigma =\pm
\tau ^{2}.$
Two type solutions of Eq. (\[eqnu\]) exist. First type is exact solution $%
\Theta =\tilde{\nu}\nu $ or $$1-\frac{1}{q\tilde{q}}=\sigma \tilde{\nu}\nu . \label{qnu}$$Transformation (\[trp\]) for this case is a generalization of the Lorentz transformation. Without loss generality we use here the term the Lorentz transformation in spate the fact that $\sigma $ may be as positive as negative. From the viewpoint of the experimental checking this case is equivalent to results of Ref. [@pra].
The second type of solutions may be presented as series$$1-\frac{1}{\tilde{q}q}=\sigma (r\tilde{\nu}\nu )^{\frac{1}{2}}+\sigma
\sum_{n=2}(r_{n}\tilde{\nu}\nu )^{\frac{1}{2}n}, \label{qng}$$where $r$ is a negative root of equation $F(r)=0$. The number of such solutions equals the number of zeros of $F(r)$. A necessary condition for existence of the solutions is $n\geq 2$ in expansion (\[nugen\]). First term in the right part of Eq. (\[qng\]) determines the frequency superposition law in the first approximation. The characteristic constant in the given case is $$\tau =\sigma \sqrt{-r}. \label{t2}$$For simplicity we use the same letter for the characteristic constant in both the types of solutions. Note that expansion (\[qng\]) is valid for small $\nu $. However at zeros of $F(\tilde{\nu}\nu )$ exact equality $%
\tilde{\nu}=-\nu $ holds, i.e., values $\nu =\pm \sqrt{|r|}$ are some distinctive points.
The second type of solutions adds to the list one further strange property of the point rotation frames. Normalization $$\text{ }\nu ^{\ast }=\sqrt{-\tilde{\nu}\nu },\text{ \ }\varphi ^{\ast
}=\varphi \frac{\nu ^{\ast }}{\nu },\text{ }t^{\ast }=t,\text{ \ }q^{\ast }=q%
\frac{\nu }{\tilde{\nu}} \label{ntr2}$$imparts the Lorentz shape to Eq. (\[trp\])$$\tilde{\varphi}^{\ast }=q^{\ast }(\varphi ^{\ast }-\nu ^{\ast }t),\;\;\;%
\tilde{t}^{\ast }=q^{\ast }(-\tau \nu ^{\ast }\varphi ^{\ast }+t).
\label{tr2}$$If $\nu ^{\ast }\rightarrow 0$ then $\varphi ^{\ast }\rightarrow \infty $. On the other hand the non-normalized transformation at $\nu \rightarrow 0$ tends to the Galilean form $$\tilde{\varphi}=\varphi ,\;\tilde{t}=\tau \varphi +t, \label{tr3}$$where $\varphi $ and $t$ switch places. Term $\tau \varphi $ is very small because of the small value of $\tau $.
In accordance with Eq. (\[tr3\]) consider the time $\Delta \tilde{t}=\tau
\Delta \varphi +\Delta t$ and angle $\Delta \tilde{\varphi}=\Delta \varphi $ intervals. The time interval measured at the same value of angle $(\Delta
\varphi =0)$ is quite determined $\Delta \tilde{t}=\Delta t$ whereas at the same time $(\Delta t=0)$ a time leap $\Delta \tilde{t}=\tau \Delta \varphi $ exists. The leap is the time of the rotation through angle $\varphi $ at frequency $1/\tau .$ Since Eq. (\[tr3\]) is the frame transformation into itself the result may be interpreted as an uncertainty of the time determination. The maximal value of the leap is $2\pi \tau $ as at $\varphi
=2\pi $ the frame also coincides with itself.
If the second type truly corresponds to physical reality then a lower limit for measurements of the characteristic constant may be drastically decreased.
Frequency superposition
=======================
Consider a plane circularly polarized light wave moving through an electro-optical crystal with the rotating optical indicatrix. The light and the indicatrix, for definiteness, are assumed to rotate in the same direction with frequencies $\omega $ and $\nu $. In correspondence with Eq. (\[trp\]) the optical frequency in the frame with the resting optical indicatrix is$$\omega ^{\prime }=\frac{\omega -\nu }{\sigma \Theta \omega /\tilde{\nu}-\nu /%
\tilde{\nu}}. \label{af}$$It is obvious that the reverse transition result to exact equality$$\frac{\omega ^{\prime }-\tilde{\nu}}{\sigma \Theta \omega ^{\prime }/\nu -%
\tilde{\nu}/\nu }\equiv \omega . \label{raf}$$However if the reversal rotation occurs then instead of $\omega ^{\prime }$ we must substitute $-\omega ^{\prime }$ in Eq. (\[raf\]). After simplification we obtain for the output frequency $$\omega _{out}=\frac{-\omega +2\nu -\sigma \Theta \omega }{-2\sigma \Theta
\omega /\nu +\sigma \Theta +1}. \label{outf}$$For the solution of the first type $1-1/\tilde{q}q=\sigma \tilde{\nu}\nu $. Taking into account that $\nu \ll \omega $ and $\tilde{\nu}\approx -\nu $ for small $\nu $ we obtain from Eq. (\[outf\]) in the first approximation$$\omega _{out}\approx -\omega +2\nu \mp 2\tau ^{2}\nu \omega ^{2}.
\label{om1}$$The extra frequency shift $2\tau ^{2}\nu \omega ^{2}$ is an equivalent of that in Ref. [@pra]. The shift can not be measured optically because the very small characteristic constant $\tau $. Even if the modulating frequency is a powerful optical wave $\nu \sim 10^{14}$ Hz then at $\tau \sim 10^{-23}$ sec. the extra shift $|2\sigma \nu \omega ^{2}|\sim 10^{-2}$ Hz can not be picked out in laser or photodetector noises.
Consider the second type of solutions with$\ 1-1/\tilde{q}q\approx \tau
\sqrt{-\tilde{\nu}\nu }$. In the first approximation$$\omega _{out}\approx -\omega +2\nu -\ 2\tau \omega ^{2}. \label{om2}$$The extra shift $2\tau \omega ^{2}$ do not depend on $\nu $, i.e., such a shift must be produced by the usual half wave plate! Emphasize that from the viewpoint the Maxwell equation the frequency shift in the single-sideband modulator is a consequence of the phase difference between two component of the electric field of light wave whereas from the viewpoint of photons it is something different.
The extra shift may be interpreted as an energy of the polarization reversal. The sign difference of the energy corresponds to the assumption on inequivalence of the right and left rotation. The shift of the second type may far exceed the shift of the first type. The relative value of the extra shift for $\tau \sim 10^{-23}$ sec. is $2\tau \omega \sim 10^{-8}$ in visible range.
Measurement of the extra shift
==============================
Schematic of the experiment for measuring the extra shift of the second type is shown in Fig.1.
+0.2in ![[]{data-label="rot2.eps"}](fig2.eps "fig:"){width=".96\linewidth"}
Linearly polarized light from laser passes through the single-sideband modulator (for example Lithium Niobate modulator [@jpc]). A electric field rotating at frequency $\Omega =2\nu $ is applied to the modulator. Evolution of the laser spectrum under change of the amplitude and frequency of the electric field may be observed by means of the scanning interferometer. The linearly polarized light is a sum of two circularly polarized waves of frequencies $\omega $ and $-\omega $. The modulator changes the frequencies to $\omega +\Omega -2\tau \omega ^{2}$ and $-\omega
+\Omega -2\tau \omega ^{2}$ respectively. After the paralyzer light is modulated in intensity at frequency $2\Omega -4\tau \omega ^{2}$. The extra shift $4\tau \omega ^{2}$ could be extracted by heterodyning as it is in Ref. [@pra]. However in the given case the schematic is simplified (heterodyne and doubler are crossed out in Fig. 1) since resonance $\Omega
=2\tau \omega ^{2}$ can be used with matching the sign of $\tau $ by the reversal of the applied electric field. The shift can be measured with confidence if the characteristic constant is about $10^{-23}$ sec. Note that if the extra shift really exists then similar schematic may be effectively used for precision measurements in spectroscopy.
Conclusion
==========
The idea of inequivalence of the direct and reverse frequencies and the principle symmetry leads to two types of the transformation for the point rotation frames. The first type is a generalization of the Lorentz transformation. The type seems more applicable to the Cartesian frames. Measurements of the extra frequency shift in this case lies beyond possibilities of optics. The second type is applicable only to the point rotation frames. An unusual and strange property of this type is uncertainty of the time determination. The extra shift in this case may be verified in optical measurements if the characteristic constant in the transformation is about (and even less) $10^{-23}$ sec.
Another question arises in the above construction: is parity violation connected with the inequivalence of the direct and reverse frequency?
[9]{} B. V. Gisin, Kristallografia, 37, 218 (1992) \[Sov. Phys. Crystallogr. 37(1), (1992)\].
L. P. Kaminov, $\emph{An}$ $\emph{introduction}$ $\emph{to}$ $%
\emph{electrooptical}$ $\emph{devices}$ (Academic press, New York, 1974).
J. F. Nye, *Physical Properties of Crystals* (University Press, London, 1964).
D. H. Baird and C. F. Buhrer, *Single-Sideband Light Modulator* (U.S. Patent 3204104, 1965); C. F. Buhrer, D. H. Baird, and E. M. Conwell, Appl. Phys. Lett. **1**, 46 (1962).
J. P. Campbell, and W. H. Steier, IEEE J. Quantum Electron. **QE-7**, 450 (1971).
B. V. Gisin, *Electro-optical single-sideband modulators: problems and applications*, Selected Papers from Photonics-98, SPIE Proceedings, **3666**, 132 (1999).
B. V. Gisin, Phys. Rev. **A 50**, 2003 (1994).
B. V. Gisin, Phys. Lett. A **209**, 285 (1995).
B. V. Gisin, Phys. Rev. **A 61**, 53808 (2000).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Skew-symmetric differential forms play an unique role in mathematics and mathematical physics. This relates to the fact that closed exterior skew-symmetric differential forms are invariants. The concept of “Exterior differential forms" was introduced by E.Cartan for a notation of integrand expressions, which can create the integral invariants. (The existence of integral invariants was recognized by A. Poincare while studying the general equations of dynamics.)
All invariant mathematical formalisms are based on invariant properties of closed exterior forms. The invariant properties of closed exterior forms explicitly or implicitly manifest themselves essentially in all formalisms of field theory, such as the Hamilton formalism, tensor approaches, group methods, quantum mechanics equations, the Yang-Mills theory and others. They lie at the basis of field theory.
However, in this case the question of how the closed exterior forms are obtained arises. In present work it is shown that closed exterior forms, which possess the invariant properties, are obtained from skew-symmetric differential forms, which, as contrasted to exterior forms, are defined on nonintegrable manifolds. The process of generating closed exterior forms describes the mechanism of realization of invariants and invariant structures.
---
[**Skew-symmetric differential forms. Invariants.**]{}
[**Realization of invariant structures.**]{}
*L. I. Petrova*
Closed exterior skew-symmetric differential forms: Invariants. Invariant structures.
====================================================================================
Distinguishing properties of the mathematical apparatus of exterior differential forms were formulated by Cartan \[1\]: “…I wanted to build the theory, which contains concepts and operations being [*independent of any change of variables both dependent and independent*]{}; to do so it is necessary to change [*partial derivatives by differentials*]{} that have [*interior meaning*]{}.”
Some foundations of closed exterior differential forms
------------------------------------------------------
The exterior differential form of degree $p$ ($p$-form) on integrable manifold can be written as \[2,3\] $$\theta^p=\sum_{i_1\dots i_p}a_{i_1\dots i_p}dx^{i_1}\wedge
dx^{i_2}\wedge\dots \wedge dx^{i_p}\quad 0\leq p\leq n\eqno(1)$$ Here $a_{i_1\dots i_p}$ are functions of variables $x^{i_1}$, $x^{i_2}$, …, $x^{i_n}$, $n$ is the dimension of space, $\wedge$ is the operator of exterior multiplication, $dx^i$, $dx^{i}\wedge dx^{j}$, $dx^{i}\wedge dx^{j}\wedge dx^{k}$, …is the local basis which satisfies the condition of exterior multiplication: $$\begin{array}{l}
dx^{i}\wedge dx^{i}=0\\
dx^{i}\wedge dx^{j}=-dx^{j}\wedge dx^{i}\quad i\ne j
\end{array}$$ \[In further presentation the symbol of summing $\sum$ and the symbol of exterior multiplication $\wedge$ will be omitted. Summation over repeated indices is implied.\]
An exterior differential form is called a closed one if its differential is equal to zero: $$d\theta^p=0\eqno(2)$$ From condition (2) one can see that the closed form is a conservative quantity. This means that this can correspond to the conservation law, namely, to some conservative physical quantity.
The differential of the form is a closed form. That is $$dd\omega=0$$ where $\omega$ is an arbitrary exterior form.
The form which is the differential of some other form: $$\theta^p=d\omega\eqno(3)$$ is called the exact form. The exact forms prove to be closed automatically $$d\theta^p=dd\omega=0\eqno(4)$$
Here it is necessary to pay attention to the following points. In the above presented formulas it was implicitly assumed that the differential operator $d$ is a total one (that is, the operator $d$ acts everywhere in the vicinity of the point considered). However, the differential may be internal. Such a differential acts on some structure with the dimension being less than that of the initial manifold.
If the exterior form is closed only on structure, the closure condition is written as $$d_\pi\theta^p=0\eqno(5)$$ In this case the structure $\pi$ obeys the condition $$d_\pi{}^*\theta^p=0\eqno(6)$$ where ${}^*\theta^p$ is the dual form.
Such an exterior form is called the closed inexact form.
The structure, on which the exterior differential form may become a closed (inexact) form, is a pseudostructure with respect to its metric properties.
From conditions (5) and (6) one can see that the form closed on pseudostructure is a conservative object, namely, this quantity conserves on pseudostructure. This can also correspond to some conservation law, i.e. to conservative object.
Pseudostructures {#pseudostructures .unnumbered}
----------------
As one can see from condition (6), the structure, on which a closed (inexact) form is defined, is described by dual form. The dual form is a closed metric form of this structure.
To understand the properties of such structure, one can use the correspondence between the exterior differential form and skew-symmetric tensor. It is known that the skew-symmetric tensors correspond to closed exterior differential forms, and the pseudotensors correspond to relevant dual forms. This points to the fact that the structures, on which closed inexact forms are defined, are pseudostructures.
The characteristics, integral surfaces, surfaces of potential (of simple layer, double layer), sections of cotangent bundles (Yang-Mills fields), cotangent manifold, eikonals, cohomologies by de Rham, singular cohomologies, the pseudo-Riemann and pseudo-Euclidean spaces and others can be regarded as examples of pseudostructures and pseudospaces, on which closed inexact forms are defined.
It should be emphasized that the pseudostructure and corresponding closed inexact form make up a differential - geometrical structure. As it will be shown below, such a differential - geometrical structure proves to be an invariant structure.
Differentials {#differentials .unnumbered}
-------------
The exact form is, by definition, a differential (see condition (3)). In this case the differential is total. The closed inexact form is a differential too. And in this case the differential is an interior one defined on pseudostructure. Thus, any closed form is a differential. The exact form is a total differential. The closed inexact form is an interior (on pseudostructure) differential, that is $$\theta^p_\pi=d_\pi\omega\eqno(7)$$ At this point it is worth noting that the total differential of the form closed on pseudostructure is nonzero, that is $$dd_\pi\omega\ne0\eqno(8)$$
Invariants. Invariant structures.
----------------------------------
Since the closed form is a differential (a total one if the form is exact, or an interior one on the pseudostructure if the form is inexact), it is obvious that the closed form proves to be invariant under all transformations that conserve the differential. The unitary transformations (0-form), the tangent and canonical transformations (1-form), the gradient and gauge transformations (2-form) and so on are examples of such transformations.
[*These are gauge transformations for spinor, scalar, vector, tensor fields*]{}. It can be pointed out that just such transformations are used in field theory.
As mentioned above, from the closure conditions it follows that the closed form is a conservative quantity. As the result, the closed form is a conservative invariant quantity. This property of closed forms plays an essential role in describing the conservation laws and lies at the basis of field theory. The covariance of dual form is directly connected with the invariance of exterior closed inexact form.
Invariant structures {#invariant-structures .unnumbered}
--------------------
The closed inexact exterior forms are of most significance in mathematical formalisms and mathematical physics. This is due to the fact that the closed exterior form and relevant dual form describe the differential-geometrical structure, which is invariant one.
From the definition of closed inexact exterior form one can see that to this form there correspond two conditions:
\(1) condition (5) is a closure condition of exterior form itself, and
\(2) condition (6) is that of dual form.
Conditions (5) and (6) can be regarded as equations for a binary object that combines the pseudostructure (dual form) and the conservative quantity (the exterior differential form) defined on this pseudostructure. Such a binary object is a differential - geometrical structure. (The well-known G-Structure is an example of such differential-geometrical structure.)
As it has been already pointed out, the closed inexact exterior form is a differential (an interior one on pseudostructure), and hence it remains invariant under all transformations that conserve the differential. Therefore, the relevant differential-geometrical structure also remains invariant under all transformations that conserve differential. For the sake of convenience in subsequent presentation such differential - geometrical structures will be called the Inv. Structures.
To an unique role of such invariant structures in mathematics it points the fact that the transformations conserving the differential (unitary, tangent, canonical, gradient and gauge ones) lie at the basis of many branches of mathematics, mathematical physics and field theory. The differential-geometrical structures made up of characteristics and integral curves of differential equations and relevant conditions on those are examples of Inv. Structures.
As it will be shown in Section 3 of present paper, the Inv. Structures are of unique importance in mathematical physics and field theory. The physical structures, of which physical fields are made up, are such invariant structures.
It should be emphasized ones more that the Inv. Structure is a differential-geometrical structure. That is not a spatial structure. The spatial structure is described by [*exact* ]{} exterior form, whereas the invariant structure is described by [*inexact* ]{} exterior form.
Invariance as the result of conjugacy of elements of exterior or dual forms
---------------------------------------------------------------------------
The closure of exterior differential forms, and hence their invariance, results from the conjugacy of elements of exterior or dual forms.
From the definition of the exterior differential form one can see that exterior differential forms have complex structure. The specific features of the exterior form structure are the homogeneity with respect to the basis, skew-symmetry, the integration of terms each of which made up by two objects of different nature (the algebraic nature of the form coefficients, and the geometric nature of the base components). Besides, the exterior form depends on the space dimension and on the manifold topology. The closure property of exterior form means that any objects, namely, elements of exterior form, components of elements, elements of the form differential, exterior and dual forms and others, turn out to be conjugated. The variety of objects of conjugacy leads to the fact that closed forms can describe a great number of different invariant structures.
Identical relations of exterior differential forms, description of conjugacy and invariance
-------------------------------------------------------------------------------------------
Since the conjugacy is a certain connection between two operators or mathematical objects, it is evident that the relations can be used to express conjugacy mathematically. Identical relations of exterior differential forms disclose also the properties of Inv. Structure.
At this point it should be emphasized the following. The relation is a comparison, i.e. a correlation of two objects. The relation may be identical or nonidentical. The basis of mathematical apparatus of exterior differential forms is made up of identical relations. (Below nonidentical relations will be presented, and it will be shown that identical relations for exterior differential forms are obtained from nonidentical relations. Also it will be shown that transitions from nonidentical relations to identical ones describe the realization of invariant structures.)
The identical relations of exterior differential forms reflect the closure conditions of differential forms, namely, vanishing the form differential (see formulas (2), (5) and (6)) and the condition that the closed differential form is a differential (see formulas (3) and (7)). All these conditions are the expression of conjugacy and invariance.
One can distinguish several types of identical relations.
1\. [*Relations in differential forms*]{}.
They correspond to formulas (3) and (7). The examples of such identical relations are
\(a) the Poincare invariant $ds\,=\,-H\,dt\,+\,p_j\,dq_j$,
\(b) the second principle of thermodynamics $dS\,=\,(dE+p\,dV)/T$,
\(c) the vital force theorem in theoretical mechanics: $dT=X_idx^i$ where $X_i$ are the components of potential force, and $T=mV^2/2$ is the vital force,
\(d) the conditions on characteristics in the theory of differential equations.
The requirement that the function is an antiderivative (the integrand is a differential of a certain function) can be written in terms of such an identical relation.
The existence of harmonic function is written by means of identical relation: the harmonic function is a closed form, that is, a differential (a differential on the Riemann surface).
In general form such an identical relation can be written as $$d\phi=\theta^p\eqno(10)$$ In this relation the form in the right-hand side has to be a [*closed* ]{} one.
As it will be shown below, the identical relations are satisfied only on pseudostructures. That is, the identical relation can be written as $$d _{\pi}\phi=\theta _{\pi}^p\eqno(11)$$
Identical relations (10) and (11) are the proof that the closed exterior form is a differential, and hence, this form is an invariant with respect to all transformations that conserve the differential.
Identical relations occur in various branches of mathematics and mathematical physics. Identical relations can be of another type, namely, integral, tensor and others. And all identical relations are an analog to the identical relation in differential forms.
All identical relations correspond to invariant structures.
It would be noted some another types of identical relations.
2\. [*Integral identical relations*]{}.
At the beginning of the paper it was pointed out that the exterior differential forms were introduced as integrand expressions possessing the following property: they can have integral invariants. This fact (the availability of integral invariant) is mathematically expressed as a certain identical relation.
The formulas by Newton, Leibnitz and Green, the integral relations by Stokes and Gauss-Ostrogradskii are examples of integral identical relations.
3\. [*Tensor identical relations*]{}.
From the relations that connect exterior forms of consequent degrees one can obtain the vector and tensor identical relations that connect the operators of gradient, curl, divergence and so on.
From the closure conditions of exterior and dual forms one can obtain the identical relations such as the gauge relations in electromagnetic field theory, the tensor relations between connectednesses and their derivatives in gravitation (the symmetry of connectednesses with respect to lower indices, the Bianchi identity, the conditions imposed on the Christoffel symbols) and so on.
4\. [*Identical relations between derivatives*]{}.
The identical relations between derivatives correspond to the closure conditions of exterior and dual forms. The examples of such relations are the above presented Cauchi-Riemann conditions in the theory of complex variables, the transversality condition in the calculus of variations, the canonical relations in the Hamilton formalism, the thermodynamic relations between derivatives of thermodynamic functions, the condition that the derivative of implicit function is subject to, the eikonal relations and so on.
The importance of identical relations is manifested by the fact that practically in all branches of physics, mechanics, thermodynamics one faces such identical relations.
The functional significance of identical relations for exterior differential forms lies in the fact that they can describe the conjugacy of objects that have different mathematical meaning and different physical nature. This enables one to see internal connections between various branches of mathematics and physics. Due to these possibilities the exterior differential forms, and correspondingly, the Inv. Structures, have wide application in various branches of mathematics and mathematical physics.
Identical relations possess the duality that discloses the significance of invariant structures. The availability of differential in the left-hand side points to the availability of potential or state function, and the availability of closed inexact form points to that there is an invariant structure. Below it will be shown that such a relation has a deep physical sense.
Realization of invariant structures
===================================
The mechanism of realization of invariant structures is described by skew-symmetric differential forms, which, in contrast to exterior forms, are defined on deforming nonintegrable manifolds (see Appendix of work \[4\]). Such skew-symmetric differential forms possess the evolutionary properties. The evolutionary forms possess a peculiarity, namely, the closed inexact exterior forms are obtained from them. This elucidates the process of realization of invariant structures.
Some properties of evolutionary forms
-------------------------------------
The evolutionary skew-symmetric differential forms are obtained from differential equations that describe any processes.
Examples of nonintegrable manifolds, on which the evolutionary skew-symmetric differential forms are defined, are the tangent manifolds of differential equations, the Lagrangian manifolds, the manifolds constructed of trajectories of material medium particles and so on. These are manifolds with unclosed metric forms. The metric form differential, and correspondingly its commutator, are nonzero. (The commutators of metric forms of such manifolds describe the manifold deformation: torsion, curvature and others).
The specific feature of evolutionary forms, i.e skew-symmetric forms defined on deforming manifolds, is the fact that evolutionary forms are unclosed ones. Since the basis of evolutionary form changes, the evolutionary form differential includes the nonvanishing differential of manifold metric form due to differentiating the basis. Therefore, the evolutionary form differential cannot be equal to zero. Hence, the evolutionary form, in contrast to the case of exterior form, cannot be closed. This leads to that in the mathematical apparatus of evolutionary forms there arise new nonconventional elements like nonidentical relations and degenerate transformations that allow to describe the generation of closed inexact exterior forms and the realization of invariant structures.
The nonidentical relations of evolutionary forms can be written as $$d\phi=\eta^p\eqno(12)$$ Here $\eta^p$ is the $p$-degree evolutionary form being unclosed, $\phi$ is some form of degree $(p-1)$, and the differential $d\phi$ is a closed form of degree $p$.
The form differential, i.e. a closed form being an invariant object, appears in the left-hand side of this relation. In the right-hand side it is appeared the unclosed form, which is not an invariant object. Such a relation cannot be identical one.
One can see the difference of relations for exterior forms and evolutionary ones. In the right-hand side of identical relation (see relation (10)) it is appeared the closed form, whereas the form in the right-hand side of nonidentical relation (12) is an unclosed one.
Nonidentical relations are obtained while describing any processes. A relation of such type is obtained while, for example, analyzing the integrability of the partial differential equation. The equation is integrable if it can be reduced to the form $d\phi=dU$. However, it appears that, if the equation is not subject to an additional condition (the integrability condition), it is reduced to the form (12), where $\eta^p$ is an unclosed form and it cannot be written as a differential.
Nonidentical relations of evolutionary forms are evolutionary relations because they include the evolutionary form. Such nonidentical evolutionary relations appear to be selfvarying ones. The variation of any object of the relation in some process leads to variation of another object and, in turn, the variation of the latter leads to variation of the former. Since one of the objects is a noninvariant (i.e. unmeasurable) quantity, the other cannot be compared with the first one, and hence, the process of mutual variation cannot be completed.
The nonidentity of evolutionary relation is connected with the nonclosure of evolutionary form, that is, it is connected with the fact that the evolutionary form commutator is nonzero. The evolutionary form commutator includes two terms. The first term specifies the mutual variations of evolutionary form coefficients, and the second term (the metric form commutator) specifies the manifold deformation. These terms have a different nature and cannot make the commutator to be vanishing. In the process of selfvariation of nonidentical evolutionary relation the exchange between the terms of evolutionary relation proceeds and this is realized according to the evolutionary relation. The evolutionary form commutator describes the quantity that is a moving force of evolutionary process and leads to realization of differential-geometrical structures.
The process of the evolutionary relation selfvariation plays a governing role in description of evolutionary processes.
The significance of the evolutionary relation selfvariation consists in the fact that in such a process it can be realized conditions under which the closed inexact form is obtained from the evolutionary form and the identical relation is obtained from the nonidentical relation. These are conditions of degenerate transformation. Since the evolutionary form differential is nonzero, whereas the closed exterior form differential is zero, the transition from the evolutionary form to closed exterior form is allowed only under [*degenerate transformation*]{}. The conditions of vanishing the dual form differential are conditions of degenerate transformation.
These are such conditions that can be realized under selfvariation of the nonidentical evolutionary relation.
Realization of closed inexact exterior form. Derivation of invariant structures
-------------------------------------------------------------------------------
To obtain the differential-geometrical structure, it is necessary to obtain the closed inexact exterior form, i.e. the form closed on pseudostructure.
To the pseudostructure it is assigned the closed dual form (whose differential vanishes). For this reason the transition from the evolutionary form to closed inexact exterior form proceeds only when the conditions of vanishing the dual form differential are realized, in other words, when the metric form differential or commutator becomes equal to zero.
The conditions of degenerate transformation are conditions of vanishing the dual form differential. That is, it is the condition of realization of pseudostructure. And this leads to realization of closed inexact exterior form.
As it has been already mentioned, the evolutionary differential form $\eta^p$ involved into nonidentical relation (12) is an unclosed one. The commutator, and hence the differential, of this form is nonzero. That is, $$d\eta^p\ne 0 \eqno(13)$$ If the conditions of degenerate transformation are realized, then from the unclosed evolutionary form one can obtain the differential form closed on pseudostructure. The differential of this form equals zero. That is, it is realized the transition $d\eta^p\ne 0 \to $ (degenerate transformation) $\to d_\pi{}^*\eta^p=0$, $d_\pi \eta^p=0$.
The relations obtained $$d_\pi \eta^p=0, d_\pi{}^*\eta^p=0 \eqno(14)$$ are closure conditions for exterior inexact form, and this points to realization of exterior form closed on pseudostructure, that is, this points to origination of the differential-geometrical invariant structure.
Vanishing on pseudostructure the exterior form differential (that is, vanishing on pseudostructure the interior differential of the evolutionary form) points to that the exterior inexact form is a conservative quantity in the direction of pseudostructure. However, in the direction normal to pseudostructure this quantity exhibits the discontinuity. The value of such discontinuity is defined by the value of the evolutionary form commutator being nonzero. This argues to discreteness of the differential-geometrical structures.
Thus, while selfvariation of the evolutionary nonidentical relation the dual form commutator can vanish. This means that it is made up the pseudostructure on which the differential form turns out to be closed. The emergence of the form being closed on pseudostructure points out to origination of invariant structures.
On the pseudostructure $\pi$ from evolutionary relation (12) it follows the relation $$d_\pi\psi=\omega_\pi^p\eqno(15)$$ which proves to be an identical relation. Indeed, since the form $\omega_\pi^p$ is a closed one, on the pseudostructure this form turns out to be the differential of some differential form. In other words, this form can be written as $\omega_\pi^p=d_\pi\theta$. Relation (15) is now written as $$d_\pi\psi=d_\pi\theta$$ There are differentials in the left-hand and right-hand sides of this relation. This means that the relation is an identical one.
From evolutionary nonidentical relation (12) it is obtained the identical on pseudostructure relation. In this case the evolutionary relation itself remains to be nonidentical one. (At this point it should be emphasized that differential, which equals zero, is an interior one. The evolutionary form commutator becomes zero only on pseudostructure. The total evolutionary form commutator is nonzero. That is, under degenerate transformation the evolutionary form differential vanishes only [*on pseudostructure*]{}. The total differential of evolutionary form is nonzero. The evolutionary form remains to be unclosed.)
It can be shown that all identical relations of the exterior differential form theory are obtained from nonidentical relations (that contain evolutionary forms) by applying degenerate transformations.
Thus, the mathematical apparatus of evolutionary differential forms can describe the process of generation of closed inexact exterior differential forms, and this discloses the process of origination of invariant structures.
The process of generation of closed inexact exterior differential forms and the origination of invariant structures are processes of conjecting the operators. To the closed exterior form there correspond conjugated operators, whereas to the evolutionary form there correspond nonconjugated operators. The transition from evolutionary form to closed exterior form and the origination of differential-geometrical structures is a transition from nonconjugated operators to conjugated ones. This is expressed mathematically as the transition from nonzero differential (the evolutionary form differential is nonzero) to the differential that equals zero (the closed exterior form differential equals zero).
It can be seen that the process of conjugating the objects and obtaining the differential-geometrical structures is a mutual exchange between the quantities of different nature (for example, between the algebraic and geometric quantities, between the physical and spatial quantities) and vanishing some functional expressions (Jacobians, determinants and so on).
Characteristics of Inv. Structure {#characteristics-of-inv.-structure .unnumbered}
---------------------------------
Since the closed exterior differential form, which corresponds to the Inv. Structure emerged, was obtained from evolutionary form that enters to the nonidentical relation, it is evident that the Inv. Structure characteristics must be connected with those of the evolutionary form and of the manifold on which this form is defined, as well as the conditions of degenerate transformation and the values of commutators of the evolutionary form and the manifold metric form.
The conditions of degenerate transformation, as it was said before, determine the pseudostructures. The first term of the evolutionary form commutator determines the value of discrete change (the quantum), which the quantity conserved on the pseudostructure undergoes under transition from one pseudostructure to another. The second term of the evolutionary form commutator specifies the characteristics that fixes the character of initial manifold deformation, which took place before the Inv. Structure had been arisen. (Spin is such an example).
The discrete (quantum) change of a quantity proceeds in the direction that is normal (more exactly, transverse) to the pseudostructure. Jumps of the derivatives normal to potential surfaces are examples of such changes.
The connection of Inv. Structure with the skew-symmetric differential forms allows to introduce the classification of Inv. Structures in its dependence on parameters that specify the skew-symmetric differential forms and enter into nonidentical and identical relation of skew-symmetric differential forms. To determine these parameters one has to consider the problem of integration of nonidentical evolutionary relation.
Under degenerate transformation from the nonidentical evolutionary relation one obtains the relation being identical on pseudostructure. Since the right-hand side of such a relation can be expressed in terms of differential (as well as the left-hand side), one obtains the relation that can be integrated, and as the result one obtains the relation with differential forms of less by one degree.
The relation obtained after integration proves to be nonidentical as well.
The resulting nonidentical relation of degree $(p-1)$ (relation that includes the forms of degree $(p-1)$) can be integrated once again if the corresponding degenerate transformation has been realized and the identical relation has been formatted.
By sequential integrating the evolutionary relation of degree $p$ (in the case of realization of corresponding degenerate transformations and formatting the identical relation), one can get closed (on the pseudostructure) exterior forms of degree $k$, where $k$ ranges from $p$ to $0$.
In this case one can see that after such integration the closed (on pseudostructure) exterior forms, which depend on two parameters, are obtained. These parameters are the degree of evolutionary form $p$ (in the evolutionary relation) and the degree of created closed forms $k$.
In addition to these parameters, another parameter appears, namely, the dimension of space. If the evolutionary relation generates the closed forms of degrees $k=p$, $k=p-1$, …, $k=0$, to them there correspond the pseudostructures of dimensions $(n+1-k)$, where $n$ is the space dimension.
The invariant structures are of unique significance in mathematical physics and field theory. The physical structures that made up physical fields are such Inv. Structures.
As it will be shown below, the mechanism of realization of Inv. Structures, which correspond to physical fields, describes the mechanism of generation of physical structures. This discloses the physical meaning of Inv. Structures.
Physical meaning of invariant structures. Mechanism of generation of physical structures.
=========================================================================================
As it has been already pointed out, the invariant structures are realized while analyzing the integrability of differential equations. Their role in the theory of differential equations relates to the fact that they correspond to generalized solutions which describe measurable physical quantities. In this case the integral surfaces with conservative quantities (like the characteristics, the characteristic surfaces, potential surfaces and so on) are invariant structures. The examples of such studying the integrability of differential equations using the skew-symmetric differential forms are presented in paper \[5\].
The unique results are obtained in studying the differential equations that describe the conservation laws for material media. The Inv. Structures that correspond to physical structures are obtained under investigation of these equations.
The properties of conservation laws are at the basis of the process of physical structure emergence. Therefore it is necessary to call attention to some properties and peculiarities of conservation laws.
Properties and peculiarities of conservation laws.
--------------------------------------------------
From the closure condition of exterior form it follows that the closed inexact differential form is a conservative quantity on some pseudostructure. From this one can see that the closed inexact exterior differential form can correspond to conservation law. The conservation laws for physical fields are just such conservation laws.
One can see that Inv. Structures made up by closed inexact form and relevant dual form correspond to conservation laws for physical fields.
The evolutionary skew-symmetric forms, from which, as it has been shown, the closed inexact forms are obtained, correspond to conservation laws as well. However, these are conservation laws for material systems (material media). In contrast to conservation laws for physical fields, they are balance conservation laws (they establish the balance between the variation of physical quantities and external actions to the system) and are described by differential equations. .
The conservation laws for material systems are conservation laws for energy, linear momentum, angular momentum, and mass.
The invariant structures corresponding to physical fields are obtained from the equations that describe balance conservation laws for material media.
Analysis of the equations of conservation laws for material systems. {#analysis-of-the-equations-of-conservation-laws-for-material-systems. .unnumbered}
--------------------------------------------------------------------
The balance conservation laws for energy, linear momentum, angular momentum, and mass are described by partial differential equations \[6\]. (On examination of the integrability of these equations it is obtained the nonidentical relation that includes evolutionary form. From such evolutionary form the closed inexact forms and invariant structures corresponding to physical structures are obtained.)
Let us analyze the equations that describe the balance conservation laws for energy and linear momentum.
In the accompanying frame of reference (this system is connected with the manifold made up by the trajectories of material system elements) the equations for energy and linear momentum are written as $${{\partial \psi }\over {\partial \xi ^1}}\,=\,A_1 \eqno(16)$$ $${{\partial \psi}\over {\partial \xi^{\nu }}}\,=\,A_{\nu },\quad \nu \,=\,2,\,...\eqno(17)$$
Here $\psi$ is the functional specifying the state of material system (the action functional, entropy, wave function can be regarded as examples of such a functional), $\xi^1$ is the coordinate along the trajectory, $\xi ^{\nu }$ are the coordinates in the direction normal to trajectory, $A_1$ is the quantity that depends on specific features of material system and on external energy actions onto the system, and $A_{\nu }$ are the quantities that depend on specific features of material system and on external force actions.
Eqs. (16) and (17) can be convoluted into the relation $$d\psi\,=\,A_{\mu }\,d\xi ^{\mu },\quad (\mu\,=\,1,\,\nu )\eqno(18)$$ where $d\psi $ is the differential expression $d\psi\,=\,(\partial \psi /\partial \xi ^{\mu })d\xi ^{\mu }$.
Relation (18) can be written as $$d\psi \,=\,\omega \eqno(19)$$ here $\omega \,=\,A_{\mu }\,d\xi ^{\mu }$ is the skew-symmetric differential form of first degree.
The relation obtained is an evolutionary relation.
Relation (19) was obtained from the equations of balance conservation laws for energy and linear momentum. In this relation the form $\omega $ is that of first degree. If the equations of balance conservation laws for angular momentum be added to the equations for energy and linear momentum, this form in the evolutionary relation will be a form of second degree. And in combination with the equation of balance conservation law for mass this form will be the form of degree 3.
Thus, in general case the evolutionary relation can be written as $$d\psi \,=\,\omega^p \eqno(20)$$ where the form degree $p$ takes the values $p\,=\,0,1,2,3$.. (The evolutionary relation for $p\,=\,0$ is similar to that in differential forms, and it was obtained from the interaction of energy and time.)
The relations (19) and (20) are nonidentical evolutionary relations.
Let us show that the relation obtained from the equations of balance conservation laws proves to be nonidentical.
To do so we shall analyze relation (19).
In the left-hand side of relation (19) there is the differential that is a closed form. This form is an invariant object. The right-hand side of relation (20) contains the differential form $\omega$, which is not an invariant object since in real processes, as it will be shown below, this form proves to be unclosed. The commutator of this form is nonzero. The components of commutator of the form $\omega \,=\,A_{\mu }d\xi ^{\mu }$ can be written as follows: $$K_{\alpha \beta }\,=\,\left ({{\partial A_{\beta }}\over {\partial \xi ^{\alpha }}}\,-\,
{{\partial A_{\alpha }}\over {\partial \xi ^{\beta }}}\right )$$ (here the term connected with the manifold metric form has not yet been taken into account).
The coefficients $A_{\mu }$ of the form $\omega $ have been obtained either from the equation of balance conservation law for energy or from that for linear momentum. This means that in the first case the coefficients depend on the energetic action and in the second case they depend on the force action. In actual processes energetic and force actions have different nature and appear to be inconsistent. The commutator of the form $\omega $ made up of the derivatives of such coefficients is nonzero. This means that the differential of the form $\omega $ is nonzero as well. Thus, the form $\omega$ proves to be unclosed and cannot be a differential like the left-hand side.
This means that relation (19), as well as relation (20), cannot be identical ones. In such a way it can be shown that relation (20) is nonidentical as well.
Thus, the nonidentity of evolutionary relation means that the balance conservation law equations are inconsistent. And this indicates that the balance conservation laws are noncommutative. (If the balance conservation laws be commutative, the equations would be consistent and the evolutionary relation would be identical).
The noncommutativity of balance conservation laws is a moving force of evolutionary processes that proceed in material medium and lead to emergence of physical structures. This follows from the further analysis of the equations of balance conservation laws. The invariant structures obtained from these equations correspond to such physical structures.
Mechanism of generation of physical structures.
-----------------------------------------------
The relation obtained from the equations of balance conservation laws involves the functional that specifies the material system state. However, since this relation turns out to be not identical, from this relation one cannot get the differential $d\psi $ that could point out to the equilibrium state of material system. The absence of differential means that the system state is nonequilibrium. That is, in material system the internal force acts.
As it has been already shown, the nonidentical evolutionary relation turns out to be a selfvarying relation.
Selfvariation of the nonidentical evolutionary relation points to the fact that the nonequilibrium state of material system turns out to be selfvarying. It is evident that this selfvariation proceeds under the action of internal force whose quantity is described by commutator of the unclosed evolutionary form $\omega^p $. (If the commutator be zero, the evolutionary relation would be identical, and this would point to the equilibrium state, i.e. the absence of internal forces.) Everything that gives the contribution into the commutator of the form $\omega^p $ leads to emergence of internal force.
Above it has been shown that under degenerate transformation from nonidentical evolutionary relation it can be obtained the identical relation $$d_\pi\psi=\omega_\pi^p\eqno(22)$$
From such a relation one can obtain the state function and this corresponds to equilibrium state of the system. But identical relation can be realized only on pseudostructure (which is specified by the condition of degenerate transformation). This means that the transition of material system to equilibrium state proceeds only locally. In other words, it is realized the transition of material system from nonequilibrium state to locally equilibrium one. In this case the total state of material system remains to be nonequilibrium. The conditions of degenerate transformation can be caused by the degrees of freedom of material system.
As one can see from the analysis of nonidentical evolutionary relation, the transition of material system from nonequilibrium state to locally-equilibrium state proceeds spontaneously in the process of selfvarying nonequilibrium state of material system under realization of any degrees of freedom of this system. (Translational degrees of freedom, internal degrees of freedom of the system elements, and so on can be examples of such degrees of freedom).
As it has been already said above, the transition from nonidentical relation (21) obtained from balance conservation laws to identical relation (22) means the following. Firstly, the existence of state differential (left-hand side of relation (22)) points to the transition of material system from nonequilibrium state to locally-equilibrium state. And, secondly, the emergence of closed (on pseudostructure) inexact exterior form (right-hand side of relation (22)) points to the origination of physical structure. (Physical structures that are generated by material systems made up physical fields.)
Thus one can see that the transition of material system from nonequilibrium state to locally-equilibrium state is accompanied by originating the differential-geometrical structures, which are physical structures. The emergence of physical structures in the evolutionary process reveals in material system as the emergence of certain observable formations that develop spontaneously. Such formations and their manifestations are fluctuations, turbulent pulsations, waves, vortices, creating massless particles and others. The intensity of such formations is controlled by a quantity accumulated by the evolutionary form commutator at the instant in time of originating physical structures. The transition from evolutionary forms to closed exterior forms describes such processes like the emergence of waves, vortices, turbulent pulsations, the origination of massless particles and others \[7\].
Since the closed exterior forms corresponding to physical structures are obtained from the evolutionary forms describing material systems, the characteristics of physical structures are determined by characteristics of material system generating these structures, and this enables one to classify physical structures by the parameters of evolutionary forms and closed exterior forms.
As it has been shown above, the type of differential-geometrical invariant structures, and hence of physical structures (and, accordingly, of physical fields) generated by the evolutionary relation, depends on the degrees of differential form $p$ and $k$ and on the dimension of original inertial space $n$ (here $p$ is the degree of evolutionary form in nonidentical relation that is connected with the number of interacting balance conservation laws, and $k$ is the degree of closed form generated by nonidentical relation). Introducing the classification by numbers $p$, $k$, $n$ one can understand the internal connection between various physical fields.
The above described mechanism of generation of physical structures discloses an unique role of invariant structures in mathematical physics and field theory.
It should be emphasized that such results were obtained due to using the skew-symmetric exterior and evolutionary differential forms. The mathematic apparatus of evolutionary forms, which describes the process of realization of closed exterior forms and invariant structures, enables one to investigate the integrability of differential equations (the conjugacy of the differential equations elements), discloses the mechanism of evolutionary processes, discrete transitions, quantum steps, transitions from nonconjugated operators to conjugated ones, and generation of various structures. There are no such possibilities in any mathematical formalism.
[00]{}
1\. Cartan E., Les Systemes Differentials Exterieus ef Leurs Application Geometriques. -Paris, Hermann, 1945.
2\. Bott R., Tu L. W., Differential Forms in Algebraic Topology. Springer, NY, 1982.
3\. Schutz B. F., Geometrical Methods of Mathematical Physics. Cambridge University Press, Cambridge, 1982.
4\. Tonnelat M.-A., Les principles de la theorie electromagnetique et la relativite. Masson, Paris, 1959.
5\. Petrova L. I., Specific features of differential equations of mathematical physics. http://arxiv.org/abs/math-ph/0702019, 2007.
6\. Clark J. F., Machesney M., The Dynamics of Real Gases. Butterworths, London, 1964.
7\. Petrova L. I., The mechanism of generation of physical structures. //Nonlinear Acoustics - Fundamentals and Applications (18th International Symposium on Nonlinear Acoustics, Stockholm, Sweden, 2008) - New York, American Institute of Physics (AIP), 2008, pp.151-154.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Classical and quantum theory of spin waves in the vortex state of a mesoscopic sub-micron magnetic disk has been developed with account of the finite mass density of the vortex. Oscillations of the vortex core resemble oscillations of a charged string in a potential well in the presence of the magnetic field. Conventional gyroscopic frequency appears as a gap in the spectrum of spin waves of the vortex. The mass of the vortex has been computed that agrees with experimental findings. Finite vortex mass generates a high-frequency branch of spin waves. Effects of the external magnetic field and dissipation have been addressed.'
author:
- 'R. Zarzuela$^{1,2}$, E. M. Chudnovsky$^{2}$, J. Tejada$^{1}$'
date:
-
-
title: 'Excitation modes of vortices in sub-micron magnetic disks'
---
Introduction
============
Recent advances in optical and electron-beam lithography offered possibility to fabricate arrays of micron and submicron-size magnetic structures with controlled magnetic properties. Among such structures are mesoscopic circular disks of soft ferromagnetic materials. Arrays of such disks, as well as individual disks, have been intensively studied [@Cowburn; @Shinjo; @Novosad1; @Novosad2; @GdL; @Castel; @Zarzuela] due to their unusual magnetic properties and potential for technological and biomedical applications[@Parkin; @Rozhkova; @Kim].
Micron-size circular disks exhibit a wide variety of magnetic equilibrium configurations due to geometric constraints on the spin field[@Hertel]. Their applications are based on static and dynamic properties of one of the essentially non-uniform ground states, the *vortex state*. It is characterized by the curling of the magnetization in the plane of the disk, leaving virtually no magnetic “charges”. The very weak uncompensated magnetic moment of the disk sticks out of a small area confined to the vortex core (VC). The diameter of the core is comparable to the material exchange length[@Novosad1; @Guslienko1]. The low frequency dynamics of the vortex state is due to the gyrotropic mode, consisting of the spiral-like precessional motion of the VC as a whole[@Choe; @Guslienko2; @Guslienko3; @Guslienko4; @Lee], and it is intrinsically distinct from conventional spin wave excitations.
Because of the strong exchange interaction among the out-of-plane spins in the VC, it behaves as an independent entity. The research on excitation modes of vortices has focused on the low-frequency gyroscopic mode that describes circular motion of the vortex about the center of the disk. It can also be viewed as the uniform precession of the magnetic moment of the disk due to the vortex. The natural question is whether the gyroscopic mode allows spatial dispersion similar to spin waves of finite wavelength in ferromagnets. The aim of this paper is to study spin waves related to the gyroscopic motion of the vortex. Such a wave is shown in Fig. 1. It must exist due to finite elasticity of the vortex provided by the exchange interaction.
Most of the research on the gyroscopic motion of vortices in circularly polarized disks ignores the inertial mass of the VC. Such a mass has a dynamical origin stemming from the variation of the shape of the VC as it moves inside the disk. Meantime, experimental studies of vortex oscillations in micrometer permalloy rings[@Bedau] hinted towards a non-negligible vortex mass of order $10^{-24}$kg. On a theoretical side the vortex mass has been previously computed in a two-dimensional Heisenberg model with anisotropic exchange interaction [@Gouvea; @Wysin; @Ivanov]. In disks made of soft magnetic materials that have been experimented with, the exchange interaction is isotropic. We will show that in this case the finite mass density of the vortex originates from the geometrical confinement of the spin field and the magnetic dipole-dipole interactions.
In this paper we will derive the generalized Thiele equation that describes spin waves in the vortex core of finite mass density and will obtain the spectrum of such waves. We will show that the conventional gyroscopic mode, $\omega_G$, appears as a gap in the spectrum of the spin waves in the vortex, $\omega(q) = \omega_{G}
+ \alpha q^2$, when the vortex mass is neglected. From the mathematical point of view the above problem resembles the problem of the motion of a charged string in a potential well in the presence of the magnetic field. The latter problem is a generalization of the problem of Landau levels of an electron in a two-dimensional potential well in the magnetic field. We will show that this problem has a nice exact solution for quantized oscillations of the string, thus providing the spectrum of magnons in the vortex in the quantum regime as well. Classical and quantum solutions for the spectrum of excitations of the vortex lead to the same dispersion law, $\omega(q)$, in the limit of small $q$.
The paper is structured as follows. In Sec. II a Lagrangian formulation of the problem is presented. Formal derivation of the massive elastic Thiele’s equation that allows deformations of the vortex line is given in Section III. The spectrum of spin waves in the vortex core is obtained in Section IV. We show that a finite mass of the vortex results in the additional excitation mode that is absent in the case of zero mass. Quantum mechanical treatment of magnons in the vortex core is developed in Section V. The vortex mass in a circularly polarized disk is computed in Sec. VI and is shown to be in good agreement with experimental findings. The field dependence of the vortex excitation modes and effects of dissipation are discussed in Sec. VII. Sec. VIII contains final conclusions and suggestions for experiment.
Lagrangian mechanics of the vortex core
=======================================
We shall describe the vortex line by the vector field $\vec{X}=(x,y)$, where $x(t,z)$ and $y(t,z)$ are coordinates of the center of the vortex core in the $XY$ plane. Landau-Lifshitz dynamics of the fixed-length magnetization vector $\vec{M}(\Theta,\Phi)=M_{s}(\cos\Phi\sin\Theta,\sin\Phi\sin\Theta,\cos\Theta)$ follows from the Lagrangian [@CT-lectures] $$\begin{aligned}
\label{Lagrangian}
\mathcal{L}&\Big[t;\Theta,\Phi,\dot{\Theta},\dot{\Phi},\partial_{z}\Theta,\partial_{z}\Phi\Big]
=\nonumber\\
&\int{\,\mathrm{d}}z{\,\mathrm{d}}^{2}\vec{r}\Bigg[\frac{M_{s}}{\gamma}(D_{t}\Phi)\cos\Theta-\mathcal{E}(\Theta,\Phi, \partial_{z}\Theta,\partial_{z}\Phi)\Bigg]\end{aligned}$$ where $\mathcal{E}(\Theta,\Phi,\partial_{z}\Theta,\partial_{z}\Phi)$ is the energy density. The dependence of the Lagrangian on the partial derivatives $\partial_{z}$ of the angular coordinates comes from the elastic nature of the vortex core. It is contained in the total energy, $\mathcal{E}$, that takes into account interaction between different layers of the vortex line, see below.
The spatial dependence of angular coordinates $(\Theta, \Phi)$ for the vortex state is given by $\Theta=\Theta(t;\vec{r},z)=\Theta(\vec{r}-\vec{X}(t,z),t)$ and $\Phi=\Phi(t;\vec{r},z)=\Phi(\vec{r}-\vec{X}(t,z),t)$. We only consider long-wave solutions that do not deform the vortex core in any $z$-cross-section of the disk. This means that the angular coordinates depend on $t$ and $z$ via the coordinates of the vortex core $\vec{X}(t,z)$. The covariant derivative with respect to time, $D_{t}\Phi$, along the vortex core is given by $${\displaystyle}D_{t}\Phi=\nabla_{\dot{\vec{X}}(t,z)}\Phi=-\dot{\vec{X}}(t,z)\cdot\nabla_{\vec{r}}
\Phi(\vec{r}-\vec{X}(t,z))$$ where “dot” denotes partial derivative with respect to $t$.
Taking all these considerations into account, the above Lagrangian becomes $$\label{Lagrangian2}
\mathcal{L}\left[t;\vec{X},\dot{\vec{X}},\partial_{z}\vec{X}\right]=\int{\,\mathrm{d}}z\tilde{\mathcal{L}}
\left[t,z;\vec{X},\dot{\vec{X}},\partial_{z}\vec{X}\right]$$ with the Lagrangian density being $$\begin{aligned}
\label{Lagdensity}
\tilde{\mathcal{L}}\left[t,z;\vec{X},\dot{\vec{X}},\partial_{z}\vec{X}\right]&=\int{\,\mathrm{d}}^{2}\vec{r}
\Bigg[\frac{M_{s}}{\gamma}\left(-\dot{\vec{X}}(t,z)\cdot\nabla_{\vec{r}}\Phi\right)\cos\Theta\nonumber\\
&-\mathcal{E}(\Theta,\Phi,\partial_{z}\Theta,\partial_{z}\Phi)\Bigg].\end{aligned}$$ Thus the generalized momentum densities are given by $$\begin{aligned}
\label{momentums}
\vec{\Pi}_{t}\left[t,z;\vec{X},\dot{\vec{X}},\partial_{z}\vec{X}\right]&\equiv&
\frac{\delta\tilde{\mathcal{L}}}{\delta\left(\dot{\vec{X}}(t,z)\right)}\nonumber\\
&=& -\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}\left(\nabla_{\vec{r}}\Phi\right)\cos\Theta\\
\vec{\Pi}_{z}\left[t,z;\vec{X},\dot{\vec{X}},\partial_{z}\vec{X}\right]
&\equiv&\frac{\delta\tilde{\mathcal{L}}}{\delta\left(\partial_{z}\vec{X}(t,z)\right)}\nonumber\\
&=&-\frac{\delta\omega(\vec{X},\partial_{z}\vec{X})}{\delta(\partial_{z}\vec{X}(t,z))}\end{aligned}$$ with ${\displaystyle}\omega(\vec{X},\partial_{z}\vec{X})=\int{\,\mathrm{d}}^2\vec{r}\;\mathcal{E}(\Theta,\Phi,\partial_{z}\Theta,\partial_{z}\Phi)$ being the linear energy density. The dynamics of the vortex core is governed by the Euler-Lagrange equation, $${\displaystyle}D_{t}\vec{\Pi}_{t}+\partial_{z}\vec{\Pi}_{z}-\frac{\delta\tilde{\mathcal{L}}}
{\delta\vec{X}(t,z)}=0$$ Notice that $$\begin{aligned}
D_{t}\Lambda(t,z;\;&\vec{r},\vec{v})=\dot{\xi}(t,z)(\Lambda)=\frac{\partial}{\partial
t}\Lambda(\vec{r}-\vec{X}(t,z),\dot{\vec{X}}(t,z)) \nonumber \\
&=-\dot{\vec{X}}(t,z)\cdot\nabla_{\vec{r}}\Lambda(\vec{r}-\vec{X}(t,z),\dot{\vec{X}}(t,z))
+\nonumber\\
&\qquad\ddot{\vec{X}}(t,z)\cdot\nabla_{\vec{v}}\Lambda(\vec{r}-\vec{X}(t,z),\dot{\vec{X}}(t,z))\end{aligned}$$ means covariant derivative along the curve that is tangent to the vortex core, $\xi(t,z)$. All terms involving $\dot{\vec{X}}(t,z)$ and $\ddot{\vec{X}}(t,z)$ in the Euler-Lagrange equation come from $D_{t}\vec{\Pi}_{t}$, which is given by $$\begin{aligned}
D_{t}\vec{\Pi}_{t}&=-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}\;D_{t}\left(\cos\Theta\nabla_{\vec{r}}\Phi\right)\nonumber\\
&=-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}\Bigg(\nabla_{\vec{r}}\left[-\dot{X}_{j}
\partial_{j}\Phi+\ddot{X}_{j}\tilde{\partial}_{j}\Phi\right]\cos\Theta\nonumber\\
&\quad+\nabla_{\vec{r}}\Phi\left[-\dot{X}_{j}\partial_{j}\cos\Theta+\ddot{X}_{j}
\tilde{\partial}_{j}\cos\Theta\right]\Bigg)\nonumber\\
&=\hat{e}_{i}M_{ij}\ddot{X}_{j}-\hat{e}_{i}K_{ij}\dot{X}_{j},\end{aligned}$$ where $$\begin{aligned}
\label{tensors}
M_{ij}&=-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}\;\Big[(\partial_{i}
\tilde{\partial}_{j}\Phi)\cos\Theta+(\partial_{i}\Phi)\tilde{\partial}_{j}\cos\Theta\Big],\nonumber\\
K_{ij}&=-\frac{M_{s}}
{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}\;\Big[(\partial_{i}
\partial_{j}\Phi)\cos\Theta+(\partial_{i}\Phi)\partial_{j}\cos\Theta\Big]\end{aligned}$$ and ${\displaystyle}\partial_{j}\equiv\nabla_{r_{j}}$, ${\displaystyle}\tilde{\partial}_{j}=\nabla_{v_{j}}$.
We want $\hat{e}_{i}K_{ij}\dot{X}_{j}$ to be of the form $\vec{\rho}_{G}\times\dot{\vec{X}}$ which results in the identity $\epsilon_{ijk}\rho_{G,j}=K_{ik}$. From this we obtain ${\displaystyle}\rho_{G,j}=-\frac{1}{2}\epsilon_{ikj}K_{ik}$, which translates into the vector form as $$\begin{aligned}
\label{gyro}
\vec{\rho}_{G}&=\frac{M_{s}}{2\gamma}\int\mathrm{d}^{2}\vec{r}\epsilon_{ikj}
\Big[(\partial_{i} \partial_{k}\Phi)\cos\Theta+(
\partial_{i}\Phi)(\partial_{k}\cos\Theta)\Big]
\hat{e}_{j}\nonumber\\
&=\frac{M_{s}}{2\gamma}\int{\,\mathrm{d}}^{2}\vec{r}\;\Big[\big(\nabla_{\vec{r}}\times\nabla_{\vec{r}}\Phi\big)\cos\Theta+\nonumber\\
&\qquad\qquad\qquad\qquad\qquad\qquad\quad\nabla_{\vec{r}}\Phi
\times\nabla_{\vec{r}}\cos\Theta\Big]\end{aligned}$$
To compute the mass tensor and the gyrovector we have to find the solutions $(\Theta,\Phi)$ of the Landau-Lifshitz equation in the low dynamics regime that is characterized by the condition $|\dot{\vec{X}}|\ll1$. In this regime solutions can be expanded as a perturbative series on the differential speed, $|\dot{\vec{X}}|$, of the vortex core $$\begin{aligned}
\Theta(t,z;\vec{r})&=\Theta^{(0)}(z;\vec{r})+\Theta^{(1)}(t,z;\vec{r})+\ldots\nonumber\\
\Phi(t,z;\vec{r})&=\Phi^{(0)}(z;\vec{r})+\Phi^{(1)}(t,z;\vec{r})+\ldots\end{aligned}$$ Notice that the zero-th order is time independent, otherwise the gyrovector would depend on time.
The approach that neglects deformation of the vortex core in any $z$-cross-section of the disk is correct only for weak deviations of the centerline of the vortex core from the straight line along the $Z$-axis. We now proceed to the study of the Landau-Lifshitz equation for the set of variables $(\Theta,\Phi)$ in such weak bending regime. It can be obtained by applying the variational principle to the Lagrangian density $$\begin{aligned}
\tilde{\mathcal{L}}\big[t,z;\Theta,\Phi,\dot{\Theta},\dot{\Phi},&\partial_{z}\Theta,\partial_{z}\Phi\big]=\int{\,\mathrm{d}}^{2}\vec{r}\;\Bigg[\frac{M_{s}}
{\gamma}(D_{t}\Phi)\cos\Theta\nonumber\\
&-\mathcal{E}(\Theta,\Phi,\nabla_{\vec{r}}\Theta,
\nabla_{\vec{r}}\Phi,\partial_{z}\Theta,\partial_{z}\Phi)\Bigg]\end{aligned}$$ Notice that $$\begin{aligned}
\mathcal{E}(\Theta,\Phi,\nabla_{\vec{r}}\Theta,\nabla_{\vec{r}}
\Phi,\partial_{z}\Theta,&\partial_{z}\Phi)=\mathcal{E}_{XY}(\Theta,\Phi,\nabla_{\vec{r}}\Theta,\nabla_{\vec{r}}\Phi)\nonumber\\
&+\mathcal{E}_{el}(\Theta,\Phi,\partial_{z}\Theta,\partial_{z}\Phi)\end{aligned}$$ with $\mathcal{E}_{XY}(\Theta,\Phi,\nabla_{\vec{r}}\Theta,\nabla_{\vec{r}}\Phi)$ being the sum of the exchange, anisotropy and dipolar energy responsible for the formation of the vortex, and $$\mathcal{E}_{el}(\Theta,\Phi,\partial_{z}\Theta,\partial_{z}\Phi)=
A_{eff}\left[(\partial_{z}\Theta)^2+\sin^2\Theta(\partial_{z}\Phi)^2\right]$$ being the elastic energy in which $A_{eff}$ is a constant. It describes contribution of the exchange and dipolar forces to the elasticity of the vortex line, with the exchange playing a dominant role. Consequently, with good accuracy, $A_{eff}$ can be identified with the exchange constant $A$.
The set of dynamical equations for $(\Theta,\Phi)$ is $$\begin{aligned}
D_{t}\left(\frac{\delta \tilde{\mathcal{L}}}{\delta(D_{t}\Phi)}\right)+\partial_{z}\left(\frac{\delta\tilde{\mathcal{L}}}
{\delta(\partial_{z}\Phi)}\right)-
\frac{\delta\tilde{\mathcal{L}}}{\delta\Phi}&=&0\nonumber\\
-2A_{eff}\left[\sin2\Theta\partial_{z}\Theta\partial_{z}\Phi+\sin^2\Theta\partial_{z}^{2}\Phi\right]&+&\nonumber\\
\frac{M_{s}}{\gamma}\frac{d\cos\Theta}{dt}+\frac{\delta\mathcal{E}_{XY}}{\delta\Phi}&=&0\end{aligned}$$ and $$\begin{aligned}
D_{t}\left(\frac{\delta \tilde{\mathcal{L}}}{\delta(D_{t}\Theta)}\right)+\partial_{z}
\left(\frac{\delta\tilde{\mathcal{L}}}{\delta(\partial_{z}\Theta)}\right)-
\frac{\delta\tilde{\mathcal{L}}}{\delta\Theta}&=&0\nonumber\\
-2A_{eff}\left[\partial_{z}^{2}\Theta-\frac{\sin2\Theta}{2}(\partial_{z}\Phi)^2\right]&+&\nonumber\\
\frac{M_{s}}{\gamma}\sin\Theta\frac{d\Phi}{dt}+\frac{\delta\mathcal{E}_{XY}}{\delta\Theta}&=&0\end{aligned}$$
Performing a Fourier transform $$\begin{aligned}
\label{Devel}
\Phi(t,z;\vec{r})&=&\frac{1}{\sqrt{2\pi}}\int{\,\mathrm{d}}q\;\bar{\Phi}(t,q;\vec{r})e^{iqz}\nonumber\\
\Theta(t,z;\vec{r})&=&\frac{1}{\sqrt{2\pi}}\int{\,\mathrm{d}}q\;\bar{\Theta}(t,q;\vec{r})e^{iqz}\end{aligned}$$ we obtain the following set of equations for the pair $(\bar{\Theta},\bar{\Phi})$: $$\begin{aligned}
\label{Eq1}
\frac{A_{eff}q^2}{\pi}\Big[\overline{\sin2\Theta}\star\bar{\Theta}\star\bar{\Phi}&+\overline{\sin\Theta}\star\overline{\sin\Theta}\star\bar{\Phi}\Big]+\nonumber\\
&\frac{M_{s}}{\gamma}\frac{d\;\overline{\cos\Theta}}{dt}+\overline{\frac{\delta\mathcal{E}_{XY}}{\delta\Phi}}=0
\end{aligned}$$ $$\begin{aligned}
\label{Eq2}
\frac{1}{\sqrt{2\pi}}\frac{M_{s}}{\gamma}&\overline{\sin\Theta}\star\frac{d\overline{\Phi}}{dt}+\overline{\frac{\delta\mathcal{E}_{XY}}{\delta\Theta}}+\nonumber\\
&2A_{eff}q^2\left[\bar{\Theta}-\frac{\overline{\sin2\Theta}}{4\pi}\star\bar{\Phi}\star\bar{\Phi}\right]=0\end{aligned}$$ where $\star$ means Fourier convolution.
For a small bending of the vortex core, the boundary conditions on the angle $\Theta$ are the same as in the rigid VC case, i.e.: $\Theta\simeq0\textrm{ or }\pi$ in the limit $\tilde{r}\ll\Delta_{0}$ and $\Theta\simeq\pi/2$ in the limit $\tilde{r}\gg\Delta_{0}$, with $\Delta_{0}=\sqrt{A/M_{s}^2}$ being the exchange length of the material and where $\tilde{r}=||\vec{r}-\vec{X}(t,z)||_{2}$ is the radial distance from the VC center at any height $z$. Considering this two limits Eq. becomes:
- Limit $\tilde{r}\ll\Delta_{0}$. In this case, $\sin\Theta\simeq0$ and thus $\sin2\Theta\partial_{z}\Theta\partial_{z}\Phi+\sin^2\Theta\partial_{z}^{2}\Phi\simeq0$. So we have the following equation in the Fourier space $$\frac{M_{s}}{\gamma}\frac{d\;\overline{\cos\Theta}}{dt}+\overline{\frac{\delta\mathcal{E}_{XY}}{\delta\Phi}}=0$$
- Limit $\tilde{r}\gg\Delta_{0}$. In this case, $\sin\Theta\simeq1$ and thus $\sin2\Theta\partial_{z}\Theta\partial_{z}\Phi+\sin^2\Theta\partial_{z}^{2}\Phi\simeq\partial_{z}^2\Phi$. So we have the equation $$\frac{M_{s}}{\gamma}\frac{d\;\overline{\cos\Theta}}{dt}+2A_{eff}q^2\bar{\Phi}+\overline{\frac{\delta\mathcal{E}_{XY}}{\delta\Phi}}=0$$
Notice that in both limits $\sin2\Theta\simeq0$ and so $\partial_{z}^{2}\Theta-\frac{\sin2\Theta}{2}(\partial_{z}\Phi)^2\simeq\partial_{z}^{2}\Theta$. Consequently, in the Fourier space Eq. becomes $$\frac{1}{\sqrt{2\pi}}\frac{M_{s}}{\gamma}\overline{\sin\Theta}\star\frac{d\overline{\Phi}}{dt}+2A_{eff}q^2\bar{\Theta}+\overline{\frac{\delta\mathcal{E}_{XY}}{\delta\Theta}}=0$$
Finally, in the limit of weak bending ($A_{eff}q^2\ll1$), we can neglect the terms of the form $2A_{eff}q^2\bar{\xi}$ in the above equations. In doing so, we recover the standard Landau-Lifshitz equations for $(\Theta,\Phi)$ at any $z$ layer, with the VC center depending on the value of $z$. Introducing now the perturbative series into the Landau-Lifshitz equation and splitting it into $O(|\dot{\vec{X}}|^{n})$ terms, we obtain the equations of motion for the $\Phi^{(n)}/\Theta^{(n)}$ terms. In the case of the zero-th and first order terms, we recover the static solution and the first perturbative solution for the rigid vortex (see Section VI). For the particular case of the zero-th order we obtain $$\begin{aligned}
\label{zeroth}
\Phi_{0}(x,y)&=n_v\tan^{-1}(y-y_{v}/x-x_{v})\nonumber\\
\cos\Theta_{0}(\tilde{r})&= \left\{\begin{array}{lcc}
p\left(1-C_{1}\left(\frac{\tilde{r}}{\Delta_{0}}\right)^2\right)
& & \tilde{r} \ll\Delta_{0}\\
C_{2}\left(\frac{\Delta_{0}}{\tilde{r}}\right)^{1/2}\exp(-\tilde{r}/\Delta_{0})
& & \tilde{r}\gg\Delta_{0}
\end{array}\right.\end{aligned}$$ where $n_v = \pm 1$ is the vorticity of the magnetization of the disk and $C_{1},C_{2}$ are constants that can be obtained by imposing the smoothness condition on $\cos\Theta_{0}$ at $\tilde{r}=\Delta_{0}$ up to its first derivative. The corresponding values are $C_{1}=\frac{3}{7}$ and $C_{2}=\frac{4}{7}pe$. From all this we straightforwardly deduce that $$\begin{aligned}
\label{Phi_prop} \nabla_{\vec{r}}\times\nabla_{\vec{r}}\Phi_{0}& =
& 2\pi n_v \delta^{(2)}
\left(\vec{r}-\vec{X}(t,z)\right)\hat{e}_{z}\nonumber\\
\nabla_{\vec{r}}^{2}\Phi_{0} &=& 0\end{aligned}$$
Elastic Thiele equation
=======================
We now proceed to the computation of the gyrovector and the mass density tensor. Using we obtain that the first term of Eq. equals $(\pi n_v p
M_{s}/\gamma)\hat{e}_{z}$, where $p=\cos\Theta(\vec{0})=\pm1$ defines the direction of the polarization of the vortex core ($\Theta(\vec{0})=0$ or $\pi$). The second term is evaluated at the zero-th order of the perturbative expansion of the angular coordinates in the low dynamics regime, Eq. , taking into account that in the weak bending regime the deformation of the vortex core is small and one can consider $\tilde{r}\simeq r$ because $||\vec{X}(t,z)||_{2}\ll1$. In doing so we obtain $(\pi
n_v p M_{s}/\gamma)\hat{e}_{z}$ again[@Huber]. Thus the gyrovector becomes $\vec{\rho}_{G}=\rho_{G}pn_v\hat{e}_{z}$ with $$\label{rho-G}
\rho_{G}=2\pi M_{s}/\gamma$$ Notice that $\vec{\rho}_{G}$ is the gyrovector linear density as compared to the gyrovector in the Thiele equation for a rigid vortex[@Thiele].
Computation of the mass density tensor will be performed in Section VI. For circular polarized disks we show that this tensor reduces to a scalar, $M_{ij}=\rho_{M}\delta_{ij}$, with the vortex core mass density given by ${\displaystyle}\rho_{M}=\frac{1}{4\gamma^2}\ln(R/\Delta_{0})$, where $R$ is the radius of the disk. Only $\omega(\vec{X},\partial_{z}\vec{X})$ contributes to the partial derivative $\delta\tilde{\mathcal{L}}/\delta\vec{X}$ in the slow dynamics regime, because $\vec{\Pi}_{t}=\rho_{M}\dot{\vec{X}}$ and so the term $\dot{\vec{X}}\cdot\vec{\Pi}_{t}$ equals $\rho_{M}\dot{\vec{X}}^2$. Consequently, the generalized Thiele equation becomes $$\label{Thiele}
\rho_{M}\ddot{\vec{X}}(t,z)+\dot{\vec{X}}(t,z)\times\vec{\rho}_{G}
+\partial_{z}\vec{\Pi}_{z}+\nabla_{\vec{X}}\omega=0$$
The linear energy density $\omega(\vec{X},\partial_{z}\vec{X})$ is the sum of the magnetostatic and exchange contributions in the $z$-cross-section, $\omega_{XY}(\vec{X})$, and an elastic contribution due to the deformation of the vortex core line, $\omega_{el}(\partial_{z}\vec{X})$. Zeeman contribution will be considered later. The dependence on the vortex core coordinates on the $\omega_{XY}(\vec{X})$ term for small displacements is[@Guslienko1; @Guslienko2] $$\omega_{XY}(\vec{X})=\frac{1}{2}\rho_{M}\omega_{M}^2\epsilon_{0}\vec{X}^2,$$ where ${\displaystyle}\omega_{M}=\rho_{G}/\rho_{M}$ is the characteristic frequency of the system and $\epsilon_{0}=\omega_{G}/\omega_{M}$ is a dimensionless parameter. Recall that the conventional gyrofrequency $\omega_{G}$ is defined as[@Guslienko2; @Guslienko3] $$\label{gyrofreq}
\omega_{G}=\frac{\omega_{XY}^{''}(\vec{X}=\vec{0})}{\rho_{G}}\simeq\frac{20}{9}\gamma
M_{s}\beta$$ where $\beta=L/R$ is the ratio of the thickness and the radius of the disk. This last expression is valid in the limit $\beta\ll 1$.
From the continuous spin-field model we know that $${\displaystyle}\omega_{el}(\partial_{z}\vec{X})=A_{eff}\int{\,\mathrm{d}}^{2}\vec{r}\left[(\partial_{z}\Theta)^2
+\sin^2\Theta(\partial_{z}\Phi)^2\right]$$ Noticing that $\partial_{z}\Theta=-\nabla_{\vec{r}}\Theta\cdot\partial_{z}\vec{X}$ and $\partial_{z}\Phi=-\nabla_{\vec{r}}\Phi\cdot\partial_{z}\vec{X}$, and taking into account the vector identity, $(\vec{A}\times\vec{B})\cdot(\vec{C}\times\vec{D})=(\vec{A}\cdot\vec{C})
(\vec{B}\cdot\vec{D})-(\vec{A}\cdot\vec{D})(\vec{B}\cdot\vec{C})$, with either $\vec{A}=\vec{D}=\nabla_{\vec{r}}\Theta,\;\vec{B}=\vec{C}=-\partial_{z}\vec{X}$ or $\vec{A}=\vec{D}=\nabla_{\vec{r}}\Phi,\;\vec{B}=\vec{C}=-\partial_{z}\vec{X}$, we obtain the following relations $$\begin{aligned}
\left(\nabla_{\vec{r}}\Theta\cdot\partial_{z}\vec{X}\right)^{2}&=\left(\nabla_{\vec{r}}\Theta\right)^{2}\left(\partial_{z}\vec{X}\right)^{2}-\left(\nabla_{\vec{r}}\Theta\times\partial_{z}\vec{X}\right)^{2}\\
\left(\nabla_{\vec{r}}\Phi\cdot\partial_{z}\vec{X}\right)^{2}&=\left(\nabla_{\vec{r}}\Phi\right)^{2}\left(\partial_{z}\vec{X}\right)^{2}-\left(\nabla_{\vec{r}}\Phi\times\partial_{z}\vec{X}\right)^{2}\end{aligned}$$ The main contribution to the integral comes from the zero-th order in the perturbative expansion . Notice that $\Theta_{0}(\vec{r})=\Theta_{0}(\tilde{r})$ and that $\nabla_{\vec{r}}\Phi_{0}=\frac{n_{v}}{\tilde{r}}\hat{e}_{\phi}$. As discussed before, in the weak bending regime we can use approximation $\tilde{r}\simeq r$, so that $\nabla_{\vec{r}}\Phi_{0}\times\partial_{z}\vec{X}=
-\frac{n_{v}}{r}\partial_{z}X_{r}\hat{e}_{z}$ and $\nabla_{\vec{r}}\Theta_{0}\times\partial_{z}\vec{X}=
\frac{d\Theta_{0}}{dr}\partial_{z}X_{\phi}\hat{e}_{z}$, where $X_{r}=\hat{e}_{r}\cdot\vec{X}=x\cos\theta+y\sin\theta$ and $X_{\phi}=\hat{e}_{\phi}\cdot\vec{X}=-x\sin\theta+y\cos\theta$. The elastic energy density finally becomes $$\begin{aligned}
\omega_{el}(\partial_{z}\vec{X})=&A_{eff}\int{\,\mathrm{d}}^{2}\vec{r}
\left[(\nabla_{\vec{r}}\Theta_{0})^{2}+\frac{\sin^{2}
\Theta_{0}}{r^{2}}\right]\left(\frac{\partial\vec{X}}
{\partial z}\right)^{2}\nonumber\\
&-A_{eff}\int{\,\mathrm{d}}^{2}\vec{r}\left(\frac{d\Theta_{0}}{dr}\right)^{2}
\left(\frac{\partial X_{\phi}}{\partial z}\right)^{2}\nonumber\\
&-A_{eff}\int{\,\mathrm{d}}^{2}\vec{r}\;\frac{\sin^{2}\Theta_{0}}{r^2}\left(\frac{\partial X_{r}}
{\partial z}\right)^{2}\nonumber\\
=&\pi A_{eff}\int r{\,\mathrm{d}}r\left[\left(\frac{d\Theta_{0}}{dr}
\right)^{2}+\frac{\sin^{2}\Theta_{0}}{r^{2}}\right]
\left(\frac{\partial\vec{X}}{\partial z}\right)^{2}\end{aligned}$$ where the angular dependence of $(\partial_{z}X_{r})^{2}$ and $(\partial_{z}X_{\phi})^{2}$ has been integrated on $\theta$. We can recast this energy density as ${\displaystyle}\omega_{el}(\partial_{z}\vec{X})=\frac{1}{2}\lambda
\left(\frac{\partial\vec{X}}{\partial z}\right)^2$, where $\lambda$ is the elastic constant given by $$\label{lambda}
\lambda=2\pi A_{eff}\int r{\,\mathrm{d}}r\left[\left(\frac{d\Theta_{0}}{dr}\right)^{2}+\frac{\sin^{2}\Theta_{0}}{r^{2}}\right]$$ Making use of the variable $m_{0}(r)=\cos\Theta_{0}(r)$, we can rewrite the above equation as $$\lambda=2\pi A_{eff} \int r{\,\mathrm{d}}r\left[\frac{1}{1-m_{0}^{2}}\left(\frac{dm_{0}}{dr}\right)^{2}+\frac{1-m_{0}^{2}}{r^{2}}\right]$$ and using the spatial dependence , we get $$\begin{aligned}
\label{integrand}
&\frac{1}{1-m_{0}^{2}}\left(\frac{dm_{0}}{dr}\right)^{2}+\frac{1-m_{0}^{2}}{r^{2}}
=\nonumber\\
&\left\{\begin{array}{lcc}
\frac{\Delta_{0}^{2}}{2C_{1} r^{2}}\left(\frac{2pC_{1}}{\Delta_{0}^{2}}r\right)^{2}
+\frac{2C_{1}r^{2}}{\Delta_{0}^{2}}\frac{1}{r^{2}}=\frac{4C_{1}}{\Delta_{0}^{2}}
& & r\ll\Delta_{0}\\
\frac{1}{\Delta_{0}^{2}}m_{0}^{2}+\frac{1}{r^{2}}=\frac{C_{2}^{2}}{\Delta_{0}r}
\exp{(-2r/\Delta_{0})}+\frac{1}{r^{2}}& & r\gg\Delta_{0}
\end{array}\right.\end{aligned}$$ Recalling that $C_{1}=\frac{3}{7}$ and $C_{2}=\frac{4}{7}pe$ and computing the integral by splitting into two regions, $[0,\Delta_{0}]$ and $[\Delta_{0},R]$, we obtain $$\lambda=2\pi A_{eff}\left(\frac{50}{49}+\ln(R/\Delta_{0})\right)$$ In the limit $R\gg \Delta_{0}$ the logarithmic term dominates and $\lambda$ becomes $$\label{lambda-A}
\lambda=2\pi A_{eff}\ln(R/\Delta_{0})$$
Finally, for the total energy density we obtain $$\label{Potential}
\omega(\vec{X},\partial_{z}\vec{X})=\frac{1}{2}\rho_{M}\omega_{M}^2\epsilon_{0}
\vec{X}^2+\frac{1}{2}\lambda\left(\frac{\partial\vec{X}}{\partial
z}\right)^2$$ and thus the generalized Thiele equation for an elastic vortex core line becomes $$\label{Thiele2}
\rho_{M}\ddot{\vec{X}}(t,z)-\lambda\partial^{2}_{z}\vec{X}(t,z)
+\dot{\vec{X}}(t,z)\times\vec{\rho}_{G}+\rho_{M}\omega_{M}^2\epsilon_{0}\vec{X}(t,z)=0$$
Spin waves in the vortex core
=============================
Introducing the complex variable $\chi=x-iy$ we can recast Eq. as the following complex partial differential equation: $$\label{CThiele}
\rho_{M}\ddot{\chi}-\lambda\partial^{2}_{z}\chi+i\rho_{G}\dot{\chi}+\rho_{M}
\omega_{M}^2\epsilon_{0}\chi=0$$ Let $\chi_{0}(z)$ be the equilibrium complex center of the straight vortex core line. In the presence of the wave it gets perturbed and becomes $\chi(t,z)=\chi_{0}(z)+\chi_{w}(t,z)$, with $||\chi_{w}||_{z}\ll||\chi_{0}||_{z}$. Switching to the Fourier transform, $$\chi_{w}(t,z)=\frac{1}{2\pi}\int{\,\mathrm{d}}\omega{\,\mathrm{d}}q\chi_{w}(\omega,q)e^{i(\omega t-qz)},$$ we obtain the following equation for $\chi_{w}(\omega,q)$: $$\label{FCThiele}
\left[-\rho_{M}\omega^2+\lambda
q^2-\rho_{G}\omega+\rho_{M}\omega_{M}^2\epsilon_{0}\right]\chi_{w}(\omega,q)=0$$ For non-zero amplitude of the wave the expression in the square parenthesis must vanish. This determines the spectrum of the waves: $$\label{spectrum}
\rho_{M}\omega^2-\lambda q^2+\rho_{G}\omega-\rho_{M}\omega_{M}^2\epsilon_{0}=0$$
At $\rho_{M}\neq0$ one can normalize Eq. to get $$\omega^2+\omega_{M}\omega-\omega_{M}^2\left(\epsilon_{0}+
\frac{\lambda}{\rho_{M}\omega_{M}^{2}}q^2\right)=0$$ Solving this equation we obtain the spectrum of vortex core excitations: $$\label{waves}
\omega_{\pm}(q)=\frac{\omega_{M}}{2}\left[\sqrt{(1+4\epsilon_{0})+
\frac{4\lambda q^2}{\rho_{M}\omega_{M}^2}}\pm1\right]$$ In the weak bending regime we have $4\lambda q^2/\rho_{M}\omega_{M}^2\ll1$ and so we can expand the square root and obtain the following expression for the frequencies $$\omega_{\pm}(q)=\frac{\omega_{M}}{2}\left(\sqrt{1+4\epsilon_{0}}\pm1\right)
+\frac{1}{\sqrt{1+4\epsilon_{0}}}\frac{\lambda}{\rho_G}q^2$$ where we have used the relation $\rho_M\omega_M = \rho_G$.
As will be shown in Section IV the parameter $\epsilon_0 =
\omega_G/\omega_M$ is normally small due to the smallness of $\beta = L/R$. Consequently $$\begin{aligned}
\omega_-(q) & \approx & \omega_{G}+\frac{\lambda}{\rho_{G}}q^2
\\
\omega_+(q) & \approx & \omega_M + \frac{\lambda}{\rho_{G}}q^2\end{aligned}$$ With the help of Eqs. (\[rho-G\]) and (\[lambda-A\]) with $A_{eff} \approx A = M_s^2\Delta_0^2$ the above equations can be written in a transparent form: $$\begin{aligned}
\omega_-(q) &= & \omega_{G}+\gamma M_s
(q\Delta_0)^2\ln(R/\Delta_0) \\
\omega_+(q) &= & \omega_{M}+\gamma M_s
(q\Delta_0)^2\ln(R/\Delta_0)\end{aligned}$$ Note that the weak bending regime corresponds to $q\Delta_0 \ll
1$.
Quantum mechanics of the excitations in the vortex core
=======================================================
In this section we will show that excitations of the vortex core can be also obtained in a rather non-trivial way from the quantum theory as well. This problem is interesting on its own as it turns out to be equivalent to the problem of quantum excitations of a charged string confined in a parabolic potential and subjected to the magnetic field.
It is straightforward to prove that the generalized Thiele equation is the Euler-Lagrange equation associated with the following effective Lagrangian density that can be derived from Eq. (\[Lagdensity\]) $$\tilde{\mathcal{L}}(t,z;\vec{X},\dot{\vec{X}},\partial_{z}\vec{X})
=\frac{1}{2}\rho_{M}\dot{\vec{X}}^2+\dot{\vec{X}}
\cdot\vec{A}_{\rho_{G}}-\omega(\vec{X},\partial_{z}\vec{X})$$ where $\vec{A}_{\rho_{G}}$ is the gyrovector potential satisfying $\nabla_{\vec{X}}\times\vec{A}_{\rho_{G}}=-\vec{\rho}_{G}$. Thus the total Lagrangian becomes $$\label{Lagrangian3}
\mathcal{L}=\int{\,\mathrm{d}}z\;\tilde{\mathcal{L}}=\int {\,\mathrm{d}}z\;\left[\frac{1}{2}\rho_{M}
\dot{\vec{X}}^2+\dot{\vec{X}}\cdot\vec{A}_{\rho_{G}}-\omega(\vec{X},\partial_{z}\vec{X})\right]$$ Noticing that ${\displaystyle}\left\{\varphi_{n}(z)\right\}_{n\in\mathbf{N}}=
\left\{\sqrt{\frac{2}{L}}\sin\left(q_{n}z\right)\right\}_{n\in\mathbf{N}}$, with ${\displaystyle}q_{n}=\frac{2\pi}{L}n$, is a Hilbert basis of the function subspace $\mathcal{W}=\{\varphi\in\mathcal{L}^{2}(0,L),\;\varphi(0)
=\varphi(L)=0\}$, we can expand $\vec{X}$ as $$\vec{X}(t,z)=\vec{X}_{0}(t)+\sum_{n}\vec{X}_{n}(t)\varphi_{n}(z),$$ where $\vec{X}_{0}(t)$ is the center of the undisturbed vortex and $\vec{X}_{n}(t)=<\vec{X}(t,z),\varphi_{n}(z)>_{\mathcal{L}^2(0,L)}$. Introducing this expansion in Eq. and taking into account the orthonormality of the Hilbert basis (and its spatial derivatives) we obtain the following identity: $$\begin{aligned}
&\mathcal{L}\Bigg(t,\left\{\vec{X}_{n}\right\}_{n\in\mathbf{Z}^{+}},\left\{\dot{\vec{X}}_{n}\right\}_{n\in\mathbf{Z}^{+}}\Bigg)=\nonumber\\
&\left[\frac{1}{2}M\dot{\vec{X}}_{0}^{2}+\dot{\vec{X}}_{0}\cdot\vec{A}_{0}-\frac{1}{2}M\omega_{M}^{2}\epsilon_{0}\vec{X}_{0}^{2}\right]+\nonumber\\
&\sum_{n>0}\left[\frac{1}{2}\rho_{M}\dot{\vec{X}}_{n}^{2}+\dot{\vec{X}}_{n}\cdot\vec{A}_{n}-\frac{1}{2}\rho_{M}\omega_{M}^{2}\epsilon_{0}\vec{X}_{n}^{2}-\frac{1}{2}\lambda
q_{n}^{2}\vec{X}_{n}^{2}\right]\end{aligned}$$ where $\mathbf{Z}^{+}=\{0\}\cup\mathbf{N}$, $M=\rho_{M}L$ is the total mass of the rigid vortex line and $\vec{A}_{n}$ is the gyrovector potential associated to the $n$-th coordinate $\vec{X}_{n}$, which satisfies $\nabla_{\vec{X}_{0}}\times\vec{A}_{0}=-\vec{G}$ and $\nabla_{\vec{X}_{n}}\times\vec{A}_{n}=-\vec{\rho}_{G},\; n>0$, with $\vec{G}=\vec{\rho}_{G}L$ being the gyrovector of the rigid vortex.
Applying the Laguerre transformation to the above Lagrangian we obtain the following expression for the Hamiltonian $$\begin{aligned}
\label{Hamiltonian}
&\mathcal{H}\left(t,\left\{\vec{X}_{n}\right\}_{n\in\mathbf{Z}^{+}},\left\{\vec{\Pi}_{n}\right\}_{n\in\mathbf{Z}^{+}}\right)=\nonumber\\
&\left[\frac{1}{2M}(\vec{\Pi}_{0}-\vec{A}_{0})^{2}+\frac{1}{2}M\omega_{M}^{2}\epsilon_{0}\vec{X}_{0}^{2}\right]+\nonumber\\
&\sum_{n>0}\left[\frac{1}{2\rho_{M}}(\vec{\Pi}_{n}-\vec{A}_{n})^{2}+\frac{1}{2}\rho_{M}\omega_{M}^{2}\left(\epsilon_{0}+\frac{\lambda}{\rho_{M}\omega_{M}^{2}}
q_{n}^{2}\right)\vec{X}_{n}^{2}\right]\end{aligned}$$ where $\vec{\Pi}_{0}=M\dot{\vec{X}}_{0}+\vec{A}_{0}$ and $\vec{\Pi}_{n}=\rho_{M}\dot{\vec{X}}_{n}+\vec{A}_{n},\;n>0$ are the corresponding canonical momenta. Notice that Eq. shows that $\mathcal{H}$ splits into the direct sum $\oplus_{m\in\mathbf{Z}^{+}}\mathcal{H}_{m}$, with $\mathcal{H}_{m}$ being the Hamiltonian defined over the phase space $(\vec{X}_{m},\vec{\Pi}_{m})$. It has a structure of the form $$\label{Hamiltonian2}
\mathcal{H'}=\frac{1}{2\eta}(\vec{\Pi}-\vec{A})^{2}+\frac{1}{2}\eta\omega_{M}^{2}\xi\vec{X}^{2},$$ where $(\vec{X},\vec{\Pi})$ are the canonically conjugate variables, $\eta,\xi$ are constants and $\vec{A}$ is the gyrovector satisfying $\nabla_{\vec{X}}\times\vec{A}=-\chi\hat{z}$, with $\chi$ being a constant. It is important to point out that $\chi/\eta=\omega_{M}$ in all cases.
From now on we consider the case of the vortex core of a non-zero mass, ($\eta\neq0$). It is convenient to choose a “symmetric gauge” given by $$\label{Gauge}
\vec{A}=\frac{1}{2}(-\chi\hat{z})\times\vec{X}=\frac{\chi
y}{2}\hat{x}-\frac{\chi x}{2}\hat{y}$$ Firstly, we define the kinetic momentum operators as $\vec{p}=\eta\dot{\vec{X}}$, so that $\vec{\Pi}=\vec{p}+\vec{A}$. Notice the following non vanishing commutators $$\label{pxpy} \left [p_{j},p_{k}\right]=-i\hbar \chi \epsilon_{jk}
\qquad j,k\in\{x,y\},$$ where $\epsilon_{jk}$ is the antisymmetric tensor $\epsilon_{xy} =
-\epsilon_{yx} = 1$. Secondly, we introduce the operators $$\begin{aligned}
\label{Aoperators}
a&=\sqrt{\frac{1}{2\hbar \chi}}\left(p_{y}+ip_{x}\right)\nonumber\\
a^{\dagger}&=\sqrt{\frac{1}{2\hbar
\chi}}\left(p_{y}-ip_{x}\right)\end{aligned}$$ which satisfy standard commutation relations for Bose operators, ${\displaystyle}\left[a,a^{\dagger}\right]=1$. The number operator $N_{a}=a^{\dagger}a$ satisfies commutation relations $\left[N_{a},a\right]=-a,\quad\left[N_{a},a^{\dagger}\right]=a^{\dagger}$ and we have the identity $$\frac{1}{2\eta}\left(\vec{\Pi}-\vec{A}\right)^2=\hbar\omega_{M}\left(N_{a}+\frac{1}{2}\right),$$
In analogy with the case of a charged particle in the electromagnetic field[@Zak], we obtain that the gyrotropic translational group is generated by $\vec{T}=\vec{\Pi}+\vec{A}$, $$\label{Trans gen} T_{x}=p_{x}+\chi y,\qquad\qquad T_{y}=p_{y}-\chi
x$$ which satisfies the following commutation relations, $$\label{Comm trans gen}
\left[T_{j},p_{k}\right]=0,\qquad\left[T_{j},T_{k}\right]=i\hbar
\chi\epsilon_{jk},\quad j,k\in\{x,y\}$$ Now we introduce another set of Bose operators $$\begin{aligned}
\label{Boperators}
b&=\sqrt{\frac{1}{2\hbar \chi}}\left(T_{y}-iT_{x}\right)\nonumber\\
b^{\dagger}&=\sqrt{\frac{1}{2\hbar
\chi}}\left(T_{y}+iT_{x}\right),\end{aligned}$$ which satisfy commutation relations $$\left[b,b^{\dagger}\right]=1,\quad\left[M_{b},b\right]=-b,\quad\left[M_{b},b^{\dagger}\right]=b^{\dagger},$$ where $M_{b}=b^{\dagger}b$ is the corresponding number operator. Notice that the commutation relations $\left[a,b\right]=\left[a,b^{\dagger}\right]=0$ also hold.
Coordinates $x$ and $y$ can be expressed in terms of the above Bose operators: $$\begin{aligned}
\label{xy}
x&=\frac{1}{\chi}\left(p_{y}-T_{y}\right)\\
y&=-\frac{1}{\chi}\left(p_{x}-T_{x}\right)\nonumber,\end{aligned}$$ so that $$\label{r2}
\frac{1}{2}\eta\omega_{M}^2\left(x^2+y^2\right)=\hbar\omega_{M}\left(N_{a}+M_{b}-ab-a^{\dagger}b^{\dagger}+1\right).$$ Consequently, the Hamiltonian becomes $$\label{Hamiltonian3}
\mathcal{H'}=\hbar\omega_{M}\left[(1+\xi)N_{a}+\xi
M_{b}-\xi(ab+a^{\dagger}b^{\dagger})+\xi+\frac{1}{2}\right]$$ It can be diagonalized with the help of Bogoliubov transformations $$\label{Bogoliubov} \bar{\alpha}=ua-vb^{\dagger},\qquad
\bar{\beta}=ub-va^{\dagger}$$ with $u,v$ being real numbers. These new operators satisfy Bose commutation relations if $u^2-v^2=1$. Substituting the above equations into Eq. we obtain $$\begin{aligned}
\label{Hamiltonian4}
\mathcal{H'}=\hbar\omega_{M}&\Bigg[\bar{\alpha}^{\dagger}\bar{\alpha}(u^2(1+\xi)+\xi v^2-2\xi uv)+\\
&\bar{\beta}^{\dagger}\bar{\beta}(v^2(1+\xi)+\xi u^2-2\xi uv)+\nonumber\\
&(\bar{\alpha}^{\dagger}\bar{\beta}^{\dagger}+\bar{\alpha}\bar{\beta})\left(uv(1+2\xi)-\xi(u^2 + v^2)\right)\nonumber\\
&+\left(v^2(1+2\xi)-2\xi uv+(\xi+1/2)\right)\Bigg]\nonumber\end{aligned}$$ To get a Hamiltonian in the oscillator form, the coefficient related to ($\bar{\alpha}^{\dagger}\bar{\beta}^{\dagger}+
\bar{\alpha}\bar{\beta}$) should be zero, which requires $$uv(1+2\xi)-\xi(u^2+v^2)=0$$ The solution is $u=\cosh(\theta),\; v=\sinh(\theta)$, $$\tanh(2\theta)=\frac{2\xi}{1+2\xi}$$ Finally, the coefficients of the terms $\bar{\alpha}^{\dagger}\bar{\alpha}$ and $\bar{\beta}^{\dagger}\bar{\beta}$ become $$\begin{aligned}
u^2(1+\xi)+\xi v^2-2\xi uv&=\frac{1}{2}\left[\frac{1+2\xi}{\cosh(2\theta)}+1\right]\nonumber\\
&=\frac{1}{2}\left[\sqrt{1+4\xi}+1\right]\nonumber\\
v^2(1+\xi)+\xi u^2-2\xi
uv&=\frac{1}{2}\left[\frac{1+2\xi}{\cosh(2\theta)}-1\right]\nonumber\\
&=\frac{1}{2}\left[\sqrt{1+4\xi}-1\right]\end{aligned}$$ and, consequently, the Hamiltonian in the second quantization formalism becomes $$\label{Hamiltonian5}
\mathcal{H'}=\hbar\omega_{+}\left(\bar{\alpha}^{\dagger}\bar{\alpha}+\frac{1}{2}\right)+\hbar\omega_{-}\left(\bar{\beta}^{\dagger}\bar{\beta}+\frac{1}{2}\right)$$ where $$\begin{aligned}
\label{freq}
\omega_{\pm}&=\frac{1}{2}\left[\sqrt{1+4\xi}\pm1\right]\omega_{M}\end{aligned}$$
Noticing that for any $n\in\mathbf{Z}^{+}$ we have $\xi=\epsilon_{0}+\frac{\lambda}{\rho_{M}\omega_{M}^{2}} q_{n}^{2}$, the second quantization procedure yields the following form of Hamiltonian $$\mathcal{H}=\sum_{n\geq0}\hbar\omega^{+}_{n}\left(\bar{\alpha}_{n}^{\dagger}\bar{\alpha}_{n}+\frac{1}{2}\right)+\sum_{n\geq0}\hbar\omega^{-}_{n}\left(\bar{\beta}_{n}^{\dagger}\bar{\beta}_{n}+\frac{1}{2}\right)$$ where $\omega^{\pm}_{n}$ are the eigenfrequencies of the vortex state given by $$\label{freq2}
\omega_{n}^{\pm}=\frac{1}{2}\left[\sqrt{(1+4\epsilon_{0})+\frac{4\lambda}{\rho_{M}\omega_{M}^{2}}
q_{n}^{2}}\pm1\right]\omega_{M}$$ which coincides with (\[waves\]).
Computation of the vortex mass
==============================
As it has been discussed in Section II, to calculate the vortex mass density tensor (see Eq. ) we need to find solution $(\Phi(\vec{r},t),\Theta(\vec{r},t))$ of the Landau-Lifshitz equation in the slow dynamics regime, i.e. in the first order on $|\dot{\vec{X}}|$. A more convenient set of variables for this problem is the pair $(\Phi,m)$, where ${\displaystyle}m\equiv m_{z}=\frac{M_{z}}{M_{s}}=\cos\Theta$ is the projection of the magnetic moment onto the $z$ axis. Notice that Landau-Lifshitz equation can be recast as the set of equations $$\begin{aligned}
\label{LL2}
\frac{d\Phi}{dt}&=\frac{\gamma}{M_{s}}\frac{\delta\mathcal{E}}{\delta m}\nonumber\\
\frac{dm}{dt}&=-\frac{\gamma}{M_{s}}\frac{\delta\mathcal{E}}{\delta\Phi}\end{aligned}$$ The total energy $\mathcal{E}(\Phi,m)$ splits into the sum $$\begin{aligned}
\mathcal{E}(\Phi,m)&=\mathcal{E}_{ex}(\Phi,m)+\mathcal{E}_{an}(\Phi,m)
+\mathcal{E}_{demag}(\Phi,m)\nonumber\\
&=A\left[(\nabla\Theta)^2+\sin^2\Theta(\nabla\Phi)^2\right]\nonumber\\
&\quad-K_{\shortparallel}\frac{M_{x}^2}{M_{s}^2}+K_{\perp}\frac{M_{z}^2}{M_{s}^2}
-\frac{1}{2}\vec{M}\cdot\vec{H}_{d}\nonumber\\
&=A\left[\frac{1}{1-m^2}(\nabla m)^2+(1-m^2)(\nabla\Phi)^2\right]\nonumber\\
&-K_{\shortparallel}\cos^2\Phi(1-m^2)+K_{\perp}m^2-\frac{1}{2}
\vec{M}\cdot\vec{H}_{d}\end{aligned}$$ with $A$ being the exchange constant, $K_{\shortparallel},\;K_{\perp}$ being the anisotropy constants and $\vec{H}_{d}$ being the demagnetizing field. Recall that $\vec{H}_{d}(\vec{r})=-\nabla\Phi_{d}(\vec{r})$, with $\nabla^{2}\Phi_{d}(\vec{r})=-4\pi\rho_{d}$ and $\rho_{d}=-\nabla\cdot\vec{M}$. Equivalently[@CT-lectures], $$\begin{aligned}
\Phi_{d}(\vec{r})&=\int_{V}{\,\mathrm{d}}^{3}\vec{r}' \vec{M}(\vec{r}')\cdot\nabla'
\left(\frac{1}{|\vec{r}-\vec{r}'|}\right)\nonumber\\
&=-\int_{V}{\,\mathrm{d}}^{3}\vec{r}'\frac{\nabla'\cdot\vec{M}(\vec{r}')}{|\vec{r}-\vec{r}'|}
+\int_{\partial V}{\,\mathrm{d}}\vec{S}'\cdot\frac{\vec{M}(\vec{r}')}{|\vec{r}-\vec{r}'|}\nonumber\\
&=\int_{V}{\,\mathrm{d}}^{3}\vec{r}'\frac{\rho_{d}(\vec{r}')}{|\vec{r}-\vec{r}'|}
+\int_{\partial V}{\,\mathrm{d}}^{2}\vec{r}'\frac{\sigma_{d}(\vec{r}')}{|\vec{r}-\vec{r}'|}\end{aligned}$$ with $\sigma_{d}=\vec{M}\cdot\vec{n}$ being the effective surface “charge” density and $V$ being the volume of the system. Consequently, the demagnetizing energy can be written as $$\mathcal{E}_{demag}=\frac{1}{2}\int_{V}{\,\mathrm{d}}^{3}\vec{r}\rho_{d}(\vec{r})
\Phi_{d}(\vec{r})+\frac{1}{2}\int_{\partial V}{\,\mathrm{d}}^{2}\vec{r}\;
\sigma_{d}(\vec{r})\Phi_{d}(\vec{r})$$
We are dealing with a two-dimensional micrometric object, so the surface energy term dominates over the volume energy term. We can approximate this surface term by an effective easy plane anisotropy contribution given by $$\mathcal{E}_{demag,S}=\int_{V}{\,\mathrm{d}}^{3}\vec{r}\;2\pi M_{z}^{2}(\vec{r})$$ This gives for the total energy $$\begin{aligned}
\label{Energy}
\mathcal{E}(\Phi,m)&=A\left[\frac{1}{1-m^2}(\nabla m)^2+(1-m^2)
(\nabla\Phi)^2\right]\nonumber\\
&-K_{\shortparallel}\cos^2\Phi(1-m^2)+(K_{\perp}+2\pi M_{s}^{2})m^2\end{aligned}$$ and the equations of motion become: $$\begin{aligned}
\label{Eqmotion}
\frac{M_{s}}{\gamma}\frac{d\Phi}{dt}=&-\frac{2Am}{(1-m^2)^2}
(\nabla m)^2-\frac{2A}{1-m^2}\bigtriangleup m\nonumber\\
&- 2Am(\nabla\Phi)^2+2K_{\shortparallel}\cos^2\Phi\;m\nonumber\\
&+2(K_{\perp}+2\pi M_{s}^{2})m\nonumber\\
\frac{M_{s}}{\gamma}\frac{dm}{dt}=&-K_{\shortparallel}\sin(2\Phi)
(1-m^2)-4Am\nabla m\cdot\nabla\Phi\nonumber\\
&+2A(1-m^2)\bigtriangleup\Phi\end{aligned}$$ In the slow dynamics regime ($|\dot{\vec{X}}|\ll1$) solutions $(\Phi,m)$ can be split into $\Phi=\Phi_{0}+\Phi_{1}$ and $m=m_{0}+m_{1}$, where $\Phi_{0}$ and $m_{0}$ are the static solutions of the Landau-Lifshitz equation (we consider the anisotropy interaction to be weak enough so that the static solutions of the Hamiltonian $\mathcal{E}_{ex}+\mathcal{E}_{demag}$ are valid for our problem) and where $\Phi_{1}$ and $m_{1}$ are linear on $|\dot{\vec{X}}|$. Static solutions are given by Eqs. . As discussed in Section III, in the weak bending regime we can approximate $\tilde{r}\simeq r$ so that ${\displaystyle}\nabla\Phi_{0}=n_v\frac{\hat{e}_{\phi}}{r}$ and ${\displaystyle}\nabla
m_{0}=\frac{dm_{0}}{dr}\hat{e}_{r}$.
Linearizing Eqs. and taking into account that ${\displaystyle}\frac{d\Phi}{dt}=-\dot{\vec{X}}\cdot\nabla\Phi$ and that ${\displaystyle}\frac{dm}{dt}=-\dot{\vec{X}}\cdot\nabla m$, we obtain the equations of motion $$\begin{aligned}
\label{Lineqsmotion2}
-\frac{M_{s}}{\gamma}&n_v\dot{\vec{X}}\cdot\frac{\hat{e}_{\phi}}{r}
=\frac{-2A}{1-m_{0}^{2}}\bigtriangleup m_{1}-\Bigg[\frac{2A}
{(1-m_{0}^{2})^2}\left(\frac{dm_{0}}{dr}\right)^2\nonumber\\
&+\frac{2A
}{r^2}-2K_{\shortparallel}\cos^{2}\Phi_{0}
-2(K_{\perp}+2\pi M_{s}^2)\Bigg]m_{1}\nonumber\\
&-\frac{4A\;m_{0}}{(1-m_{0}^{2})^2}\frac{dm_{0}}{dr}
\hat{e}_{r}\cdot\nabla m_{1}-4A n_v\;m_{0}\frac{\hat{e}_{\phi}}
{r}\cdot\nabla\Phi_{1}\nonumber\\
&-2K_{\shortparallel}m_{0}\sin(2\Phi_{0})
\Phi_{1}\nonumber\\
-\frac{M_{s}}{\gamma}&\dot{\vec{X}}\cdot\hat{e}_{r} \frac{dm_{0}}
{dr}=2A(1-m_{0}^{2})\bigtriangleup\Phi_{1}\nonumber\\
&-4An_v\;m_{0}\nabla m_{1}\cdot \frac{\hat{e}_{\phi}}{r}-4A\;m_{0}\frac{dm_{0}}{dr}\hat{e}_{r}\cdot\nabla\Phi_{1}\nonumber\\
&+2K_{\shortparallel}\sin(2\Phi_{0})m_{0}m_{1}-2K_{\shortparallel}\cos(2\Phi_{0})(1-m_{0}^2)\Phi_{1}\end{aligned}$$
Asymptotic expressions for the $O(|\dot{\vec{X}}|)$ corrections to the out-of-plane vortex shape can be determined by substituting Eqs. into Eqs. . In doing so we obtain $$\begin{aligned}
\label{solinfty}
m_{1}&=-\frac{M_{s}}{2\gamma}n_v\frac{\dot{\vec{X}}\cdot\hat{e}_{\phi}}
{(K_{\perp}+2\pi M_{s}^{2})+K_{\shortparallel}\cos^{2}\Phi_{0}}\frac{1}{r}\nonumber\\
\Phi_{1}&=\frac{C_{2}M_{s}}{2\gamma A}\Delta_{0}^{3/2}(\dot{\vec{X}}\cdot\hat{e}_{r})
\frac{\exp{(-r/\Delta_{0})}}{r^{1/2}}\end{aligned}$$ for $r\gg\Delta_{0}$, and $$\begin{aligned}
\label{solzero} m_{1}&=\frac{M_{s} C_{1}n_v}{3\gamma
A\Delta_{0}^{2}}(\dot{\vec{X}}
\cdot\hat{e}_{\phi})r^{3}\nonumber\\
\Phi_{1}&=\frac{M_{s}}{\gamma}\frac{p}{18A}
\dot{\vec{X}}\cdot\vec{r}\end{aligned}$$ for $r\ll\Delta_{0}$. Computation of the mass of the vortex core can be made via $\vec{\Pi}_{t}$, which should be proportional to $\dot{\vec{X}}$ in this limit: $$\begin{aligned}
\label{momentum2}
\vec{\Pi}_{t}&=-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi)m\nonumber\\
&=-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{0})m_{0}-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{0})m_{1}\nonumber\\
&\quad-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{1})m_{0}-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{1})m_{1}\end{aligned}$$ Notice that ${\displaystyle}-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{0})m_{0}=\vec{0}$ because it corresponds to the momentum of the static solution. The last term of Eq. can be neglected because it is quadratic in $|\dot{\vec{X}}|$. Therefore it remains to calculate the second and third terms, which are given by $$\begin{aligned}
&-\frac{M_{s}}{\gamma}\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{0})m_{1}=-\frac{M_{s}}{\gamma}\int_{r\leq\Delta_{0}}{\,\mathrm{d}}^{2}\vec{r}
(\nabla\Phi_{0})m_{1}\nonumber\\
&\qquad\qquad-\frac{M_{s}}{\gamma}\int_{r\geq\Delta_{0}}
{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{0})m_{1}=\\
&
\frac{2\pi}{\gamma^{2}}\left(\frac{M_{s}^{2}}{K_{\perp}+2\pi
M_{s}^{2}}\frac{1/4}{\sqrt{1+\frac{K_{\shortparallel}}
{K_{\perp}+2\pi M_{s}^{2}}}}\ln(R/\Delta_{0})
-\frac{1}{56}\right)\dot{\vec{X}}\nonumber\end{aligned}$$ and $$\begin{aligned}
-\frac{M_{s}}{\gamma}&\int{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{1})m_{0}=
-\frac{M_{s}}{\gamma}\int_{r\leq\Delta_{0}}{\,\mathrm{d}}^{2}\vec{r}
(\nabla\Phi_{1})m_{0}\nonumber\\
&\qquad\qquad-\frac{M_{s}}{\gamma}\int_{r\geq\Delta_{0}}
{\,\mathrm{d}}^{2}\vec{r}(\nabla\Phi_{1})m_{0}\nonumber\\
&=\frac{2\pi}{\gamma^{2}}\left(-\frac{11}{504}
+\frac{2}{49}\left(1-\Xi\cdot e^2\right)\right)\dot{\vec{X}},\end{aligned}$$ respectively. Notice that ${\displaystyle}\Xi=\int_{1}^{R/\Delta_{0}} {\,\mathrm{d}}x
\frac{\exp(-2x)}{x}\simeq\int_{1}^{\infty} {\,\mathrm{d}}x
\frac{\exp(-2x)}{x}=0.049$ because we are interested in the limit $R\gg\Delta_{0}$.
Collecting all terms for the momentum, we get for the total mass density $$\begin{aligned}
\label{mass}
\rho_{M}=&\frac{2\pi}{\gamma^{2}}\left[
\frac{M_{s}^{2}}
{K_{\perp}+2\pi M_{s}^{2}}\frac{\ln(R/\Delta_{0})}{4\sqrt{1+
\frac{K_{\shortparallel}}{K_{\perp}+2\pi M_{s}^{2}}}}-0.0014\right]\end{aligned}$$
Notice that we are interested in the limit $R\gg\Delta_{0}$, so that the term involving $\ln(R/\Delta_{0})$ is the dominant one. Furthermore, redefining the exchange length by a factor close to unity we can always absorb the small numerical constant in Eq.(\[mass\]) into the logarithmic term. Magneto-crystalline anisotropies, if they are sufficiently large, destroy the circularly polarized state. Consequently, materials like permalloy, used in the studies of the vortex state, have negligible magneto-crystalline anisotropy energy as compared to the demagnetizing energy. This means that the above expression for the vortex mass density can be reduced to $$\label{mass3}
\rho_{M}\simeq\frac{1}{4\gamma^{2}}\ln(R/\Delta_{0})$$ With account of this formula one obtains the following expressions for the parameters $\omega_{M}$ and $\epsilon_{0}$ that determine eigenfrequencies in the equation (\[freq2\]): $$\begin{aligned}
\label{parameters}
\omega_{M}&=\frac{8\pi\gamma M_{s}}{\ln(R/\Delta_{0})}\nonumber\\
\epsilon_{0}&=\frac{5 L}{18\pi R}\ln(R/\Delta_{0})\end{aligned}$$
Effects of the magnetic field and dissipation
=============================================
In this section we study the effects of a magnetic field on the excitation modes of the vortex state. Arbitrary directed magnetic field can be split into two components, one being in the plane of the disk and the other one being perpendicular to it. The effects of these two components can be investigated separately.
Consider first the case of a spatially uniform in-plane magnetic field, $\vec{H}_{in}=h_{x}\hat{e}_{x}+h_{y}\hat{e}_{y}$. For small displacements along the disk, the magnetic vortex develops an in-plane magnetization density given by[@Guslienko2] $$\label{Magnetization}
\vec{M}(\vec{X})=-\mu\left[\hat{z}\times\vec{X}\right], \qquad
\mu=(2\pi/3)M_{s} n_v R.$$ The Zeeman energy density term is $$\begin{aligned}
\omega_{Z}(\vec{X})&=-M(\vec{X})\cdot\vec{H}_{in}=
-\mu\left[\hat{z}\times\vec{H}_{in}\right]\cdot\vec{X}\nonumber\\
&=\mu h_{y} x- \mu h_{x} y\end{aligned}$$ and thus the total in-plane potential energy becomes $$\begin{aligned}
\label{Potential2}
\omega_{XY}&(\vec{X})=\frac{1}{2}\rho_{M}\omega_{M}^2\epsilon_{0}\left(x^2+y^2\right)+\mu h_{y}x-\mu h_{x}y\nonumber\\
&=\frac{1}{2}\rho_{M}\omega_{M}^2\epsilon_{0}\Bigg[\left(x+\frac{\mu h_{y}}
{\rho_{M}\omega_{M}^2\epsilon_{0}}\right)^2\nonumber\\
&\quad+\left(y-\frac{\mu h_{x}}{\rho_{M}\omega_{M}^2\epsilon_{0}}\right)^2\Bigg]-\frac{1}{2}
\frac{\mu^2}{\rho_{M}\omega_{M}^2\epsilon_{0}}\vec{H}_{in}^2\end{aligned}$$ Notice that by shifting the origin of the coordinate system we retrieve the original in-plane term of the total energy density except for the constant term ${\displaystyle}-\frac{1}{2}\frac{\mu^2}{\rho_{M}\omega_{M}^2\epsilon_{0}}\vec{H}_{in}^2$, which is field dependent. Consequently, the application of an in-plane magnetic field does not modify the excitation modes given by .
Consider now the effect of the magnetic field perpendicular to the plane of the disk. $\vec{H}_{\perp}=H\hat{z}$. Application of such a field results in the precession of the magnetic moment of the vortex about the direction of the field, described by the Landau-Lifshitz equation[@CT-lectures], $$\label{LL} \frac{\partial \vec{M}(t,\vec{X})}{\partial
t}=-\gamma\left[\vec{M}(t,\vec{X}) \times\vec{H}_{\perp}\right],$$ where $\gamma$ is the electron gyromagnetic ratio. Formally, this effect can be accounted for by adding an extra term to the gyrovector. Indeed, integration of Eq. (with no potential energy) on time gives $\dot{\vec{X}}=\alpha\left[\vec{X}\times\vec{\rho}_{G}\right]$, where $\alpha=-1/\rho_{M}$. With account of Eq. , we have $$\begin{aligned}
\left[\hat{z}\times\dot{\vec{X}}\right]&=-\gamma\left[\hat{z}
\times\vec{X}\right]\times\vec{H}_{\perp}\\
\alpha\left(\hat{z}\times\left[\vec{X}\times\vec{\rho}_{G}\right]\right)&=
-\gamma\left[\hat{z}\times\vec{X}\right]\times\vec{H}_{\perp}\end{aligned}$$ The vector identity $\vec{a}\times\vec{b}\times\vec{c}=
(\vec{a}\cdot\vec{c})\vec{b}-(\vec{a}\cdot\vec{b})\vec{c}\;$ leads to $\alpha \rho_{G}=-\gamma H$. Consequently, the precessional effect of the perpendicular field can be absorbed into the gyrovector density if one adds to it the term ${\displaystyle}\vec{\rho}_{G,\vec{H}_{\perp}}=-\frac{\gamma}{\alpha}\vec{H}_{\perp}=\rho_{M}\gamma\vec{H}_{\perp}$. This adds the Larmor frequency to $\omega_{M}$: $$\omega_{M}(H)=\frac{\rho_{G,tot}}{\rho_{M}}=\omega_{M}+\frac{\rho_{G,\vec{H}_{\perp}}}
{\rho_{M}}=\omega_{M}+\gamma H$$ so that the eigenfrequencies become $$\label{freq3}
\omega_{n}^{\pm}(H)=\frac{1}{2}\left[\sqrt{(1+4\epsilon(H))+\frac{4\lambda}{\rho_{M}\omega_{M}^{2}(H)}
q_{n}^{2}}\pm1\right]\omega_{M}(H)$$ with $\epsilon(H)$ given by $$\label{epsilon}
\epsilon(H)=\frac{\omega_{G}(H)}{\omega_{M}(H)}=\frac{\omega_{XY}^{''}(\vec{X}=\vec{0})} {\rho_{M}\omega_{M}^2(H)}=\frac{\epsilon_{0}}{(1+\gamma
H/\omega_{M})^2}$$
Introducing dimensionless variables $h=\gamma H/\omega_{M}$ and $\bar{\omega}_{n}^{\pm}(h)=\omega_{n}^{\pm}(H)/\omega_{M}$ we can rewrite Eqs. as $$\begin{aligned}
\label{freq4}
\bar{\omega}_{n}^{\pm}(h)&=\frac{1}{2}\left[\sqrt{1+\frac{4\epsilon_{0}}{(1+h)^2}+\frac{4\lambda}{\rho_{M}\omega_{M}^{2}(1+h)^{2}}
q_{n}^{2}}\pm1\right]\times\nonumber\\
&\qquad(1+h)\nonumber\\
&\simeq\frac{1}{2}\left[\sqrt{1+\frac{4\epsilon_{0}}{(1+h)^2}}\pm1\right](1+h)\nonumber\\
&\qquad\qquad+\frac{\textrm{sgn}(1+h)}{\sqrt{(1+h)^2+4\epsilon_{0}}}\frac{\lambda q_{n}^{2}}{\rho_{M}\omega_{M}^{2}}\end{aligned}$$ The distance between $\omega_{n}^{+}$ and $\omega_{n}^{-}$ equals $\Delta
\omega=\omega_{M}+\gamma H$.
To conclude this Section, we investigate the effects of the dissipation on the excitation modes of magnetic vortices. We consider only the zero field case. Derivation of the corresponding expressions when a magnetic field is applied is straightforward. The way to introduce dissipation into our equations is by adding a damping term of the form $-D\dot{\vec{X}}$ ($D$ being the damping constant) to Eq. [@Guslienko2; @Thiele]. Therefore, the elastic Thiele’s equation becomes $$\label{Damping}
\rho_{M}\ddot{\vec{X}}-\lambda\partial^{2}_{z}\vec{X}+\dot{\vec{X}}\times\vec{\rho}_{G}-D\dot{\vec{X}}+\rho_{M}\omega_{M}^2\epsilon_{0}\vec{X}=0$$ Repeating the procedure of Sec. IV with the above equation in the massive vortex case ($\rho_{M}\neq0$) we obtain the following equation for the frequency modes $$\label{eqz}
\omega^{2}+(\omega_{M}+id)\omega-\omega_{M}^2\epsilon(q)=0$$ with $d=D/\rho_{M}$ and $\epsilon(q)=\epsilon_{0}+\frac{\lambda}{\rho_{M}\omega_{M}^{2}}q^{2}$. The (complex) roots of this equation, $\omega_{\pm}=\textrm{Re}(\omega_{\pm})+i\textrm{Im}(\omega_{\pm})$, are given by $$\begin{aligned}
\textrm{Re}(\omega_{\pm})&=&\mp\frac{r^{1/2}}{2}\cos(\theta/2)
-\frac{\omega_{M}}{2},\nonumber\\
\quad\textrm{Im}(\omega_{\pm})&=&\mp\frac{r^{1/2}}{2}\sin(\theta/2)-\frac{d}{2}\end{aligned}$$ with $$\begin{aligned}
\label{polar}
r&=\sqrt{\left[(1+4\epsilon(q))\omega_{M}^2-d^2\right]^2+4d^2\omega_{M}^2}\nonumber\\
\theta&=\arg{\Big(\left[(1+4\epsilon(q))\omega_{M}^2-d^2\right]+
i\left[2d\omega_{M}\right]\Big)}\nonumber\\
&=\arctan{\left(\frac{2d\omega_{M}}
{(1+4\epsilon(q))\omega_{M}^2-d^2}\right)}\end{aligned}$$
In the regime of weak dissipation, $d<<\omega_{M}$, we have ${\displaystyle}\theta\simeq\arctan\left[\frac{2d}{(1+4\epsilon(q))\omega_{M}}\right]$ and $r\simeq(1+4\epsilon(q))\omega_{M}^2$. As ${\displaystyle}\cos[\arctan(x)/2]\simeq1-\frac{x^2}{8}+o(x^4)$ and ${\displaystyle}\sin[\arctan(x)/2]\simeq\frac{x}{2}+o(x^3)$ if $|x|\ll1$, we finally obtain $$\begin{aligned}
\label{ReIm}
\textrm{Re}(\omega_{\pm})&=\mp\left[\frac{1}{2}\left(\sqrt{1+4\epsilon(q)}\pm1\right)
-\frac{1}{4}\frac{(d/\omega_{M})^2}{(1+4\epsilon(q))^{3/2}}\right]\omega_{M}\nonumber\\
&\simeq\mp\Bigg[\frac{\omega_{M}}{2}\left(\sqrt{1+4\epsilon_{0}}\pm1\right)
-\frac{\omega_{M}}{4}\frac{(d/\omega_{M})^2}{(1+4\epsilon_{0})^{3/2}}\nonumber\\
&+\frac{\lambda}{\sqrt{1+4\epsilon_{0}}}\left(1+\frac{3}{2}\frac{(d/\omega_{M})^{2}}{(1+4\epsilon_{0})^{2}}\right)\frac{q^{2}}{\rho_{M}\omega_{M}}\Bigg]\end{aligned}$$ and $$\begin{aligned}
\textrm{Im}(\omega_{\pm})&=\left(\mp\frac{1}{\sqrt{1+4\epsilon(q)}}-1\right)\frac{d}{2}\\
\frac{\textrm{Im}(\omega_{+})}
{\textrm{Im}(\omega_{-})}&=-\frac{1+\sqrt{1+4\epsilon(q)}}{1-
\sqrt{1+4\epsilon(q)}}=\frac{(1+\sqrt{1+4\epsilon(q)})^2}{4\epsilon(q)}\end{aligned}$$
Conclusions
===========
We have studied excitation modes of vortices in circularly polarized mesoscopic magnetic disks that correspond to the string-like gyroscopic waves in the vortex core. This problem was studied by classical treatment based upon Landau-Lifshitz equation and by quantum treatment based upon Hamiltonian approach. The quantum problem is interesting on its own as it is equivalent to the problem of quantum oscillations of a charged string confined in a parabolic potential and subjected to the magnetic field, which in its turn, is a generalization of the problem of the field-induced orbital motion of the electron in a potential well. Both treatments rendered identical results. Our solution generalizes the expression for the frequency of the gyroscopic motion of the vortex for the case of the finite wave number $q$, as $\omega_-(q) = \omega_G + \gamma
M_s(q\Delta_0)^2\ln(R/\Delta_0)$, where $\omega_G$ is the conventional gyrofrequency, $\gamma$ is the gyromagnetic ratio, $M_s$ is the saturation magnetization, $\Delta_0$ is the exchange length, and $R$ is the radius of the disk. This expression is valid in the long-wave limit $q\Delta_0 \ll 1$. The wave number is quantized, $q_n = 2\pi n/L$, where $L$ is the thickness of the disk and $n$ is an integer. For a disk of radius $R \sim 1\mu$m, thickness $L \sim 100$nm, exchange length $\Delta_0 \sim 5$nm, and saturation magnetization $M_s \sim 10^3$emu, the $n = 1$ mode is separated from $\omega_G$ by a few GHz. It could be excited by, e.g., a tip of a force microscope or a micro-SQUID placed at the center of the disk. Such measurement, while challenging, is definitely within experimental reach.
Throughout this paper we considered of a non-zero mass of the vortex. In addition to the gyroscopic mode the finite provides a new excitation mode, $\omega_+(q) = \omega_M + \gamma
M_s(q\Delta_0)^2\ln(R/\Delta_0)$. The gap, $\omega_M$ is higher than the gyroscopic frequency $\omega_G$. It depends explicitly on the vortex mass. The vortex mass density has been computed by us as a coefficient of proportionality, $\rho_{M}$, in the kinetic energy of the moving vortex $\rho_{M}v^2/2$. It is given by $\rho_{M}\simeq {1}/(4\gamma^{2})\ln(R/\Delta_{0})$, where $R$ is the radius of the disk, $\Delta_0$ is the exchange length, and $\gamma$ is the gyromagnetic ratio. For a $25$nm thick, micron size permalloy disk this gives the vortex mass in the ball park of $10^{-23}$kg, which is close to the experimental value estimated for a comparable size permalloy ring[@Bedau]. Our result for the mass gives $\omega_M = 8\pi \gamma M_{s}/\ln(R/\Delta_{0})$. This is in the ballpark of, or below, the uniform ferromagnetic resonance of the disk. It would be interesting to investigate this frequency range experimentally alongside with the low-frequency gyroscopic mode. One can also test in experiment the explicit field dependence of the vortex modes, computed in this paper. So far we have done it for the low field that only slightly disturbs the vortex state formed in a zero field. However, the statement concerning the existence of the additional mode due to the finite vortex mass should apply to higher fields as well. This case, however, defies analytical study and must employ full-scale numerical micromagnetic calculations. When the field is sufficient to fully polarize the disk in the perpendicular direction, we expect the high frequency mode to evolve into the uniform ferromagnetic resonance.
Acknowledgements
================
The work at the University of Barcelona was supported by the Spanish Government Project No. MAT2008-04535. R.Z. acknowledges financial support from the Ministerio de Ciencia e Innovación de España. The work of E.M.C. at Lehman College is supported by the Department of Energy through grant No. DE-FG02-93ER45487.
[99]{} R. P. Cowburn, D. K. Koltsov, A. O. Adeyeye, M. E. Welland, and D. M. Tricker. Phys. Rev. Lett. [**83**]{}, 1042 (1999). T. Shinjo, T. Okuno, R. Hassdorf, K. Shigeto, and T. Ono. Science [**289**]{}, 930 (2000). V. Novosad, K. Yu. Guslienko, H. Shima, Y. Otani, S. G. Kim, K. Fukamichi, N. Kikuchi, O. Kitakami and Y. Shimada. Phys. Rev. B [**65**]{}, 060402(R) (2002). V. Novosad, F. Y. Fradin, P. E. Roy, K. S. Buchanan, K. Yu. Guslienko and S. D. Bader. Phys. Rev. B [**72**]{}, 024455 (2005). G. de Loubens, A. Riegler, B. Pigeau, F. Lochner, F. Boust, K. Y. Guslienko, H. Hurdequint, L. W. Molenkamp, G. Schmidt, A. N. Slavin, V. S. Tiberkevich, N. Vukadinovic and O. Klein. Phys. Rev. Lett. [**102**]{}, 177602 (2009). V. Castel, J. Ben Youssef, F. Boust, R. Weil, B. Pigeau, G. de Loubens, V. V. Naletov, O. Klein and N. Vukadinovic. Phys. Rev. B [**85**]{}, 184419 (2012). R. Zarzuela, S. Vélez, J. M. Hernandez, J. Tejada and V. Novosad. Phys. Rev. B [**85**]{}, 180401(R) (2012). S. S. P. Parkin, M. Hayashi, and L. Thomas. Science [**320**]{}, 190 (2008). E. A. Rozhkova, V. Novosad, D.-H. Kim, J. Pearson, R. Divan, T. Rajh and S. D. Bader. J. Appl. Phys. [**105**]{}, 07B306 (2009). D.-H. Kim, E. A. Rozhkova, I. V. Ulasov, S. D. Bader, T. Rajh, M. S. Lesniak and V. Novosad. Nature Materials [**9**]{}, 165-171 (2010). J. Kin Ha, R. Hertel and J. Kirschner. Phys. Rev. B [**67**]{}, 224432 (2003). K. L. Metlov and K. Yu. Guslienko. J. Mag. Mag. Mat. [**242-245**]{}, 1015 (2002). S.-B. Choe, Y. Acremann, A. Scholl, A. Bauer, A. Doran, J. Stöhr and H. A. Padmore. Science [**304**]{}, 420-422 (2004). K. Yu. Guslienko. Appl. Phys. Lett. [**89**]{}, 022510 (2006). K. Yu. Guslienko, B. A. Ivanov, V. Novosad, Y. Otani, H. Shima and K. Fukamichi. J. Appl. Phys. [**91**]{}, 8037 (2002). K. Yu. Guslienko, X. F. Han, D. J. Keavney, R. Divan and S. D. Bader. Phys. Rev. Lett [**96**]{}, 067205 (2006). K.-S. Lee and S.-K. Kim. Appl. Phys. Lett. [**91**]{}, 132511 (2007). D. Bedau, M. Kläui, S. Krzyk, U. Rüdiger, G. Faini and L. Vila. Phys. Rev. Lett. [**99**]{}, 146601 (2007). M. E. Gouvêa, G. M. Wysin, A. R. Bishop and F. G. Mertens. Phys. Rev. B [**39**]{}, 11840 (1989). G. M. Wysin. Phys. Rev. B [**54**]{}, 15156 (1996). B. A. Ivanov and G. M. Wysin. Phys. Rev. B [**65**]{}, 134434 (2002). E.M. Chudnovsky and J. Tejada. *Lectures on Magnetism* (Rinton Press, Princeton, NJ, 2006). A. A. Thiele. Phys. Rev. Lett. [**30**]{}, 230 (1973). D. L. Huber, Phys. Rev. B [**26**]{}, 3758 (1982). R. B. Laughlin. Phys. Rev. B [**27**]{}, 3383 (1983). J. Zak. Phys. Rev. **134**, 6A (1964).
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.